From 6bd1d82c7f647611f4fc2c0e68cf107c9fa1dcb1 Mon Sep 17 00:00:00 2001 From: mxz96102 Date: Mon, 15 Apr 2019 16:14:16 +0800 Subject: [PATCH 1/9] update to using git submodule to maintain docs --- .gitmodules | 3 + _config.yml | 4 +- _layouts/index.html | 29 +- _layouts/single.html | 4 +- assets/select.gif | Bin 0 -> 71982 bytes assets/sqlflow-logo.svg | 44 + assets/train.gif | Bin 0 -> 70492 bytes doc/submitter.md | 93 -- docs/assets/js/search-data.json | 111 +- docs/doc/build-tensorflow.html | 339 ----- docs/doc/build-tensorflow.md | 135 -- docs/doc/build.html | 316 ----- docs/doc/build.md | 128 -- docs/doc/close_producer_from_consumer.html | 357 ------ docs/doc/close_producer_from_consumer.md | 161 --- docs/doc/database_abstraction_layer.html | 293 ----- docs/doc/database_abstraction_layer.md | 92 -- docs/doc/figures/sqlflow-arch.graffle | Bin 3985 -> 0 bytes docs/doc/figures/sqlflow-arch.png | Bin 99876 -> 0 bytes docs/doc/figures/sqlflow-arch2.dot | 50 - docs/doc/figures/sqlflow-arch2.png | Bin 96987 -> 0 bytes docs/doc/mysql-setup.html | 343 ----- docs/doc/mysql-setup.md | 138 -- docs/doc/pipe.html | 284 ----- docs/doc/pipe.md | 91 -- docs/doc/quickstart.html | 294 ----- docs/doc/quickstart.md | 91 -- docs/doc/sql_parser.html | 265 ---- docs/doc/sql_parser.md | 65 - docs/doc/submitter.html | 288 ----- docs/doc/syntax.html | 358 ------ docs/doc/syntax.md | 160 --- docs/doc/walkthrough.html | 272 ---- docs/doc/walkthrough.md | 71 -- docs/feed.xml | 2 +- docs/index.html | 33 +- docs/pages/about.html | 112 +- docs/sitemap.xml | 55 +- docs/sqlflow/cmd/demo/index.html | 368 +++++- docs/sqlflow/cmd/sqlflowserver/index.html | 13 - docs/sqlflow/index.html | 319 ++++- docs/sqlflow/server/index.html | 306 ++++- docs/sqlflow/sql/index.html | 303 ++++- docs/sqlflow/sqlfs/index.html | 341 ++++- index.md | 2 + sqlflow/.gitignore | 11 + sqlflow/.pre-commit-config.yaml | 9 + sqlflow/.travis.yml | 37 + sqlflow/Dockerfile | 24 + sqlflow/Dockerfile.dev | 45 + sqlflow/LICENSE | 201 +++ sqlflow/README.md | 24 + sqlflow/cmd/demo/README.md | 82 ++ sqlflow/cmd/demo/demo.go | 121 ++ sqlflow/cmd/demo/index.html | 13 - sqlflow/cmd/demo/slides/201812/automl.sql | 7 + sqlflow/cmd/demo/slides/201812/pred.sql | 22 + sqlflow/cmd/demo/slides/201812/preprocess.sql | 10 + sqlflow/cmd/demo/slides/201812/regular.sql | 20 + sqlflow/cmd/demo/slides/201812/sqlflow.slide | 40 + sqlflow/cmd/demo/slides/201812/train.sql | 12 + sqlflow/cmd/sqlflowserver/index.html | 13 - sqlflow/cmd/sqlflowserver/main.go | 59 + sqlflow/cmd/sqlflowserver/main_test.go | 40 + {doc => sqlflow/doc}/build-tensorflow.md | 0 {doc => sqlflow/doc}/build.md | 0 .../doc}/close_producer_from_consumer.md | 0 .../doc}/database_abstraction_layer.md | 0 .../doc}/figures/sqlflow-arch.graffle | Bin {doc => sqlflow/doc}/figures/sqlflow-arch.png | Bin .../doc}/figures/sqlflow-arch2.dot | 0 .../doc}/figures/sqlflow-arch2.png | Bin {doc => sqlflow/doc}/mysql-setup.md | 0 {doc => sqlflow/doc}/pipe.md | 0 {doc => sqlflow/doc}/quickstart.md | 10 +- {doc => sqlflow/doc}/sql_parser.md | 0 {docs => sqlflow}/doc/submitter.md | 3 +- {doc => sqlflow/doc}/syntax.md | 0 {doc => sqlflow/doc}/walkthrough.md | 0 sqlflow/example/creditcard/README.md | 149 +++ .../initialize_dataset_and_database.bash | 127 ++ sqlflow/example/datasets/Dockerfile | 4 + sqlflow/example/datasets/README.md | 75 ++ sqlflow/example/datasets/data_to_sql.awk | 20 + sqlflow/example/datasets/entrypoint.sh | 8 + sqlflow/example/datasets/popularize_churn.sql | 144 +++ sqlflow/example/datasets/popularize_iris.sql | 136 ++ sqlflow/example/fraud_detection/README.md | 101 ++ .../fraud_detection/creditcard.csv.test | 10 + .../fraud_detection/creditcard.csv.train | 10 + .../fraud_detection/creditcard_data.py | 59 + .../fraud_detection/premade_estimator.py | 115 ++ sqlflow/example/jupyter/example.ipynb | 106 ++ sqlflow/index.html | 13 - sqlflow/scripts/deploy.sh | 13 + sqlflow/scripts/test.sh | 9 + sqlflow/server/.gitignore | 5 + sqlflow/server/.travis.yml | 36 + sqlflow/server/README.md | 14 + sqlflow/server/index.html | 13 - sqlflow/server/proto/sqlflow.proto | 60 + sqlflow/server/sqlflowserver.go | 130 ++ sqlflow/server/sqlflowserver_test.go | 139 ++ sqlflow/sql/.gitignore | 2 + sqlflow/sql/README.md | 13 + sqlflow/sql/cmd.go | 54 + sqlflow/sql/codegen.go | 213 ++++ sqlflow/sql/codegen_test.go | 59 + sqlflow/sql/column_type.go | 157 +++ sqlflow/sql/database.go | 36 + sqlflow/sql/database_test.go | 33 + sqlflow/sql/db_test.go | 96 ++ sqlflow/sql/executor.go | 346 +++++ sqlflow/sql/executor_test.go | 125 ++ sqlflow/sql/index.html | 13 - sqlflow/sql/lexer.go | 163 +++ sqlflow/sql/lexer_test.go | 102 ++ sqlflow/sql/log.go | 44 + sqlflow/sql/model.go | 65 + sqlflow/sql/parser.go | 1114 +++++++++++++++++ sqlflow/sql/parser_test.go | 96 ++ sqlflow/sql/pipe.go | 65 + sqlflow/sql/pipe_test.go | 59 + sqlflow/sql/python/.gitignore | 1 + sqlflow/sql/python/README.md | 13 + sqlflow/sql/python/sql_data.py | 57 + sqlflow/sql/python/test_sql_data.py | 32 + sqlflow/sql/sql.y | 422 +++++++ sqlflow/sql/test | 0 sqlflow/sql/testdata/churn.sql | 143 +++ sqlflow/sql/testdata/iris.sql | 143 +++ sqlflow/sql/verifier.go | 164 +++ sqlflow/sql/verifier_test.go | 113 ++ sqlflow/sqlfs/.gitignore | 0 sqlflow/sqlfs/README.md | 56 + sqlflow/sqlfs/index.html | 13 - sqlflow/sqlfs/reader.go | 73 ++ sqlflow/sqlfs/sqlfile_test.go | 151 +++ sqlflow/sqlfs/table.go | 49 + sqlflow/sqlfs/writer.go | 84 ++ 140 files changed, 8500 insertions(+), 4901 deletions(-) create mode 100644 .gitmodules create mode 100644 assets/select.gif create mode 100644 assets/sqlflow-logo.svg create mode 100644 assets/train.gif delete mode 100644 doc/submitter.md delete mode 100644 docs/doc/build-tensorflow.html delete mode 100644 docs/doc/build-tensorflow.md delete mode 100644 docs/doc/build.html delete mode 100644 docs/doc/build.md delete mode 100644 docs/doc/close_producer_from_consumer.html delete mode 100644 docs/doc/close_producer_from_consumer.md delete mode 100644 docs/doc/database_abstraction_layer.html delete mode 100644 docs/doc/database_abstraction_layer.md delete mode 100644 docs/doc/figures/sqlflow-arch.graffle delete mode 100644 docs/doc/figures/sqlflow-arch.png delete mode 100644 docs/doc/figures/sqlflow-arch2.dot delete mode 100644 docs/doc/figures/sqlflow-arch2.png delete mode 100644 docs/doc/mysql-setup.html delete mode 100644 docs/doc/mysql-setup.md delete mode 100644 docs/doc/pipe.html delete mode 100644 docs/doc/pipe.md delete mode 100644 docs/doc/quickstart.html delete mode 100644 docs/doc/quickstart.md delete mode 100644 docs/doc/sql_parser.html delete mode 100644 docs/doc/sql_parser.md delete mode 100644 docs/doc/submitter.html delete mode 100644 docs/doc/syntax.html delete mode 100644 docs/doc/syntax.md delete mode 100644 docs/doc/walkthrough.html delete mode 100644 docs/doc/walkthrough.md delete mode 100644 docs/sqlflow/cmd/sqlflowserver/index.html create mode 100644 sqlflow/.gitignore create mode 100644 sqlflow/.pre-commit-config.yaml create mode 100644 sqlflow/.travis.yml create mode 100644 sqlflow/Dockerfile create mode 100644 sqlflow/Dockerfile.dev create mode 100644 sqlflow/LICENSE create mode 100644 sqlflow/README.md create mode 100644 sqlflow/cmd/demo/README.md create mode 100644 sqlflow/cmd/demo/demo.go delete mode 100644 sqlflow/cmd/demo/index.html create mode 100644 sqlflow/cmd/demo/slides/201812/automl.sql create mode 100644 sqlflow/cmd/demo/slides/201812/pred.sql create mode 100644 sqlflow/cmd/demo/slides/201812/preprocess.sql create mode 100644 sqlflow/cmd/demo/slides/201812/regular.sql create mode 100644 sqlflow/cmd/demo/slides/201812/sqlflow.slide create mode 100644 sqlflow/cmd/demo/slides/201812/train.sql delete mode 100644 sqlflow/cmd/sqlflowserver/index.html create mode 100644 sqlflow/cmd/sqlflowserver/main.go create mode 100644 sqlflow/cmd/sqlflowserver/main_test.go rename {doc => sqlflow/doc}/build-tensorflow.md (100%) rename {doc => sqlflow/doc}/build.md (100%) rename {doc => sqlflow/doc}/close_producer_from_consumer.md (100%) rename {doc => sqlflow/doc}/database_abstraction_layer.md (100%) rename {doc => sqlflow/doc}/figures/sqlflow-arch.graffle (100%) rename {doc => sqlflow/doc}/figures/sqlflow-arch.png (100%) rename {doc => sqlflow/doc}/figures/sqlflow-arch2.dot (100%) rename {doc => sqlflow/doc}/figures/sqlflow-arch2.png (100%) rename {doc => sqlflow/doc}/mysql-setup.md (100%) rename {doc => sqlflow/doc}/pipe.md (100%) rename {doc => sqlflow/doc}/quickstart.md (88%) rename {doc => sqlflow/doc}/sql_parser.md (100%) rename {docs => sqlflow}/doc/submitter.md (99%) rename {doc => sqlflow/doc}/syntax.md (100%) rename {doc => sqlflow/doc}/walkthrough.md (100%) create mode 100644 sqlflow/example/creditcard/README.md create mode 100755 sqlflow/example/creditcard/initialize_dataset_and_database.bash create mode 100644 sqlflow/example/datasets/Dockerfile create mode 100644 sqlflow/example/datasets/README.md create mode 100644 sqlflow/example/datasets/data_to_sql.awk create mode 100755 sqlflow/example/datasets/entrypoint.sh create mode 100644 sqlflow/example/datasets/popularize_churn.sql create mode 100644 sqlflow/example/datasets/popularize_iris.sql create mode 100644 sqlflow/example/fraud_detection/README.md create mode 100644 sqlflow/example/fraud_detection/creditcard.csv.test create mode 100644 sqlflow/example/fraud_detection/creditcard.csv.train create mode 100644 sqlflow/example/fraud_detection/creditcard_data.py create mode 100644 sqlflow/example/fraud_detection/premade_estimator.py create mode 100644 sqlflow/example/jupyter/example.ipynb delete mode 100644 sqlflow/index.html create mode 100644 sqlflow/scripts/deploy.sh create mode 100644 sqlflow/scripts/test.sh create mode 100644 sqlflow/server/.gitignore create mode 100644 sqlflow/server/.travis.yml create mode 100644 sqlflow/server/README.md delete mode 100644 sqlflow/server/index.html create mode 100644 sqlflow/server/proto/sqlflow.proto create mode 100644 sqlflow/server/sqlflowserver.go create mode 100644 sqlflow/server/sqlflowserver_test.go create mode 100644 sqlflow/sql/.gitignore create mode 100644 sqlflow/sql/README.md create mode 100644 sqlflow/sql/cmd.go create mode 100644 sqlflow/sql/codegen.go create mode 100644 sqlflow/sql/codegen_test.go create mode 100644 sqlflow/sql/column_type.go create mode 100644 sqlflow/sql/database.go create mode 100644 sqlflow/sql/database_test.go create mode 100644 sqlflow/sql/db_test.go create mode 100644 sqlflow/sql/executor.go create mode 100644 sqlflow/sql/executor_test.go delete mode 100644 sqlflow/sql/index.html create mode 100644 sqlflow/sql/lexer.go create mode 100644 sqlflow/sql/lexer_test.go create mode 100644 sqlflow/sql/log.go create mode 100644 sqlflow/sql/model.go create mode 100644 sqlflow/sql/parser.go create mode 100644 sqlflow/sql/parser_test.go create mode 100644 sqlflow/sql/pipe.go create mode 100644 sqlflow/sql/pipe_test.go create mode 100644 sqlflow/sql/python/.gitignore create mode 100644 sqlflow/sql/python/README.md create mode 100644 sqlflow/sql/python/sql_data.py create mode 100644 sqlflow/sql/python/test_sql_data.py create mode 100644 sqlflow/sql/sql.y create mode 100644 sqlflow/sql/test create mode 100644 sqlflow/sql/testdata/churn.sql create mode 100644 sqlflow/sql/testdata/iris.sql create mode 100644 sqlflow/sql/verifier.go create mode 100644 sqlflow/sql/verifier_test.go create mode 100644 sqlflow/sqlfs/.gitignore create mode 100644 sqlflow/sqlfs/README.md delete mode 100644 sqlflow/sqlfs/index.html create mode 100644 sqlflow/sqlfs/reader.go create mode 100644 sqlflow/sqlfs/sqlfile_test.go create mode 100644 sqlflow/sqlfs/table.go create mode 100644 sqlflow/sqlfs/writer.go diff --git a/.gitmodules b/.gitmodules new file mode 100644 index 0000000..0b2a778 --- /dev/null +++ b/.gitmodules @@ -0,0 +1,3 @@ +[submodule "sqlflow"] + path = sqlflow + url = https://github.com/sql-machine-learning/sqlflow diff --git a/_config.yml b/_config.yml index a9e28ec..a7ef609 100644 --- a/_config.yml +++ b/_config.yml @@ -2,8 +2,6 @@ remote_theme: pmarsceill/just-the-docs classes: wide -include: ["pages", "doc"] - search_enabled: true title: 'Sqlflow' @@ -16,6 +14,8 @@ aux_links: "GitHub": - "https://github.com/sql-machine-learning/sqlflow" +exclude: ['*.sql', '*.go', '.gitignore', '*.sh', '*.proto', '*.y'] + # baseurl: "/sql-machine-learning.github.io/" footer: diff --git a/_layouts/index.html b/_layouts/index.html index 1453f01..e2b096f 100644 --- a/_layouts/index.html +++ b/_layouts/index.html @@ -685,18 +685,28 @@ border-top: 1px dashed #aaa; } - section.feature h3, section.feature p{ + section.feature{ text-align: left; padding-left: 6px; padding-right: 6px; } - section.feature:nth-of-type(2n) h3, section.feature:nth-of-type(2n) p { + section.feature img { + width: 550px; + margin: auto 20px; + } + + section.feature:nth-of-type(2n){ text-align: right; padding-right: 6px; padding-right: 6px; } + section.feature .gif-holder { + width: 550px; + height: 275px; + } + @media only screen and (min-width: 40.063em) { #home h1 { font-size: 2.7rem; @@ -713,7 +723,7 @@ } .main-nav .logo img { - height: 50px; + height: 30px; } .main-page .intro { @@ -723,7 +733,7 @@ .main-page .intro h1 { margin: 50px auto; - color: #ffb214; + color: #1BA2FF; } .main-page .intro p { @@ -746,13 +756,13 @@
@@ -795,6 +805,11 @@

{{page.subtitle}}

{{feat.title}}

{{feat.content}}

+ {% if feat.gif %} + gif + {% else %} +
+ {% endif %}
{% endfor %}
diff --git a/_layouts/single.html b/_layouts/single.html index 3f48868..9bfea2d 100644 --- a/_layouts/single.html +++ b/_layouts/single.html @@ -704,7 +704,7 @@ } .main-nav .logo img { - height: 50px; + height: 30px; } .main-page .intro { @@ -737,7 +737,7 @@
diff --git a/assets/select.gif b/assets/select.gif new file mode 100644 index 0000000000000000000000000000000000000000..037549c7244f2b0e5509335bca12e89f71e71219 GIT binary patch literal 71982 zcmce-Wq2FSmbNQ1b8N>DbIi=lF+0YXnVDl|W{jDcnVFfHnVBt9tdnJ z*U=w!m)2TU)q1Y#lKOtcCB-bx-+^&2i`bM?}#@v67|7vC+A^MLHCrdsOp}#&L`m0Y7TL)tz zHhOkCLnc->A`UKkCUzE9E-qRkW=1AX21ZT>CKfs-c5YTiZf0hpf4oTk^yXk>!mTJO z_K&{)obi#EIXT&JGcdTgxX`<>(AzqgGB9y*aWOD5GcYsL{fVG+bhB~Nccrs&B>lUB zsIjA=gSnlPxvdS+UlsKYY=1iOk^C9ye@?;L?w@6C9O?gKI`oFN)(o!tb_`7PjDJn( z??74E|2fp!`k&E`PKw6=HQxWv!j8&rcE${f#*Vf>9Sn{Cgp>Z&l^wT;gR#Dot%I_y zt<~RGQQpkf$=1=#){aO-Dw3@ON#Q5{Ha54 zZf?ZQBFfDAP57G_8wUp$6Vo>?u5awZoUF`jj7-8Dj6$Mff9HzY8ve94wsHD9*XWk`fuy~BiHEPj^*38f1AtjXD|$Z zeZBwkb^qP;M?-%d{!_evPW~x>W1BykclaZ2Fwn>E_qW%VXW-N0!~Nau&Gpsg#rfIk z$??(Q!T#Rv&i2;k#`@ao%JS0U!u;Iq%=Fac#Q50g$nenMK!0CvPj^>mM|)dqOLJ3W zLw#LsO?6deMR{3iNpVqOL4ICtPIgviMtWLmN^(--uY~xx*qG?3$cXT;(2(GuzyNPL2-tcD6RwR+bj#W~L^_MurCZdb&E=TACW_YN{$fl$8_}5+|ujjp~&{%JED0?wuM}8T$0; zlLq%moA&LS7SHe1>(|d6-qY7_AYVWKfWV;OkkGL3h{&kunAo`ZgkOnC$tkI6=^2?> z**Up+`2~eV#U-U>l>R}+dI2^`v-?d$0w&}=NFe(*EhF!_YaRxz~`6OxA)&4AOHkBzC=w< zZvZR?wbn>YZhr_WsZ@qUZQfwSXBLZ%k=p#>7$T7;YQegK(F95*P?^?fUEz2#qv2SF zWPQ3iXVxsWR>HrponZJIlLQ_1vn>Hva3g&GF{y?QUNP1Ol0sn%(|T3>uw@mfHQ{ zSW@Y%3y$N8j9)C4TNADI$J03?5d^Yp_miru(&ajnZH?#4HHPEPNRF5jt1XV_Ta)e0 z*W2N`No`0SEw}q42{gJkEye-hd0Ncb-#gnL&sT)1-+$c+0+4VaLkZ=&I$rKi=F4@b zyE@;1H~Zt+a@}3O-=6Q!x2L4}3{lf0qQL^pav~ASsLDCK z&&~nC@zfRI0_ZtswVzPoNaMC=Pj@PK*wu;}An5N#EG zth$W)_%m}H<~EKHcs#+3K79}FXTF^Q0$_&W3_MQ;*&Heh!egRLvE<#7Ymtn} zj9gvCa^ZE}LUIS!l+1lJ2Li%l7f<=?J~#?3Dqx=RXwn=2!-&rs$w1!&=wk$*GT##` zN62anqp)L+K@O&EUEmT%_1!`O&m!_ma9jt1ntVNg&k#jQS2->^Pm^{56_&m(XB_Sy zE<5AgRxPLSwo}}#JucD@{=PgmJbjqSSviZ01XV=+Gu_V*zJ5M1u3J03p&pGWYCeD` zPK)dv$kpkq!mVHyGtN^Ovn@u*O&@Ldrs#xrE{F?}ox%Xu{Sgqv!o2P!RxEda>$DWu zP5MgJPrzExxQeHcCb8lkcGK>Nbr!DLMqndp+ajGhpGME|XTgS8eQEnT77a*9wc1$q zi$I(R-|fC*-ql8AJ&^xy4E$0f@uZy@_QQU{>QBQk|KVjKl_l4deqNwnv_%Tr)p*$OSkXk zY4BjpM3^L2j`(jIU@6ezaoWgEh`3o;erg3*y;JVrf8|B%zzAZae|H|wPs6k<7M9_^ zi$3GoPgW`EBB`T~qzTZ2+=POKe;NeGy1;~n$>~LkE(;)R(2NNl?$O~#)=XDM{lJ2i zxcOBRvU$yUY|CdqfW{|7we{tq5eGeL-vO3R=t!|6`co|W3fpYwHd$Z z(w0nWYQQY6y}v)rDmZZ+Sa1L2xkMNTOW$c66XmRh+VlJ@9a0(X%JsUkJi#>x6=&^! zb8yf|WS74iM;A68QOuq_Dr=z#T1a1dD}jbg$xry<;0MmP?ed;FW6~Py{$Q0kzd==C zvGysg(Xn){FLfYzbqrk4I&~a8m%SUnmQ?4YNyow2tW{MeF4B%Z%dWy@6kh9>VWkJL z2vg_ORTX%8aq5o+@MFA-HN0`mh8=|`03=;!a|jpRb*L^~58U&>=?OwSau4QPTjmt7 z>_4&I%&eU@N^2+V`fq5+^*M~i2#lwYXaaq4Z5)ch*v#}!&HJZo%k*4506ZSc{FO~q z)swF)yx1OevuP@k_I!i^=n)eQiVe2LlWPwp+)Pr6w$X@om%C2q0+~8ghI~l4vVNZc zuq#a!Gm8c+T!jr;^c)hPM))%4X?A6SPm3irzfu+bFzGarX~A1B)Olgfe(2FxnIoO4 zFjfEh{(8ikYS@V;`D0m}wDm{M^P@?eDExOFw^}cf_miO&kMDMumDRlNr|#-_^QXHE z9tQVQF02&1(Z&}>u;cmChATQhhZz0v>n-cwS9#)IBL^|kZ#)eV^zo{mdOxV`9j8ZW zlEf0bdJaBUpP3G3*UqQFw_PQCSJ|`AVlswTgFA9su;Ya8ABCm5bJ6V8~m8nl}w1}rO@-SAq$tIm%6x{Qq_It@uC1zk= zb|swQLgbC#z2IBhxotC4{%cc1>Y?TK2du>CYunT)(3_=Sf=A82Yr)>Ivc(ZdJ~7e8 z1Qw8(n~V&*ZN{vnA560iApy@Bk6{;&IFsF`FwAJJf#!}x=2FdO)w!tMd}xzpmUla( z7`d(59Wr^w3__wWOnOW8+o{jx23_P|IQXtF4>h^eR$x~UHk$63pt|`jAk2bs*b=8vAOUrM{&=UleQ3$x*&bm3#xtTOQ#P_X!xp(nEn#KGeW-DLOzc!zSw$T?WWewem+Pg zKHAIP8O?qG2|vsduq`*^UAs;j|AEFfxh8FJ>Q|v`(Kj;=e3|DBfUVpk0Fb>!N zbYcLUuRm*x)K>+V3mD~v5HPnyfB+}NqyZR34n6h^V6qO17d7xvI8f|!fMknbKZV!E zpsxd}SGtS8B4^MKN(d-BFkIiDuX=QDY~YS+A_K0 zEJLH;Vf-#&9ePd#$aciOtbXqeB^?bVg<*_bWl0)vrJJ4w*Lx0EK1Ee=OVm~mh~i8% z*^xpojplm})_{s6i;Z|PgYfyB?5UojaO(e)Dgu)W3^B}Ioyz|>C1AtZ-yaz4{u2;b z3htu}MUz4r@sa}a4Mbv1MRJSKfP+C!NlkxAM0!Z2ij6}eOT#itL-0$>D@+Tg^o(pt zjtNcGcJs%CbEt>&xh+nIBnD`FiwJe~1uh3x;lz1|ByfdgjE-h(P(gz02`Z%o?x*-3 zs-+nUX85iK443}MTFtzAf!Kb@=t@oOQ_uW5`*Wo=6>~Lp{Uz%o72=jE8)V!P$SRqc z?3TT(o-vfFdLfw(ddZCvQgz^lEx-B|nCGX2?UJPzQoJM{*jxuLjF;r{;Ol%eP zWzo`QnUG#dP>1h5uEjZ^zROe0JxYbdb$uZ}JaOKq-8ueRaC0ra$$?35s zJUK2A%@^TN24nesQjHQkZ4e&ij2^WV?KKh_fVIXFY`ZESjm8@3DhA6+Xb=BiA?_H^HBGn!6D?7^L&s@#@zsz%22gxk@@{xWoEi+h%Ee7O6WVT>)l1nWY|3 zIAhI?uN_~`+IR8V>&DyhM>Q>g79J&;mMxM}UcIlK zzt6j!;d;Qdx>3}!f+DivCrX;vdXOhj_oefWqtFb0Hq;TAVb^|Mu=aqxCPSknWUtq5? zqJHyilGW^S^XN$|&bdCT&BGfcU>YonNK|s~;FIY~DJxAnA0z}D!lD^!EN_;1Z|!&+ zu;>_~mKh|{$~v`#+8*rGizvg<8rJn3*7vNHtQd&rX{dkgH}mZC$8QPT82TVO_?05|g?J81 zSB&_CjVhgw)bsi@R`e5xc9Oh9nIMjL(hPLJ_snb7x<$4zL}IhI_c(P-^lVJHT8)^8 zj)zr@Cr^wtosBPf2KH7=uH#Qu;|JkejP3KHWxY>)+8Be)$f?jA+uN85Qk$~DAG_-q z8PS@2#z)(i7NkL(X0dGFdY{ys$O7We;8#r3dCu5n%*Y~EBVEqyxDLm^7t(Kx)oV0# zZVZ0@J+q=Uf+jnvgEy?BUanO!LAsfNs$D=i+06{eABCW~y&u8us^W4gMT` z*=@5_cB*A=0WP;em9vCaGhqCE9^(-jbeqs! z^NX7sH&L5+bX#aTC;ty*`4yrJsSSq zuR6PMxLZ{8`%kj_N7nmq+U1|C_E~gR;3v|C#&^rsn+h`ag|m=pZPw_c*JQtJK5QQB z+#IyiL3JV>$k-en`5q2#4wPB1|M`8W=e@~+yl??X~x=y<-)PtQT3iB)$8B zwD2KyILdcC(KED2lh8ZbuTZt!BUMDVdF)+|o^hzaxQ?N zxi#zfcJr)h>wM$pyqw^|p8sN?3h7YxgsO5blxE>a*Tsx{HKcae*Ygus?+dHUwYjR3 zg%5;#3)ZLKXT{^+-W7i$)^a(x6seH|$VbwiP3?jS0IhF1? zi@-yl{DU{r!)(-*)HOUvMqu?l>wcc#v5WEXOY}W6!-JXtqDJ`=&>F3>5c!X z!{*6n>t5gHN@(gy^!mx{W8Vnn(ZL7aY|_M>q&u&D+pUHNp4{OovrYmRic@%;=P z1Pm5H@Ml>6o_!MIgBUlx?&Ska)V+5C9Yxwcx^}|4MNsA3%*_ZWdgeCsPAG{~q0?%-jGqgx+le!=IGr!Zv zTl*P4viK|37(k#n0C*oW{cv2)XXo1FM~y^U?4O%EtzGw>Acz2NO>6Q1bq#7nYI~w6 z$Udq@i%x_3HM-{W-~Ro3Os!V@`SZN9CvFy18uTe~YV!G)P=-+Cx_K$Q-*z!^2}sDP z>6zKN`Guit8#})aG!hL?tbDULy(Ok9g+xYo&dMt)tEy{i>r#|WQZ5@IzP0ar^~4RT z18}FtN}OWh5t35SFtCiJ^=MB$E*?^^`wfx(X5$h z;$DGw;`&vhl`nqiM;0{Cus;<2^E(|H<4Hp=l%X@Wg;M2Zc? zcj6R~p7RG%L7b}xz6!xlGS-5eAnZ9Tb?(+hFe2N#$z(WF*Hr)9Fp8(a9goRDiX8oe zC|81TmN~Pe@SVZKJx^V0b`&a@rOySivG(E-Kk(16c9NBZ_atDk)OUMlpDm zkUH6dC4UN2RKY1qdDbdj3_zWui17lBvKOV>(&z6QjkBHTCrLA_tI7&A!4+rwv?-E% zq73C1xgq+&mE=W8{W!?D+AS$ya$Djc_Kq}2qayLIb2Z6zzE#gF%8z%4E_QATJ&w@j zax=>=?8BwOE3AY~jUp8F_is42Xaul61|kIbk$nIMNn`C;nWk>4&%;w;iU) zSF26Jb}^q%9?8S6_orSTbErbKrLCWWozuFiJPw95?&ZPg)Sd1VPM;*CUxsj;l{0i{ zpzo*Wv!0SSjLKN3k>9KwA~*fgWUbjfhn76x4p*>Zi7nnlqfK#ixo9 zeR&h8<5C~r>{a{Yq4&kwvFzQ^z+Yp*xw zy_-(Y=OvpEALV#h4#~%N^jJ8npA5#@=6Tv@A^lCdknp|h)talkiF>)HCjwxxdU122OdM7GawPDbUeB#m2J5L6FI&lnIjqf2@e z&rn?`7eL|(34^oSivn$GOR7Z-LDi58h79kxai;R9X4W4LiAs4L@vHuTLx8Wa(wPH4 z4@as6wgpNdR8-60b7TqZ2UT2{V#W{Tb^k&ubs0T-%dF?3lR-L8CN%j5&^GpXNjx=u zq^~+F5t4}65J%Q*P|Rx#WUE-O%3^hd-9|nNd9eu47xBlHzXIOk(6Gn{d=w+1Iv9(K zAiV)JVu z5ec!FmDsvO9o|CXp%QWJwD^SBwmk}pXAxc!)Rf8tV`9ZsP8n*J1go4e2wUQD^`#2r z`T!GFG%i^fc)O70fg)f-il~ZnLfE#aK82=|IM0Wj`R=$ex9#(oI1)>6Kt}G?f!MTr zon5A5MS?{S6RpB4Movh_F$7rYkl5WlcG7F1a0um$#gb$!bSEh+2uXdGS%)>_#c4(? z=0$@2_kBK2205L4PN0ky3;wTo3XUIb(ozLPP>GrfLdFs~5Y$0=>|cnbw{M1>b*c&k zl#Fi`xs^e+pR$n~BRL;xDjm@(?Bp`{lv~iE;+W@(N$QN{Gusv)zcuE{-WAAP!;1Rn z-dDhX5^_nY@h=ZesFcz^W8qN(2U8QQlCBWc_y#|cqLP^ZMcWq|oJfrXS`dQs2aQm` zw{ci1p*(?bUuXgzQ6$lvDw0W2BD(ZR7o2cz@=z;C-f`h6Xu$^V0f71DYCoIFmB}V2 zz!VQpZTVvwJPxXo@hm)Ua$=x??M+7<3Jov?Pt-P-N&mCl60~`4z|p?8N$=8rzIpA> z(Xsc1!EM}f>pqX8^Gut;WBq&!*ry7}iK;Zyex1AQ#Q2r-e2UdA>JI>L_JAYXLBd<@ zK#+3w!s{>w;9cy%h;a6yXE6rRSnVPha`xkHF^2G5?4ks44v-Kq`NLQl0C>h){HQK% z#(F!Zh%6u}*lYc=rT6gmIfuEom|{FG_6Z@lMudxQ9HH?2aIhttR5k|mw0M=|!^or% zQkbGFc@Jm|xyCiNn3LNt4jJTHM#RJ%kbkC$b&@X?P#{_08Oa<{2XIX}WU*wPTOISj zs8866*+a>W3fMbJp`?eb%z)V>whLOznx_ zcPNjrzAz-^SrsLB4m0dLR>yk;(3nMN?!<48i%B^8wHQaTXj=ea)`teLfTg3_SJ2(= zy{9T1^?(r-6noyO=63V+vLJNi_JwjNLaA z=2s3V%e;piAiF7?F2|nLQ%Ip?hh z!qwbMtzC~entZeR*dEK6$xW#=zvuG?)}z$}+; z559BXVfSqqA0P&(jo@nzCtY-xl0iJnWNsJafHNa`01M2Cm7C2Rb43)+ zgRN|V-{-cVw|xQ7>$U*s4dnw6u{|&e0>@V#>{Cv1Xv7pR9j!u&*2UC)9?cF5(t>=| zh0@r8GSGvz*nxJ~`7Nt=b-4BRn5KZU#CZD&oP1~cCNzOH8P8V>fJ+D%(|eyJLdyFb zznBAgl!LsLGjK_?_nQdyNiWZ9X9kKm5i@t>Yc4v$ZarHr2J-F}UG67-{ZiT8%OAOD zV){qgyE{K~vB8NDpH^vO^}t9*!Njr>etv^^hgPf5{cf!(pXmRFBn&IIgYCbIl}13| zppOMkKxjii1f56tluPiDO-!(NJDE*XML^0am$nY<8ZipB!Edg-i^4o=VGzk>WgbnF0f{2fN2kHVVjkTP37V9_-sRqag27jf zyb^`1IrBYhzqD386%%#&{n=dvY8|RQ0xB`mD_K&ukK`}m#MyBgZ2Y7g&iicqJM}qk z{)R%{hCFMeUxd@M$e51?q(cVHZ$QL9z@X6hm7|WN$ zEfCH-U?(_`swj_-rxeN!Yf1d3unJxQ?-kMxm+IA(=(t-p{XU2FB@L2WlpTYHft-Sh{Nb zM4C;7>fHINQ@g@_#{6x@Kb8swwv2hNzUm$s3q5_+E;>{PKhljh(mN{@!7coH_DQD| z5dVo3p1=^TMjwwH%&uc7uBTVbqCf#WAC^5$SjN80@*|>&B&a<#A2$z zvWeVkt-va;zrB;6s!kQ!yKpnUUR?Pn3~g%sk0U{Pza&7*G89OFYF&0>7U)s+2(d zF!MAp3l^pfQ7j43pbWL440Waq^DYSsrVNj!3{Rqr$SR2_qKvGfjC4NnZOZkjf(iDTz6bq_QhxIskS;OEUwF_KGI4YssF@hXXpVwl-NK9q^QDH_3KQT?5 zKH^``r#Gdgv#XTx7Dm}k+o}9!xnk)C%2WXaXKCbRy5GwDsj2gXB^M-16+f5eY*7~5 zoO+s27urzg8&G>$fX)h&P783U{n<uucg(%GIJ1$i*E>_X~op!9Qd`v^Q7T0Qmg!V0$rgx0CwU2fZx^gC*X6A{uRbdrK-v`dOwr8 z%XOW~^*p+bIqS_SYr7olwX@5uHS6tH>#eKHosY{M^r{_Pn;m${9desp@~d6?t9{%* ze|lH#v4ica&>yN)9S&U{Ib3bJ({G8{oP4L>i>E)(xjJR9IxV|8HmEw-Dqh-B^{!K+ zcp6xChjQcn+LTp!5$eK%C)M3%`GX0_~4@xzN&GPF|l)0$c$-B6SH+^p?yor>!w zD(g*#k0Vg^%T@L2tL+Q2-8DGl$6fUYmK``@4T#weZ21~O#0~(whQzCZ60ZRm-Yluy zKx5RvDAj-m+<>C(pk3_X)~jLjYal&pxIJq??2;MoX*_Q6Oj!eO?oe4pw(J$UupD-9 z^^EXuHK;r_&=|L@vqhjgP9XG!hU|%;7;dSPV>9x=i@ewOHd!pO)>>@;@luFmXmeQvJ|3@~Bt@KGGTzS)08s3VuNBX?vb@~Qh8;y}@TOKHoD zm;j|C zPDI}xHIJ*bDIRgEueAe0e4g&Lvzc^N8+DiO^+}&}fsY0hO*%a6hBVGbcI>v zqi9PMXLBICEeuenpWP0{#S*327Q)%#jm?3D!yfcxA<%3m$Kk}_@{`8J(6ITZCC58A z8(lFp9rm#GzOxM;(3uD5>q1Q@8YZE8nh1#f^ZGCe3k+}Mt^Q`|NQL7 z0psuTY#{d>X4~RZ{v6KZ8X@B9-~a4C?HVBN8gEKqR+ZIpAT@1lpV#%Ez!JUgT50{(lo+syCX6RlP z)0WZR7VG0)5yD;B{aUf!R(9+jZ~I#1;$EHXt_6ZCduyx0eydJ+t;J}sljg3!cdz4V zui$WRz~gD$Zp+&T_*stCX?m1vzBN|AHgmMM2=G)Sv^RvjwR*g@VgzeJEVuW2bd0}s z%zJdMzjf|=be+F--FtMuy>&x)_8`3XV0iZ8z4wxO_R+lev3U0Ly!VTE4oJTbD0vQQ zz7HCD4q3hrIeHFzyblL>jzqkVBzTUdzmKL*#pHMNstQ7uypJJxj)(jU=ZX*b4p0Nf z{fl$mp@{xBoNMd>4z5>JzQezAt`*Mz%DFzUyx^4lC+C{K^7J2^Yn;!&a;}U}QU6QM zHPyBG|4+^}9AEPP%DD!>fBEm6t9-uHe{rs3jphI3TvzMO=Ktng_s24&o2xe4UGM)# z&Q+$h?r=Qizj3Z@4X5)ZO8>>VX34fUU9SI=bL|O0ApHI}=Q`EVdVf5VFZ~bBwcc`j zs&adO;}GS3BVD3;{xaamf$@9O`%4+SXTcB zi5)Xi@INT@e53Fx16Vj zK#2mfm+{}MZ*77S=<>t7-Xfn<4D98<^<$viW`2GaO+&JyUpX+Ls%SG|E5 z`LrhmvSXzT%8N;qrpffU zWZarz5-hgwCrzaKQYu8^x(Ln^u7G7|N=qBJ1goYeXq&Xjoqri#!F~ZgI_zYF)`=mI z;fND-1KU41MvMifCf_wEkDLFk_WEc&n)*SwQf%e=%c#>I5}Pj&;vWj-jy_%bLt_7} zP!aQ<^yaJW7A8H3`SupM9cdj(@hQ)s+r7yIvy|_h+7IU>S->G1qgP+Bh1P5b4Y79H zTPErm%ADPHkMonbbb;?AJx-5ow5+o}{vcRg!sq5ba5OVJA!S_OB!7@Eo0;Gi{6#D9 zi|6fc_H8&QwH&Yj+M40O5OP$tzTn46C?S6s0H-LVDHzxu62cC)NFchR8nI317gPdU zIQix!7_x9*zZhW^mtH6+XjCvY5k(ih1YAzTY$$9QpkFkUg&;h%B(yw=2TXX5kRMb$ zDjB@Gm>seQH2I5A%siLUrjvf!jc=U4vD?nCAMQ$EnUV%vn*v%EsJ_XB@2f?*w%kpXD z+Xcb*oGkaNwTJRgz$ZT3eW9eHbM!NhqUA4OlE0^vdLqTP_CO(RxR-cYLQ&o11|!o{ zcr(n+e+SGp1yuNjU0J|~;-;Zq`zU0+_*GL|@%jfmNzo>8P>Mm`4Zqf^)uzCFa51}~ zX?fCjeTv^uPFhmvh@)Z-DYqD8+v>Q+hQI5@gMoC%sL9p&;9>TGVW796~ZD+8?ucqD--)Rytu0#YJpxgdUET_4n13vpxdHIx!>f z^iGdua23yvdqBZAn8hnT`yv083bTgCAtjjtx2IjfyXOyyPY`-41SXovSiq9+u_JU+ zeth~U78Vfp%1^?hta2qd_T0u&j+e!eOrbpL4%y#|?=4jW3+E36zxBTyGnVF9i3J1K@ zD~`-xlvYp^(!Ci~%9*Y=*IrU%=~GfU-6%D-9`lyysR!8H68bk^nz=*0pa*+h3(RG# z@}|JSpM}s?hFM?g6kkRJF=KiMc{4jk0P3-v(GI@+Ixrj!H#Nm2YrCYEH8ey2??`bD`2@g`VgA?K7+=H!A{faLVcfgeP?-V}?9mZ1?z_0y$O1A7$x$hdyZDre z0&pp1wTB|k$K{rc2VV8$5P8}YmZf~ljS5jU`PCaTJ*SbFnJ1K1|LGFj!^JidA54s72+s&*Z=)Od}>hCAZ3m0Kte ze5eC<-b7V<9!qb(t^vEAC#$`ICJE?~g5uyf5Q#YwUEz+1qYKL)CxD|AQqP$o&JWjKoeDpUZ9^0%^hC%L})ZLRi+ zErVZ%1D96=yx>TjarXB$I!P!Apn1VR4>fB+soJ}8eltf4Ax;h@a)+DK8GH>I(AaZ= z?;|a;5XAW@VVj~ZC(1sX#2%MPu{Zlke?#=3iR=^dfQGXz+&RB|(icjM^MFB?=IESF zUh?o*5w7wRj!*Au{Vmo`ns?FLQOS*Wq`SBR=3(26f`W3AJEz0wXvkqfUqHwZbj!-p zhkRgq^0TSUIw)#x#{urmNNCx}XdkKro>U`wB3?Y67=heC?ZQ6wTN%oJfA$ed5)(ep zhuxo?Ev698=wJ2l>=H-iRh96k1otDD1p)bR{Bx)VaW*fDq?ig0c_i3MIE^X6RQC2O z(Xza;a*(5DzL;qOY4Y*@gou(NN&ADV>H-y$ZVPa1vN-*A)QE@2iA9%G_$X|@F{X3G zJR8W+$m`+`P!SpiFx5q5jPvD#UIdGfe5dS)5c&AT{1qbeocmQ958yBnb9GXLg%Z}t zu!~4}aYUtK2;|Ep$=(eYAR0$jOT_8da^j)LVxRr$alT38e5#eAfd|zunT!Y2a?F%g zGOu{kT(umm?O3;-6qR4KT{NYmHD3+URC8goYKjza2l8BZy%+hDAil=S+jc`@BaSkG z_84n=*Ccvs7-#rtt09p+YYGR_1jYNhXZ9kP*uZ=uV~3gbZbSz5KB(`;c+4`u$1rMX z?oOMF6Bod#_7Y292x)FUkOSRi^= z7lBiu_r8hjdQyIO*(Ck-TXHD^B-mP}^AS5#@Sx`3r9K`<*Du_4_>^Q4ILSjA}Ri zHQNG3=i^3mmltq{rK0xK!Yan=`J`D|t);E_`eO|D`1i*9y*^zqK{zz6qy zo((W&XFEf>@x6QAh6NApfxeUWMvf$6ZzAu7e~R|Qn%Lo&DC$K{iU=S;%xw~+>_ayR z3?!Gy<(KCKhb_Aa8rRE(4TR~(9jbw5ZOBCqBJTh4T7$#)o{JiWHb9Kag#C>-4_!5G z;43%UCt1(DPaJ69@MUkq!%OmD1w#huhlC<{Z}UFGfenF}(rQtRdGm3PV~5z0!5Px>$6V%5J^Svs-d%foqrK64r;iAI@@fMJT#EyL9o`Z?zDgDtdiN0WUeTQ0auRb6x7x1Q~0-`KY;~Qv6a!5ovv?>^Fz{^!A6Z zoL2(+^X5Z8>y#p|JWa9psix8M>a)rg!1!)Ydp+)yqCVb>$$(Tdr05Mf zBD|EsJ|#0jS2elFjl`m)@cm(m%0XD2C2%Zd3TO_XhJ1o|N(nZ~xj2oQ0`f#6DW$WX z#JW#@^j@a`L+bgD*@mLGcB=39#q*g{)y4cj3FQOAyK>PnJ;nH_f$?PvXH1Qy(;n2y zsT7N))YWAwUkFs{*V-!`emd*?rateZUK);TtkC13(MT;?YP_qcw8_NL+8?WHo%>ao9qo0|Mnbddq|j7HI8Y(PW;rnK z@T8gb+p0;JM#DU=sewZDQshT@|9qWuKy_uVr6Ens>Ri&7rp`((hjLAao2KTqXF9uD z;*DL!{Fb3aXom!Zs6*ytoGGs=<~$m05*1(pXt|Q;XMK6Zr2{th*6%91J{leJBsO4>)s>m_niy+$8TFQaM~Sl)RvwwPDX!}+#M+F zyk2Ai$UxFaF?bi-KCC5(V3r9nWS+(TnRUj{ZK6H2=G*~t9f)ux8F4J##lay>rpSJx zeH@?MA@(cqXvYb0f`G;0%5|pLVxj|L@!S!yVA=TjHY0M5)JnNvhhN*=e~d%wRjNY{ zN#||G^e#Bny3md(KyDM}BBb`$pI)Ee7h#|#u6Q|@K;T9+ZmKlHXKKM<$R z2ihrzg2zl4d#x@+%qfqh-OPcwbv8Q?)X+jecdDd<+mKA4UnJgcZmIA(mtEyl!Xs

Hb-6+V!2@W;(y3}G~^X7mQrG5XpO#Q%CQ!8|?kCMMKlmtPG znRH>+6Y1Ay(}bBc8GtY4f&5&K3)63GJ*t(Cu6r&v;%Op?wqoeGE!Ir3z;^))x#OV4 zI)t&eS!#P;YAU0~R#Wq2Fj*=`lIZjKPV z$;^V^GX&IM-;&rm03+N#**f(PPu(R8c+T;3y+|tP9H(5rBk~hEPR{#2)P;i6+XVR3Z8hYpPxr0;%T{{%rCE)xLx>^_hS z1iVf-wqKV``Il=@AkR#*AEj=((9vU?V&T@8ymT@3g70FWT72O4uK9sss34++?T>qo zzTZJLJ|D+~aKD?~oQWAhQi8&-D#X5~kcVOUNsS8yW1!#v!`@p)#hI?#+E7U0P`CvMZo%E%Lkg1MNeJ!`AV9Dn1r+WCcXtc! z?iPYWaCZnA!l`7f-rc=d_g-@L+TT87e1m@se$*K6TRitO=RL1khloZe&Wj?!*$&|v z_ls5t^y%}Dg$90z_tUHkOumAoUIixf1Sjh(kKIgS(0ShUkIj>j7VKp4=y?QMeFSbGYQv`p|1khwJsKOB66c-e zzJj8sR|H-Gatj!RFd;IHDhkOIHOtAgviyblSQKwglp+q`R{@O*HsIU+Zon7f`}+ev z3}x_piTA$}(Ef@t@Ft@nkctlffie(8O8hgF0hIJ33}s*oYWy3NK|RLkf1nJ^aONej zmzP#6YuEk|W#B?iczX|*Vuu1l8A!U|;?j#iHv!UySky|6$K{!_c(`m+Cwb$~)o^Xh*{yly#% z;FiCj434>@{QrV7FjTXkR8Cb9+*lldNxX#b-c^}OdWr`h;=QZh5lDIRKP>Uy|9>m- z+VwDPCI1FxaIv$M=7r3>oem+i*v^2`7j9?%f->02j?u8#$%!{9+{sOL+TF=Z_h;VC z&yKU$Ey&L;+$}8rE6QN6xO&QBucU6XaIduKVt22s4Vh)Xyo=CszoL)6XuooZZ*RY9 zOpfKCdQ!vkVC>_JNzp;gg45nX?Xv%m0YA?25H{fNshaH~x7##|H=)~NpN?1_HD7J+ zDL3C->>ahj0a=gR5Q(gg+l@|wk2~%!Rv{f2(nH5x_-j-ZXjso#0S21qxZa&@eCa2s z)Zuzk{q%WW_5;&w<=q1)o$fLKR-zTI7VXc84#QCLa_&)y!HCbJbT|jOqhd)>aKmY& zqVhe6ugK3Qn+-TBKd0(H=B>%OL*wm!W&7x2#%O-<&8%5ob4N3QLJ2zq(n4(FoFk>k zWtYtsOBK<>-`wH_;P~N00iwU&;y)LZTK4`8ujE@!&_At`b0d2F_wh=)DPXtwH=`4U zNYgV&pXV1|U@b1M=zLk-*i@g}+TByw**`iKIXpT4%5`>eeM5V7`~BA2`5%DpUqY0%URoF)$X7A)9bq>YuQ9xDgv*EFJ|= zTW42DrcM(~p@Q}fjgGaBj8A^9`ZQHb3Z0)_8t8bxSiAB?e9d)RcEfFd^N0@N_{;8D zUBAlVh0M+E)T!h#(hh&fGO@20&ILH%FQYYpfqaTe3p@k&goN zC|RkW-baY^SJK7Auo(`-tYU2r!B{*}R&CLFsuF~rrMrsHU@RUUD6{N9wK-}IIAyh+ z!>90>0yy~PL6v^HhOD-Sr|75W{+~fFdk?h&>5H{K8cobaDEpLP1#%t7jiLJu-1mHr6`WkZ6W+*Q!8G&m(;1@t;^>!%gea8 z|FlO*DeLw4o{W?gy6j16MlBD)zA=@AgEzyGghQ=+1Mg32ueIVyjELjy4X=y#;aqje3S~Rhq6rmhaDfkIq~I`kkM9YFk0dqoYL9x%7wC(rupG%|91ZO|+V>V6?P+F|GdzvVF#R~x>BB|9sl)k<>yy^BnDbeq?B>Y1{-6s`GxmE5;w%_ z;ev^!LJ`WjFNGqc6G_`v)5Bwp64X7l=@JY!ml{9l@7m{i}EHqXb+HyuTP!6llETyo4An+Tq|o& z6VLc*k!&kAPZbHOOHb=}kB#@DZC*v&l^n$+$IK|}lKHCD;-5HZeXy%hkdN$HUxhBq zEDca0HcC{*!&_<)p9e_K{Dyw)sWg6Ce&h(ro`Y4J9<)%tJH*}0G1qc2v}-2oe!8Ab-biLK zObaMXy-^nYsQ$(ACR9X-a5Sub4>zF@AD%6vK#L!erI-svmO{g&*3*-w#{P#Kpvoso z0>|UC!iL#r9~UNU!V@Joh{_mhL2;`eEn&jeDPs=*%$~HDjibFPe>rhxGKVkV zVb%M`O45qTih>Lgt?mhy#snzOi}7U$0Gpo>yF#`^vR*SgdctmpE+&a)vTcD;A$n-Z zQ#W$@)q~IO;Zl3_wz*k3#;lL`t-Q$7#qy1x*dygW!&f4;!X;IYWLwiz%fc0Ul3kEe z@=U7bGzidaPn~r)$j>vJ`#4xFC#QShdE$0dqycd}GYQkPHFwQIEkfVy0f&P6aztKC zR9^|?jUw7(X057|$;PiJ@6!I8nh|OPBT%o zbkXo%IY2;i-3wd+EM&1`GTh>2K90I}OBW8vC#n_Hw2TfyHkd*s;*aTY)j4&{0d0mU zR1PqOO0^j9?%-8jH?~mU~{`3)W@od?^W64x|H{0LV^GFdMXc_TsZ+G zaBxUPunnYjtyr_d+{N8LI^F;2$@TA&_8_2)zmoR6vVPg`5G{dSPjkp&r+ao&Ye#2% zdslB?SWo}ZFl2CKeBv$9r>W^lnuouB{vV`3c-Rjd62dakZ?4^cJGIRITc;KR*whmD zV`_;_6iWZIspXf=DmsG9j-KAOuD-#c=5E+TUN$=MD?v~d;mc2g;N}W!B8L$Kt9Oo0 z&xDE2FRz|oTz|V$xcv@5JBNo&YDxP(v!vLEG zxyYqrx+kn&Dy;Y|mvM0%yY6st0`u00y11`DvupLIiQI%t)e=Wu%k0uu|FO67SAyWb zXS3=bzjiZuj#7McCXtXzsyskdpW`F{TviRX81xpdg7idn{M@}O;^ zAhtQ#tS}pR^K5Mur5%E}mpZb!i9NG_a_#2oH7~Bl;<79eSO_Y@Y&cYsE^8TpvnuDn zP_{zrp>UI`9EfYPs-h7=KLAch=~_Rlnr2QqsQ#O4_fKDK@38T>hvlg0eCp@5y9nm1 z{r=jG!cctNfx*9j+=(O4`X&xt<(p(T*Ke-}f1q#u0|xfLqHp~jRHrp*=`W~GW%&9g zp2p@jUhH;Npkyl-sw#*VG}Oi2IrgzfcVglhhS6NijPigfk_dtkd=F^pGt)9h;{r_| z^DfsJPtV!h&{w{*?U`F9ELKMNV}xRCfT0Kt3PMMKH@aGRZ~#ER-J6eGOc*?xgz^@L zLSF==4?uOICBOGX#63lX_%NPxlAjD^1$r>cL#~`oeb8YmG3?Ac-Brggav*fkowo=IOGGd|@L_T&+Era! zQdU(``^$=WkCWFnDPe!-V6ocJCH7G+uTp#qox%q7sY){%(6+T3%7=v5xdhfWP=oWl zKH5^OnT1i#c@;jb3~6D0mXyxyVs7OkXkWN7bmVN&!N8mZ_xLSspi3+sxg%CW7ssL^ zq1X3g%SOthC!0(_M*9q|fHDWYY);&cwsRo?b%r;iub1#AHt#hapEz!j?51~WyJC;Y zrit7S)B4;SzTy@?c9kj5=N%k<2-?KCQ5XQ$R`&7;mJ<=$p^n}O<#2L(x#f##l5B-L zIY%BS2oVx4ag#aa$eC@Ji2HV49+@6(TE=lpaPVafOUG-)yok_se?gZs496!gl)i|D zy{pr|FEf3o8LpJlV3eZ$Ar?pNC6?eNuIQxDtQqRV*hC-C>J>pBZL5Rd)(j*Vg{14>>urFGq5pq|`E|*m>67DN%yKdt5RL^?IoiX~66v6u^eY78} zPWu^)9@-DE)~X;QgS}V*Lu}Jl4is={5@*9AmnyQsqQ!bXLt;eMJUGHxa6BL)9FOyd z%Hj`#$J8~gKVm744oOcwoh*cWg6!)kjA&9uy`AE>6YUt#AN9Jx3R1I{&ZXTSqU#!J z3I%l0ZI`Ib^F~-BAiFm{{)9C%yj(qPj4gFG;`)A^Po4t?X|(_%C`e5TJ{`V7wcqI-=>aS|4>41Ey2!6I@~fiIYCD0_3&E9KJ>x}OD4g4wU2 zLZz-13d>_tkdVR*Y~0{tNow|fMX;G+_ytuq+QoFgddEYrIL;jAH}C_`8=^2vk|m44 zcJSxTr*WjoQD5!tT0rOa4@ZeH%$lgnEMLpLZ0^e8GUb(h&}E-6Cey{8rTCZ?!!Qx0 zLxZ<=7TGE{C~0}$1TT?f#2(O*w5hTM8hb~hR(O{3N}mC39d*?8&?}X`El>24sf!-6 z{aQ9ilIz-B)dTlDU7I7H&?<1uKlkB#EWccm==3q@j|H(5BXWfVHG0+ibCmV{e1#(L zc;wX6jPlidMH-!n*ajYf8`lDrTfvcpx$02f<{t}UPuZ%ge=dl9IR2C|SDl}P3(q~X zl46SStYDU<{b|PTprL<%K1s2eLIv8B68q1&5lrleDp?vAhpu|2K9!l{}d2{Nm< zmb#J5n!&CgOG!fasJ4SN{Zc-HbPtM#ub5|icX@;Tj0LtxhfqEAIj%$2pxkKNhb-GpHaV$;o+1O@ITUftiOs#?FlHSpqW_yQkZ z-loo5PDg^Hg;WHt#dv#6$6EwJRwfPzI3D!tCZ5N9h1pJlY#Lz7v5J#))%CvI4j(RL zRU<#1;>7$PydF${?N(KN7?YLU&h%8!TchRlzOr~@=S}@!(Jj}H>%nTwb*O>Z*{6wO zldOvRVKF!Ez0;p8gWQ7w8tk_=nDlYeZ%32B(jSN&M1{ATz0PXn7*K=;v}=vc#4 z-N?sM=i$AVw>tQ>-#?aDknO)FyPfG{p&GYb7uY^g65sG`4y+2dAuT+WgkVFih z#H6`n)$yD&^c@_d9)0^FoPYgn7w$}qzkNigmBB>DWuqrT1eKLvUd&$bBrvbRPYwHg zqPLEUT95STDV)I6tyqjCam#o@IeNF{!W9on=IKX&p)p^w15uppo##}yku2D^qITcT zjUxz7!_kg4D8w zAoyD!L-%T-@$rF)F!y>U5awQ|69;9&+-t3%T$pTuR6G*|DZL6RCl0O@2+kk# zdI;nK$sl;gBgf=BKaLJwpYm#qM@;HMc%h&i=VDK22jw{i^fI8wK<&fL;Kvn^KNG(t zhzK1u1J2DLbk~JGw+mfnMEN2R_G&CN)HH0@1--W}Ono^-U<__!J4`tublWWCJU;y9 zO75U8{P9@$cPNZV;%4->#`BBc==T3MDUhg7OV485(FucvR2 zA7iK#o)><6^z)SQ$n=~HruzW=tmFXQngou(mh28S{Kf_+d>1t~=L!?)_$2iO(~Ze@ zx~l{n0o%|fDgY`UD$$B>2P>uklTMz`T!S+;N0crgV+Hhn@@hj`lGCpzghB0IBqY)v zSpt_BFVPIuS4Y{EIi`fR@=UrPB>r?pcZ<_3#*CsiWt#+%;Xw)q4{5Aq7332$y9a=z z7r$qc3V2D>a?QY6?FMZ7PJa~KRZ>y)eP`%GS=Z)%B9S8CuFfRK44-yCf* zP!w!D;F<|)ZJwAAkq}|Nz(Bd7033SP`=1bxy6sPaJU)h`YrbV)%=gLUnA%su0Ro!7 zgt;&jCWRosJPbUQpFpw`(%v!TIR3=`hEG(^6Q6;^_o)zM(LYjbCUiOe-iP6Q8UcZh zmtV?_TIJ&>!AG3nSn*O!T0h_1d|)DFZRj=-NS}L!=*K+gZ4f*KMKNS$AAx0yitY5_ zVGMl*rBE?sCR!RI3>ZT{7(ehAhJIAJA&jA4-xH(pn{3fFG(0iL!at;v6JWU{)h2g@GVa^1@fPaXH{mkocJ3T^W5aQI`h1A8(-u6 zY#%<0UyNtcoxIW>3-sd3pT_g9gk@<{X{BXZoBye0d6xy3RYjkez18oH=STP0uTg2R zIp;ki<|My#@BwP+j^7>pvRV4KH(=}Z{vXD(NU#Ik(557+82i$d68c36p)p73ST?|kszqpvU z>wR>&;5co4x#+T0a=GMwd2so~3&?i03?Z_)T7fc@Uaf}kA6~6R$g^Fq$80ez2e94TZ)bn7x!K9jE4|q*u06cjEAM5y-LIauxjm@cD!n~yx;(r+Y6G%= zJMJQ~{dUsF0NX4b;)iXPj>)s%olk21s@oskT`mW(f4^G)VEg@gJFo2f&3^6C_uJ!M z_WN(=)3*0_S3fsPFOOg~4jAchK!r8{>9H4zN;-h3uMJrU{MOEuikzBJk{?{|V^v3m z{24}^&&=?p;6XrJ#jnC8^sWf9JG(N=!!npkk(u@U>90CMClN$yZWO0ZHnek9X zW-G4fbQOQ>oXdW&fkDi)G#wgNzr13iI%4U1Cp?w})^ktxxUwN*Lal-}#49d%x8)6N z_iJ(eUEC;Qg-y=LH(a=S%aP-5o7}f!{l+yXAyY#p#LR7vUt9b~a^ow(a`H?yEQa^1 z&M5#V0M`uVFOWlcea6MwH*>1b!#8XjlyVX(Ngr2ofOb|1*aNlUa0vqJC*a zp^HgL&k+8N;QHf({8ez(2mMBH{q64w{^a-nH-Aq96ROn#oWCAtue|@L5{4xo`Y#IA z9&-LaDq$edpHW#W`JX`v^|+)|0Z@tPI=Yi%96b5zf{7slxeiL zKI()C)%~nJkwXK2QVEkyvAMGwy1nJ&`3c9~m!lXv(|Fsh4e@j%he$D z*V<+ZH9h$IY<}AB*WLu+RN&%Zhf>tPdMN$()pW^Y+K^WMsOb`{F|TGo`puE`BV*?X z(Es)5F#AU`cE3A2|Iv>r{9^|Kl=x@gv0TJoC<|5f?XWhnCQu>v^Fiw2 zUYHjM)EJ$pO?985ndd>ZoODB{TwCBFHD00XXZk|D%XQ2+zi4%uJL|nqzrlSex4*lA z!;%_-N5zY=$_*PAX^51p>N-=B|PqaplI zOn)v>;J=BPL~|n>alJy(7|pkyf#;2fIvB%qsOEbO(NQBoPK(>u&iNvqE55K(p6|0U zLHSDXy9-(;m0U4aExD|a|6alyyJ(1XuoQ)MG)}>SAxDe_ZWY5qi}vl{F3MYEBo^vq z*#|Fh!|N9aR`~SiSoU+V0#&-G5GjK|Ul#|6X~X(W7+bhcNPr?iUR6{Ock11+2~VJu z$)zp%))YS$qE{XpV0Vx%%vbb{1UK5U`e_O3u!il|JxS91axWNV;l*67C*_kRqOYfa zX%iC-M?YK}3|M~+YZD9EU4ylW5ng_W`$<^{%X|)_ENnyX>Cj-51q@&i93|+Uq$B-d z@@Kmx-|jl@*M6c3dA*{bFuDQ%zBdZN0nx?qnK&eg#|B88g|MDVIeoC2B(vCRGA{d4 zz$ydX1u^hRAx1qUN{(+CCTccQc169&un6L`HSJuGm4;F!Cf~)ULp>QxT!f(s+ga5jMvNFC zd}up43f|i0sGfz(jJfHhndWGATtvGCJvj1GdGhW-yG6(pR2#UZ9TqH%A5+_QGionf zl}jbBnXR&)!wK(K5EBOPR~8fN`c@6e%>%1Bz$^#%ADso5O*QLZ@0HNc0d5ZJ5+Fq# z*%(^NhYibc6$dDdl`QNuBuWgcO;?_{gjf^Jd*&^0=%;I#m&jJftam|f#~n0>5yw@F zT&ySLBsWQ2vCR6(K0SDby0+N6POO#0*wP}W{TDKN((R13LqJp3MplO!p)QfLVW`$V zXhh^aJa?Zcums6e=!o@vd|*?*N?d{eQTM12WdD59=~Nh{_t`w!+bO;2igRMh*NY#^ zb<+edX0_ZyFXmq7-`V0AREfTswVxLCpU<5waaQL&I;hq2xPRml?=3HSwelP<h-2UcmMTP;5^m!{R0p!9{=|3c&N~3VVUYXqpZAbq3QB7 zs;~QOy{dqWx<);>qv#{c&W*PDMb{Hw%J8P;NEpoNRExO`cr>=g;PJIb*oNss>ecAo zTjO44c$X9|lj76iKV~4HbRktUD*cNvex2fX0S4Z!;?!H^Tf)ku>g&!IA zN?=)bWb~D|W-Rx{3DM7NVrXyJDfM}fMTM-J+m8ADi35LnDzG2912j?wp-$VZ9e#FrEQ< zyh8VD^W`8d^5OBQZDeK*H7!zNFJhow_kF>e3e1<^sHuYJWyuAwk#m?O$)13{IKm=B zOeWjOB01#AvnvAWQQ8PDiRA^QavV=qy^+N-y(v@EBRI^mF%b*A$rQPwfIjO$W*oS+ z^z<-iRzv*SUVm}6$S}e&FM4G#{G;6TC}VRYzy%6H%A@7bx0CIU0-WKW)Az<<4A5aH zIm2Pvnd3U)XE7!@^{DeL!+{vraD(5;`d+Cc61N!$Rt9#X>r*Adqrf>%z3DcyDNgP; z%R#Q2jX*RzL(pzo&6ZnM#Z*y9`NETnU8Y5EW&lFE#@G^=d=-)ADI7eegc>4=cRZa}{0evrhEVl-hTI0g5QUSj(ry^bu2e0UYy*oLjegiwh+w=j zmm36q&S2YsOJ3-GDvvYlE{7{)NY^VwrXvK(wwEHt$(DExmQUe1N1(zqr})SZrIY(4 zKx76d-N`il;k#-13J<-sgB0rP!t_#t$MEusIJ4OLv0NapTp6fRN%rsaOrIhq3YQbv{aA#!inKu@)v?@Wmh$3>DqIbxwjPf3>hnPpA^ufT(=z*S z*p&ebhs89}9!$RR%0osoRWvKLnp@m>zUrDJ{70GC}@qnZlY(pD~?v+6j5 zpD8JJ2)Nb-trG;de|(5`S1#jL=osCcjm(htM(qQkett7~01`FFIC>wC+=>yqFF=IB zPz>(DUe_UNHTtNVd2L}yHR2PdJ^f4@qL;B zHdj}`No6(_%G=M zLYT^|UP`TX18$K#DB&0P$=Fzrkrt=M!^GvtbH??Rv^ZyL@QwAdW8>vm{9WH71P!7V)oa5k+VxC(S3(4GWd=B*Nb!S3Pq zObc~z7o)g$ahGB{y@+U9-xCxZ_>^@T`yp?GjS_pv@GvFe6S**7>$|u1r4HH0OHoI>hEHjy%}hY>Q*QZEs*>W#R>cj5NjnjH%Bp`OGSo1BIKTUk#eG5@0y9XP% z^(LjQ#iO+Bj>WF^_W70ARUXy%6Zwp9o8fvlv%cTIa%bEfY68BkjmqC-nSQ@u2Yf#h zhuocsqkO-7FOK7N4=|%gf;s@9UZ`~r!Z^n0E@W6 zD{m@dpDk)1S}h-X7aztx)}$z(;Bq)z1twNvUoHV(9xdOtV;??Zv@?7JmMg@(XkQJg z2clPqy@(J~2`{2}1zIYI+!c!OHoNo;BooI^T>(Y74?*F|tAGzs^vqAc4IeBZ@>0Ox z+1`)74PtbKN-gDQj)IDb!>xVg^@h=py3%j&lfO}$zl*T~Q#!;YAIan@Aj}J1X~xg7 z8^N0qsu}@x5|9tfhbkRIRjH_>Py*Yf5S`2dpR6HxDj+4G5Msx;T%84m(;>v>2R_k< z%HhxwjRgkfYl#)bBeFpGOJ@WK<4yd)9Ci~yWi!Dp41nzI;0_m-V*8*@MnqRenqDYU z-Xx+jUdW&VLgOZ?=vBxVaVSe0Vt+hfnlWT?8<7wvtPmQ?eH@yj5T>mk#+hy=4~9Ry z3Oj^`y+{u`jSt}B2>2QwZXkhrlpk)k9v)C03Ks|j2u3_LictTpQ9%kP0cQ^o`OSF# z`?6n=#Ph$D{f1Xj{wKs#@KgIAWxsmoQEo78x@ziUK{)p(*0S*Ji*G7f3_a#90=DnW zJ>*c`)x7=1Lc+p1f^lf0S%Ya}sS_fSogAalV^TzuvU2l8a&!yxOJHEpr{xuS#k@8A zA9$<4?;EP~-bWnWw z#C$wYF^wE~Is8?AF94gydlmG@Yf@ z#lg$qr5Fzcbiz}9*_C$a2E_||feORA&?Avlm21C2SVGSBZ4s>OmnKp?xQARvW@kJf zGgWms@yM9wh3=q9qifBY37tsb=-s$XmCz{Iy}hIBOrM$obs%BrFM zJz*2=b~<^ZDC0xsi^paeGz2g6UsI5TfmZynoET-?sp5zOAJ8V|Q)BxD&U6H@sLmLK z@T9I7gbFZ6{jTiS)yeQj*)J_W5$jI2k}x?;7JqgTZ}3+L8XB$f2U$ zW>h3s?9e~8?P2PSwZDOyz{-B>jp1D{0s(2B$U=a0&g?gN>FXd@>dY6+q@5Wd{#*Iz z{&IyI*>OFA=DG3gNScWA7x9axqxD|(^=tSVO@4O%&khzZYxOhl1b-+vr?5nSRhXH%qCyF?1Ksp|+; zh9%@qjcgifN(M`<4qFCwnocw2Y@0I|gO6I&PGJeTPXwS0$BqdSahEh+4S=i_qHevT8ekA```s^TT3lT>n#zshR7dZ0F?WBU1{pWgc) z3Aqo}7qEm}$;F)25t-B97!}wjPG3&GVO|OOVE zHl9;0;xe|QQS|zhrc)vF_qCIom%D+;RDpu(=^*4{3Ce|+Xpj3k2pf+5s?mqhiWpdk z;|%@Lx?aP-Rgj<%vgH-S(8sY>kZHBfg9@r(QS4IE)uWaOQ^#$ns6{`h)BVa z#&uhD!3%P36XZ3xblnp@iwsGDTf@(Wwsz}#hgvTH$PMFVcqphb99UCGO~>Gb?G;Gg z4JDHq+VwCXPS|$~(QaG_Aj>M*X#1Yipi|%|igtBHdg)Hn(Y^D@;)ntzt<&~16ArMs zRz_8z(9wT%Qu^dx`LM(>SVNT`2qx0`{qdsM5p8J-d11?IKj#Vn@BLi;!I zbJ!euXT^75gkZBa<-y8-^u@vhfr#GSrF^x=dsg+ojiVX-ze7dM3ncjMk~Amm{S<;H z`E|tE;?pqO3)*KHb+rECQhMZmbpO{8=kJ+U8d!73b?&NV^G8=rODeOs_kWbKkn;E~ zdaLXciTh{6f+ELHYV6jR2#3b;CxSei-Fzc;rvat3QMQ+Tnnsu})a$Rz8T`OuIFPJZl{*Z}Q0^}DSLKbSN&F$_b#5;d zrRbTZ{;-fV!Zhj2)&Ufj8Y@inM}3K+8BW%|A0Lf%#8!Omoc{>{7_2S(()t$&K>121 z^e+&AwcZFC7zE(o%XyW+`i2Sr>;EHm*z>{1?xA2M?+m@Jt23Tll%~C|74uNZv9K4pmtMJ=k{% zSZv*7|662;V*CgH86=C7T_-s;Dp#I2cs49ow;mXYKxxOM}-kYDoCCgfxi7pP;OugBQ zWVix`Hum!DJzDQ6jqVthWrk+ProB*3arFOg`UKZoGLVEB_yYBoxR)A(jg4V3i181( zz9b9Ya`XkT-5VQ-Q8ZYNX#=)<^Cw*22kP=`LC)Hj)L04Yne>VO5cDBY4k1Q^dB-r; z#MRs|L8pPqAjLIt1%SRemPU~-!la%bVaM>aQGUCH3tp=)bS#6A4}{k)h>5uMtv-#i(5{s6emt$)-ap z`DjCC4cV6xmMx7*==N=Xg$XKaJd0yVtLbQ-qb*9Km$Pe6w+Q+*eewspH^nin)VKIW z$ibpMl>h;1Sa$5#im>`jBKPQ+TLl z$U>_p!emM$$DiRDe4W&@x~rCi=Q_}P1H=a}%o4?WI3b#s4cE+>X)Zg>TYd`VTj-um z;a{>H9j;ySnD&65ke3J7d~q0jarq^vH|1(An1YQ~=%tA6b*f6L`&+EWUl z`w4H&CBbd0S1+&NQ$M6S?-8|qzub>0Q#APUcCU30>k6qb`VkMND8DBU zvOQ&qH}a71z5DqD_fhiMq~VuuXCT`0yR)S<567#|npC3aUo0hBF1FGbrA`l$-izKW zJhAn>9<2(0|9vs9RqX0?@YO{!x5UlK6LZl$iEpR;EpT(;-k@>{gl0`~gw3P(yNzkz z*|sIn`RqLJiqdaK47l&j%e}vyr30m1+VD)k5H`J6Zz_Rp_!Bs;gG?*mUJA71RYv#< zXX|5=Kszt*iYa;5W&k?~V%cUMk8DYd{VF!;s*%cM>jTTktA>etTS1f_P z51n!U0GLoBHbpDb!(e#;+;bXNVr0bNMWXnkv@)($#hCs{TfbZ(R~%)&af3e2V451& zhpNeJURiU1R;4~_D~1UjiFuVCArv_$djnK|?Tr{8g*YsdCxvnCgQQRfCl^goK-K57 zH%pxca)*~_jeN58s7)}PcpjL2(S(2>;|6qDLP`xNtLkb?o$h1+UG6lYoZm-hQZq6V z)z^+~Y3m0dLCwBPc&mLcaLm;7h1b~cZ$-nO6hzl&RJLBD=TS(C)Az$G!j>U;)QoJIYJk!DIFZr6&j<<`nqJfD1#NS#^AGa$W zmHB&xsrdX70)IkKU$D1a74tmYP-adA$DEzv>sO%aTEtpE#Wzh&A!gPV zmhiS*+Jn;%!gEi%qXdZp_1Zs=@I|2Eo$Cs+ELe@|KH1rc9o`+jPx#Qw>cjPZd5Bh_ zDs{g2nSgv~M__4HfZE{JV|iD^XtojG;bg$syK6H))$p^e5d-ihU7)|`8I8)Uf! zcItHR>$w1^I)X7oblq9=52~xYxdSIiyY&cbeV?%I6BLc&_qVhGaEL9Ix`px1_~ht| zAJ|!EVdUP0R!trbYk7VVLK;REU_@McDMTjHolsuwvHN0)CcJS9pTW8HG0xhcr~cB| z%j^nnfisZm(}()i+vxOYc%LrL(xU0Sk=CD0M@Z)kWj(%-y1x$wY;~1Zrk9p>%h#Ws zCh|opweC5xKXyFU(4iPD+AZi^{35Z}xXRyi#S>HN_~r?I?aTM|A{I%Gy0mV_s?KkX zo|Z1gCh6}@W9=kcUtBcPJYatP>U_4@v^s3C;fRRc+x)S+=DVuUZ4ujI{}M4-5~gW@9i6-nriDfN+bku8;IhJb$@J9s%;2}yfpx~! z$h<{7es|*K>NlxqU$25OZ;vd+Q$E}Bq~W{YA!dlXZL57j92$MauXT?xyTzVg{2}0Y zt*gzzHtqJB*1pBI6mU)X$YTqg`LSe>+sC=@R685*go#Y=jmFRj^;H^h5X^ne-*G5B z<4}cUbMeAG7Kl(Vk@q4Zen-CI<=Eyu(IFqP?A=ENg1-`@V)Pcl^D&_Ev0A5~oAD7n z@!_b$gNe?9CqC?1*xU#xg4;5}eXQaNNTM^?0>lu14v2~o#Lo*7&<4?2h7hjs3KF5H z(D1%8L{Zm5)pkMA%}06>5Bg;Ctk_qmLfXH@K`4_(B+=JD_mlrS8a`D#F>?WtYVLry z+x`~BBAwjU@eTr3@xsOe0j4g(RtUoW`2hip0%dZ1fzSYN1!y=5)EPzOeSE;%&VXh^ zf2e~0dxQXmr@qK1Tx{z$GLF*OeG<85I{5-YoY_J5fb}5k>Y$vNpd1BD&aR-U_~6RA zU;rezNkO9g%C><~qDCO3VkWpJUuq;SB=0JypHYVVJfvGdrj;=S4?kqsC1kEox=%}d zRzdvCFzC}vDA_aVCKRz5V(H6G5ICP(35t3|jFM6dK`~?d>H>NiFRP1>cD)U}6F|CG zKu1r?ZX81(f5kXTaHw>10%t^ zoaZ6)eSN`!qXy5xO?8dvepCvO>j0{{XS76LNg8=Z+B>>VHZ3|~UrRMl70a?M&fvG%} zG#}Zl-(FrDN&ou&z83JT(~IWV$<{koMqdnG%1D4W&I~6^FG1X~;H90uPL4smrywFU zAzD=(IGQD(r-*T8=3|RvL%b!TsSWhLh@$}}gj|2Xc%5K(jhLw&^}zeXL^+BygX3~*DI#QJdw;yJ{ZRkZ%Z*+FxqmN95{nc zjN~jC7iSH|(SRsVWkmZcV>ImM<2ygvJYp+bN^h~R^OjVqS48}~bOqNQu z4}lH}7E&~_5Hi0u^GmrW_T|wWIN2`?fStVY5GEu=*X%cnC`tv{SaZoT@ARI|W|>Dk z1^TfA-{wpT0-c`AXfY!O<-iNXXMsuKOav3p+HmHzGo6leUya9k=)X4HB@~p;q|rg~ zdYAXQJl?w?-A)NOLW1D8oA)FkPbxoOw18KXH~l3zap8070+W1<4!}<~&wWf(^#(u0 zH8lY2@=6D}MezgmR;rzZ5v|0B`@*1n_mcuO{oE9tSrMO#770D_ZHAyG!&6ZmIC4L+r)=~ZoR<2N718d5!6OKgt8{fP)myM=el z#ZTZeRvVBp2a+8kO2rjQHI_>cT_JpPkTd48*EnU$T4mP_k(91w^fzU2WaR)i#sh)O zlNm1{I)rttEbT+NC%9Z>yZkM4MVMkmqd`S2*zb)ND(N7v-V6^#BM(($3}mc=)(wfC zC2MUtFrg~YDPG{^rmB3j6!iRf`Do$6p27(J7C*1P>|HOl#H1AcoH8*AU7` ztku=D5Qk1~hg1@WnJ9;CLTjDS#JBRL_`7OyiE9z@Yu^QjneBzz4u;yf)!EP2%@fy| zv(y<3)VdDV#3|QV3)d4XW8x%1(A~I`$bcy7Hf27_{W;UBO$Yw#d>ht6vV4}cB%foTleb*%%Ooa-#izENIauYrUK@?gm%1IVBCO+( z%r~n#R826ylNlB|@yaDDl$)~iTeVIVbP1EJ2N# zqYQ!48q3+Sl04 zDnotALltj^1b*)hvcWk4Ui~7an5UX0{+sxjf8h(Lu*b$HB*H)*sein@=ZE-%pV`>t z32J5}wZ&{J0;<98?KJ(*TRDci-*ind4H`{{&wdsfUyvVVUZP&3+o0T8~9yqkkR0bXofP=nF?5iP%AI({I} z+yfLUE7U#%hQ)iUBzU#a0c>=mCR(+~hY^T$Z@a1mXmpfZv{ymoYG+?c<0LxnFeZ*9 zRD6ZqJC3z{n3XA&1+fy$+2uv^C4(m7Q8m60SI~`3^%V|%fhO}rSfIi2A?dNoe$e4< zpXz-<$xazGb49;Ug`$^)3K4$Z7!6`3vG$R`f|84adzPz%NR{CUmnz-f-BeaZKQ7!w!-_OcO= zhstuVMKJn8G|3$#UNHs<^dJg>X%Hix;atNbFka*%I*FuT`o7Vz-p*o1o!Rz_*7(-F zjzVIsEgw#)nh-26r=yn3$FtlAju(vyq$O1fLIoyRp=u?EZbcR_Hw8iRD>iioik~kR?v-V3&sgBL04(`QcpwHkYP$L^xmr)jGY@LI zMkZG*`gTJW)rP{)#p;8~&pOb-NW6dC6(gb2)r@Z%BF9CD?OoABl{8)v5a7>x+?T?7 zQPE$Rx!^0$(wifd9b3s-I>@)Rk23OjviNLNdJ_G7EdEsGeEj_tE6#*U(nB6RPX0%4 zK}t`oKaxIS2Yvml$8E&Z@FGA|c1p?0Ln=EaP}KG_b9~9ge9?ndx&@Nx6a`zePu2yC z5&OKgOAV)@zSTz4(2x5ItD)*BsqzCN7`sDqZZsM=kn4De^DPGi;;o5#!U zG-$OO%on?F_nEyG+g5^I_FHfF*YEaqT)KQ3RDN=a(W9qqdo=l~4=Bigi!8l651cZH zs^j!#&Zvz`Tw6AGM6W(C?IHSnKJ35kc6Bja=5}p;_R0zZj~fzG6lmt6O2-b z2Pz>EG{tExg6hCignIDtZjPvqZK)Nv49Nn2XFOU}n77C<&iQG1LMs?uh~8y>!;rT1 z!(A!{O!?*=i(EFO;~+nq0|QhVD2XTp47fXt>80!CJt3v{pfS?)W2bVDdulkRYpkUb^vGQdwkn_w`P+u5VF_sO#)4r zTAwaqkekCg$8z9o2jkw^Ng&mmVMw`P3>SdnFZ1Y*1UKd1LNoJM(V)Y6y#1K=B_nUg zEmpPk%ye3H*2Dt8?Q80ELi;eO7Z5|?I~rI`$p#`3sY@8j(dv!+Y) z@?#f<@<-Hik`~saJl5YOlt<)B#w@Bt(C0>i1`d9I2DgiO$Tas^2#Ak7?jgP%Ssj=*ui9Zle^{3qXi7 zg{VS?Lhy$rcTaEYhA5%Jdv6fqh44DGsoY>1)dVtD6T_;i!tk8YjE-`0FmkjKi-E}k zKXdW}J$yxxpbm`5i&4|&Q6x%FMtJ2_xQR9#E{cIZ4%Z~_m&JZH1{(NwcT1--eAf;2 z%(ijNJ<3wqMXRru(LoMSmfl-pkV!n~rbC{r6LY%r_*MJN6!~vPXt9tU48;)BE%ZEW zUQfHTF6m<*wWsZKNcxuZ$ou}lV&lFcWc?Hwq&rWuZlrS>)t1OEI;G1j)&MSHmy4uV zwTOX;V_XgKnhhm0V-QNt$m>FTtiY%CEE~EG?~p9mf(@t^@eA_XxrwnupVl+uzK!p= zo?r{7dbE%?)E_C%V22@rOvA=@c1w8%RgAqW19oyOhu1$bwxQ)?B|cd5&3kL1ZLB%M6{V@FzVljox4mQ%{sSig*Q*MXJajWEmm_RHRE@NQ zO@B$h*r%RX*Y`OHw;1I%;s=BUy|Z8%%+K*~(_&v6W}kQ&Q&SNjzH-=LUzYplUaI3%FG+$Hd@=$JH;QUA*jp<`FR=UU$r5n`_o7K9snFDvEWiuN*l4AvIlGlr)@M|mYq3+f$cZ!2_L?s2DVSdR0LzqMT{y8}OKVhS@k ztUG_NyaDf#+Y8|(ROGVZX@|?kJIsy6C(?j&tQEFJsg#rJszyS6PwH9w@%eIE8-r{u+>}-5AK&S#(lIK!A(27TRph;YU7;8 z#T7}i*w^~@V(E|Gqd02(d8V>jl?>&hgP}XZQiaI|k zdwc1DG;=wH>if)DBO-^M%d)QKQL0XN$%8))Nr6YJKAEGo zxjYFA-X}{y3*(``gmI99|LL@|ZrtG}gy7KIU^&FCrKz^&z{(7GXgr_|)K>}6 z5s|y*8@J}j34Ev@$fKg2;_uF>3Q~3iLAv{stYW2`K79=jjeQiD%^46;jq9laBW@fV zj7PP$5TuIdzSW9TWdnf)T|c0B9KdPFCz#hC#0-iqA+Cy?pAD+QS>p78vE0OB$k*g=cOaY-j zt8Rp}RulFB>%^lZK7!>q;;x;K7#-o7N;IW=v}}N!Qd_iQSCsOE``s=wC474o^O$wj zfEdlRX!N)k;s-G}YcV?b)DQ7p4Dn-QRAcdEVQ+Umto8`h$G0&xkCn@aZQzN6^NDkE z;&liJaCVAY8;(=Pj1v`)HyH>B14V;8Ak)d?lWb$8=He9$;?=|xVh|E4yb==5;wGT) zMam+Uzz3|=AY8C(-vLwOxl^R`B^Jt(#^plePsA4CBbEnvRV2U!b_Jm9CQ{6WR-Yy9 zmPIt4K^C_sJrL2ZMM!?dn%s>5vzQCjt_jt`m(n|t{5S@t-#kS%JVgjSWilXT7C)8j z9>k0}42-@`XF~Eq9%yMIYNHl$>nv6G_v;T6kX`zI{mD<2+^Dkr0e|z;%8}|nr6hk( zhy|47Z_XA6?08`P3BXQF{YTD$AaD!d9JmAY7+OL0`lR(6!wr4#9vqDt2|x}*n+W_k zN4pT%r?h0dvbd>&L9wQ?=P{Pv*k?BrtTv)Z*W^d}D1`^~PqS}!f1p{KRpMhTKI z=DeGp8fdeAq;Xohh#O2?p$Y#7LToW})9;)E%GeZch(z5Kw1XKkQS>^2@RY+jo5^A+ z`O!vPd1@8(&H1k_2?}&u`=BjD3N*@$O%nx?IJQDSj~CUSGE`ej*Et;Qw-!`C8GY>v zB^*gtbK@M?fLJc$EnDOxEju#1sM#lOp{C@h>P^60#!P&^yQNS!U`i1!UqBo+DYfwd z;BQ_z5p-@$)O6wL>@PZO&+1!z-bJZDC0=NcV$cJeEl)Atypn(7<5FioI`&0LmN?qB zee3m=5X&u{akKu=9c?e_VX)k+KUS-YvuA`$|_cE+yphjAC^IHL@8L^3CrgyV7gq(jC&# z-_1eD#A)Eht#imEuhLdm_d=|m?x@EkKnYQg z$e^4hkX;(bRJE$=m3htd1s}jW8+u!zc6lKR;Dgam0zHPuxGeWp`z+-fj+f;r8_$nF zQoOx_4W?)Um5A9i!;%%2M#Hf@lxy`A>7`A4EwNHs525oA`5nP$a@&q%7vyp(@-T(+ zF0Pp1a!UFV5i9eW}bs|N7jFs6NEokr$iw(P%@o_4TI zsg`rxQB0YwI7F!-;zkszLH8|)IGqpH8{w4=65f{j6uC0yu$12!;?#0arO~PX6~!l~ z>(ya3HvYB9cAT@dyJ?ZkT{q$&O4IxhQ@`ofjR+)rOg9KG*Abv>P_aCvzaDiP*-zJavp@;NLRO?cq| zio)$u^i`M}S?x~P!Oj&D7=nhosTw=n`uq-<ucvof;tOsCDKxYE7>PQ z@~H8`Q}3e_#aj5_p9@7HhZ-OW?RCq_K8F_7An+hR9TP6?osrd^U+-iuZfktgHCtd> z5N{LtB6wVaBR$Q=n$wGb3%-MQvW&#O-UyGCyn=JxHjHm|k%TOxP_!Kl>1nSvImv0J zWXt;qluS&@2i#qt*t|%J)=Y>sPEzrFR;0v--{|Z|sI* zAx(6Xx0Q$I7E+-|<@Hd&GsN|5h+;O3%j=~=#&2N4-A=i$@Uav!VUd4{BZ-gDVqreE zJo6DDn2yODb0{$d5uJB!Kz4!(8g0t^KL4}WK4~UWVqHs`B)fKW62qP9)8IEj9bu&6atE~KnM4@W_pbj*bDG@i z{0{{h6&BO~mq5tt|96-tR{;Ng-b$*=-zC%pdJZN0F_tb2!UJOI>H2&Bv#ykXn;Tda zx2*DYQ3b#(&!_>AO9>4CvEs&UsYVQRTEFZe`7Wk@^A5lL+5f|emw(@%P56^P8}~2& z?En62_wW4Jk{kW4Q0GUwg6iM@*CqgHEcSoc2NZw24`lu!EcGidIH~`qJwRWar8hM= zzwQH6H;o2~u|N8=ztt^&?<9VR@s1A7f}qp#h8L{aJw)06Ap&D06K?b)Z!Z@*Hxl_G z$2|QS`v&KV;4IA^#WeR0<&xOh9k|b&2P6>e6F8)#+zc>QBNQv25}F9Ngd_^fcBNXV zR$9y*SC!g4OmWQ_1)4TfG2Lj1ua_53n==t&;nLKwAUUKvX}E%F7;V|6m5A2XFV)!D zV7hOOT^-fLamutMvv`T1aJ%=2yWf-Tt!LxRs>K9p&Uo28lNlt@lz+@rrk~VQL5XQ{ z=`bwe4reay%{6!=Uh87)(^hbMaiO~L4$j5&1McPECpOL9w?)U>D_%38&vt$s*OE`s zL&+x^+}5Zq-W$b6*;T3t%q}sK$j+;7+oXhe5>^u{@;oY_)XnFkLlKdQZIvV?m`IBo zV^*fR^0uy_X+s~wvaLYk7_&J+bJ$XwtGMmy?aZzLgbG#N4V~64Zptz;9902C)nm?RsHSy9FW7J1b9wU2s|NS6~8)5 z5fp-fRyUx;DpjQ^R^ONICM;#XoM3hnmU>*()f1<>E+ZCZ4XRBhup732IiW5VW0uspdz1hX> zbwM!L19BVih28lKOx4?8aAGfD3P~90H_P(TEg8#uW78j347B*Sln<&i1YGCpw#XK} zp5pc_Lh!NvxK-otqAdnF8Cg(Fxf)2iUHAK5b4-kj+V62DHteZnk8heHK2FOS%C!UVY|pRg#N%zSHgqpMA2nI`HO zCz1fQ^@Z1^ZQnB3=Qt?%H7s@0>Sk7SxJTwXMD78ZqhFlXOzrk&fPl9R3<;09>v>t` z2@_G;tY*`T2>(f(xAB9hd~GrEaI+WmCHVEN^)&AUF9S|v|7LHNNVj%NC;u2=FE|OZ?QZiZ}*s#ax&&U7I-L8Vqh7*5kZN} zzan{6A%}kqpWVEi1996WM_@&^-+A&N)xQEy=%lOP_59!!#Pod#dZ>O1^k*^Nkb1}q z=21rU&tP;6SCLV>mDSCM$je>KL<(#ML!6zlP>3#bWu)bX#LOx)X+vNo-@tpIWglj# z56~!>T2MafjLzmI5(6c|OQkL!LcuvUL0X>=lh%Y|D)U?1eK1#-$$uYO$w&2GM*axI zxf-^(R{rf*`5%hgDFF`rBbcW3518Svfx6`atv5~YU)LQKjS_BXU_PuAX~U6gq!hVNs^ezO4o_-2_;Cf7<2)Tja%4Z_--FtxK>bX zo1j^6J(DBcv&%(9*|t9QmgO`H&s2~F-{Z^DOV2VmMj&FV==4m#YujW5!QH7IK~5`n zvBAPZr2rp@Bxctq=R-b1XA@*Kr!#~hzmC}B%c&Smq2Vw2wzQ`S)wXM&LZdm)P~_8F$DJ5o}CtR~pgLEM!9j_M zgmFVYdjRspASa9Vyb6g(H?))xNe{pRYI0%N7jTS1c$Zh(L-^Cn4T1C289W>(;Cl*+cPN7UV~=c!tS2rY-TEZm3b15I(?kE6BaB!F(Y7;kk&YZ{3^lb4g^bZiEU zBg9WoDIPe1#;-Ws4Oi2}$myZdf~5j8G9uOMD6^<(EdaVDedqU>jVFel>-q0( zTU9wW=D>MMOVuNP*4L7m!XT7Etahd9&n2UOz-+LS724pXGH=@4LJ?+U!^BX&DH)AZ zwf&`JRJg_Hm-hdkj3ZRs&4(PYe?5-4t^)pg9Gy%_nJM0kBPNziXO5odvz=Ecyk(wf zsE_A;R{|XuN}2n33xMpd#iz(5Ij5ye-O$?H#3!g0E2&Qrf$VNg%v^EmR&hpg$%Vp`kfLD?=x@;q2#@ssT}BpB9j5x5yg3jI_S2?N2|oQv4K za_XfRHI<*ZMEMkH)xwjeU0S$WW&7+B50WKr)9_b-`|RruQg(xZaTFlDHE@u6+M5H6 zBV|8HaDb?ziDz^kF$MfHa66+J25R7dYqtB7xo5CU*y}1h5|!zAR;`j8-Giad2hkx| zz&Lu1B}>K(&ZJPxyQPvp6sx|-Kx~(%(AuS{y7}%lnxlxYowIxDekF-Y*$TY%DKU)PA5 z97zi`D==g#P=Cr#UQK_$P*ulL)f}we5II-4WkWA~UN;I3{7^hh#;A(+5jCIuxYYQA z#Y5Ro_p5P^a-u>kb#ja*dWc_>+_PsimHs>i4R4>V#9KtE7BorFnIEj2Dfw7Cs7q(? zxJs6M&Wa~rduknA-H1|b<0n70KyXrH;@C{y-Zo(U`#C5EVFi-<&ZYMI4htaQ$pQPj zdR7X`FM+>Jdh!c}*FcAb!Y{k~vfxd?vu|hs2za`F*(TXAz6Sojtz{;-5nr}R_Nas# zms&s)_+}28r|0~e1D;)Z0B-6h0Z&dtgLKK~*eu@yfAgD2vgMPVk+f;oDGFWtnm)0A zH{kh)8`{SS1o^7Mi3+pfe4U=EqL0-!i`^;m)j)^E4dwY?4S2qg_LvlZvWS}h;-r73 z@dXI2@hR_Ka(mdULUBFy!7|4y@i4lV&OPg+_YC`W?e3g}&%AX*W@)THJNQ~L2e{OB z7kmETQu|slhw}>n_M1#2MLX|}rEc~+nZ}oP;hRj8yAa4&wYu-v9v79W(TTGIC>k(+k?D`O7QXI|!+?2~^R}Qi3Re&_ua|pmq*%1cP5(YAN zOJ#)S1Gp(O1-U%@Slz#*JXeJSxT)_c&kUp}H_OJ)D9_P9p*-KQC+}6C3|zV)rLH>tEVT|KbQ9MKJKSgoXmo0s3hqJaCud;sE5`yHwtW)B!^N$}e z3!DB-FHl0$4;Bw1R;;UB8_vD=J*NH6J+DccCyUH^FHSmV8pT0qSGK%+q3vN$6p95X z4wi9WiI>p6Ja1bcr7-Njvbl5OJM)(1KlOl7KNX$I1u<8FjDib#WsK;#XB_21`D0s* z!ZI%4jcLVD&txN{96XI9lNh@)t(5x~jH6=}&J zxxq`Z8tVh@2zm}q;_+1BbibbE*RsV;8tGP&p9Ny%z&{^N@QkJStN*0`{So~iE-9^n zME}7h<^S0q(VL!jdyww8A%7F~`cI)`S+TVuf!3TjBkX9Cr1+miy?*-!*GmLg)NKmQ z_Gs@~FuMIy-rGQ`bPasC*Kdb0!buP}d8d~Gj&_=RkrG{Wo?=O5je28+e2C*CilR&l-yFb>AkKlg#VAWrEGUmSE%|0B=sm&zKei>uqIkK)L9MB{_rUHb~^? zj-4`-M2^#>xJe<3{zB17gtzq6Yp7<^!B-8VnnjJ+2g-E-9gp=rli_m->bv}s7Uc0n zz8vEO?d!LWv#$j1w3}8;@I3%MCa6d^>ZwnwZ^w8ey%{wk7F44nrl&ty!#C@%jcFrj z$IGb_Rt6@2JBF5OA$uE~>1-f3EpTs`NsFIFt*lR^}_-Rl3zQ0nEVFbn9h(s*G?c`nAkD4bqm3Xp>2b z>_mrk;Fuh0XqKC)WweHn5u0zpmXmlB^*SH2KqzUquw<^7$!7?52eeJ5yT`y1T@Ee3 zSq3oUHp_E9IWNJturpZE_634XuPcX5UdX;44RkN6n)BRmfdh`oKoMm2&%eR{q;0+` z?b^LzHt)4%lFF2|;OO0aga4{+enrXtc1&hSGy7HB{3)^ONs-bplKsfOk|I~Vy55%c zs8&Uf{n)xI3U{4hn3eT}iC2)r_khHtaN85d0Nu(>fMI7PLv_*TBByOO0f<%a0AiJ28qm|8?qvj#7L`gShJR6V zkIGxE-I+8N+j-aLH7huJy~5|#x`(!nT^gb>4Tz{a$dA!G?I*;l6q8oEbTbp!s5I;^ z#Ht@?oAG=E<%8JPE))z4w9@wn%0hE%D2NpSZ8H+kHnCtxp%4N^F^6vX3Vfo%%JZpHFOGu^uwuayEGyn!2KQshhTU` z+QqB90esfVH1v>M2^zVJU7SWp=YL$j>=esv_oNi;6NVQJXwmnpy+my zECxxBv`{kW`1Zr>`!;k+&+Fh)_74G+EIl(wL@15mgDf|Su}f&Tny7wdp340peP*l= z`RQbk8jp?DgFHSIJ0BHj6hUfqOx`bxv@bLf+;}jYKUzKnF0y>d@Nk;nt9Ixw-hn&a0+Zuc$q`{g`&2D_9I;`8sJm5D!L_Po?>%oy(uvF|l9z!xXSg`9t6OAU7|2 z1VgQFd=wjhhdjS98xwo|v|FqUt+`EUmrDj$im!^IP0F}1V?(8rG%t&k5d>PJ4*e{r z$yzI+mcyp|Ae-#;Fenz9%N5T7OAq-0Njap%odw9%G$!5>nIgOLO`H?=3>7^0Oo29v z`q%1A!Ec^-DOWw`-|phTM4BD0#sJFUe!UgHH-aAqG6t62jJZ=NKfuDXh~GP%yBKQosYVP&0G2KjN2pO` zF?_1QKa*5=zNsrxMGBz}a{Qf5X0(QhQ>);n`Pxz(tJ;^oW5B{={E*bs*Q{!d8tL(u zU18LG1wmvc&otIzp?2lf-D8ARuuFP4O`87iBcu=JZ0 zY*kG;Y;0BQCczt06h*ou_XyN%2m@%;)=?2WW4W9G;0%m+wY5{;5t*hcAH)Nfb9CSnZQWBpqo35n3 zK>##UGe9#<$bO|{vYzkN<-DQWi25S7yr22eR+V_bi*1?hzBl2GH=5}eFylUPB0`aq zbc62!vwk>TS&A4*YpFj)$>s6X^I!Xp`_o*~!*rRT4`5AS5oK)0u?V}5!kOR3-rU%$ zN!}|c6~@0r5{LMV-gLcW==v*|@uu$>3{4}`jHtEiiPxro#V$kcGhyK_no9yFEz^RT zxa&=;kcRl+Yu_sik8D<%alz^dNWNzDC z*0O*!M{=w?Wbf!q(gW2WL*rD`_%qqDVd!2gkhK7s>Gu|%y1fvaUo1S6d!dfK8N?@@ z9YW`OVeZJ8q;Rr8qIo~uUm=qmx2sc%e?KBT_le-RAf7779p0<78%qE#fzd3R zBaIF!+G-cZ*n24^FdWOm`6<%aMK(znXVT0gF2)79LQy{V>?X{aY>b*PYxp< z35W~rboH|gRPbB|i?Tu`Len>}Ii93Nis5#v_%wpk&~eY^wNm)dvcz;ecG4T*u@SVA0 z0oDk3H5yv-Xzi{C|Gi{aQb?7TVQ?XL!=a(E^Ax#K;v=0sjiy=)6s^qF<##cHZ0unk zG$1K=Oo9qjxe6J1H|~AtR!f+~hAWA()Ew*{#YN^UVuRRx_=t3@cnCX7Ns$qy332p| zwK+J&pI+1noNc6GT;j*&ivWwL*@-$z!La+-*8v&uMSowO+e)AWJvlo!AB-XRX<=n`O>asBl75qV zdHvw^k1W=wk4mHo^p-U=q!8#}HiU5$NkwmskX!RyJD8wq&;9PbjeDULu7FUZw#FLV|YD>=Ak$SiDyBf}ZSV z$gr??KVfN+(~D!Ls@4}Ss~7$>@u5*_VVwNe?&4+}9N zm%;pznCqpkS^y}kt%7sw7=|!bS|`yh@5(!ZZQD2~?jHFS{zQRhWe=$gmd>d;dp+-= z9i-;5ts;t28lscSJM)z`uNkGmkVEJvlrAfIQfVb!I~d+CKF`T&7xgak8p{)_PMm3B zgDB9dTowqbT1^ZgZZk$Y5{?o`8S5#aGX;QyiV+nn%8*mZF#< zcs%P$L73}(ZID1lLxMAz!qMdY;J%E46V7iDxZmbw_TSQ0iSUT?j4@@I0d19@8(}G6 zK@liZt(l%f0;HJ$a2$agf(h>ZQ1j(z9V@5hM6+_hV}#NNrcVA_Wvakh7_DxxmKAUE zRa;$zN%f`w)$$6U0b&BX)(eDcw6K4#tpf8BmSOXU+G-6k4a4{Ia?Y3GmwEY5t%b-u zo4>4uKLQ#5aV`Ab^7?Hp)FyBK3CLK{e&7xkDJ19@kntbqWz|PVK8T&qE z@|fa64uOF93NrrFS_sTbMB2Hgh2S36O2~}gK*q|&5TM)~)YkRaUqQws7hAP)4ENPeo|!L@!l%<=`oa^JUaWy`}vz;qC}!mxw#B-Vxh)&TY;)Qtvid5#~Uw@uAt_ z)azZa`-`qb^n1czq{1P=*J3E`xQ{P_5oF;Z)&kHWK3UGgP{4_uhYY-+tt5tfbSi|7 zxcegVsR4pwZY%yit8aoOF-loZQ|u)T1M&rgs7Waoj9PynO(8LaJgwMgb2dLAP!RbcQxIlp5@NaU@9cV zyq>GO6+Sp>ROm~*-4?`{e)|>E$B8tGPpt2`4!04t3xs>k`g^C)4k%Kt$jLROS5+N&Q5DPag5Ux0YR?w!NOCpSnmptffFC}dPOp({VJM~hR|Bf zwNeW4Q+QJNTsp;1%v8Q;@YWvBCwlVLtGn-y>+EK{8zw?jnfy?mcvNWC zV7R8l!;b9z0RhuSWNle)tZwqC*ztXV?n(Dp5SJ-XT7SuxqA4@*S22N6&AY=uK3(qlyg_orXgbGmma;MT*jjVC(j8I6H@zR8$9`3@( zhN7Z8rexH6QL4!BXZgH%7$)M-0s>I5pF7)6_}OlQ!s~H`#XBCVvf2p2etH?N-zL#y zixFK9Lrt-UsHCnIF^v9*gv790eHkP6mGSJTHD3?kIeYvQ%4N2(++LGZjKpEW46ceZ zNo&`qK?{s4*n_owbafm-=)o&I&uja~a@oOfie|UgY6r;pIn#wRS0$h44u*zaS7y>E zuF2sN48`igXLF>4D?U9ZNxga|XvVmX7g0CD!q1hjowr zBY(T14fMUbF_Lqx(mRx!22XOw+xyVVqjb_9aU_xTxjJxaFm55=eKUz&$6bY=zGcOb z`EdpozNXbG%r+{Ccv+~Lf2hd%S?QbU1;=Xat6mQ0Co(g;SFnu-DIqTpJ`-GBJ?8>3 z?|R<8n1ffYX~FH=^}Ta3j~-UjMx(SB$l_Uv(*Zlk*s~X`CtlBX%*#Q!n~KIO@@a&( z7C|j^-z?l^i75`YN0#*-`&-9Q@gQIx8nWd!`$hXb%btWEu*Hrku<4yEJrq{~e+e zMsN#2%*R2mR-^er!}DHl`gtBdDgAs(l6FDT^!9QZe0{06u>AqnEsWItTF3zcsZ{$Q z4tdZI8gASY!3+h7)^?i)c~{bvW7nO;1_CeFk|5p#uf_xTQRjwm|GeaSgYSa(M9cQX zp+&&Pdy2>EMy`Q;I_5^HVa^c`x;%kr(SYKs(HpArikmZ}AoZr#*F^veN%>o_;dybf zx_&72KJx>quDZ!W_%J>2Va^2!uG--2`baqnXqfrv?fUSu`gRTYVEOqDfqiw&AhqC8 ztYv(k$$%cqKuO>sKh^NFQGwKs^uveqA0+ihSV4MWhT=kM;};Lof0KGsDh@=f5 zDTPaS5Qy*ei`w-|91oaK57-z5`K$&Q&4I#Y0(0X7^J@YN#{*f4PXmE#F<_mp)Cj6_ z45|U>EHy#(<3Ww5K}~qUExf^P8o}+W*a61Y_tMIrZ%+F8q>Kcb zF~&=s-+%Mh6D{5_fZ$E99e&gWu z?B&t<@BE4Hm}CI}1=Cjm#o$0gARb3cP3~79LH8UjCSq44Es*hrcCfnvQ1A?y;orh~ zpUAs1JyOZ>3xHz1gyw6FM0aK0aE?lbT*^0p;-$f}f-xl>fUzYXYc^i~Cx2pq>5aW9 zPIIQweJ}ja$nE;U!yi~HpERQI*n2OkORwX}Pn87$nRcG4ua&%S2~ipbVfs9Fn6 zo*nE8O_`oI4^5qH;KO<~aYly)9z!6&O7qehBu<}aMU2ki!dt(UIf8Be>o>H;Vg&g2 z-_V6s>+9duNYs3}Ifa^ZC`BK)B$n!_Om38QQWtEN_X0t=iUBDGCcd^-5FCItTd|{*BJkAa-o+O6%Ewqc!|*xh~Gh5q@Y5l16*j zBDFcX_{wW`@TK)R;nwU)?zO}9NwN-Un;Slc%xzbQb(6VkNu<6}eMz)-oOy{`-KNSu z#Rl*ZzJ<1{;eL5!xfuR9RpR6p1m2Cex5qwnd#WEy|%jE z^=S7%hfr4|wK_9@Mtbded07RPKaCoo)L=^$q?H6~+@7|$;P_^NYBO3e?$~Vb*6D8$TAUXditLsZt5wPe* zE&!NO0(&5$@egWF&SSM-X#HhGquqHizI@ce@>Rmjud>@8&xmhsKkrMO*Qp#)mzRB3 zY^-m2BfGr_s+z$B-l1sl4QBMlT_P|UpI^KA?|Fwgh2<|8Zj$rVtx1pJ|N0^R3j?wK zpBac+Uk2ho_YevH`w(y7Xg~AHmQnubA@=JE+9KS3ZS?A`lB@J-!`9QbSF;q?)7S6d zCSk;&<0Cw3)@Z=jPvCV&=B&`vu?(zkDXEdYJbg%;?xWQ;ME%?G0YDB%Dz@-wIG_1>sqPgARsGsvx|sny)n$2X|de-*9PLRHZQS!q?-zh-qPOw?+q#c zos5Kk=$QQX_=5jC83Etre<$OQe3w6-j5i@fU69fDtKv_}US%f#sSsjZyz5U^#pgs~ ziSfUtoM<4ke~yWj!M;P#YLMG$(%r-Mu5YjlW5{O!xYDU0PX?hIj>L2bEb6|en{`5m zqeNbt*4~NBJ-jqmXJ3QSII^^wQ9Z{!y`v5^I&M1ELhKFJ)F;r76Ht| zxzEbKh&POeuO{K+>}e*Baj$GY$nNOg80PzEIxEmK#-c#wSKXa;Esyc5NqDU#$wvaT zupYckPsqjXzLP9N(0D6Q4Q+}T1D8`iX>mKSkz)KxpOtxD5I$5V)}8c$go%4h+(#Br#j?Y}Cu|AQlb3-E0rQ-!sB)Is>ef3V@rM?V# zzo+weVW5AwDn8IlYdVzpfR8Pr5B{EY@!ORS*6{iJ>la})G@ULG5XGYTIX{+fkQH?_HikS}Gg5~v9jVKSiuEn$S1bH))W?J~_#5=!`4{?ICO z^AWPdk%lDlu4a$?p&#NKBw&ZqktJG|2h2xfnQ0owiq|2eCllFCP=y`Pv1Ub?UCFK_ z-a@3$0Y4%SG);6LGN*-5zd|t2uv6tvOA0Z}H^??MS<8$KGvTMraJMk0OO6EXuH^*k zj_2p6@+0PY=P7B?7ilK?yNAAj%^*M-qYK}NZw(EkgLECLwTLaz>ZGqwZppLoZfm#L zxX!6j<=x0~P(58qFHA(VN-k5?oUd5OCZJ84Z_z5Kj&&fQd5s0%lbyJoc&8w>fqfhV zIz|m#uNqcG+RCrQOiW7zmuoMB&6ep4Tc#|6SPGNs=onKiitF~?!O|4w$KDEap?;T= z8f;LVp_B3lCyOKm3ZN^F4h7UvK|~u6wXw1?aq?-!$%30%}=iD*?Q$V=&U-x%+BTw zPzObut)0v$RT(iRThH2Wc?#yueliqvXw9}GVw)3*(_t-)Uq*3Xu&LlXNnpGEkox-l zV>nd~g3H4thm7=l3FfWUlbGj z!Q{`{G-MRgG2=ShsQLH))TS{Wf@1fZ0a`R!8QGNaYq`@OoMD_Gi~UY%uFbMAcZpUK zDA`UnLNJ`y$6j8M zV&p<2&`R6yaknry1{z2f?g()@^{lE{)UIzoU4%6|z8lbs2ujevIQSxre zogxztGiep`B6qphb!tefBL%<@vPyzEB-WHFJPV3DEGMaYtS46?J&Fi4 zrwt_7f)tXjVMBiQT(zM}t>`g!o}M!QIPC80OqlQukcymQOf(lf3sVyer_Fe&g?^TY z;2LPVP6W&hE~2_Dlt;3eh!3pFip(mMQ7+Idi{#EK=W~Yz*$nA`mZ9)yTa4swJj8v~ zsw|n9U_IYxgy$b;wCY&eM!gwD($_DBgcq@uC77yBd{%r-5UvPCUpo3}h#Rtv%F5=A zw(6W`%o_s7?Z>wvW)`_2YAl)7V89I1Tim#HJ1EvV2QMB@7~+> ze*Emi9ynV39c2q7mO_?tYZcWfvG%@^s?wcQe1oX@j=$;qfO?1Ew9FL%4$bVkAf|2kFJH+Ajn)ck7oV}0$ly>D4jwW1%Y*t zN^SNacA>_t(76dQtx50uRj*G3EnMA3-ixpkFn=bBecNAzptIa+c zf+pUHew})h6~?oA26HFQi^04&Oz_PflTFi@^U|ieb3GUH?m&^y!^{${RYkWKHYtM3 zDT_mb(2`;@2FR1JO?buROfL7Zh1J@naSHg~mCsTko^@+q?Q=hJED8#Ez@UG107h$& zi=aOrN+LcisV^c>>OUWq$Bxb4S;a77`Ak04cr0Eoxb2nudHTBaq;ZFA*Gu_gLEGT8 z{VmyklwRu+hnsJon=5h#%S%+%x95j%gpLDlIbtX{ejZvAI(3EZXM6ACcaHOL?8vx+-62hO1b?B*`oB#rpB zsVf7f2j(h^;OO*RaFyd+_~o2b8M6sxyXu7u1ujx0Q)d_1f$o;19Mr1H3T z*Igf;5jO-MA9WeRhvPmidp`P|w+(rHtM2-m#QVbd_*#skJRbL**!8v1K(RBEWsySSg zXB~9*_w~KrpZ8}jq6WBhPlQPsN9MXjHltWt zOC!l$kc^toqL0?2a!!q(px6*m_(1cAQMX7Gn#$0wNIqN1ZH(Nm?6 zndRtk7=GRpvS5r~PR4&K#jp0^*Ou`hff%sT_pS>v=@?sCD9hKbi|a!%yk4>V{d#;U zvB8wsJ>krIS7O6paUw0u2$MMRl(?LR7zrHqK-wqDohSn-@bN|8F-;7(Adit&m(kcj(V9 z)7@T9lC@1jxtt3!p~srU;fk4AyOIN;400#H#i7Xr0S0Yxpzeb11`3?hmm)itl7*x{ zjZ8G~N_Byy_S*1NJxP^CC!qoYp`f%pqgBru> z>%$pKOy~=s!Et=VXG+9Q9WXzd*oU;0ui@y|Au4tepV`tggEG3yh@%1-ay7(p6C#3= zF)*1vzCt&p!#!V?F*cboXq!2AC)0qEE(%Lm2a$^3OcFlWGS*I}t?3XL?Z9Af2KI6= zryUWQLNcZJQ$-2?ewx0YM7{_k!!D$%yTl!xB%eBrLC^@)Yjk{bBrDM@xCH%fJ9-Xo z0@5W*S0`KEJX@_L22~)dNXka`XG=>2dBAft%E2>d*bbEw1U+&tk&c}xr`K;Kyc&;w z5kk+tLU%AFm;Fu-v?Z!iCKsd3%nk=zDd*YXqJ&-Y+&~cJmOT5^JW(206m(nb$I1>b zkP#>g!~*E2j}9&Evjk$Se`YPtmiy(5EhZXf21Be+~nR3G5f z;@?vXQg7hg?Pa_t-ZlPVXy5!kxo2l|)ncH|>vF*cC3Vhn`TDM;8W!%FK|d5Ip1J%)ce4Z{h$9;JR-4c@5v0kvXZxL z$+Gri?t%4Vout`|5I|aE%fq-Q-A}~g+FrJz@dwfRrkb>YQX|cb9OXx|1{G&{(LJ9Z zy&uuUZBQMd4D{jrikWJYAGs#c=gm1E9Bp!-eBPd4Ap`W#E< zp;zFtYe4@O87y8=V3aJjs-xqPgqgjJ~X;_?uzga2X>>y4(GYZx6 zlYukpfLk=p;=*od6SBm|M}=xdE#D?cpx}rkJOUfSuxVdZ%nKMxy;2a6$L>%By_$ZN zsxwY>dVOa+O}rT@Cd9XjoFE1B*-enc;^wWgBI|f2AQ6BJSQZb2h;ouKXD4%Gt!cK0 zll&xTvL1$V{C>aDb*nhZC4=wOH7RB`5l zL>;b4|8(%H_zwC=1y?-R5OT}4kpIIIXJ5YpD4wa@} z9EtW}^#ZyP)v(QpSBqAj1Uasmj%SQB0c7-pD2%s5US{U=vub*K`p`Rrd2Mq+9f=FY9J{+J}{j{bst znsk5dQ=$Fi<`sYOG>NgPUgz~4_C*}0AJai>i)u&|1YL9$`a_baoy4Tr43vXx?2E(X zkD@@5=(Hn3m8@kLhT|=n3)$KiHT23T*!a?Y9HQiUKv?Ug#wKE$C3;*NRvZ?M65$pg zH;77Pg=3C0@MhbDAq`?-=PI+{CHb0Ym*NP&WfA@rc+>h38adLsfnKoNCRDy8mc!}h zmJIlhNun$Ld(hhIEb}o>`?y60NzIxMkHtoLjPk>gI`XGW;%rJ0VjZLw{Wu`q%Nr!( zzN=MAT*f!Iy$*R+N#|U79Av)RU|%eM8>)I9@}Sacx8)rmNd)YRqx=#hH4w+n)iXIZVm)yQ<%J- zZP*vb_kmAid_T$Vn_0b$4q9ms6y%j<*0wpu+1e$7Fb|Xnw$achb%kIIAq#gIm=wI9 z3SE&i3}TD1kv!3TMz{`}4gA5YPK9^Mp9{#9kfPJOtPAXma%?Xufqk*OC8^qwtG>ln zMg#PH!0Zr7&3HviotEtMa2)h3ng#TQfL=-a;71n5?Ek?kuf_%Fl~AlEbg`~FPSKl# zL?JnIGh!2!2*VU837XuKNyU`N6y3G|9YwiaS*7w2QFEe9uTHt?K!tLvZ9{b@zFD?i zuE&*%e6IRTsSiU?eA%V{+W4*gfLKB6r1VtC^t8&n!@K<-gYQumM#5In;|{AIVvlLo za?&#nzamv8JC?~2m&*O2^j1LKa5e-licW-jYFtV7h3T)u_6sni*NHQ~I$WP9v?kZWcf43w?w@d-XP3rg)sQNTtOiPYgiljCz=4|C= zjAc3Oxvxu9!!9sLKV*~la zN8@gaxBP@r`Lo9NHV+@w_9_Se%YRY<^0SlN-;UVqFi3QXZ-7W@4s$x|rE;3v6Q_!S zDpu+1^Hlrasy^z5I#;((x7~MJTeH7d1Lo0ksjd4Z@~|_D^g+{ljvyx2h75hZ4Xga7 zdEwpgW3TI*k<<-FW$Ps8py_gGBVv|UFGecamtM-Y>MKfLA1qwLdCeZx9rW`DdkzuVFw0_C;k`O{W?AqtKC(`R*HNM{-9)CAm_5&%=Nst2;gy1lk_~W7F8G1H05`*nr4fFzqbywx) zAW=I`34&Kf)bXL~V5u(b!c(M1Omm$~Peib&g9QsWyk2yVd5C2_8Ol?mA((gEF7X2y z>1kZQXJX2pU;B+wfbsFCzYT?JIb`vEDS;!gRIpLH2uDeN<00?uhsRPzRs_5jjfF*m zKkxp59g+g61BaPE>s?i4=Cyh~*EGFnA6E*~#1O}DC=oZ0i?~#3zK3q%8tM5ig@~7K zCtIBl3t|%Ef)$2!7eGdRGnH#)VXez>WEV;d`0vme-Mn>V~BoZ{pLw zQ}KPW`2aqx#HarapDwUOZd4={Az4Ui;6n3yTvh+Id;R*o)G-Ou*uo11%QAja@{giS zgKnG|^%atzsV%=r&uU!{pB~!4r!Tc9XPfz-6S@&`u*^-cHPPLI+MY7j(LvnV4*7T4 zp&bu1TP5Uwsdw$W#l*B-LOwAR1p9}2*ZsaM8-otB5g4!m|_)Y19{3{v?I Dw9Je@ literal 0 HcmV?d00001 diff --git a/assets/sqlflow-logo.svg b/assets/sqlflow-logo.svg new file mode 100644 index 0000000..dc7b103 --- /dev/null +++ b/assets/sqlflow-logo.svg @@ -0,0 +1,44 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + \ No newline at end of file diff --git a/assets/train.gif b/assets/train.gif new file mode 100644 index 0000000000000000000000000000000000000000..4507b6882897167d21114da6d8cc7dbe0649d26c GIT binary patch literal 70492 zcmb4~Wmp_-wyqm@O#;C?xVt-q;1Cj$;2vCqySux)ySuwP1b5eN+@TMdJ>Sgib7uCn zPyOkt>UX_TwXW4qRX?|clsGq!PAhadR4)mjvZ691BqaZeR7OIoNky2?`OPMz1oSd7 zIy$;MY2fVS{Pc1Ofj|cP2A-ZkIXOA4t*yyPNnNd-hQm4s`v*U>RZdP$uCA_jc6OGR zmwS7AXJ%$5$0m*r_DA}MUIRZsP?Y5qfMPnfI&Zc@iKS_lIf&7lVj^KiY5mR8((JFRs9Qwbrq{_Z{kgsWa=ZUJ^_oL}cK?yQuLu7~e|?MB zoVR{WZYaph^V8$Q{oU=&_0{FY`572=dUAYpc(A{>yR*Huxv{>sy0W~qxG+EWdv<1e zYI0(HY;EZK zEj1-MDKQ~FE;c4QDl#HGEHorIC@{d^&)3J>%hSW%&DF)($>a~ULI~P zP7Zc9Ru*O^Mh1F1S{iC9N(yo^(hooqVj@BUe7yI#IPc!R!N$VGKu1GGK}JGEfQN&H zfrbJAKHS5?Apl4K^nlmxPgMfIL9~Ce{4`ECDlRuWZnvVJur*(IEC;)SA)T34lJxjq zI8(od%y=^<3vH##R4W#CsYgzoJ2omEWf@Z@4r-mpELzsj8eQcpmce(e?i1igh^Lpg zkFTG9KwwaCNN8AiL}XNSOl(|yLSj;KN@`kqMrKxaPHtX)L19sGNoiSmMP*fWO>JF$ zLt|5OOKV$uM`u@ePj6rUz~IpE$mrPk#N^cU%r-mE9Nep3P)s zuqtOD0{HQ!S-3iPD27@QQmj5yoi~!etT&u0Rg*uK!fkW3GE`GAks<6-jv`%KIF%zA zL#r`dTQpN3ks*>MU03|OM7hRfHJAf$zCybvm`J9+R3){2yjWwTzHGU{%5;m?B&&S2 zSrB}*I?_;2`yOo`S+d)wpf+4e}r=Z^O0$Ggkp^@)xS2o$nd zrigS}Cak9elvw5l#2c0KV8aJfoO;vu?GK|(zxM`(oBo9M2b%#vf9kD3@_3`IAnN@0 zA^>D+0F-n}uCH(iz4PsQ824V`b~yjd!FGf&Jk3rxdp$QWG`{qz6F`rCxD%rwPO}^P z?T0Z%EWLPS8achnWOkU6Kh0j^&v@g#B%}HPQcYO#X^NPAEMP$@We6QfIl1Fx|3je*_Kk8fRc5OF(P2TR<{?p*A8U5B zW>EORLGdr~BNJ%ta3z}jT7%{yXJo+SnwPvFrbr>(P2#ZwGJmFNB^a%RxB^BDAQI9% z96%pF43Fmt$(-@WC#ks&b54-XPj^---YOd|bb@VtFe%*_Q>80=w-THVOE35@xpGGUQ~-K#g=>$-26GWO=IFG#cl1-z`?hE@yCRVk3HLg=I6c9x zYwpXGq+8j+6-}kl3FXaE$h60K;5pSUXZ(hJ=2)JTqOQRWD9ke#Ogj21cvhFzl2J4l zT#;J1&-}HH(gjL({RAkB=d(Y8N4LjU(TnHf19JI6!}Lt5#G~cb&l*td^Pz`hd%H^j z0G8y*)+#4X2^4^p?<6}DO1Sg39pedgtviT`z9vL=s|*TOHetV{A5mc~9Tr-!^KKnV z9=#Ug6icwS>HuwmS-Dx9g1`(YzBT8rY4R2X94iK(b5~W83|e+!I!^NTL5o@ecVq=q zo*alBl^$U@Dbcwr*eKDRXBnL{d1{i^!!sR_0?@lEUM~b+&A9dc?dC=r2>_v2G_omd z8(~8CtOFX6Zy)>#2<0G3(0NQgDNrmBt~G|N*=aH$1SMDowdq6zi|*8Cci=(A#ALV8;Qx&^GRTRN!K2rWBTvU*!M zp0L$(WqjqRr`DNP-SgrgK3`3Yp8a0x`a&b5M?EZrvjJ+-FVFz&q>xDNN&?D4sB8Qk z7zG~B`*%#j=rC!Xw1ITQ=n*ig{8t#2k9#OE6 zNNg@1Srao$eSohEiwXXhK>kqAAQmTO%g?+a}J|HahP3(NT@7`tBr#BnXi9?@M4Ir~38$PO+ zusS0& znQYD>_6Pk19BRb4sU`{8Acr2~_kfzK0tfr}FT*Ooz42U(J|@N1r65Pfe^jSpp(^|e zhsfy3`R?=Q4(YF;$=QIM616aVlc~(ziafS!8rN1IuWlTBvpi^6C8%se6B@Z8}=rtE4JNdBu zROi}Fn3qi$f5E=}CQDXH4(s~EaqY`t$IC-aIa~MS30h^N?IDi;MupX@i{~P5o6i7%rTz?9YLXxBm$R@X!_eRe>0Gue zC^K%$ni-a+uVpMB`ekQnZ^%#Vrk^0dr)E9@7$L?yVvpX2DudT_S%zX`vc<6=k@{k$ zN{Y@Ej}4Pm1#k6L*NGvzV=I-AO?-NmR{j^SMtaQ_eUrg;j7f(m(b|mkt;&l$<)xUzDbf)^@;M^xBcpIxH!Q)4A(4>M+668toVkY63k# zRnRVs^n23CFyIu4`D*_2xM!>jSSG0P{- z6f^)|eGmb|fZ60V>TUM~vO=x2TxnHv(u8=jJq44B#6CW1%YuJ;|JmrOGz#-t4f!=E zQbQ-N?zGC}1DlJn$*}J4sHFvS;h2gf=4TMA*UQf@jv0FcdSBkwO@>&|egf4c+J9HT zXP^Fgym!8NcndWPhtX@A&>D>a+-t)A=60&qm54h&%OKqt1F^IfZ~0)f?zPj1gLRWf{%n^(?=7vayM$^ zR0}`9MNmzR&t^AoOPt@dx)kN)tzwMD7X~Es3F2H3f?aUcUQGoI#e4s z+%k4*%UV&-b2oYEx5@S~or*<_ifc$1?LkTn4fk)f3rSzBJmvHXM-=sXF5uMN7o0WB z%)GkYiB4b}Y*Qhe~Dn!;^2@B%A& zv)<9XA0*-6_ThG7F%;r59r&s_`Wq$sn^04Hzw;GLw6kIM`zGlN1H`s3 zgojb^!|U`BDD)=xLU<$TZ<7c`pb%h!>E)1!<_8>KN#rO2)`2* z_?`t2f8h7|Aix1LP(1)D&M3$~5FQucaa9n6TNMcLUGP8I5AcvgIGT6YknoK!{1{Le zy#5QSR2i!-HYD9quUZ))Mmd;3TfKlAX1N}Y!ipcS6EJoMQ&h+avk}t4jnyR?hzJPT zLkNl*gwHk#i46>CDh%PUg3=6tqCBJT=YT%q2AsTy$=eI1Q3$h{4p}S=>!%1^lMKz( z4b4{e$7zHbI)F-Wgx>uGO#l#N_zHcY3>}dJpu~(KuVe#g%YWHMud5X zJtAUt9Yplp`G1iNfpH3fOA38%3?=V0M708#_{APHbgMdkK>iYel@cZ%4z4s$z92p)p^aUYgRjSvu& z=!+N=LKANqL~SP`{T@hdqLMWI9?AkQHn%sbqv`wlhp>(&XcwvEyhTippk(jEz$(OW zKbjy)mSD#FWNoB)g6(+ThFC<_q^8y`dBpr-;_T)S)I!m@^&E zp}_FQfW3lf9qP21p*UET$bJshMJH%Tdt*9VQ3|Ja1XoffvsEhWLpt0SnTy4s9z=h% zWOy~Ha392!42pQVPP#oQxTJ&3vmgN4WQ3Pg#*LHm9Z$%E3ifPJVX)B`jgP*q3; z?gMQ$5>!@*aU@l8wi{v^{SkbbR5097SO||Qp)`D8VGf^kHX>&RC|5HXrlH545t03rA^R^A-T zr2zOA?!k@h5YBBInxC z7CR#6m?CGs*cCS~6{b0-bs(2^;*?bLmh3u)tI?`9CYLrhmulY^P2d%8HD*jpmog9K zhXj-@6qkLK&QQAdxxX*l2+k6JNM=6NQa33d!Y{zvDWN_pHI&W`IEoewPP+;&y*bLi zLoSM<&Gl1-2jiDB1?8$am%yr3YC2bb87^DJPcxmYtO*WvdH6UlT{Ub{R;n8JI9##T zTtyg?_#U&8grJ)AxB#c50zzE`k}h)DOs-Qc!zQRn;jLMAuEr~=!g(l%J*@6GDSLBV z19sx8e^>Y(CGvzf_Y*-G1l_dgO0`^eq_9DcUx6;GVyS|bpiDxoq<^VKjjpDj3<{u{ z?O7C{lTz)!REfu1q0&-oOcy6VQg5D8t}mS@r&>q;rHIM6Au(9mX}LfnrGlL=k&~{G zn=d%mq@*FaDCjXS_v)kWaOq=nxtI&QQgB{lV*##jV?sz1@Uaewu0Fu1w$`LEE~KGg zBsj@5$Kp6g5x-Q=v<@#h_a(2crX(k1x#4oL$pF96My9@kt}>6VrOPFtXStyRzd=!^ zZit{sNhV{vxUloFRnFyOiF3=0O9t%%{I+rf^y5c_?bbB{?5X7vO0}8@mo}c`War_w z_~rJz`}Q!k{I@3U5IUC-#*j44l#YjzhN6*Tk(3r>l+M=WmiCtV6PHe+u8!r0V2>2< zO>TH4X}?+3vLn8(x2;%pEybhA4PVvT(CJHXGqcZ!1D79KJjlAqN&`0@xsadg;fk^& z1EB^{nm3lZX@f%;2zokDYB)-RC{QzTSHi{@-PgBzPLZ3wsHgF{^h&1oPNKAP;`j6u zku}p zt1k?=btzelALwWtaD6Hv*l1F8ZPPRBz+Bb>rt}A%w5ykP5@bSY(br{|!ta}QR66y> zyAF|54NyJxLkx<$t?2`^&3fZxd-BbqlST({={tU#ws}2uGNl#*T?T8+0t#1p*AV-3 zR)(WZixTihy68vMSNa=5N1EvSznk^NqEuI}jH0rPex)5NZEf-4E6#E0_fPFxGV4Ja zZF=+2F_=2)oZSAwbsX3=Hnvn|-7>PtU!rf;zhXLmVTRSCK2hR2k-5^9ZPwtRJ`6#f ztScQ~dm2JLZWd4;s3e>~8N<4wuN^>{l5LI?l`U0Wnpkw5!dRL_Iq~Zcjhp2!eXB9m zDLZ}OI#5qPsm(uT9olBnJbF|z$xMWWr%^-KJS`O1UY|TkwmJiKK2!4811Hx|A=}C4 zHY@pz0g-+laQx605js&WJJ@bErHB?vxjOcvZ2Ae)P`l=p*?az3CzJYJ`rn>u7{RX7$F8Frt-m#; zn*2u>!rU+%%_n5r=I@`T+)n2;(-swJ7BJA}buowC&|*?j=Mv?Xt}+Kh2p4~%jYmDr z#uEOnJWa@#Tjms41f68?n)92qEyJ8I@T!k0tu7KVOaaSg9L-kD(#9jx=0}O<$KB?F zpXaB`FjLxw=hK!JH3kp`mR8*`*H4EEiQwUhmX^)e_BE!F_-Ck7r#n`ghf@|lp{<|0 ztw7oY=1OJy1kA=pO=q9OFb|*SRLmEmP?l^~C;Hqr9wRqS%$8#rR(gmwu|W$E^VMTC zc!s8l_iGa`t($MwHdCKBwG?vWG&WXEw*VrWo}(MwXkBGvtIWjP)aYxpAh_byX`uQ% z4+vurt%q8!5QT2-zN}_cdbmPhX9u{Araou8vV|40tuT)9m3ZTdUp?@ zVZ#x5XN6!@w|uv*br3&nN7QX^+jaM;ZSO{SZ*y$THEd7bV&6<~Ju+=eU=2>lV)yQG z$9)a$li&gyXpuvHLqTpOF#TZJbKhqi?sRNFu4FT|9X-B%BiwzH?4{*+`S9-S5aLI4 zq$xO@JARlSjvf-;eUx@I=C+^|vMC~XSYd%yn7-pYK9XF%+(Nv}CwN>NzFJ?t&Zv3R zBZ$#Py#HNr|L*u`Ob~rSes@@Ye^m4I%<8n0VSjqUe7AUZD@c6D+i@2PzVO%DyAizgp16(CLJR+V z5Giz#+;KlMd7t=lA1rj4E%XqVvG0U&;=*)a#Dtcc0hjdhP)PE)+y0Os|Cm2V-E#YRt zPjisxLClx$3I~0V$Gr{|ZH!Wpw5z4EJJGe5p|h7?BM+YIkB|{`Xc)HWJg4lU=IXM? z?kfJ7I=Ahn_RBU1jF0GV0D#9N8nM?P#>s&BYiOQxmwNT$hlH%LA0>)3>@_Oc91@7a zvu-`x7d&ki9ov`A>c7g=Ipf0oRNPJ%T{SROW#f|aD2SkHiGHYR>Kd3>Slb_6eK5|1 z6Gg+ObS@`napeU;Pl0Ub!wzYQz()1R6V&RFCXUM)N$F3J`Bkh%*z7UuBtf-{6 z{s0{#nBtrDym@ow)4<1UdRsec#$sIX>fu3R2ukWGmc z2vMHMmMfWqrA7nnm`_$>(!9&Vl@>=mKKib%Vm4ZiM(5H6LE$^T(`>MB11pauWLEA2R zC*Q~MYyI#FQj2P@=cvOmFNs{0ii_Tmxj=1^r`lkv8c}?!HlSeS@b{aJuWsDU%ZmBB z1ed+>oNo}o$9b8b8|RIGgq~nfc)+kpeOMT7>_}T&Qt={ulpf+7UmRazeMnZ?a$(|A z3Yi+tD$YTCFs>N5c`)_q%1$KZiKq`Qt(~dnfb5Y-GzIKg%1a?g)y{f{n5C_-cu|FQ zrkh}=RM|a@WNvm76j$T3F7G>fRxjzdKDJB$bTA-pGPg1?TLi^Ck*XiI<}Y1rWQ;W4 z>vX@FIAp4S9!oAaRH;6mZ+YuOpWDaAnE0y)V|stH>q8CU?qKrwr2PEjcmxz|l7>_p zKPCdan$8TK<{T}$V|ClW?X@6Jlh8ZYEz(<2`dY1ie2r^6@gx^eH9ks;weJs~uuP_o z6K*9gj{JMufQq*S^~?9?s8%yvyt{VoVjXF>?sEonHrd7)>KR>*#~O{lP+$ZzRymJN z8@B`0=*UJrAgHztp>JLq_u~4-R#sd=myoj>QID3U!`Kg+_l_d#VV8#$8NZucDCc#f z%gpx(Xii#C{B4s^S<RjrZ<8v!SIKU$f8Puxm=dWxm-7g_u0xY=CKj$H zepm|JY;%#^gKKOukc-*h5Vn`^^k66+$#db?HG~ahcu2HELfjwHnQ5(PZPH$l55n)K z%G_^yRvu9gI=z9Iw#c*Kfl7}in2Jzuwh1Uvy;(3OtDq0)zh#!|cQqkLGOw5XsF zYuR-fBNGVoB#@oNAX6g&wCX^VdC6fV?5@~kc0K5K0ZfdC=<&^?B6vaxqJ3W2k>x8C zEcXs?bNRz!QiT^8PHlX3uoV*`PjboI7Gk8u*-{u1_gG7s2G#2nQno_#NH`^{LOVpGasx} zDju89p?}~SEj`!4`L)7%gAN@Wtc*g-i4l5wqz9++O@BBhiSoONeh{rp(8JfbH?SsJ zM9YiNnm@{C#fq(z59Vj$nX4TWPb73q7blSC3QO`#g-lJCd53S%Yy0*rYu>B#q25;Z znIDghO2rh!(-zy*C2$zmwsOpOW(IT52i;pF#$uHiE^KAJGZWiy-@xUg$p zFiy0BtjpKSGnWtJYwP5lnOq&Hw9qUZB8+0-Wh@)lU+`~R3;1)L=Z_^|KBIMwrdfK{ z_$^xW|MF_us_>PgT0i~PNcrnKgS!OErlI{^5M*@&J3cB8OGc^qoq2`4qFMJl7-&qD zkA|T#EAOYG?0Yzgw}L4ew#O7c);xrrMcM^!J9*k?xuNRFqz!FS;(~^{G7)GA`1>g4 zerqfBTd$=P;?^((M!+%>&lZp7vBWHz!vyx$r6TKLgK=bp$u+RyJ-xWKCZSagzFM5~ zE3t3Mvws5A!YzB^QA30pTCksfW+`X8|Aix3Dv~`{*prL-hE!@z=Qr z_(10oKKwhRBs@bh78fx&{JR87jnK;DXFeGG061L7P6$yGFQ2GIX_I)7SD@E-l?S{U zz#R-J=22Yb1*uJ5I2&9+bvUEu7&_c3uVmtr1cFuU2$ket0Bh)-)?tUpe2A5Ru9%+I? z#k8wPM1j+zlN>R_wrhm{(rn*l{n^59!h`tA4%cm2O3Qg>o#@&|$?ZVJ!+E}f_$DVo zpabp0mZt2^OC`IqgT44LDuRjYUnwqYd;6ZH$Kv4DSuZrOYXf z^xe~R`&?j_z8ni`{27a4;wR!qqRm^pRZQefG<@x+MeRMb54iRp(A{-#G&6BkbvA-F zalCc#lC$tlJ`iAM64ZYn)XX9*|3K89MYQ&Tcsz?ZI*S;TMRJox@{$FF&&Ee41mSIB z-2_&CNN*CxwR*ptP$gPs4nVzFMEweUr@014tot6B^y%Fe)m#>}?gtugUHoKS;(T4g zatJA1eKu`7DSiJIa9o#qe2aQmhl-SRljs9BFH93Rl)yJqr3EE&XC>AZl()#sV^`T1 z-?c&Cv)&zTK`m{tJFioAXK=&=Iju>#>~jcA^f;WiG5yJS!gJWu$#{=+c<^<3>$j<@ z$+(+y_>*<{_qO=g^yrqzu&_6gpb0NNkoXPx6Q9DrS(^_#z{B%QFbeb*p~zr!&p;9- z69kbF{mA{aMMkWdOJqSVVUSB=pDXU33-sUN4JYSMCzoE^;!6H0w5Q2dEy6A&|@VBU48Hrx<99GCo(joLq)dpELg_EwR1;=Ps>ao}zf3;15XN zcU^rkQljZo;$5hn#08|%z9pK(_50^&Zt|S!ccG}9^K9%nN~*+P`tvBqcNOM#H7qDJ zj`B1>6q*lt3O9LTCi_$eH$MJAXax3(iwOAlks1lTx2@|0k=VzO#XCbiR!^mF?F8TSh~BQYr_W0xf4l zBXk2RXMM}&y_}#tW`m#%?hW3?Oi@udEy>Kx#=H(Wo!vMpr!^{55S7)9p}ReV!ktrZ z7wN$Bh0^neRGrkw8>i3>Hs@;r7U~B5 z@P2xN-|cK|?(oeHV&v(&BBMpH=aRI3io!3lR6R_zE=q(xN=lG9w6bW`+se4aG8yY-qlg*CHnj7DUXhE*`JEG@|*`!M|m5=iR4n}?sC z*KShMU+h**TQ*mm(R}3DU0gm)Td`E^1uFJ|FR8pS@kM`qk4oo9TvAP0;>TE0!)fX- zSW+u~>@QzZ_ro+m&{XXPF>s1n%l0TLC^5IGVEfa6DE0^*{WfW_Ndx>(Ce4z1gX4gH zy0&q;pgFqswd0^Yx(<+O@C{w(i)j)%eb)CnGoX<1c0-FQpUc43juzlQ;}hlx0(l4AYEd)7>Y{=s&}XA(<)A5oY6*=DZ)? zZw#KEsLk5a_73~azK}x4md*$=Oyr+-XP+*ymMy=!d)muZ`pZ_v%U0*gR@chb_R7{k zW$QO(qtT}gdkeC=E6iI)b4Lg4lgWiNJ4>}RpHq((&@C1P&FBf&haXNCr%(3s%{H9N z$NkGkyg>)?<%jm5gJ_GRV$goI#c}n?F|qkh5(6x%#M@}OMaNQg+ao+`HIY53xGUu* zPQPTwGQEY*F@>h;nMaX(ho=?@rn_NZ#BV}7EF5vOg9#mr=k_N{@|-= zy!#*EyKKt`d+>vM#bbQM^;f2+cBbcQ@Y7Po)iC&e&+_>OQUQ6Xcqz67z%WDcoxGSZ zLE%QcgK%0s(w;qESR$oZLFb%7 zSDYbkgOLlE@4A^$CMr=f&QJ%yfbGiLrAlP56)NQH;ieK5;rurH2*y%j=K%x;ElYJC z(%SaYE}$T={c`eUz>W}&?Ct}JGXJ(exu{-1h%m=FW_%xW&hZQUr2A`gtLU!2;=KdQ zGU6FQ&pG@(Q-Q1nAq)%iY*j-p3&eUD-=ma(?;}agFVqAU#Ef6)T1>7Jbv{uZjcE>`X=N~8l7R9aC!@)P2LD;GOM1y zL%GO#UyU1c!LV*j8o@#Uwr1S62JZiQzn@Qudr2J=LIJ5}WMX3*_{C((%2d$B{%Fmv zRKtN%!)|Ir!*$8*z{Wvh!yRJ79Am=@bIBIK#>>aX^QMNnpyu^=n98RnoF)ux;THvK zNe5mz8!i%D7)f5)+ZKguPh}NMsNrp4u(Zvil6JSG3vg5totFvp`^v3 z>|pm*sZKedPKAj>H0N5G?OL^_PPN5O)zVIVp85~Z^A3B{0#2SnR=2wwAUWe)w0dgM-?wlc9rC zgoASemvc>neUF2y&8_o{gUfn@>#~Ep4}{Cfnz`lS71t)dx&FT!kue^qgRVVaE@beNn=RK zt#{8|$i$sPgJYP>U2sby{dQp?l$PziggnM4Kea|(h?GB@!&X>h=(b}xms6<6UCeDm zxX^u=OjC>&Ppsm7oSIX-UQ>c8Phx^Ygx;MG7pxvzx4a&lH5#hcz*CgtT&#<)$7AVwOUp;T7O+cB z!ed9yLvspWU57+2ooEUlr-HUiJ;_tfBRlIazDAVRMkUv_w2mWS7 z*TIn1j+oZIjMf1J*Wm=-u7H*$MSH1H%2J$z5_A0PUc`9>}uhcEg2nI0#OUHef4 zhH>2{Y@Vjbp2z8)ClTCcxFCF^Lan0{Pu*&MW9zLsIZkz--B@;>x)82%U=Zd5+7`&v zM){g&6WW#v+U6q!78=@?Gu-9}o|9VzmgWVP_ub~e0_(SJYY>4Ag!V-?v^5OD&6wxS z5x32mwk=D66}p!#D)${3!JX~e+>QWaZ9a16iFpFSU9$FFpZ1Li_X7|2!{z6d3V~gr zm%|+QBes_#ANOOfm*W)o(`CVv&+Z_#v(;@u@J#zjhx-}H%igyCImWsomJqn%`GU=3 zi4SsV+O`jAzdCL|Q1bwnxZe!8UDkJUzBOe}p`4Lm+pLf{5yKlJR>%`6kXf*?uZpEps<+^_KkGFCEiij@8qK`on9W z7h-7!4Mq}qp>lW2^ooq}(g-j~ryDfF13}{vP*ffnH{1B2a;+&4%CM^?kR_*|7W@7S z`5lW+X(VY0B+8Zhg1je|_DizaViiy}kXg@vrIw{&JPm9%N6!P^A5IW;!y&@IDjdm` z4&&4?0TY4@ zwYfe%Jfm<|Zglyg5D12HS8er$67adB@>K5(#nP#@hw{|yjivIryrA;d9!%xRqzH!b z)*byWRcmoa73p8IH&X%Zk zgbTFXoUXLEK+pwS@6LBdQiLJ|+a9h@mRme9gxa6(ua1{HB7{2rSl}Na7}}oD=#tuA z@Wgn5+TO^F+{%=wf{0oAIAR;xzBrnN3%(!z#Vt+%U;7?zZ%~E*af=6QnvE3y zzi#n}!CdQqaf|1p*#G4g7i1X^t^RLr@%)m4vA?^;WtjZ_<`(C6vyKS+;}-XhqA`yD zn_K*2qN0(+TekP_$%;tdmX(tf{!ec4l1~BD$vKo5ng4W)lOdA=R1~Jk&^BJ(;`3WH z*%JGuq~Oz6x419z&BGtJc=X2YhqbpTo3O8LaVr#&ckl$SZt-7VdV^ox;yeCUWDV{M zqzjmbEbE@pC&G|$U=X&JXS^IBP|?ZT00ZgTOwh>+8(E(8PnMp0y&*TzYiAX7?IqMD(|SUX|domzw8K9 zHHa-adwxFU`hi6e5N*S%&q(6mt?5oB?IGFB`wVkInvl<%-LLn=okr?WBXZh8{~G zbjbCg_IG!-NwEb~7*(S+T=rqiO#~r)uVA1P1@~kazPJ7-AE75WJvnw+YWXNiWjKJ! zVAVp7sq!c79Y?94jZITSWB88>l~NSwLEOImSwCd6mHbS;Yl;(NicoHj;!uxwYFHEG zdn##%Bmuje#8hsCBXm=7zol}6I*sg2qm1Ien{@*;U=CYhIzQ*=#E~JzEH!=1=iV4d zee)YQkdEo}1FFdILkB(r%Pg>cMi;|{CHDZ#nA8mM`Yw2F!tHns=UT8pF9-NI!XBH4O% zbIgSFknLH4;kIM~(*dDGjcNOYii+5cEU_|OPkK^hfm{_Yd>$Z&D*Uz=J3t+)hi=k} z1+!)@5N43H!R%xRc3r;4E5K^cDp6aPl{1$W|-l!}0F%$_b@ zFg?VMMkTbqmNv^+K3^{wRy^OVIDnsTHv*Vm?k+_?zuZG7MZ7$mG=N_o&jy$vPgnDn zkmtMo3J=V2Fc<;>d{c%%OLjuzpLrt6W!qh@0U zNcNKAp9f0HW@9G|_EPem2PuSRzb%mLqg6c*Rvyj9X&CHdbUqK!M$LIYAlVOLO+F7b zlFh-NAMEFBJ`b}B%^}>E9N=9#4|g2RA-)|P5PUd~@I=i8B1jF2;$K7t%I1>d4h@R) zUPMKP=8{uM4N0qBL?@2sQgRIq$va=fWTNI#OGpiWO}>aNl+B}68XEr5d=Xa}nn!Oa zHKMw75#KnP$LKILqWN%<(21JQ93VBSi+`CoD4WlkFf?kwdzmyDn$KP!HD;oEnY=KX z&)G0EX5oC9vVmH_Js>r1oqU;kAX~sYKQwONe3^C@TEM?AHQ~H;nSM7~Ab2}8;r?)$ z0YEDhMv$KL#=pu$lq(d)9iH^(y~@H2D-@%Wo(fjI%6>mqD9$xJ748hV$^oJkNlHjh zM<-w9Qp*)dD-BP_H(%v3hZV^hO3x(Ads}g8bnjX_!={2_3xv^%6<%!zS#!c}oOeZE zOcmPlc(02NeP|nUi=n^Cy%W{QDaMY)CoQ|^)Kt7QS{B2F$#n;o;(YI%u&Sg=P1Y{M zBLa+NZ-2)Ea@ew1`JX8mWY#>y0iP13vzCHR*8XMqD8aQ}H-ex$hWElw4Hj{OjwImqxHD+m0jlDq+%R8?! zE?aqtdDUvj&*wG4!F>6}x>FBNJG)RP^CJ2A307DyX-dHgQj0|{U@ZAKJe^?w3uBVs zUo_>L6AibOhCcA9I%-2b-Ty$HPr&5uT_aG11FF#Qpe|}RDb2y<6-#uVmnNLC%GS|8WI#4RJNZp!vMB3d?ba}m*Ulb5napnu{lL{NUoJFP1_bo1%LN6GUk9k=Fc6qIo$4vp3i9Md1lnf8(QnAG4zDK_ixCiP>eydv)T|A*s zRyfaa;G^)NYGgogd>VK3rz;?UJymA=g}U*Z`uVpXO!j>jS6`s=XP!ayZjq~p7G6!b zA7KB<;t{>7D;sE?e^FN^tp6KzwNUyGb@jhw@mz6nBmS^>j!~3Y3IAa6+@kaTuy`~D zw8-L>m8GvNo==T(Qv+K|t6jT)mt&7}pY%ZPP)Fg&=)@%7_|)H7JU$rL8T<9+AFx}y zdyV@NnBURJR)1h^eg?CqPyxO|Q6AHra$hJs$yIT6_uzQJ&R|IeijrV?>7t580mQc6 zhLC9Mj6$&>^t^p~$--RlBu4Wml7ncY7GZ`m1%D^J2@<{}3)L7PKDldT@so;>;2b-0 z^3i{j&H6RJ)1g8eCxyYyQJ7YA4>OUi@*xM&DnQIygMW(()dPXHziwxtzh~aVcigBLAEp! zrhcOwU_e%2cL3K7qK*m2*}--sor&a5lq7n0PNWp43VdV&{ozh5uJj;zoHAv#NW6N# zVJ@1QJq@Lc)*R+uq7kv9L9$tVj$VpYe}YJ=?VRy`n&aLd7Ekbw5=sAthMLU--QfYVEm_K-G8-h_rGPP6!sD#_vxi|0>!RG1rYa#U29Uwl+tiq30TQrS;? zTv|J4a$MH9SNy83ZjO%ssjjL9DNFuRSK@T1wX;7=Pa$;+1|_HUEB42y4IBP+pvIke zQ&7`EehH}gr2ZJxa@J1=ZoQf_1-ISpm4MryZjQkn00dWmc<5zmtFBt&(8^L|CREcN z^cv(}y~XmfEX6olW+wgk4ks+GgdV6SgQPCf)mg zEc{F`zgaXYD7#rQYdF1Gwi;l#U9p`vzg>0QFS}iHy*<5M_e5a4+wjG;xZ4b*D!@5Cxu-0vnDmf!ECI)Lu?GXoeOAP2b#77vGo1?3M%r468mn-RJ+=;;Sbk>V&`E))=Rq=E&$_0M9oRnaCzM56~f7R7ag5}HoK|#g8 zQ&*6er`uO`#ma{PC;@gt;ag$C@&8?2Rencu&4k4`FGixw^2XtZhG+YSx(dUmcvV-C zNUMt?uo6}jPkEK_bZOOH!76HacihlvY*l2GgymI z#DI0ClOh{SmV`S8QO`(%9JmpI*hmP3ISiz**obtH&BcBQerrHacFHO&B~A(r zL+<*ju8c*9{kxzuLqAacLtSZ+vqt?*U9oHLz7zeI=Uc|t3M>izF|9=Jk|D~Zx0P;|oT^;=c!Ng>k8N(XLgF0wBlj-AA3yXoo_!7T;SN&&}$kr`3 z>-PG$Jx}uczRgEmd_Uki0cqV!A2FWo{5XMHsb*3V;X*+8w=^`%Z|?{azF@*3vYo)! z#|ucY@8(9YcO0``yCcnir=fk2iT+7LUrC8_!z4@pp`jDwDUDQDRd;Fd%hY?4yLq8Y z7ix@V{-U9`htp&l%2!()FaJjxTDGZrcO>z@(a_DchcktW|D>VQq!Q1&^a?f;90jugW&eyzZiMLW^5oc}^Y@5U(`6z#^V+aK;G zX!|z>E9sJM=_Un71{r)XEO!b_wm_aXNZpuR989%`_cl!TI2ke|aerx&$n?fBImim6 zEI!B%gfM)_m*K*=WX}*W&@j$ z2I-P60qK_RMv-oj7HP>%cXxMpNJ~qDfJ!$Ah;)Mj&&F{^opIut`+wd4=Xs8Ksm=?I z$BDGb_?4>8HakYccaJtg=BjFIAZhGRJD%Nh&OVa)ii$*@0rd^%DiY zd+qm1$3>)edR)-FPDm2Kh&3BVAkwhbvS-=Imocz|u+yU1(Ri?o_Vgv$NbpnS)8(0` zFR%Lv_L?qscstv+ur2l)ZsBBFcQTTuly$8MAnm`xeNwXDP3X9_-$UxpaL`K;Z*kB^ zlV1W+;JU4YzfoW^GjQ7&ylX|x8(E3Fz&DDV{KxRh_laJ-RWZLjdiRiX`*>W{@$Sil zetIg$I|{|2lWE2BP>ylKH`^~}AJ=KuO*uZbbbO~aGsih$42{AuYxo-ZY{sZ~+Y!Mk zd#--P=cAj$lu^%RJwl{m+WEXBM1heZ3Y@kDn`ax+IQabr>_#DYzgGyV?EjZ2L~C0M zGC3fGqkG-{X>R?ah(W-v3c8h9pzFT+E$K>fnmtTwnJuUxa zMhot!-H8J{Z6+U-*@Zh9OMcp1Hb2wVvICqgYumzTE^FT;ODXHvU~w($TocwR>sp0O zaIahx`O3PNH9{(@N&bOCe9Tk*s~OEL2J~e{V`r;4p3rbaKAAKKzqFq+_LuXXFfHbX zp0P9xb(poS({`M8OityTGVe)!@$Q!yO_NAvNrYtB4*uJx*&V=LfF)4(+hAZ)q+{{1 z%IpuA6v$w({Z9ECCgnf=H2XFf2w{`prX)fzDIqvCs986d6dqH-+Ipx^3_5v#s4b zIx;yGLPpj`6!}1VFhn5ydZ8R0sEZKOv*%xmI=R#SX1@Yd;VXK!4>Z{H7Q!S@;PU}1 z1pwVSq_4XLp9|-9Vgc^g`*1@iiK64@h_CVZ1Y2c0qX() zT3|sK#L(OA+5`~|P%&gQfOcqJp+e8coMgFkF-{oENe0x5j^qF>heF8TJM#i3dShR- zLz6@T+vpXI6)6z}>5|rS<7Q1^{CUZabAcgY)*J8yu4sS)P-+k`nN}-fH7=oB8i^zP zo*onop?ifsAkP9IRLBv;Ojk~F*4Q3aoD?7|pWDA#!iQAS00o_+H}d#e0;-uMzv3!| zR=Boo#|SEML-c&BVRfpXJRF@Ft-X;?7!A-Y`SR8-y*a2sDj%A`f2VcF8h6|C9Yjl> zV}P9YI$)%G_d1aY;P<-{>?HPI;UpJ(b<10a?DsfFoMpn_X*u(R#S7C*?Q04>4IZF% zvOpgsoe1eRm88NG%H>jhNUI!-0;U?F1!OQFiZAH3r%KIi*%1Numg$+S(wZrO&}Rm# zV^@Q`I{8YCJZ&JZq%5fLwP8F8#3&TNI=xJ|cI`7cj%3>yHO|7RH?IEeG!fJ;M}jjX zXgt77oY@V}<$ImdEa3oB04gX<e^Fz-Pu39ED&Z|7} zfGX6VU}-#2OqKgYlwmkFIp#*MOVZq$@{n|t+mwcyyu-kWRl zIBWB^SCvJ>SX!h|U;wAjABpWB`?93qAWuBJ?evhtcW-PC!4fM6NzYOnyPE!CW8J{Zm0))}VlNS}Cr4~lH9+a~t9! zGee^r+mS7NO0NZ_YN48oF{_(z1+Cb4l`xKx&~TJosl`&&1;$o$QVE6G&k~U%D2D3j zsGz_X$!gPz4C0S>WR4rXE$~`bVrE}WG?UhTqi!4t#T6H&=$!Q-RX%m)XhYCb=%pq) zT9qBKXO0~qRB-Ajh)d(TAPWxH)4QrrH`p5k)(7Bl^%oWHlNpaq`a-cU22Y6 zBV6#Gih68%l2GGvHBh z-7Dfq_NPu$RBqAt+8eK+!t3+I^&DTwZ98$=4|#@)ezcn6f~^zr;CYmWeO!==dryHC z2OQ=&OBqZ71w!14w7L~Dh}I5B)XufZ)SrETM*2!hq0IpE+SW_ezK)fJv^+ft07ZPw zrETqPy>KjzM;{Q#K zna@mlb{y+#JV-8FD6D0P<6Bp?|EbOB^A*&<`0F5|5T$+P2FJmIic&5kUzO${^4gtcU0x7>3IeOU)D8$pzHEy4Z=GVlobI(&Z#0Y2=;3l((bj=Q<<_NTM@_K`5W|8FwS1q}?M5UI^E`5Th zarp2-fl%?Py*rlctc}F`6KWQa;A2VKHDx0mKS^a=)ifVo7Yfx~N`LCL05)u zC?>Gq)Ah$e*XLDG4R(Q-aFh>gxhZ|I00F>A8oxRps`uk2(Hr7Voq(ryGSe&~(IPav z%YmxK=5kd&WEB!oYC&2(vO{M9fEo`nQjT&(T|Bhl=ryU{qrl;S;Oj9%9w)9VX1dl4 zb^<|b&Mb%ajsT8#C@`brGff~E_PPVrn^{5*p8 zIinyR?=|U^-DoG>vs)K; zb*kO{I1{{3v*&yk>OlW0HWW%7XmxJmk_oio^hxF*PDKtddZi8i03xn`hg zahO#sSf91E310?0aES|(W-;j3v+ml6vwCA(*k!GOle+HBXOoMBCuoBYCnR}C;v@Dg zskH0eKAzdabVO=4KR>?JKAJ-VvXUyBfIz--YyZj)bA5JxqQ?=%g;I{59<_yh+P>s3 zVd-VwB;)6hoB8vqWm_PLx}KOODhYGjuq%8*P%7|3Qj7|BVbAQ{Gg4qlr>Ryog+s5m zPF>E>s9`}x6>WIzEq!g8nUy(f;$Qb%CuIf}x zP(?Q%hAq{ADVRMZiPBjuPnsQDy#bDCyN%^23N!X16ndS)qo2nglj0**k*J{yvyl3A z#VN6aO&rTAMJgRZ+n#bMTwwDPBt{tWJxaBooM^vrw4O0-oGG!95%*qk`h|MG zBTK%4JR^+APHYhjOn%37(JQmC;`oxhNsJfH>Q!|zt~{lEkDNDbG$Og7sp0)UI_ogu zi)Oi5cL%RMfm2YU*NOD1y4 z1}6s7dSYFTp)z|*;tjysFMHVOS7uOGq7i_|#TD2qzy~jac=kfMO<)$_p++QG=;?Jc z-`zuGtESbY;23~rlz~~`DiCdBXLqTlbg7X{u90r4k);ReJuq9pP1b=)rfjZtjW6jV zSFEq(WIMw{9&p6#8B1}g#wmRmqn@m(@qyRt5f^Kg&}(+8^0qKKMWZ9ihy)YP9JRtH zGoq%&_Bxz22R6$2UR)fw&@s_Zcn{1?vm;mE#(4SJxFm~eoD&z@(h!V#L>ki-c%hM? zKb5Q!A24Q&?#n;$eL!$z5m_uM_avEqfU2-hM&~3nw*k^x2F55><3 zE_At99Lz8NgwC;&-w%esm&WL%$KK$P^}XpVL%1BmKGk+}QPkMtxdWl2D96*Nml|9q zWsXbM3S)It6rE({#Y{*;2Y94Rlk@QWyZ( zZ4KbDa}wnFX(I4Mt74^+_dD3=&oz}E|D$JC0jUKt^+;>AOjS$u~;vI85@f!V-o`T zKXapw*r2?MA(c`C)hoqdLf!Ru3ew6KP1N=^<&L#*_Lu1`GeGoyL2#YgEJTZE>c_xq z_=d?Rr7sF8mQpGbWZIL<6X`75;kMN?5T^)@MM+G{Yj;5X+!Sq`ATRhpbz|!O=39NH zm?Ut6+(*%(334mVtaufz#>t+>38biv4ZS<2T|tS7lNnMIAvT;9#*O+(spHR| zcSy)fj%lK6XIt9j-LI1)SR5fiF9o&RbQGqm5O$ml_MDY~E-dISMI-Q%v^ijNYNDZT zy&MyMtfEvUIoE6NVuqIl6P#W1xtfPT#I?k*THx>X1u4iY#y1~#srFFMXMz!L3DMCFWzJ5^el0Mt_#+XMB_W()TQQc6LOu_jhL21 zW3NbpNAbl%okfL>Z*AL#?Qny+S6t1NQN{lWR*09Iogln2bx!LR?D?(cp>62bZuhjX z=GWZjH;Cp}Qs=kk=65dV_aF!j{)Hp$1>=hOle-#csS7eM7Op>QD_mVJ00kCdbQVGG zi|}cSh%Jjq^NT1~i)h447y?UJI!oB@OSoxE_$^C>^Gn26OQgihWCF_+I?Gh<%QR`r zbS=va^UI7^%MdgQi@*w-&I*V73Rl_+Psyn zMr?JpwuZiLBTsCnvp)Z9-Kho7xn&HdPo`~?=xj~5z|GEYKi}J4941;R+kB65YuSCrHezR!7;bBR=jq=FVOs+8zul2;{M+6S0T*8Ww3`Ph)QnEgTl({=TXrB;W#z zo4&7Hu}^rtPfBt?cK?7v_khadfF}KbuJwRn;ehe_fSKfw<^Ca??jeWAA(z78ar_}$ z&tYT+Gau>@?|o>|M=*lBhms4(66r^>3rF(TM~WoJ1ow*<}{cHf3 zkjZpzBf3m!txb>ZD`QKi3xrwN;`4h2IVt-{^hXeBdhH>3c(EEjP?c7g4 zX8!iYIO``3wAOv{Q8e)O8iTV$Wj-raCCWB8t}rL$rTlKt3A>eB-RLP_S{ zRxfySBt^T5-X=eUR>5Xg^dxT6+s`w+qn>GZK^WJNo6X-cuK#eeQOy2w)%}HW-T$Ka zH^%kUKe+1Zm*emdePdihuDXwj2pGR-Tz|XjvVZU`GaUHoRrkzzvOD@?5sqW_!EaaH z!STPi>V9NQxPSlZ9qdUR50~ft`){wV=1!rYp#E^x-9P|iT!}8h_d_(vK;f@Gp_%`n@wKfT9QTWhc$?=hIaisUkxR54>i5} zRiSi-SCX(vW)5Lc0ZkAiqI7ncGte|AMpn9zB2EWUDldtr!!$p{6|smcD=bK|Fc06$ ztf-`R6QV(X^wm5lzPb>v=%^L4s5B=Gw!rUabP=x}j)StSF_e|Dyv7~Z`6O05?a*#n zr|Ejj3VS)sMXX^>)zhj`rDWhwd^NmFR&6{;vdkzT5(|+I#5Zl$oh*8A_pnhFN`zlw zXM>-06S~Y)Hj=1RKI{nlnRfl`oQ+~U*-`{tCrWKF=i(2uG zhp&Q`{x>}Q@XYovJbb%O`T_Hy($}n`e3@?s@!yR1Wlg`WofCvcj$RZQBBp$AvzdY~ zwiSk9ywS2Qv@>(=`7U>ldg$y3fe@Y+@Dly1i-Mvz!Vf5RD^L(Vfy_XJ347a!4`emf zO>~gqip6q3*6vO9S3Laxvs&>x9{x862%enZ@E0EbH?8=EhyPQp_{zio=ME6|dl0P< zcmz6WU*@d5fB)Upl_jj+9~~ez_w;;(<<2~Pv95gd{lt$YNPghqf6c%OPHi`zD3#`{)A1xd_A|l;g%=9GRHmPRmq}JAL;nU@m_U+bP86}QX|AtU z&)%g-)RL{x){H9K1jT*F{$0W+~_l632q0|udm zv6?XrAXKw16VgV)Go^^#*Ym-WwrQr9_|*X-r7dWS z0VzcIsRyqRmEo^Ec+Pk9ZajFQf(Yi-{9e=6>2AU^R9!@;h;|-&oU=5mmnZY4`HOY) z;B5ZRXB9WQUWVwz-3ke;O)tJ-J21NI#@+V(Bc(*oklJ#sq$fpZ@e+d4W zKKwUliSP6Q@bC12&-@s(u zOBce2#=1=bB-IP@g?lMj!@7@K|M<3}QFG-z$)14!8?(gk_2K-ZU;qEALo8Hl_@*!8 zA2`H*(wFgX^??#|U#meXPQ$q4aeTZhf-$8*7`IfSvACCUQg}(9F~tAT0P%lNLHr;7 zvw=?_{ts@5{{sZ^e|SI)IX`uXB^XQlWhP~{o96F_oWts1FH1{fg@F?!+VgRxI#FSnMH%UaeBc*|NBfwQG;(-yLGPgmmb|H&=-&H@(AGgt=xnl`#V)#WczUq-XZ%5 zSr6peN$u#;7gLY2Ltnhrdo<)QqeTDJqdQ8xux&F><^P=z6d&U9a`++VnDh^Fj`NJ; zKe2mA#GSePu@y%V+aAgPyH=c-Kq@TOM6%R0a;#^6)rylY#6*h}{O}bear{jyjz`xQ zmlvO!Ix-pJyPH-VXK0e;MMx`7fbPUb*y`4<{|rf+_VSMRsr>oIr@bpls4E0QwQQ1h zZ%Us#Y*=?>I(>PFg?(%+;5`G)D1<#nd^!UxSdcIfxWrc|VC?9w=g|W9Wms;b+Ob8f zi+iI-^XtgBBZ#kKq+5(fW^lrF$i$F?foSd3FcRo8o?#hMk3R(DliE_)=0g?0;HgIo zl%m@*ASRQwBni@YY)uoYce*ROtU%4661(ERcVfy5s+1Hnu_jotLJo76m?RowGuCUQ z;G+=NtM669h!}#u6gkL!5JpjWV&Y8BjErhmy*WrIvh!e+0wY`RGV%k`gi~Za?Z&lWaJ!VIf?DS@VA%%h|qm<|b zIJf&QYC1tiQxqgsE?_9#>}?|iRGXX}14=&9c{mF<-xV(snl^Mzd{3Vw%c5c49a zk#^YwymSH9>sp8r?iDc_*1}Shfm{jVi#gmd%mv7Eg=mEV(}N#iip+YU;3W;-l6D9a9`>FPVvf^%w!d=H)?0VD?s zQT0pdyx=w{*W>}u-O!kgxq#MY@d_V|qO`_<@7_gHkNAQ+Q#2B*%GME2x45)BkVFNi zL`NdrRs}1pgP3~_tGfhEGK664#7J=@Mc3Pik&`d2M=@ZJJ|LIQ!&Dsd;>PURq1+a; zJQh;BHM%u97Xb8WrPLc?pZ?W}Ux#gQx_2`7)rl`KN~N@Q7vjW^k9LGO@rmn~!XQq3 za$N<86aO7Rg&U1P58}jkv~U->22t4wbiLMRK)Kj_AHe9fNh3(y=gm@O#mg!Aa+~KZ zZA6(^g|wZ+r@g(iE6*)a+}5tCG`?OAG=(!%0`zj|KYkcDpCqASve-wb%YJn1*060|d!Aeuiy#B4dE$9JyeSq$7Npb94|x z7u##Q_WWu;K15dFA|Sk|aCX6*J7hKiCd+v~zb~7`5>wZ(KGhZ=9kRTXkl^=z15SW9 zCJXz6BM{UrkNASOgZixQ)iq04<|E+{PiS2=KZd0JPzSvg=C=tzZn`@mBng@5d+T8mLN_+hpW}Z-peYwoLwJ%ArY;Wlh z=CSD3L2{0maM-SSY^eZV5&6TUt%AJUXKG$<)|knQT3#Ic?VZ+1hcKf)Ks>j-A=^LA zISz6e`%{hbx0GULSkRcDPWQq4xi4kFr*=Vah;;cUd1j(%`-zbED0?*#9Q&;^<4x(I4M08RNJ>@-kmZ zY9u_05z1cPZIky5_uU2M?FdOl6IRKrcRZqc`VZwcF+1pBJ)@Qk)=BB*%}m1jv;b;0#Q5^PE&h#Da*=@u&JV zPO1`yN{l&O{L5@q<6V^+V3LZxw?2oB{{HdeZI>znB=_@7K?a29&1J1(cePC{HUp4 z^!@v)TRQZvbEBv~{Q=T90ZFjn2lT_;Fq&ZkdX&W6g}PX`vRE>2qkC>3Bh2H9)$ z`gD-Mo&aXaqAN>%J6c$U54G$PqUKo#?r?=4=im~G%ULHWN@akA>@r63Sr?6bWzeI+ zW$dQ2SIl9RAx~si@Mq4txrQsl90ym3Kb`dmqEtor%Nm(c)b(=lu|&lWu2S)y_bG-| z#pKJrr_(&|SF=)%^C(!o>vBGzhf_RK4cwMow6>w&NFj9 z>@-}Rb~?B&@agD<$Knimrsi_g86j#K0k80+zQtiom!b_{7UoCdR0D`wkf1U zsrS&dv14Oe;F}4@bx7KWknbeN2J@>k=+=zo6PJZowL31~b%}k@_ORA`kb|m_SN_Ny z*m6p=tE>>``kDBwv>m)>2j?1w)#}tf0oV`a{Jes9|9Q^};2<{s^LyIX=L66ohlN15 z4gPfKhG+#?a%$mcEN0Y0A2W^}u0Kz?pnOVOa61t$xJ0vS{e(4@2GdLu2-%8$Cc|1d z>tyPXBOlhegG&T@6knsOS`MAVOoF1QC*wukNR)ahF#3kc({xl=vObP*b2ZTUV#tZrB9?kC499-DW8Y)A$k*IG8TzzwC#M@%pq9-suURh|iEee< zyAle9?aX6?+FRJ#8&=JOslfR@ypKbV7GE!z7uQEz(MK}hM-d3%o)CVx=el;{!vYqf z>xCxJ!>4*=F+R%Xr|-tTXD3|YI~AufApv){z|2%lNM6sKzrrs-+_%r(Pbq`1-}~VM z6Enhf8Ff8LCL>IjNmJ!BcyY`C!&-mx_kJ=N0UNba#9aYmi#%1{qJHou;wsR~)u3MJ zKsGQSO%0TW8E|G*T+6XF`46LFFDpd$@pKajmE(BMdnG6Kr#L15^(g_5nWuAqOCxFr0DpSIk=+{6QfZR!};){@193v(Fz}Kz!c`g z)E=l>t%TGXyKl9}jO#(~b6Ism$EI1w7>i}_s3uwlGRIQ)z-k7D$yLN?o|~(k$Bx*> zI)XzvnGp}IX%jDnk?52X<$gR?a6&#k9)5qU2qS80$%FlUA*VC-77Cfa z6>_>jF`N$m(?U+LbUdVxlN&YFJ~1OZMe&yZLEs8-kz}a@+oCaF-P2YIr5PioKS<5W@f+N1rQYS?hN#zFfC|z#=8tv}H z5zS%v%JIdgTK@M=r}FTJQBU8N*e{ap8P-M&xACF&efGwv?5SpkGJP=XX=T`XbF2vQ z(f_fKQ^8RD?F2O3pRA-Fgboceb~3l+j;cN!V~`pm=xJ>Z-Z1c-fG_r0(|-G~LmX zpB8dL77X=jkAT=`FrH8EnA}9rZ1Od~7INy4tNyf*Gv&?*sqjOZv!G!2wt(f|B4~y< zTibsgFnLUiBTuUrs|pNw9Cu|II1kkG>Z3%`NvZ=QJPAt_PJGX8Eu3VQ)VAhi9jEvt zg?x~9KGmrx<4O9#CASzpSeVKrbB8FXz&VI9NGH3Be}g75dK%6&H-Y8Vlf0GoTH5?2 zH^j|&axT)ff?qujag2+)C`SMFJrCjM@>b*t{=HsBjCtKe|QvaF_tf4dIQ-g6J`oUOK;+ zl1aGN>4FN>)yAqOy23g6L!HBO>cv-nI=a>2Es|(&?_Z-IOfY$L;UDRaTmoW5!hu}g zP#Ak)`3Ni^HB@UBUWBLUBp8PJtPNH^%v#Yr0~YmJBl3wL@60~wxG|Y9I!(ndh?TP_ znWekm@eG&&oB%Yumym8@hekSNs^V7d zSDlb<;qt)|NNC1jY*rayKkF7|26_4YtXueh|L*$xUFknh)9MD>Y z@;qi|AsvZH;E5gj5dRYoo=XvrbIMCNQkj-i#YOO+rD^~5uGEOnCx=&>u}?l)z6knp z%!aH~F1MRM=5&>68crj7}A~b3VM1#-F`ppEI9#+K%rk> zn`v5sQ|=#|yeBm;L%M!ov;A|IgS@2wO_R50VAxHQw`oLdLgFLTq|o@(j7)wsniS%! z8+yoHKnbMDJ0YbC(&X)wPgzs{QX8=)6#wROnb8&m)$`wL^8WVP%n0)Sq{;iImrFAy z=zr&O_y@0zEmk4#ah5!?ExUPqnQn22c8urtOrR;|%OQeFtITF^k}D!MuAL=e$ro(y z&CJg-1}({-n6dMCe~>FRd{Q*GcbGnxdLLQgtV!XZN+C<(X%zRT_uUIk7L32rL%zRE z1dPLClbWPmd%{8@YFjv4_|oM4zxvwj{tvvi?+1s<%4V%U&#tqG<8J^( zzRQa80^|Vvz{h}Zki??0-#^4fI&ud5|8-_7x0`1Qz_i1;^1;^I#u z;(zvQuOhS4|1W>-=@e=8{U;9*0+IOrLsZ2ASA7qWcr0xMfk+f?E)${u;PB!E==`94 zCO_3a^zVP)+*p9KXS-N(?4){{hju4 z^y$?MkL5kflDC9tUs9RwK*(JJs_kjUrz^&VE_MC$Eiyn<0UNIUi0*I?*|KKx!!?5R z85p>ijeo}NKuz>?Io-NF7Ds$de0gzcx6s3+aR1u#>Jmp047?K2^E}V2Ai4J1BJ9l& zKoJN~@WGfC(f2*l3efkcyIY^hP3+F407>PWECP+*_0`od10tI+Qh{V3n(fmV%-6(=rjej>{ZlF^k^h#P_1OhuPp1S&|h}Y6wwZ; z^mE}ew$qPEV7x6W?Fy zsHf#Net4i?L{|&Cl^K#U2)sRqmwvhi?_AO~O`LMf=dP4j6VO0)Wq|pa>?p9T@ZPZG zCsB3InFX~HY&i%U$2F3Lx`TuVYO>RU?&H6ncPqI7j{YlDS8q5XH$yvCD+vR_=3 z{W=-7KckcFY|fXAQr@s%ztYJ{uQ6$L{v$fsEm$7xgb%yRtxSfu4;VBkg60??zZ(WvrJu^_= zM-gim@QT(fGf4Fh?AIoiq)J14-fWz)f}gQpxx2(kaZCI>jO2-+`+V4_EW+Z;AR?vxjs5B&b}wQCS`!0STCD5MG-g3v0!o?nzZ#&jWq7bg2LwH06f7k}0BZ%C8GAZ9I`BnZ+Z zOA0YN4}vtIg(Ax>`Z6=wmO0nfG&D9fvz67q{BBGIx5c6lj&_!C5s#%qPYjGt&v78n zQ_WJYI4shkalKt8M|;l>*ron(53o&zp>!aO`icAM0Wl!6j;gJ}nzkFtH9A}r;G&#S zs#doN-;NIyV`x?rCJPPGNfA@;8HIUukMjO?j?uV+A#ym?;}i=UMOmLD#&B>b{WJU0i52nll-gNW0i2Jip0`JhHAYY3LVvXwQV>lDL@6hCr zcx#fKr!8-&qs(eD*vPDE2fg+gBNB%uJ?6Mn@T{WRp6%NUO*u zPb?osoZ9EK8kD-zaI&`P47lCL^^@=5j+~4g|FX>^;qseY!f|l7~)@9XP$U`G+$_w`dkP)1xtR~M!ie^q#%=a_%xqyGZPIw zG)3FSmqoK7@%&stb`bZ@hG78z6-|zxkTJ$m#P3*Vd+N+&v1&UK8L<>G3Gd@^gW4zq z_2Q;N{f#y&JrgbBJl9A76%NqQ=C!8l=uhVh*Y7>w`H;`-c125*p^aizz^vm=x{?{v z5w(#OD(>8$9d*Mx`?{Uc5Y767)74ozT!Ct^cc@D+x;foHhwB)=jtG@QpU9 zHZJueZSuSAOoev_V~SZPN;-r#nJX{5Ux2Mkcte})f7X-ij%1TZ7Vc`=NA)P^AWA3~ zsjHun714IkZ&pRd(acMXk?ty!;do6Z-^$`ewuJ?t)V7>+Vl{UHe4_Mw?YCshe@7q-oAV$|r zBm8EiIU*3+(C=B^mkvpkJonJYS%40MaH&-IcY>6=NE}PlV9cScgI0=PV>AcA z1TmY12P%S@LZs1vwjn53>c}_+w{PmL-0t1G8(n93<=95+`^=6$%)!EZO(3%E; z0lyd?Wyg3sF*!9oGiyISx40Ba41odruC7C>7PfbG_x2ACkB|w90W8M zbR0nMdBKy&Xcdi5coxk(6U&Pv1}@D^DQY0#^Z-YgErmcK$#CUga? z(n~W=Nk_})0UyU)DoG8-s0MjGj?1vrGAv*& zQZqItXleI?MJsq4aoQ?bxNiQR!hk=*lr=!Z{mC8B0SXL3QJ!Wz>!EN(I{0DScI|Ye z>iqolhwt7SFYxzN?tkJ1)+qmVw^ZRI`}e!0UmhYicEk;pn+iKB{!8Zxgc?t}rBk+&0ly^}MO$QTxAB*U|QOS7o8P{gZS zw8-z87Hkf-k9F2+D0j(^?p>0gU1^I@cN3C%bDd1x^0E$Y^F>P+4cI8_>EuL|(#qj7 zS|;=*2!uaMAOpV)xetFmuXIkHgC3Ws6%^H1nKzWoWxF}s_jBUoPt${!q?W(iEr~@f zLHjlQlnJY<yf=K}MUoOh%iq`M=+nq&wfOBX2k5%;TP z1#NbjWOMN&7Rp7*_bcVbiTjx5-LmXAz6CcGw?PmdRx+8#8TEGEn8BB zC}LjP0wF#!HSQG2l(#GcEh=h}WJ(yiiEqhN^*44{RF5F8o7Oblqz4(tbzEd1OxO;~ zx=PnuJM_z8Gtv!fWnNZ|g^=_hErbax-Fy%x{Pbn!Z167i1*F`s^$PUeob>%c9&9T3 zj1F7ZRk|BLzT;U>wkwhi4O#eSsXi)4FPr}4lEErU##ULWK^75V+o7=85LR-Y-4e;y z0^~ru5oz3j*REQZ+c=y=?foHT$Ic9?Qc z3FRO$D=(Fpu_`yVo7FSoWt+2`4Xq<|-bI#JaBqTgTC_ltugCWz2@_uqN^W;rG2%mM zz>QKs5q}@&;`w|{%VM|zJIy0ZEO{gA_f&3q(zYs0A(y|Pa<3q%{DjKwC^GMl>iY@< z1}ls&M^_>G4AObhLX4jS`PgN(5WKz=35gaqJUeamzP!4%h-pLdOXta!7YD{KohOnB zGe;Eg?TGWc4g;Nw$lo)p@d4bBwDbLMgF}wjANO?Dl7p~8KXx?z3%&X>IDAVe-3$)D z)2piL9|ni-zUzK^eo)0(T7_g6Ha`6A;1K2o7W}*O!wNGa zF5Hn>viaG)m5?wJ-4KQbviJR%Oom|v&ks9?({6jAgj%zidgm!%%9d)7ejEqrY7PR_ zIJVqtpMIPlYO4FWUIl9irZyhg-;y#rJovz{=dBoHwOZK z>1g^1!+IP31q;JZItJPu7~fsEsK&S7BT z03-m0@BZu$|L$fn{_ID`-X9;_9~R?38xUGhruXVyZ^A+coIhENn?QwGh)2deL}Wvb zO-g2qPD!VPIQGKebL5_8Wpbk#lu`otib$%N(OytjHc*sSHd3?-ms8X;w$q`%=wx|C zi!Sw=ZIle(k~sfC(GcYu(}{TDIkrXe2>Kq`h&MAN8)Ksby!+(K+q64!t37=b$8D1# zXKo4k9VvS8X8#_Vl4$nmsz$Gz8L&G(c< z`IH8%VA~zDgxLH*>H&nK8Tkd_EEF*tutsqsX^(ueOcI~LzrmNT<2!y<1!^ZM;usF# zfy=&CCTI8nnV!KM5q{Z($~o7b@Ewufb0JLqY8q{vLsF-QRkEmaFFBXV?l~QsiMJ-0 z?QD+LqQG?0iI*i@WKIjlQ{#8`_A&TP-wyBq5MgNe!nD&7##PAHLL$TNq|f|HyMWs> zPr8@Ce{eE>%e~R`oVKpNLgVL#{WiZs;~hzBTZ2hR`1yWWjQ=vJ&v}F?vaB4E!m>MwWxwoVg4`xrCjv$#F^&R>B>&;C~ws5Hv%iIn>i59ICdBBLihr%`5GiRq*`!CR zzM_LXxCEknKxy{8`+!OPGM%(lLqJ+^S>|S$j7ifDBI46#@UTpI z`UzRX*h`Vsf%3GgtKcVXz!@p)c01yb?+1imB7whi>^)WP>ZA1j&<9VSpYnKs(F36x zp0#JzXo!6#g<_aTv(NUm@)c4|v@mL^(Wp3ipxu}_;}Bb<{C#Ajx60RCw!;rs0_?{{ zZU6UcI~d`Dc`X}vuQYaB>Oii0`p62d;@u?SatHm<9j=ENe5m}?V+!R`$KxtqZYR0c zs52+?OhEUuzz_v@0?lH*t8??5bl~Hyi6sHVk3HetpHFzbS`qfvNqVj*aYL^^0|X6_ zAhjKYLYF8selS6Lo=7m>(DwoD@Xq>Pw`)jomsehH*vayG(_5o7nSM9VWfz1+q09%J=0%c3y9`j?|TR1-Ys#y<=asOU3AQqfhta9 zSlOhnXa#<@7@uG%lGisb0Y(debJ?u)L#_Cr30KtoxW~ShqHS=n9sH~mjqUs%2;65t zVV>`@P_xy_e1<#i&gsR47&4-bi|F$#{T# zjvhQ{tKpH9)z6(UZ}4D{>M>KdLe`1ZX@4n23-N(Hk$TO`#}e`V?~i|`^>z(E96?j* zelr^EF_8Ylk3oL)@TcNYu@=N}iAncjk|8LWSmaFEM+sEHA!;d2yU43fVy$R3j%;)|qQXRquY zouH3=jXt-)vweSuWU)!_{W%3x;aXvPT__w2dh7o~-COuowXSQ!lSxe`-J#OmQX(SK zB~mH^f|PVOBGMq;4bt7+(%k|g0#YI%-6hC7FtOHJd!N0}`M$qk@MDhgJokO&ffurc zw3@FWQZ%;;8)LihMf^ZU6v4toswd!PjsxqomNqX&*;7G&iCLw*AiAAhSWOfiq_BF$ zwaifq;U--|)fVW%@~|kOm?hYov+G`T0CDhQ8292LB=eSQogG(?4o>cKzErKo5=o`2 z3MnB3i9)Wg(Alk>t#2FWqNKE&q&`nFT|*5cZ5P#h`QpuZJPGCKv!c;fw-7$O_>*9? zsRw0C1D%4LcM+xqu#Q8_vS&BAv2K^8FOD2|G*o_ImiF}1EL!Yly)zVIj(pLi2UrFZ zu(NhI-gXe4NS}*`IKo0Xi;^)xxc0Rmgf|Xl9w0$~L}i}RDm9!FS(V$uz*1Pre=P{v zB1HUB(e%SvKkG&LRcPQyNt-$AOI{pB>c`Qh=<3Tke?S(%v!|G38_ini0>L`cn4&KZaxs z{+ks|{Rqf_I(vNC=r7gTMF6Y!r-G0Vm=^!*6-|?w2q9XckKW0t20RjoF_o5Fh<5|q zpDxDxNmo;+CzwbqW#4SFUk<26yGBvmpD(^zPT~4^&2y#i8d=?GUt0tN97`=x3yF;Q z79BysF=5XQbnzV;ZkFc+!?EfoyXp`7pA%1p1}*RFxZh|xnA1{g-g|GV@lk)ZX7eNB zV*7gQ@yfUw{Kun;MyFH50}{|_P(4g`k~4xZEErqQ6P`^L1$&EM_3=yLBzzo5FQu8KZ%V!Ew182P?aMu4?fay<;i%~*KfeA79x=TxrRRNHo zSTFvn05hu4nL=4k{TeF;;S=??!|N@+#wtG*V2UYp4)~TWrI2cg`)Ke_2U!gP;o%)( zqG(&eRNMrVG=%Obl)fRfL6p=t%_QcHrb82t=cpxKN$_x(hh-Yc(Fdx#D8yTY)7oh- zt3*p0G=46?d@SouPAllIgaY$@IA;Qe`Ul`!UpC%3+i<;K1$T*N;{Qp(ofImd{(7c@ zZipEVY`lI%l*#Er**W9AS}(r78=5tNxH!q*w80gLWAZ*7E z0ySjM`Ha)*KH{G?c?QWe%tct~GGr;-65z*((5Pw54W9Fc=BKXFk&9L9fG(c1g3$Scoz(*EfOHdXKZ z7a!QKwP>afZuVCHRVbFZ(X#*swR2O?XL;ql7>8>`7~hZ3v$^)~1+*O{xj+2E)QIQA zeFU-ne<$v{De8O1HU|5LoPmb6V$pHW6fonH=#x|3K*||3iD$0EPhyZW&@eSix>}=_ z4wWt^v$Cbtr>>2o!??S%uRrhX7xw-93r7-60KYKbB)~QE4oOXXrsNM~o-qE3dYu+|l=Eu4A;JX+EI6}3V)DvCc(o3Ds zsz(*&-O%#kG?1`G59B;3)b+^3h#UuIJPDsb$i*-DaL)=;#V^YF&|7%9dNc90X!?=W ziKmFNPTB=~UR}zR4dfdk_=aJ8n`=es-i6&!(EqXc?yu(B&()*e5gdPa|Ca&{2MRa59--( zAfb$>Vw`5!O&!G}Ugg*k6ArrkCPr^?NABDW88FUqQ|^)A5_Wj|1ZlrY;t4&K_(!=N|WuwZVoFDHgv?Smii(e58>- zU>I6I_^mb=I6j1ILKtDf?L?zujZotfREf=F64TP}qi3i2hDGhTI)liIRj#TvK-I}_L8QGs5soXWrJ`6KU z-tGuHFKh}1$GA7TJcH#~rG~lx?j~8bdXGB42*<`wu}JHsO;-el zB_EIWK%tluIC)H}&2=Ns90n%o&GEu94!P`ZncW2wX00vh$E%gO(H6 zIu^#_jvz|HOz==#NKth2l_tH&+f`wVlSLVjoTk+DtPo4cGRiPctk}2P{&FmalumsK9l(*2EFE;h!-MtbE(xAni$foV z-Dx{pi?h$x;(!Y6^J{T_M+5v`+OG+#|HylEq!RpPEk06-{dFzQ36%iWV)RLCp7Fdd z&Bd{T`Av9+Mkp80*5V$#CjD>P;NNS>{r$Hie7+E?0aCPErJkZB{u+alw>zwcF!)f*+yIVL9pFd-2o%Aj z*V-9BXp7TRGw_ziYaJmkn=!Qf-f~xydVlo^iwokmZ(t_<_nqw5->6MFE+E=X@0Up( zOgyECuYBWsWi9_zKp^6yxtUwihqe)Fb*l3*y6uk^UOhpxKebK{BvyT&9vEZqkXewg zvYbzW_;{tQSNfWH`4qFtTJ;~Bi%)}$cG>`Ku=$7@i^|^EhqfpC6;IF|jb{UIC>_ox zcsL%BWfZd%0u{g zT*%Oo+=bb&&;lG%DQP4CZSbI}Z!OSV+}SB@zZIH@l1BTrx!CU>^?q`Dmas;6Zfy#K znvyHaOZ5n&+`HI%IyY}Ev4uCFq_aMi=+TT{j%<@kXVb@#xWwPU0pp%d5Ys7!${5N! zS~$;Hl6WggJTbQJ@jQFQA^f$F?KpW>8lG3tu$o0BNW+Z6yi>2;9*vN~5;W){>`3%G zPBz5fh0DBg!Mw$V83q=Dn<;<~?RvnrlZ4@)DcrB=;={w?wQznPKTFRu2S51EKO!~kU{lq}0L^ZdGTp*Aae9|p2PB< z51YO1^|brd(FC@#9+~dG5#B-Kl(}Ks7M(os*)<6QPBIP;5r~|d)OB-~ zxezO21O*ZACb!ploVDqYE`> z6p|OHT08kRQ1bTO>6k&qo`+(>sFTVFy#wvKqydfBMBdiDy!EhR_Jc6L+wG<2H_KG> z54R;EM8x_X?<{^TOFa2t&N@BSe;WGntjGSWML9|c;hR%r`s+UK;2-!!XgK51Kz!J5f+q8GwzmiA**zqbD^@pFnLj08W<+ojGVuqEXv3C7OX(Y1-IW(z_Pz3D5l}02I zhlkB*|9*r;ILa~O-}Trt3igx-Lh|GFjAJwhx+e4+lACQ#t`1wz82kU!V?R|fI~yzDc=#|bu#RA< z0A2LM+G7}}W0fA6dc3`kXH@>Dd#N2^F|zI3$lv$8m{ zqlKDh^AkwwZzDvBgXn#?G#R@DQo(>1?5iA?OwcagCc+l8>8B@^LiDQRvr|MR0%6`O znS1ODAtvstwYeFXut-LVDA7T#I3485m`Z-8Eauq?CTxM@KnZT8Jn!+R1)2=>z)I&% zR-QlPl8VQ4{|7|D;SqQ7xGymXMn?ibi~?SK@|kBN`Kny2rJwA0u z_u!CI@5BCq;j!_~w27%<)afa#z}dH=FpD!`=*w52>x-}p0Ys12wvN1Z{r5kBU|o-B zHbSdi>9*d>jm}MDctR*7VU+3(fe6Gb!4FT8YJFu)NK<~9A1ktMGgULzt=fouG zs~(Eo;**dMllKwU4kuOhP7_l$>6D9nU;w8RMc^+%Eus^D6*ELSn@}#+M2GMP58nU! zn%?w}*7Om>KYBLKOARXQgnpW{i*FML0R8mWgq_)-^aZ(soea$@Tj$b;Ge98zl7&si zQW6ZD>sCk@L9$kZVkJaYL(JWXob8YXQh@hNh}$J8Kq_i{rS5c{GePwb{uB-IrA%!1@HIk^FNjt|6a1^1M_zUB^_r3 z1`A>De>pMWOtKf^7ZSm$)fM-pRQ_L+82|EFcRjmV0VD?eR$}}^LCL%B|J6|OAJDgN zpLLci3Ky_GWBh+qvS&ZeZq7zT1t)m^?LCXce*otE{`UUSsUkl1pKz*#>*E8Oh}dBK zsM`r~XM$5c*iUaS^elMhSHY=6d`o*rCuiGbxUT*IpG?`L7YsuquS;Gs^o^ad0}T0Z z=Q-Xj7!xnWykA{d+1TC*+1fogJQ7C_S^oeE7l8ouTcaBUn_i{kTlhjS^lvVu^KY;B z7D<1?mrLnIig@VRrIa<4@l_y?gMp#)&)ESIF3=z8w_n+T+c^*ZnjHWxrO-2W;Oc#Z zK7YYLX)7M%#H-u!1Y`)fS)ZtZlAwY^*O`pGgZme(Igt%+$ju$^9yIM`E~#fS34 zIUm)nMyg4I(Y}H|e|me#h;HH>kO`1Av1$ra`Pe1%(d9$re3^NGdL5Kr@f1O>NiW#| z9-fHC=fTnh#f@ZO(hkUYdNPQ*xc<}Y$vtIh^*dw*Dl?bIPP&WH6HXQ(8Px@l6H zKfnM+fZ?!rjel*SESN>YOD_m_8jeAm10R@3y9RjE_$^5bHm<&7LxiSmHA`_Yxl7g_Wg%uZ>e>a8;40{T z#rtv4R_rn&+0^6R4j3Gi#vXJdLu(X|m`ki&5zv(>5wm`#rm^AVWV;4^s1$w2VRoW3 zf?JBSz^i>UmA=z?^{FbIPTL128e`cyM_GAsO}(58>O*;jy_pmK`;+^E8@Y1{VA5le z>YNSxLdHl?Hj){B|Iu5J`iAWm{o^jy<_~o@1-mFzFdA+=ax*TkpKhr)?0eOPUcp!ppCIjcd2t#sO z$$N3`Mc`BSAt-Z@m` zw#$1w$r^-Ty<%K|KbZ68pequ2M|J&4W6;#rm~Fk1cmaAZvN{u^It=WZp-YyT`XrBY z1@b=$%f_4;!3DyyeHAG_g=Gh7#&H2*+5XcR1m~}49>5dA2cAD8@DsQEAL_^dP9al> zOYk3vq!JeaJZv-}XJ%5a$R7xNqi5OY=-qE-(lbv;0AcH=0rS5fN%ey#mBtJo{VIV9zs&k*vGc7SS(s)H58ilH$) z%|9{030kCiN4-iiy}5RI{p#$v$Uf#W59%)cV36HS%)$#us9260LV~FGQH_tCPbwGQ zO`&>IvxX34Hfgr|QgAmlUoLbJrI@5WP1e&Bde>%k4X+JDBKZ<`0Gmiy%3)&~mOAkz z$5EQ51ZJC3_ngv+O6uk6Uqjo zR~2UGx}s>+&v0<{Q_4@qkWXtJxe@;tV~F5pJMh&x&S(H);mm0-Jio6vxH%=_TR#pq zvqlKPz^C5H1F`#j2y~P|b&E7}!Lk_Mb0IABgpvVQ*?t;~qt(N>?df!!)qP~=qx3>X zrdjmMmArj|a^KI#$uwN;gtr*X%D}Q&`tiE3O9gkq)Y3H3-im9IyHq!G3 zKZ*fea-$!UH5U;oh-_DW?TT`WG3l#sHJ#d1d4(B(t36 zYLUigU2=JQV@4ILd*feQWd4=JB<7sVnEj4N3)dV0R@6TQna=(_OMHO6Z+1bQ=5l@{ zyk~l{TT1J9 zvNvJ7ezHIPl;z{WTx0@+)jW;cG2&tdNzc)0%@p$S7PgSnN!>)?4VI(zf}5uwo8dvQ zdcH7NL3lvI1|wD`zak(w&$lCmQzC0b%1(C06K*|X`%4_$FX>0Iun>-L!6zIXv%QGS zRq(OsL=!?VX%-g8Wp4@f2CO7;2%Wq+wnD8Yb#Z3vO|v53q!@C7la1EX%aX}{I__{p zLyKOV#le^+4k_fbv|_>oMFF0b6vR2Sq68=@fsa+dq}r?<{6rykPu;c1@#Q*1ij#r^ zP?{G=2_-mWi$heSD5$>45(U#$fS1_6SU-Q~)9XTNAeJL%J~1>CJH zSeLbdMD+t*Budu-A}%Y43dJHM?zSvDxi*6gYX+2h8|3G6I-SH=yNLU6uLb%@7m*0& zJ>;%ET*y2~9J;DtUkDaT?O5bhJGghvv>`Hgr}Ro5Qnu_<$>C%ophf0ZwtQsQa5~R! zmaKpFoeUpvL|bC&?FKqkK4rL^zS119gzG)EK|*T3@bp{U%Sgl33R2=bh{S2h3Qo>1 zNovE%Etv$wu8qlDsO3a_v}%ar!|hfHwLx^cC7_NT>(;`yhh%86q7}|RIxb}cu|DC~ z(Y(YDZb?jvqX$dQLy5!g^kh?PmxN#(Z;Y+1FyUwjKe-Yee$yu#-e`)tct~CP-`zOl!=M2r_q!QQ4ef zwk#@Yr;O9=sy438X{tYe|I)!nbR` zJbTo>{>!!B0Qw=QkheDH*Puc`jCAJFiHMZLjIxuY3X4rNB~D6ByZS6XJv%8TC!Yk> z{??Ne>N08rU}&o1i^1TlAZcR5XraNPZJ_8T>80po>gTGVY9B_Ye2LyN#QwtWHANB2 zWVs@wK=~E>Eae)_OYM2~yb<9&@*TRSjXrjR77_J*SZD4>d>C~M3TcrLzEg3yN)g!v z$W!H>)!T^fu@Ey&(lydZ3C|`(Qjya;b2FxGk_MbO;1?XzB4v@RW&sTlA!;#eCTl*4 zh!7DsJ>it#V^K{yA1y%*qX~24YzT-L

HGf{m1Ff(9-pUVK6L1KPz=e1Ro61$!g z0-X;nb5MoT#BS65&g_Z$Ykq0kv~5m)HRKNcIU2B!I1Zx|CPiI?c#d{iAA*}t9> z!<`tyq=9dSiheur(0~5*t@1891fw3YduU69-Re$ewGWHm1Az_W12_slSGeZG!~Pfs z+E4mRa>{A;5aZEzlBDU(o|A|*?edRe$6f6bT&U7VXT0glE0cX$BfK<$1_f>i?1jH0dcG#iON%K%qV)O}aMIs{ zt-ks(amq!PvMz^to?W|VbyE(xU&3${QP&r?!)wtG)Z?j;Bm2)2^^$b~EXh0uI!CSU%o< z8;}G0Xf9t$`tV(pZvNqXoGnn^miUzU=yWmlEU-bO$@ggKK%$d}GQWOJe5IHjpr3sv z0T$c6@ndu3s8aD>&Knz-st*?V5f( zT+Ar=_(M>k6W_qzLc!@reub`j{Qb?T)6&9J81R6z7A+|l|8rmihTnJkSxV(A{Y-*I z3UMNB!rbF`f6YDvkkOnvcUOFKVy+0NpJ)5)$aoF|1*zG zJR|1W-seS>1~AW0G60WFp;6v?7Hzmmw$=w0;F z#f`e&V|qAN3uh``)TY0tw^!8)$}Q({&XiQ)Z^$hv_My)e>5ycADW@lL0YfP{!Jr17 z9mB#Ki;{)~p2VE|S21XvL0}q{pvKb@Hhxfp85QmQWl%*#mJ2wSo)ipj-~+8dMtv~& ztLBT5(W-L=)dilDs5clOyr3_1flh8gO=4h4H(-jWQ84&RVHx)YYDf}ufK>dl98)+v z5UVo+4BSQ|-BJLlKoKom&q{Oy^D%PMAsQ)T3kffVB+_c}F-+m&WcOt_ih<<~2sOGN zH3G)*(NY2V#0QJ3Y$*dK9cwYyKt<~x1zA9#b$YW(B{tC#p(FCL%Lwbtw*-ecgi=`g zFGxMHY|M4Ua2aQSJ4t&*S;@oRtbJo;i7vlx$u=iLuV<-;s5YCQRwXibw^K(U$F-XW zQqzYRQGsFg))4DjLysIBaYf9t3$n)rILXKX3|JZYYe&n4?aUumN?$oD%$G0Ze^{&i zrtl2C&VRh=ykSJU*wNomi`&H_13&qgZ^{8TPOV;er%JcrI_88EY2!Y#y9{us>cW6V z-UBH(nDg~c702uZQl}?$vKQO>5B5znPj+BQpr!NYeb*3X=rcxXBHcf%FKvH5h zXBVh)mMf~1;Q8a~Om5E)w)q5j0{Zvg1D@ZYpZ|G!-d}~_Xb}R(Reni%Rdwo_B{1fB zLvu?|)BhIW>H2@_SfU{oy7)upEH{u%B=vnXpY`}PD#c8E0`+`0kt*Fc`}{vf1JqXI$Ne)er{2*Movk2g;$|OwPIN zdq+j9&Y8`K=bm%#z)R!ikt_9sum4RWS)VLfQS!eKYA68<2FL$PIP_(%MAiS6&1hRI z@dEw&q_h*PF1ub4L_5+;f^vl3#{ui0E-ouRF zGQ>S!FtQe?D1tVs0Wn0SzUSRdjg3_PShyL{9c+5 zBK0U{qZ!~g-CD3w)YRR`CF8vO52aOlf;ZKm(r6l|q*nJ*9&!>?C>CBkbe(UB6|#6? zRrNOd7A96|)`w_vE$(Cn%{^{q3R<`zP8{l%3S?y@cXPUlXT!fj@4*u(%A}>Vnasa8E@eDE-Rzqn=Y6D%awo)3tstHI!{*`? zV3VG@b9}ZJi-l>N1^A!)_(oe!5uo2jxO;}=S0bnq_fVK(R; zqGm+ae2fNf`+O`o{Vc$r=Y^^?tqQBwGlwlQ-7j@quI9Q)Uh$i{$$nhGUcAf%z|q8# zlHUTKfxS48%?!pn@oRQP?9*(kiR{8NkDmh)CdQxP=pHU)jhw!-0RMyCOt{lhtx5@| zfU&*HzGWHeHLKoZD-un4lvX*d6>rOznG0s6;nr%3LnMFMi`TJIm>Vle9jJn#Y*FjB zk1q1I)s2O=SGvB5!!e?qXatq~J z$ldOPi|O9(=Rezvh4GC)?Zu_^;I!JEAz3jd&Tvi{ri(>QXS**D>B-ImX_@Phz+SAq zH>TD3WM9D%oiBVgkmVhGYSay#qov24b*^+DwU0hfEj=ID-MIl{`2lTy3NlR(B0Y_t zu7^?1a+Xo|0-@Sr!?DRq0@)RzKz0R{T%oIIH7UZA)JD|VOgG|EOo))UU`Kd?J2MX% zR4KIyH#^gVnFA9=>mZ46uuzDrR>RXGvl)4_@G15Y83qMq3mHIziW*^Hg+?V(v2F58 zJ7~Hnd$H4%XZk3pVB(e7B{NwR2|qYae%eGUdWsfYG{n_3FjtxPf4&7wfKBZ@Owwjj|~keAdqS$2@+1;C^uQOQ&o61qW1{c1Snrl5s; zEFAigAYr0h$kvT?-gl1jSzX5f9KHL9igAIWTd6K1qFe=?*^(|^bxS@h@mQPAfv%?v zfTOQmrDp3*>e0?7_v!nh(@@@y8SP2us_5)BvHuK5_agS1$9#sPUv>7`NYh<4fz0#F zyzjF=qPKg8qRqFetG6fx<*wGRb@!xc{-&$XxYk~19eucFas$ETxg6&@#`HDV?qV3< zwWPeqZw+Hs_8Lgb->o|u<0`K~73(BHg-%oV)whTb3K8N>-B!E7{?!^z>>VsDFUVn# zn_-$A;GZIsPHTk3RKinvx+8Enbm?B?L?Cr+@zt%J#oxc&N?TZ@_7Ou=U~I5mL;SX& z2&oT2toqe@VbGNoy@$8CvtT%hDT6jQGRRUaVQ+~FN8BjXR;=cP_rD|{SEcT9If1k% zX~lwqI$IJU>8g}9>oyu2|354>A8{>#XhOuXS0{_l` zF)m3ICJVsBP%4RrCK*eR6>fRR6^)^6#aTz&GA*k#DZ`frfWdD5(wdR>@ELWx7e;w+ z>O5|I{j_wKT}j;q!(VbNbdD3r{)v}LMhQejw%om8)!4u@b7D4qpe*T@kVOI6xt35O z5D&;2eEa)Ou$Yzlb6wiMnBL+NTm!@>?(<%UadcCMxIgu};_U6`FYX(l7DO3J0wQ;{ z$KVUX2&YAlph`&Lhz-NK5y_gt^(Z2V8ahs_V3g4ho<}8*xBh)6E zf>xS#iXw)N9_G+$^3Kae1<7pnOrRvt3)6m{j27-0+3|+ECFJwW-T1yhwr@R z6P&SpffnB=(0)(_QWPffv0zzzQFt$V*3T^ZBA0`BvI`>U#DIShCqrt_@~%=LO%JU=Z@ z>jVGb%GUI3=42hB9xWqvJTiZO5KsE$@v1~2-}%s|KpRlzZI;d4KOjqP=e)x|B1?{p zUX>dH*Y<#3p|9yJCwhzVUxLo}If3*RQRDS*JrqBsw{R_zl1%zHym7-(AP{mNSNa^E zNQR2-_@j5Ye39IMjBQ%Pk!1uTNDM@3=@AD4EcqX!2eiF|RI|Rj#pa`a?k#kT z)Q+;>xBpg z7QO)nOIZ!8-Yu|cvk3e-D?6QVIhi9+QG~C(LJXD!$6q3~zQXk%l#mK}fU#s)p+@Y~ zYlO|mxZD$nPfVC}^wq7L9g{GrvWoX*j$6uK)~Q)5Ml!86=)1`hvo~p|Q+E{7X`21Q z2Zk`Iaj(0%9>LAKP!ERYzg>7-=7s!ffYT;<@3_n}+8d)0fiv*B8yQE^3G6*VyBNV{ zqMYk)+GUsBXGSiMK8FQsXB65JvZMA@U^zbSMVzo|m))3QrQ~~yu92e7J$F}JTKMg~ z4*x357^ephP8X8{#j5d)B25O~XFIjqFTZ+@_Sm|(pHFqQygAZ-?Rd4X#ex+<;!8QJ`1IdNxo-^>F@YxZd>iHZL3kJGWQ2IV;zNV^R`9TEXi< zh3+=oWYG7CO}Jf29?b5rC|Zl2go8n-T(wC=7MVgEKo5n^0U1UoX$u+ArkA)>GS)e_ zScbR=5%aWYX!ZHPA!B5#oo1*@iW|=}8<%L7p!T z<^#WQGglG9ghhhm$NF(B@%!Q(-U^B8nKNZ3W4P7j8H)E%lU5^9Os0<1Rf6^%hMu-0 zQ6?J%^CA_7QkDBXC``C+D%B-3i8jTK;_xd2R3vxqiabEuj%<)hCw4z{k*2qVw4wO1 zClJD^xz4h^Iu_*`((J{v8j2lN5$A5r>cw}!MpzJD;I1l?wu!Jrc<-)z?cHG?QOi*L zc8l(Hx9I-*6>P-)9#r;6tg^1;*i?ujnV8TcSWhmFBr3Da2c-1KJ_@BsSk`El!wcmC zKB{a(E~yHNiyXb6^s`c++RYSsStYL^YYB~ioF%0zITSO7V*u%(J2L4SiXR*{=0l~w z#OBP9bjFgGg~CS|bff|-IdvGu9beXtEalzom|Fh3MN0?>cHtT4_h$+aJ183b`1^&U zXa9ilceT{#By4|>g*njzdA%Rw$LV&@E}UPA>;Qh4o0%@mgdIOJ=5wIGI0nEEpUJW! zVSf}uazH^?RyiC4p15LQ$jCj zR~9?hq~0lGu1lis>7`JApgf^JZR(-8EDHruz}&_`3BeF)wjsv@_sqZuKEH@%iqazv z=C8%VzrouehS4fOKqGc)c3c!fbP;E!nz!Xei0Cbf3*qi9;@fbYA+$DwQQouZ&G_%+ z<7ItzkY6~HSYz}qzy?w$(PM70`phl_$DlWr&)K^!jy}4QVX#B5N%9)qz;^g8JOwi* z66cD+BjgorH1LB1A9ZComIs7SXMDtY$qnSN{ahGPx?z`toKkkUjB6h0qclWWGIml zzJY1k-W(o$qF=;Ny90hhSiM0Os)u4r=HT5-_RTUZl>S}2k)21CM%&cc-og^MDH!e<;gQy)?f$)9s3^bT+lUy_uj$6)LE2w#hKrKwkL<( zsoQ!>b4vlZa4H^lo@F?Cy9Pf9kxP%#OGiqLF)BZti`HqV@{BV9E}T#*p!g`BeV`M_ zaO9a;z;{a~qk-kYAzDl>r0oLu;q=*r6qR|x_i1s8mD-P@Y_hvkf#M^2Iy`xvrL4@h zXn-Gv>Ko*I+Jj4vwhZzrlOGx4)XXoP@x$W=1wdE@BT`Fs@Jdk+bIA(R;+uvBgkFVl z@6)^P#QLk{LVVq;KwKC@zW8BLbCXf*+Ds$199j;i$mq%F zp#Oy*-tObwpP-fILwtkUA&BpPQ-JZpiDYLsYu?Tf$-##P0SO`FGThMzdApaTO;@e? z?mAuKNY?z5ENc-5hbczm?W(P%6T@4om&oJp%=qclFFEmq(O=k3_+A*>VGKqD_TULw z-!pzVUgll=cYgTm9_%h0So4X>INO7Nq%vCYO^-J_Xc8n>IEym@dva`(!_F`)3=dk} zE!8LfN~WJ=StLt4vys2!hyV7%`KkDbl(8dSz}vjg4ebLd1e>}E6D`-B`WPR2yRrbc zm{pD0f$X_vlqbAirzcOt42(X+6aL_@?8zD8t@O&9K8l};>_nJ|U7o6J%)qpXw|%>C zzS@)he#&tC5>`P;dMQze%-I|n*{lHurQwr3Iq3X_^V_frz^@=9(IyC<2~tvQAi(Co z2{Zc7{R+C&AQ(imcj)Tg@Jd9aFfY$!SrUB~=l0;)kO)LG5N72gc+-)=X9qnQeEXgK zjz`-GAQ4^g?A3nu(a;D8rw)v3)qpckSR^(rKfI$7jQ6Ak7`H=@nZo{spY-W)1VP@S zkeK9PNKEN7*l=B@TBR4@%HgyYsRpqpbVHV0JE<*xdLp{n-OrTofS7UVm^BQ#6Z>}3 zuC!)-l4T)SCXV^F0`gR6uoP+X-7*7NcnrB@iOd<4hgfIpTG9f9!6LD^7{_rR=*NM|WR#g+~bNW;5r}->wiXPu6-v zr)vA?Zq@AF{Ck{6LV07#jWH29nM}^=F37{sUmG7+C z?A1)87F@egGbrL(k{u$#a(`b^aY@iNdCg;08!1O6fXgOQsVvvTR#Hv9JiH)uH;?It zq)K9ZMT5Jsp)t+PQ6FZ4>iX0?%}1S+o%9vieIgsquey{s?GKts1E?SAYseg>-)pXT zo@MU~o$_o_D8|dnexQFgpxLUG5H11zwun%2EEvU{@ev*}`uR0^} zw)h23gmKD!`{BDn`_*mvh3S#KMw=EqcvO8f;WfvLI`=<}jWREkXpnWoPTcOC#w)ID zUtN4$9$9!B!NT#`%{BLkn&cxqPW6JI{ zV!@Thspn>%R}Lpoofg8V@;9E!@J6Da+OTj+-#(^3j>LOv*Kv7r)8FZ06ou^p>+Srl z@Ef~xEViYWl^$$nXtHNd@PWX4z1#60ju-gV*ufwS7!cgaqPR@;;AteYRq{PjuvB36 zaEA15_WhG3#s2D%ioV_a$0y5o#hl^LoPC9%TPu1pHDj;(_R4EcR?PxyCKdujt(1>w*WNCa{Aes&r8qbP?TfA+-}y<#(kLr<%V4PaEYb zPwpx*;K~f4VikvD<93slcM~#mk-pw)}=^ey%@g3y{AcsgJ` zwaBEgQhaJ-@zidS+P?Cs-;-2gLV~z5FVj{qJtVcg6^4J7w3O8z8K^x7%Zh3Ob{JRR26*jb+ENO zHRNruYbcm38Xki#gw8C4BQbdj#53oJ!NrKk0S2=}KZUA@>?T-CeZNPqkslNKKIuf71QR(`dDUKy zs+zvwS@m>0I4WjC}vNKHfYPdpGxR!T>2~wnGbfn*F^2e_u zovucCMo0N6#iV6jlHQ1MgMlw5q$;}_D`yrVZ2UA@TEc)t@lg^fQ8GDCGNp1dwS6*eOfr3SGUHe>^HDM@ zQ3@MR3WstEmwgI%ObSnR3h!77!cxk8XT>c1%!RJcKzf*DTg42e^8M1_5a614&Ii)BlG%PlqgQ^45ht&J6nZL9@HKA z9|}33#^>=kr_?+9KXZVK&!jKY8-J9(r2Z$oem^<1{_p?KKNd%h{hvx-{s(saGq0Z* z&jIW8&tbjam3pQHJcrxJmwr~e+ko$#Lo@!s>i>Xd#DKmf_x{s*!fP-A=0N)pH|8zW z$3eUhg-9P_)6gf+!g-?8wbJe+c;{r@K~8Li?_caDs?3s-xctjTcI1zJ5)elb7)N-Mc)2Hx%9E zE45is?XwA7xYM8j(n|CWH$yujDA^+U6g4B}T#}vQszigWc=2_9?G3jUEE>1dXJ%If zFLn_nwX~#cZcb;pa2!R~a?Z_w;<-`s*%XTveeZG#G$mIS!r#WDbHZ5JFNKRpC8sT3 z!F7Mj9ZN-N^lWoQXb?dzJed1x!F@NWyK0L)KDcc;(x!Jo;@rYbDo&EHCmxvh;MC}F zkQdCVwBlc37BIi;r5Dm1Ll3-gVXTJS6047@4FS!2Ib;D{L^mw1R(Qud%D{R8TceO( z5C4Ss#&ZXXZGCJLsxbvQ3uoXQj{8C_3tuv z%!Lw?cfx_?M>a{q6g3souEUO zRdSPg<{)i%#_IbZhMP#bWF&e~OWuTiN)NF5aWWP^WzyVBViRJTj zNLnGJ57D08AzI274Y<1m$(T)~&nf-5IF?g3pPdtVzpc6@+-;411wSRmN6bgq zF=*dWgcP@z23#yA84AhWLZ*I-2f4juZn4$XBv@;PPgf$f-N!~UxP@7w(z-psF!vX$;T^R+P6|ec&VlOqrE55xC!TK^}b{2dIKWM z@x~`Kq0HXdMyXqtJtegdnwqSw`30`-Q0cCIoZAo>5oO)w|#tC3r%X)aI98< zdxMD|p3~BDWH1n@cZA}0x;~J$h@N?R1vGt#rAw=t7HAK92A@M92YaeZj4;9eXF3)h@N=z-XhLU%uPj zvn2&_JxZLjRSrGSh?1O~uJfXNaftE-O6uo!owpo&LK9a6XoMt6EVrwa@7bI^H34y#}9HeQNgBc3R^Q zCfmjP-dEZY`fzH^;?x*upncxbd0E?VD@GYFC-q(Jy4dx17+tQfUoCYAk&HVIi@9+d zGT@A5oJc;Zo9UxIKuUT?3f_3Jk2qV_+uko3vrI1&ZCV!p&gmVvi!T=&SS4KreTIWM zN%e1MK9WaDA5LM1=-&*taYxDKPv^&S#-^9Og(%*bt~T%<0+n?`Qgq3%;5XoSLIj^! z&7XPGnI1cA$sNwGA(Ow59-7KCM8*9QQRR}LVi6Hyu}wIdn!k&}?Lq~@sBk!A?Z*2e z7$a=MxbSzb=MW5Ldsfbp!Z`Les_sfc>Y8W^-9-#FhjbK!_V_`99xG7FZg5w&a}JM( z(Zh9mXE?n*SlP4{T^ZIf926V45L{!u<{0Jfn1+&IT$j6AzRG=r5pcP5#SX_2#3OE_iBtR?t-iR2GOlq14#DCbcM-@#9tn<#8pB9;YDdT#q@FQ9y5|haGTBzS-H=&i1yu{p}Nb0h-~SiYtyG`3IS;1yZm47|?@u9%Q}-<+pkT|So3`WfiHYml!B7E z@A&b}u!`h6eEFty{=zR`yW+<{_GXZ_HmDDK)+|2`gt9&-{`5fUPy73$)c*GO%mQ?X z-zNDEKlyKqa2hB%;6$pY5*n%uLf&>_TA?OY- z=xQW*gEPo023Ud>NDnF)fKVgD!X!KiGnRvxHSrgg{s^PKYg2NQFX} zgZom2S$Ks>C^A~Og-8g4Me~LDMuug`g - - - - - - - - Build TensorFlow from Source Code using Docker - Sqlflow - - - - - - - - - - - - - - - - - -
-
- Sqlflow - -
- - -
-
-

This site uses Just the Docs, a documentation theme for Jekyll.

-
-
-
-
-
- -
-
- - Search -
-
-
- - - - -
-
-
- - - -
-

Build TensorFlow from Source Code using Docker

- -

To contribute to TensorFlow, we need to build TensorFlow from source code. The official guide is great. However, it interleaves the native building process and that using Docker and makes it confusing because packages needed by the former are not by the latter. Also, we found some useful tricks to start Docker containers in practices that are not in the official guide. Hence this document.

- -
- -

Build the Pip Package in TensorFlow Development Container

- -

On either Mac or Linux, or any other OS, we don’t have to install and configure the building tools; instead, we can use a Docker image where all tools have been installed and properly configured.

- -
    -
  1. -

    Get the Docker image containing all the building tools:

    - - 
    docker pull tensorflow/tensorflow:latest-devel
-
    -
    2. -
  2. -

    Then, let’s get the source code. On any OS, please install git using the native package manager. For example, on Ubuntu, please

    - - 
    sudo apt-get install git
-
    - -

    or, on Mac,

    - - 
    brew install git
-
    - -

    Then, use the git just installed, let’s clone tensorflow source code:

    - - 
    git clone --recursive https://github.com/tensorflow/tensorflow
-cd tensorflow
-
    - -

    By default, we will be on the master branch. Feel free to do you change in your feature branches, or switch to a release branch, for example:

    - - 
    git checkout v1.11.0
-git checkout -b v1.11.0
-
    -
    3. -
  3. -

    Then, let us start a Docker container running the tensorflow/tensorflow:latest-devel image:

    - - 
    docker run --rm -it \
-   -w /tensorflow \
-   -v $PWD:/tensorflow \
-   -v $HOME/.cache:/root/.cache \
-   -e "HOST_PERMS=$(id -u):$(id -g)" \
-   tensorflow/tensorflow:latest-devel \
-   /bin/bash
-
    - -
      -
    • -w /tensorflow brings us to the /tensorflow directory in the container once after we start it.
      • -
    • -v $PWD:/tensorflow maps the current directory, which is the just cloned TensorFlow source directory on the host, to /tensorflow in the container.
      • -
    • -v $HOME/.cache:/root/.cache maps the Bazel temporary directory on the host into the container, so the intermediate files generated by Bazel running in the container are actually saved on the host. This allows us to interrupt the container during the build and restart it later to resume the building.
      • -
    • e "HOST_PERMS=$(id -u):$(id -g)" passes the user identification on the host into the container as an environment variable. We can reset the mode of files generated in the container to this user identity.
      • -
    -
    4. -
  4. -

    From now on, we will be working in the container. Let us first configure the building:

    - - 
    ./configure
-
    - -

    Usually, I would simply choose all the default options by hitting enter all the way down.

    -
    5. -
  5. -

    Build the pip package builder:

    - - 
    bazel build --config=opt //tensorflow/tools/pip_package:build_pip_package
-
    -
    6. -
  6. -

    Build the pip package and save it into /tensorflow:

    - - 
    ./bazel-bin/tensorflow/tools/pip_package/build_pip_package /tensorflow
-
    -
    7. -
  7. -

    Change the generated wheel file into the usual mode as on the host:

    - - 
    chown $HOST_PERMS /mnt/tensorflow-*.whl
-
    -
    8. -
- -

Install the Pip Pacakge in TensorFlow Development Container

- -
    -
  1. -

    Let’s now try the new pip package. First, we need to uninstall the current tensorflow pip package, then we install the newly built one:

    - - 
    pip uninstall tensorflow
-pip install /tensorflow/tensorflow-*.whl
-
    -
    2. -
  2. -

    Now, we can verify if the new package works. First, we need to switch to a directory out from /tensorflow, so we don’t import from the source directory:

    - - 
    cd /tmp  # other directories also work
-
    - -

    then, we can try import the newly installed tensorflow package and verify it:

    - - 
    python
->>> import tensorflow
->>> print(tensorflow.__version__)
-
    -
    3. -
  3. -

    Now, let us quit from Python and from the Docker container. We should see the tensorflow-*.whl file on the host in the current directory.

    -
    4. -
- -

Install the Pip Package in a Clean Python Package

- -

After we quit from the development container, we should see the wheel file in the TensorFlow source directory on the host. Now, we can start a Python container and install the Pip package in it.

- -
    -
  1. -

    Start the Python container

    - - 
    docker run --rm -it -v $PWD:/tensorflow python:2.7 bash
-
    -
    2. -
  2. -

    Install the pip pacakge in the container

    - - 
    pip install /tensorflow/tensorflow*.whl
-
    -
    3. -
  3. -

    Try TensorFlow by starting Python in the container

    - - 
    python
->>> import tensorflow as tf
->>> print(tf.__version__)
-
    -
    4. -
- - - - -
-
-
-
- - - diff --git a/docs/doc/build-tensorflow.md b/docs/doc/build-tensorflow.md deleted file mode 100644 index 5360f1d..0000000 --- a/docs/doc/build-tensorflow.md +++ /dev/null @@ -1,135 +0,0 @@ -# Build TensorFlow from Source Code using Docker - -To contribute to TensorFlow, we need to build TensorFlow from source code. The [official guide](https://www.tensorflow.org/install/source) is great. However, it interleaves the native building process and that using Docker and makes it confusing because packages needed by the former are not by the latter. Also, we found some useful tricks to start Docker containers in practices that are not in the official guide. Hence this document. - ----- - -## Build the Pip Package in TensorFlow Development Container - -On either Mac or Linux, or any other OS, we don't have to install and configure the building tools; instead, we can use a Docker image where all tools have been installed and properly configured. - -1. Get the Docker image containing all the building tools: - - ```bash - docker pull tensorflow/tensorflow:latest-devel - ``` - -1. Then, let's get the source code. On any OS, please install git using the native package manager. For example, on Ubuntu, please - - ```bash - sudo apt-get install git - ``` - - or, on Mac, - - ```bash - brew install git - ``` - - Then, use the git just installed, let's clone tensorflow source code: - - ```bash - git clone --recursive https://github.com/tensorflow/tensorflow - cd tensorflow - ``` - - By default, we will be on the `master` branch. Feel free to do you change in your feature branches, or switch to a release branch, for example: - - ```bash - git checkout v1.11.0 - git checkout -b v1.11.0 - ``` - -1. Then, let us start a Docker container running the `tensorflow/tensorflow:latest-devel` image: - - ```bash - docker run --rm -it \ - -w /tensorflow \ - -v $PWD:/tensorflow \ - -v $HOME/.cache:/root/.cache \ - -e "HOST_PERMS=$(id -u):$(id -g)" \ - tensorflow/tensorflow:latest-devel \ - /bin/bash - ``` - - - `-w /tensorflow` brings us to the `/tensorflow` directory in the container once after we start it. - - `-v $PWD:/tensorflow` maps the current directory, which is the just cloned TensorFlow source directory on the host, to `/tensorflow` in the container. - - `-v $HOME/.cache:/root/.cache` maps the Bazel temporary directory on the host into the container, so the intermediate files generated by Bazel running in the container are actually saved on the host. This allows us to interrupt the container during the build and restart it later to resume the building. - - `e "HOST_PERMS=$(id -u):$(id -g)"` passes the user identification on the host into the container as an environment variable. We can reset the mode of files generated in the container to this user identity. - - -1. From now on, we will be working in the container. Let us first configure the building: - - ```bash - ./configure - ``` - - Usually, I would simply choose all the default options by hitting enter all the way down. - -1. Build the pip package builder: - - ```bash - bazel build --config=opt //tensorflow/tools/pip_package:build_pip_package - ``` - -1. Build the pip package and save it into `/tensorflow`: - - ```bash - ./bazel-bin/tensorflow/tools/pip_package/build_pip_package /tensorflow - ``` - -1. Change the generated wheel file into the usual mode as on the host: - - ```bash - chown $HOST_PERMS /mnt/tensorflow-*.whl - ``` - -# Install the Pip Pacakge in TensorFlow Development Container - -1. Let's now try the new pip package. First, we need to uninstall the current tensorflow pip package, then we install the newly built one: - - ```bash - pip uninstall tensorflow - pip install /tensorflow/tensorflow-*.whl - ``` - -1. Now, we can verify if the new package works. First, we need to switch to a directory out from `/tensorflow`, so we don't import from the source directory: - - ```bash - cd /tmp # other directories also work - ``` - - then, we can try import the newly installed tensorflow package and verify it: - - ```bash - python - >>> import tensorflow - >>> print(tensorflow.__version__) - ``` - -1. Now, let us quit from Python and from the Docker container. We should see the `tensorflow-*.whl` file on the host in the current directory. - - -## Install the Pip Package in a Clean Python Package - -After we quit from the development container, we should see the wheel file in the TensorFlow source directory on the host. Now, we can start a Python container and install the Pip package in it. - -1. Start the Python container - - ```bash - docker run --rm -it -v $PWD:/tensorflow python:2.7 bash - ``` - -1. Install the pip pacakge in the container - - ```bash - pip install /tensorflow/tensorflow*.whl - ``` - -1. Try TensorFlow by starting Python in the container - - ```bash - python - >>> import tensorflow as tf - >>> print(tf.__version__) - ``` diff --git a/docs/doc/build.html b/docs/doc/build.html deleted file mode 100644 index 64e6dc4..0000000 --- a/docs/doc/build.html +++ /dev/null @@ -1,316 +0,0 @@ - - - - - - - - - Canonical Development Environment - Sqlflow - - - - - - - - - - - - - - - - - -
-
- Sqlflow - -
- - -
-
-

This site uses Just the Docs, a documentation theme for Jekyll.

-
-
-
-
-
- -
-
- - Search -
-
-
- - - - -
-
-
- - - -
-

Canonical Development Environment

- -

Referring to this example, -we create a canonical development environment for Go and Python programmers using Docker.

- -

Editing on Host

- -

When we use this Docker image for daily development work, the source code relies -on the host computer instead of the container. The source code includes this repo -and all its dependencies, for example, the Go package google.golang.org/grpc. -Code-on-the-host allows us to run our favorite editors (Emacs, VIM, Eclipse, and more) -on the host. Please free to rely on editors add-ons to analyze the source code -for auto-completion.

- -

Building in Container

- -

We build a Docker image that contains development tools:

- -
    -
  1. The Python interpreter
    2. -
  2. The Go compiler
    3. -
  3. The protobuf compiler
    4. -
  4. The protobuf to Go compiler extension
    5. -
  5. The protobuf to Python compiler extension
    6. -
- -

Because this repo contains Go code, please make sure that you have the directory -structure required by Go. On my laptop computer, I have

- -
export GOPATH=$HOME/go
-
- -

You could have your $GOPATH pointing to any directory you like.

- -

Given $GOPATH$ set, we could git clone the source code of our project by running:

- -
go get github.com/sql-machine-learning/sqlflow
-
- -

Change the directory to our project root, and we can use go get to retrieve -and update Go dependencies.

- -
cd $GOPATH/src/github.com/sql-machine-learning/sqlflow
-go get -u -t ./...
-
- -

Note -t instructs get to also download the packages required to build -the tests for the specified packages.

- -

As all Git users would do, we run git pull from time to time to sync up with -others’ work. If somebody added new dependencies, we might need to run go -u ./... -after git pull to update dependencies.

- -

To build this project, we need the protobuf compiler, Go compiler, Python interpreter, -gRPC extension to the protobuf compiler. To ease the installation and configuration -of these tools, we provided a Dockerfile to install them into a Docker image. -To build the Docker image:

- -
docker build -t sqlflow:dev -f Dockerfile.dev .
-
- -

Development

- -

Build and Test

- -

We build and test the project inside the docker container.

- -

To run the container, we need to map the $GOPATH directory on the host into the -/go directory in the container, because the Dockerfile configures /go as -the $GOPATH in the container:

- -
docker run --rm -it -v $GOPATH:/go \
-    -w /go/src/github.com/sql-machine-learning/sqlflow \
-    sqlflow:dev bash
-
- -

Inside the Docker container, start a MySQL server in the background

- -
service mysql start&
-
- -

run all the tests as

- -
go generate ./...
-go install ./...
-go test -v ./...
-
- -

where go generate invokes the protoc command to translate server/sqlflow.proto -into server/sqlflow.pb.go and go test -v builds and run unit tests.

- -

Release

- -

The above build process currently generates two binary files in -$GOPATH/bin on the host. To package them into a Docker image, -please run

- -
docker build -t sqlflow -f ./Dockerfile $GOPATH/bin
-
- -

To publish the released Docker image to our official DockerHub

-
docker tag sqlflow sqlflow/sqlflow:latest
-docker push sqlflow/sqlflow:latest
-
- -

Demo: Command line Prompt

- -

The demo requires a MySQL server instance with populated data. If we don’t, we could -follow example/datasets/README.md to start one on the host. -After setting up MySQL, run the following inside the Docker container

- -
go run cmd/demo/demo.go --db_user root --db_password root --db_address host.docker.internal:3306
-
- -

You should be able to see the following prompt

- -
sqlflow>
-
- - - - -
-
-
-
- - - diff --git a/docs/doc/build.md b/docs/doc/build.md deleted file mode 100644 index 774dc91..0000000 --- a/docs/doc/build.md +++ /dev/null @@ -1,128 +0,0 @@ -# Canonical Development Environment - -Referring to [this example](https://github.com/sql-machine-learning/canonicalize-go-python-grpc-dev-env), -we create a canonical development environment for Go and Python programmers using Docker. - -## Editing on Host - -When we use this Docker image for daily development work, the source code relies -on the host computer instead of the container. The source code includes this repo -and all its dependencies, for example, the Go package `google.golang.org/grpc`. -Code-on-the-host allows us to run our favorite editors (Emacs, VIM, Eclipse, and more) -on the host. Please free to rely on editors add-ons to analyze the source code -for auto-completion. - -## Building in Container - -We build a Docker image that contains development tools: - -1. The Python interpreter -1. The Go compiler -1. The protobuf compiler -1. The protobuf to Go compiler extension -1. The protobuf to Python compiler extension - -Because this repo contains Go code, please make sure that you have the directory -structure required by Go. On my laptop computer, I have - -```bash -export GOPATH=$HOME/go -``` - -You could have your `$GOPATH` pointing to any directory you like. - -Given `$GOPATH$` set, we could git clone the source code of our project by running: - -```bash -go get github.com/sql-machine-learning/sqlflow -``` - -Change the directory to our project root, and we can use `go get` to retrieve -and update Go dependencies. - -```bash -cd $GOPATH/src/github.com/sql-machine-learning/sqlflow -go get -u -t ./... -``` - -Note `-t` instructs get to also download the packages required to build -the tests for the specified packages. - -As all Git users would do, we run `git pull` from time to time to sync up with -others' work. If somebody added new dependencies, we might need to run `go -u ./...` -after `git pull` to update dependencies. - -To build this project, we need the protobuf compiler, Go compiler, Python interpreter, -gRPC extension to the protobuf compiler. To ease the installation and configuration -of these tools, we provided a Dockerfile to install them into a Docker image. -To build the Docker image: - -```bash -docker build -t sqlflow:dev -f Dockerfile.dev . -``` - -## Development - -### Build and Test - -We build and test the project inside the docker container. - -To run the container, we need to map the `$GOPATH` directory on the host into the -`/go` directory in the container, because the Dockerfile configures `/go` as -the `$GOPATH` in the container: - -```bash -docker run --rm -it -v $GOPATH:/go \ - -w /go/src/github.com/sql-machine-learning/sqlflow \ - sqlflow:dev bash -``` - -Inside the Docker container, start a MySQL server in the background - -``` -service mysql start& -``` - -run all the tests as - -``` -go generate ./... -go install ./... -go test -v ./... -``` - -where `go generate` invokes the `protoc` command to translate `server/sqlflow.proto` -into `server/sqlflow.pb.go` and `go test -v` builds and run unit tests. - - -### Release - -The above build process currently generates two binary files in -`$GOPATH/bin` on the host. To package them into a Docker image, -please run - -```bash -docker build -t sqlflow -f ./Dockerfile $GOPATH/bin -``` - -To publish the released Docker image to our official DockerHub -```bash -docker tag sqlflow sqlflow/sqlflow:latest -docker push sqlflow/sqlflow:latest -``` - -## Demo: Command line Prompt - -The demo requires a MySQL server instance with populated data. If we don't, we could -follow [example/datasets/README.md](/example/datasets/README.md) to start one on the host. -After setting up MySQL, run the following inside the Docker container - -```bash -go run cmd/demo/demo.go --db_user root --db_password root --db_address host.docker.internal:3306 -``` - -You should be able to see the following prompt - -``` -sqlflow> -``` diff --git a/docs/doc/close_producer_from_consumer.html b/docs/doc/close_producer_from_consumer.html deleted file mode 100644 index 0e44e38..0000000 --- a/docs/doc/close_producer_from_consumer.html +++ /dev/null @@ -1,357 +0,0 @@ - - - - - - - - - Closing the producer goroutine from the consumer - Sqlflow - - - - - - - - - - - - - - - - - -
-
- Sqlflow - -
- - -
-
-

This site uses Just the Docs, a documentation theme for Jekyll.

-
-
-
-
-
- -
-
- - Search -
-
-
- - - - -
-
-
- - - -
-

Closing the producer goroutine from the consumer

- -

The producer-and-consumer pattern is well used in Go concurrent programming. When -the consumer stops, we want to gracefully stop the producer as well.

- -

Problem

- -

When a gRPC server receives a streaming request, it usually calls a -function that returns a channel, -reads the result from that channel and send the result to the client one by one.

- -

Take the following code for instance: upon receiving a request, the main goroutine -Service calls launchJob. launchJob starts a separate goroutine as an anonymous -function call and returns a channel. In the anonymous function, items will be sent to -channel. And Service on the otherside of the channel will reads from it.

- -
func Service(req *Request, stream *StreamResponse) error {
-  result := launchJob(req.Content)
-  for r := range result {
-    if e := stream.Send(result); e != nil {
-      // should we signal the running goroutine so it will stop sending?
-      return e
-    }
-  }
-}
-
-func launchJob(content string) chan Item {
-  c := make(chan Item)
-  
-  go func() {
-    defer close(c)
-
-    acquireScarceResources()
-    defer releaseScarceResources()
-
-    ...
-    // if stream.Send(result) returns an error and the Service returns, this will be blocked
-    c <- Item{}
-    ...
-  }()
-  
-  return c
-}
-
- -

There is a major problem in this implementation. As pointed out by the comment, -if the Send in Service returns an error, the Service function will return, -leaving the anonymous function being blocked on c <- Item{} forever.

- -

This problem is important because the leaking goroutine usually owns scarce system -resources such as network connection and memory.

- -

Solution: pipeline explicit cancellation

- -

Inspired by this blog post section -Explicit cancellation, we can signal the cancellation via closing on a separate -channel. And we can follow the terminology as io.Pipe.

- -
package sql
-
-import (
-	"errors"
-)
-
-var ErrClosedPipe = errors.New("pipe: write on closed pipe")
-
-// pipe follows the design at https://blog.golang.org/pipelines
-// - wrCh: chan for piping data
-// - done: chan for signaling Close from Reader to Writer
-type pipe struct {
-	wrCh chan interface{}
-	done chan struct{}
-}
-
-// PipeReader reads real data
-type PipeReader struct {
-	p *pipe
-}
-
-// PipeWriter writes real data
-type PipeWriter struct {
-	p *pipe
-}
-
-// Pipe creates a synchronous in-memory pipe.
-//
-// It is safe to call Read and Write in parallel with each other or with Close.
-// Parallel calls to Read and parallel calls to Write are also safe:
-// the individual calls will be gated sequentially.
-func Pipe() (*PipeReader, *PipeWriter) {
-	p := &pipe{
-		wrCh: make(chan interface{}),
-		done: make(chan struct{})}
-	return &PipeReader{p}, &PipeWriter{p}
-}
-
-// Close closes the reader; subsequent writes to the
-func (r *PipeReader) Close() {
-	close(r.p.done)
-}
-
-// ReadAll returns the data chan. The caller should
-// use it as `for r := range pr.ReadAll()`
-func (r *PipeReader) ReadAll() chan interface{} {
-	return r.p.wrCh
-}
-
-// Close closes the writer; subsequent ReadAll from the
-// read half of the pipe will return a closed channel.
-func (w *PipeWriter) Close() {
-	close(w.p.wrCh)
-}
-
-// Write writes the item to the underlying data stream.
-// It returns ErrClosedPipe when the data stream is closed.
-func (w *PipeWriter) Write(item interface{}) error {
-	select {
-	case w.p.wrCh <- item:
-		return nil
-	case <-w.p.done:
-		return ErrClosedPipe
-	}
-}
-
- -

And the consumer and producer be can implemented as

- -
func Service(req *Request, stream *StreamResponse) error {
-  pr := launchJob(req.Content)
-  defer pr.Close()
-  for r := range pr.ReadAll() {
-    if e := stream.Send(r); e != nil {
-      return e
-    }
-  }
-}
-
-func launchJob(content string) PipeReader {
-	pr, pw := Pipe()
-	go func() {
-		defer pw.Close()
-		
-		if err := pw.Write(Item{}); err != nil {
-			return
-		}
-	}
-	return pr
-}
-
- -

Further Reading

- -
    -
  1. Google Form: Channel send timeout
    2. -
  2. Go by Example: Timeouts
    3. -
  3. Google I/O 2013 - Advanced Go Concurrency Patterns
    4. -
  4. Go Concurrency Patterns Talk
    5. -
  5. Go Concurrency Patterns: Pipelines and cancellation
    6. -
- - - - -
-
-
-
- - - diff --git a/docs/doc/close_producer_from_consumer.md b/docs/doc/close_producer_from_consumer.md deleted file mode 100644 index 0b506e9..0000000 --- a/docs/doc/close_producer_from_consumer.md +++ /dev/null @@ -1,161 +0,0 @@ -# Closing the producer goroutine from the consumer - -The producer-and-consumer pattern is well used in Go concurrent programming. When -the consumer stops, we want to gracefully stop the producer as well. - -## Problem - -When a gRPC server receives a streaming request, it usually calls a -[function that returns a channel](https://talks.golang.org/2012/concurrency.slide#25), -reads the result from that channel and send the result to the client one by one. - -Take the following code for instance: upon receiving a request, the main goroutine -`Service` calls `launchJob`. `launchJob` starts a separate goroutine as an anonymous -function call and returns a channel. In the anonymous function, items will be sent to -channel. And `Service` on the otherside of the channel will reads from it. - -```go -func Service(req *Request, stream *StreamResponse) error { - result := launchJob(req.Content) - for r := range result { - if e := stream.Send(result); e != nil { - // should we signal the running goroutine so it will stop sending? - return e - } - } -} - -func launchJob(content string) chan Item { - c := make(chan Item) - - go func() { - defer close(c) - - acquireScarceResources() - defer releaseScarceResources() - - ... - // if stream.Send(result) returns an error and the Service returns, this will be blocked - c <- Item{} - ... - }() - - return c -} -``` - -There is a major problem in this implementation. As pointed out by the comment, -if the `Send` in `Service` returns an error, the `Service` function will return, -leaving the anonymous function being blocked on `c <- Item{}` forever. - -This problem is important because the leaking goroutine usually owns scarce system -resources such as network connection and memory. - - -## Solution: pipeline explicit cancellation - -Inspired by this [blog post](https://blog.golang.org/pipelines) section -*Explicit cancellation*, we can signal the cancellation via closing on a separate -channel. And we can follow the terminology as `io.Pipe`. - -```go -package sql - -import ( - "errors" -) - -var ErrClosedPipe = errors.New("pipe: write on closed pipe") - -// pipe follows the design at https://blog.golang.org/pipelines -// - wrCh: chan for piping data -// - done: chan for signaling Close from Reader to Writer -type pipe struct { - wrCh chan interface{} - done chan struct{} -} - -// PipeReader reads real data -type PipeReader struct { - p *pipe -} - -// PipeWriter writes real data -type PipeWriter struct { - p *pipe -} - -// Pipe creates a synchronous in-memory pipe. -// -// It is safe to call Read and Write in parallel with each other or with Close. -// Parallel calls to Read and parallel calls to Write are also safe: -// the individual calls will be gated sequentially. -func Pipe() (*PipeReader, *PipeWriter) { - p := &pipe{ - wrCh: make(chan interface{}), - done: make(chan struct{})} - return &PipeReader{p}, &PipeWriter{p} -} - -// Close closes the reader; subsequent writes to the -func (r *PipeReader) Close() { - close(r.p.done) -} - -// ReadAll returns the data chan. The caller should -// use it as `for r := range pr.ReadAll()` -func (r *PipeReader) ReadAll() chan interface{} { - return r.p.wrCh -} - -// Close closes the writer; subsequent ReadAll from the -// read half of the pipe will return a closed channel. -func (w *PipeWriter) Close() { - close(w.p.wrCh) -} - -// Write writes the item to the underlying data stream. -// It returns ErrClosedPipe when the data stream is closed. -func (w *PipeWriter) Write(item interface{}) error { - select { - case w.p.wrCh <- item: - return nil - case <-w.p.done: - return ErrClosedPipe - } -} -``` - -And the consumer and producer be can implemented as - -```go -func Service(req *Request, stream *StreamResponse) error { - pr := launchJob(req.Content) - defer pr.Close() - for r := range pr.ReadAll() { - if e := stream.Send(r); e != nil { - return e - } - } -} - -func launchJob(content string) PipeReader { - pr, pw := Pipe() - go func() { - defer pw.Close() - - if err := pw.Write(Item{}); err != nil { - return - } - } - return pr -} -``` - -## Further Reading - -1. [Google Form: Channel send timeout](https://groups.google.com/forum/#!topic/golang-nuts/Oth9CmJPoqo) -2. [Go by Example: Timeouts](https://gobyexample.com/timeouts) -3. [Google I/O 2013 - Advanced Go Concurrency Patterns](https://www.youtube.com/watch?v=QDDwwePbDtw&t=111s) -4. [Go Concurrency Patterns Talk](https://talks.golang.org/2012/concurrency.slide) -5. [Go Concurrency Patterns: Pipelines and cancellation](https://blog.golang.org/pipelines) \ No newline at end of file diff --git a/docs/doc/database_abstraction_layer.html b/docs/doc/database_abstraction_layer.html deleted file mode 100644 index ac6ff7b..0000000 --- a/docs/doc/database_abstraction_layer.html +++ /dev/null @@ -1,293 +0,0 @@ - - - - - - - - - Compatibility with Various SQL Engines - Sqlflow - - - - - - - - - - - - - - - - - -
-
- Sqlflow - -
- - -
-
-

This site uses Just the Docs, a documentation theme for Jekyll.

-
-
-
-
-
- -
-
- - Search -
-
-
- - - - -
-
-
- - - -
-

Compatibility with Various SQL Engines

- -

SQLFlow interacts with SQL engines like MySQL and Hive, while different SQL engines use variants of SQL syntax, it is important for SQLFlow to have an abstraction layer that hides such differences.

- -

SQLFlow calls Go’s standard database API. The submitter programs generated by SQLFlow call Python’s database API. Both APIs abstract the interface to various SQL engines; however, they are insufficient for SQLFlow to work. In this document, we examine all interactions between SQLFlow and the SQL engine so to identify what SQLFlow authors have to abstract in addition to calling Go’s and Python’s database APIs.

- -

Data Operations in Go

- -

Data Retrieval

- -

The basic idea of SQLFlow is to extend the SELECT statement of SQL to have the TRAIN and PREDICT clauses. For more discussion, please refer to the syntax design. SQLFlow translates such “extended SQL statements” into submitter programs, which forward the part from SELECT to TRAIN or PREDICT, which we call the “standard part”, to the SQL engine. SQLFlow also accepts the SELECT statement without TRAIN or PREDICt clauses and would forward such “standard statements” to the engine. It is noticeable that the “standard part” or “standard statements” are not standardized. For example, various engines use different syntax for joining.

- -
    -
  • MySQL: SELECT pet.name, comment FROM pet, event WHERE pet.name =event.name; with keyword WHERE .
    • -
  • Hive: SELECT pet.name, comment FROM pet JOIN event ON (pet.name =event.name) with keyword JOIN and ON.
    • -
  • ODPS and SQLite use either INNER JOIN or OUTER JOIN.
    • -
- -

Fortunately, as SQLFlow forwards the above parts to the engine, it doesn’t have to care much about the differences above.

- -

Metadata Retrieval

- -

To verify the semantics of users’ inputs, SQLFlow needs to retrieve the schema of tables. For example, the input might be

- -
SELECT name, age, income FROM employee TRAIN DNNRegressor WITH hidden_layers=[10,50,10] COLUMN name, agee LABEL income;
-
- -

In the above example, the user misspelled the field name age in the COLUMN clause as “agee”. SQLFlow must be able to find that out.

- -

To do that, SQLFlow needs to query the field names from the SQL engine. However, different engines use various syntax. For example:

- -
    -
  • MySQL: DESCRIBE/DESC employee;
    • -
  • Hive: DESCRIBE FORMATTED employee;
    • -
  • ODPS: DESC employee;
    • -
  • SQLite: PRAGMA table_info([employee]);
    • -
- -

The returned data format varies too. Our solution to avoid such differences is not-to-use-them; instead, SQLFlow retrieves the table schema by running a query like SELECT * FROM employee LIMIT 1; and inferring field types using the mechanism called DatabaseTypeName provided by SQL engines drivers beneath the Go’s standard database API.

- -

Prepare Prediction Table

- -

A SQLFlow prediction job writes its prediction results into a table. It prepares the prediction table by

- -
    -
  1. Dropping previous prediction table DROP TABLE IF EXISTS my_table;
    2. -
  2. Creating table with schema CREATE TABLE my_table (name1, type1, name2 type2);
    3. -
- -

Most SQL engines, including MySQL, Hive, ODPS, SQLite, support both statements.

- -

Translate Database Column Type to TensorFlow Feature Column Type

- -

After retrieving database column type name through DatabaseTypeName, we can derive TensorFlow’s feature column type via a mapping such as {"FLOAT", "DOUBLE"} -> tf.numeric_column.

- -

Save Model

- -

SQLFlow saves trained ML model by dumping the serialized the model directory into a table. It first creates a table by CREATE TABLE IF NOT EXISTS %s (id INT AUTO_INCREMENT, block BLOB, PRIMARY KEY (id)) and insert blobs by INSERT INTO %s (block) VALUES(?).

- -

Note that Hive and ODPS doesn’t have BLOB type, we need to use BINARY (docs at ODPS, Hive) instead.

- -

Also, note that Hive and ODPS doesn’t support AUTO_INCREMENT, we need to implemented auto increment logic in sqlfs.

- -

Load Model

- -

SQLFlow loads trained ML model by reading rows in a table and deserializing the blob to a model directory.

- -

It reads rows by running SELECT block FROM %s ORDER BY id, which is supported by most databases.

- -

Data Operations in Python

- -

Connect to SQL Engines

- -

Thanks to the Python database API, connecting to different databases follows a similar API.

- -
conn = mysql.connector.connect(user='scott', password='password',
-                               host='127.0.0.1',
-                               database='employees')
-conn = sqlite3.connect('path/to/your/sqlite/file')
-conn = pyhive.connect('localhost')
-
-cursor = conn.cursor()
-cursor.execute('select * from my_table;')
-
- -

Insert Prediction Result into Prediction Table

- -

Python database API provides execute_many(sql, value) to insert multiple values at once. So one can prepare the following insertion statement. Please be aware that MySQL and SQLite use INSERT INTO to insert rows while Hive and ODPS use INSERT INTO TABLE.

- -
-- MySQL, SQLite
-INSERT INTO table_name VALUES (value1, value2, value3, ...);
--- Hive, ODPS
-INSERT INTO TABLE table_name VALUES (value1, value2, value3, ...);
-
- - - - - -
-
-
-
- - - diff --git a/docs/doc/database_abstraction_layer.md b/docs/doc/database_abstraction_layer.md deleted file mode 100644 index c5d9821..0000000 --- a/docs/doc/database_abstraction_layer.md +++ /dev/null @@ -1,92 +0,0 @@ -# Compatibility with Various SQL Engines - -SQLFlow interacts with SQL engines like MySQL and Hive, while different SQL engines use variants of SQL syntax, it is important for SQLFlow to have an abstraction layer that hides such differences. - -SQLFlow calls Go's [standard database API](https://golang.org/pkg/database/sql/). The submitter programs generated by SQLFlow call Python's [database API](https://www.python.org/dev/peps/pep-0249/). Both APIs abstract the interface to various SQL engines; however, they are insufficient for SQLFlow to work. In this document, we examine all interactions between SQLFlow and the SQL engine so to identify what SQLFlow authors have to abstract in addition to calling Go's and Python's database APIs. - -## Data Operations in Go - -### Data Retrieval - -The basic idea of SQLFlow is to extend the SELECT statement of SQL to have the TRAIN and PREDICT clauses. For more discussion, please refer to the [syntax design](/doc/syntax.md). SQLFlow translates such "extended SQL statements" into submitter programs, which forward the part from SELECT to TRAIN or PREDICT, which we call the "standard part", to the SQL engine. SQLFlow also accepts the SELECT statement without TRAIN or PREDICt clauses and would forward such "standard statements" to the engine. It is noticeable that the "standard part" or "standard statements" are not standardized. For example, various engines use different syntax for joining. - -- MySQL: `SELECT pet.name, comment FROM pet, event WHERE pet.name =event.name;` with keyword `WHERE` . -- Hive: `SELECT pet.name, comment FROM pet JOIN event ON (pet.name =event.name)` with keyword `JOIN` and `ON`. -- ODPS and SQLite use either `INNER JOIN` or `OUTER JOIN`. - -Fortunately, as SQLFlow forwards the above parts to the engine, it doesn't have to care much about the differences above. - -### Metadata Retrieval - -To verify the semantics of users' inputs, SQLFlow needs to retrieve the schema of tables. For example, the input might be - -```SQL -SELECT name, age, income FROM employee TRAIN DNNRegressor WITH hidden_layers=[10,50,10] COLUMN name, agee LABEL income; -``` - -In the above example, the user misspelled the field name `age` in the COLUMN clause as "agee". SQLFlow must be able to find that out. - -To do that, SQLFlow needs to query the field names from the SQL engine. However, different engines use various syntax. For example: - -- MySQL: `DESCRIBE/DESC employee;` -- Hive: `DESCRIBE FORMATTED employee;` -- ODPS: `DESC employee;` -- SQLite: `PRAGMA table_info([employee]);` - -The returned data format varies too. Our solution to avoid such differences is not-to-use-them; instead, SQLFlow retrieves the table schema by running a query like `SELECT * FROM employee LIMIT 1;` and inferring field types using the mechanism called [DatabaseTypeName](https://golang.org/pkg/database/sql/#ColumnType.DatabaseTypeName) provided by SQL engines drivers beneath the Go's standard database API. - -### Prepare Prediction Table - -A SQLFlow prediction job writes its prediction results into a table. It prepares the prediction table by - -1. Dropping previous prediction table `DROP TABLE IF EXISTS my_table;` -2. Creating table with schema `CREATE TABLE my_table (name1, type1, name2 type2);` - -Most SQL engines, including MySQL, Hive, ODPS, SQLite, support both statements. - -### Translate Database Column Type to TensorFlow Feature Column Type - -After retrieving database column type name through [DatabaseTypeName](https://golang.org/pkg/database/sql/#ColumnType.DatabaseTypeName), we can derive TensorFlow's [feature column](https://www.tensorflow.org/guide/feature_columns) type via a mapping such as `{"FLOAT", "DOUBLE"} -> tf.numeric_column`. - -### Save Model - -SQLFlow saves trained ML model by dumping the serialized the model directory into a table. It first creates a table by `CREATE TABLE IF NOT EXISTS %s (id INT AUTO_INCREMENT, block BLOB, PRIMARY KEY (id))` and insert blobs by `INSERT INTO %s (block) VALUES(?)`. - -Note that Hive and ODPS doesn't have `BLOB` type, we need to use `BINARY` (docs at [ODPS](https://help.aliyun.com/document_detail/27821.html?spm=a2c4g.11186623.6.577.768231deoru03E), [Hive](https://cwiki.apache.org/confluence/display/Hive/LanguageManual+Types#LanguageManualTypes-MiscTypes)) instead. - -Also, note that Hive and ODPS doesn't support `AUTO_INCREMENT`, we need to implemented auto increment logic in `sqlfs`. - -### Load Model - -SQLFlow loads trained ML model by reading rows in a table and deserializing the blob to a model directory. - -It reads rows by running `SELECT block FROM %s ORDER BY id`, which is supported by most databases. - -## Data Operations in Python - -### Connect to SQL Engines - -Thanks to the Python database API, connecting to different databases follows a similar API. - -```python -conn = mysql.connector.connect(user='scott', password='password', - host='127.0.0.1', - database='employees') -conn = sqlite3.connect('path/to/your/sqlite/file') -conn = pyhive.connect('localhost') - -cursor = conn.cursor() -cursor.execute('select * from my_table;') -``` - -### Insert Prediction Result into Prediction Table - -Python database API provides `execute_many(sql, value)` to insert multiple values at once. So one can prepare the following insertion statement. Please be aware that MySQL and SQLite use `INSERT INTO` to insert rows while Hive and ODPS use `INSERT INTO TABLE`. - -```sql --- MySQL, SQLite -INSERT INTO table_name VALUES (value1, value2, value3, ...); --- Hive, ODPS -INSERT INTO TABLE table_name VALUES (value1, value2, value3, ...); -``` - diff --git a/docs/doc/figures/sqlflow-arch.graffle b/docs/doc/figures/sqlflow-arch.graffle deleted file mode 100644 index 1a9b5418a59a55332c7765e59ae259df44f9b059..0000000000000000000000000000000000000000 GIT binary patch literal 0 HcmV?d00001 literal 3985 zcmV;C4{q=uiwFP!000030PS6CciOlX{XF>3YKiib|>wuyxOFbHff;kZ5M0Z zT-S!*e`>#ZVXpu6dTsMxyGJ|iug81Jz;nY$IsUYLc(9|azi2kM1_O^Zo4f5@ z<@oTR)mB6v&F0>Rb!GiLiUu#6&CAQnhApic4)06Duz4Kt0Slt5Ly_4Fk)Y9yy6d8t zQSO=Qh;O^D6TM#hWAlPty_Utfjva~XpGRh!P5HgJYzKiYKm4&7MuDjRYf)T-_kFh) z@Zq3wBtE?j?6WhE(amOJlCnnBYG6&&1sjx7oHE&Lj+4=CZ-3CYdu)df zMVUddHigO@y8gwi+>4cYZ})r_-Nrr?)o7hFktpju&54OXz**vFH}x}msp6L!F7M}2 zMr^^3{=g%)_Ha`|!ur1u~B3!^${3AN>&-lbN0bx?l);5nK-hXcgq6TA| z6a*((pLyJl(r}wx(NPA4%_)3GCG7LS{l$IT6ZLlexF_ZnBsfbtZb?DgThXuX{yXpM z7o@WPZhPMj_W73nar@&zYwvvP{OFTP(CL?Vp3~bA|Lz`rLf_6$_dk1E;`+`WX}9sa z@87}|Ea3!67Ty~~}q^Q(Wz&Z4_C4>ITHe{2f|ZMnbLyzImHdP(lTDMAfP zybOX>U8ja&Y$%3>uWu++y2S~T1kHr^35I!e{c_^=&X?LTZ9Z5=_M6T0FnGf3?vd|Z z-HhZinu~KD13DR3oyh&c4(%%z9Qa-Kb5i*!O=rf8aWXLiZFT<{hEXa(6bxB&YO`%S z7jg*pyYblMIw{jhB@$C<6;7{D{q?uHp{Umz%5Ow;YePXuHLf$xn&hzK*L?& zYhXJfqf{C(f_bksA>rZ=6Z2g+9C-Fs%dx$Z$|1$|l~;;i%_c3RMIJGi+!AT zvq?chxv)|QM!FWCi5$l zfk&#}G%JUYlH9XtCrJq@B#o7yWT4UnxYYibi!ZZ1k?OaUcN$b}kPU?p^SaP@V;c)m zX=7mZx&Xp-uBALs(rn?m-Q}0rrb`>m;)2;^kZs>td{b7+4i9`5oY-A=m@jPPB~rVl z+$7~2TS1V!i%A8=%*113gKSNc{$xGdxymIzv%N4GAtr`x+x6xYq?Tv@#qCDtWw(>{ z`J7u8d{h=*4FOo2hk(py;E>W5)NyqV(HJN-+~%6MkqI4b>!; zMkysk(=8(w(O6v43kv!MUTU3-IcIksGr3Wl{fx@eW}gX7Q+V0V8k?W6!qoK!D!}(b z1(JHJbs|Ce49DaWx|uoh)W6h*ZV}D0utl+{85mi4kBg^zd^E{UKLE@Yt=>$akBn%l zmr{buyhj%5l((2-s1rnIsAKz~8~zt{bGv5{O z&bb|gEK-%pSo?tu*`L?K*5Muxgz=x9fORn;onFA0FF%}mLnhv1tABUc&w(xG;oXV7 z_`Wl+gKj7KNvJCRB81~NYRWf^_4v(H<(sC(CYlwS=z45I)YwF%mTwfrCR7YmalUTg z+tDzvz0P+>ei6y%B+NgvornkO+FE>fp@x-?bEYZ-cfgdcd{`VP!O-ulwcjX%fcFBs ze?4dZnh>sFkTTHcyB5^b0_sVtuAb0p)zfTcRIo8k4C)Eg(>=@|Dyt`xkY}Wx)SA_k zZh?9N^#tk()YEF!lUi9l6&Y)}>S^|~>4FWYCs0rKvw~AuJy9Js41_Fd8k%O9hFloL zw4ov)s|doDX=$n^wUl8#F)5{+bxSFuij<;ALk2cf%hHiabx=~Eq#lEk+HbdyJ8M3V z*eT~1Af8q!p70&SldNGA4|fXNsXEYJr%GA*n7T%UJSfC20hIJ-~sTS^`J+HxuIyQAM-Q=pa;-< zfYAHGPL%}c+U4kr*`?H5iUzzKxaE(DXx z3&BOkT1oVaxhD`38#sXgE5QpRKwrvAuo#Y|Cr)6&Tj1~tqlxpPj_>XuI!cFuu% zf}L|9m>%xVIStn&m|*7|=%xG9OK3@Yi3&foP{Y)SCH^BSHY^-3ZBpy0sU=jmT2d>h zC8BDEhABprS_ZYS1#$`G(qoWI-&nxkf9$L|QQ$RQ4ti;&dZ~JNRZXo$FO`U)f{kWT z5KJJL?qSAohvm6A*oscnhGiN;E2)Mq=vC833v^HI2@|$z)=b!}@`Oo@SSBTgNmYXo zSf2yQ^r$wldfc%+Mf~px^~5~!Q8WefBal!ll~C1VtYWhk2~{GG7KvgH5(*?#g-OI+ z*XStH8yc}pim6GksuA)e*XXcOw}PTo6qJnMhD_CV#RLvvz#)w1d(t-sUL7ZWu@1lk;63Y; zzA^R)#;eC3#B>7eJw)t@S*?63rF7Qb+dbFW>35e0JaI3p2Oh+90^mJF;2k7#Q;d8* z^Ogy^)ng6;<^Xfg@R)rJIAoctM;t^?1H?T<#Jw91t|BqFTjHD!8Ad9$g+ZQXPzzM| zo;hSz4r-yPBzaL-@`WW|2x68Y@0?fHm{RpcaV@1hs(m--QvdoDFlx zva&%fsA)XS&2p&j&2os|A)tkxs3+Jh2hs^5T7X)be;B;7SVFWWv2>$Wf`AsFl@|Wd z#id9ktm>(F79*APq9ZC^5cdn@5(KdT`E<(_iR$tR z*CL)F3FH-5$JJx(B&KrG#UwoqLxnN~gFE8yp{;iSbgj$RStu8!yxCg2Wm z_pF0g#L%lFcm@1SJwWdn$F7KRS4Zp$__cb#-7^hckzj7+kb@9u0J-NFxgy3}9g!;_ zE*b#tnFg*Px++(Da1J~M5QhM}30?HL%y8#i=pG%!U5lAB)d8OZx@Ja=%5of>IR{fK6=%-@8Zn$yqkGOASFjI8LN@mY1U+Nw1XH9MdwgoK(H*ar|t$ys; z4(qe*CX$)mRsDxO5q`|W1zlMn&sFg%RwplVdtSXUbx<)i@G9^s@ap2cYRQ>1r6V?# z=OZb4pcHCoPmMxBngWFah2|-=&HRuDayj~;bEhFBDK=v2Gz5)Ws`XGGR=~a!poS+v z4Ttt1;sZ$s{lPGLU~Hfh>(NLl9cH628wI`uDfLju(m7=@(js9yM%EwfJ}V8+A&XSBvLwgii`#??YI&~1 z!o{of*=3gtWj~kv@3R>&G=ppg^jMmqz!IYVd>GB=i7nsl+Yu{4d?DbCv~b+O;qP64 z*A3IRpTr2sjAm2&789&$=XN-bg{N9bnmlVB1+Ewlaw%sipsC{>`ApL79PT{fmpLgg z-Gy3`t1xd3Y)3R{DnVRpEge)o8Ce*M+HREL?xa6AWh=?Q!rJ%P&g#MJjkb?nR?VTjZo!{k`X;@A`IT~MvK_F@!xAUY0W(=Tlg1zMN%l}5rpK7<%P%(-`8KkK$ z^`lYLl)tVSD;%s+P5#5k#qgaBBpE`Ps;$ZOlcAS-M3^@a&3S6ybU7ObaOx>)$=$d4 zpyk=8`%FFtnx+|fuQv7GgjYlN)MF>Cmxf$X<{_fXGv-JwYlMunNvJPZjx_mfd&5Fs r6vwW4QoNXZ*;g6jpPev*K7X=@wBJ|Ru2IHP{6b- diff --git a/docs/doc/figures/sqlflow-arch.png b/docs/doc/figures/sqlflow-arch.png deleted file mode 100644 index 0a5a20236049dee1546fa2f1357d6f3320edaf8b..0000000000000000000000000000000000000000 GIT binary patch literal 0 HcmV?d00001 literal 99876 zcmeFZWmuG3-!}|6bcaYYlpxX|-AFfzbax{SQUgdxry!k*l!SDr!T<_L2na|?iXdI@ z8r^m6=epl7_wgRb`{CIK8{uST&huQe*8l&jHBlOB^0-(OSV%}nxQYt*wUCe?QbLGFdXIyTgNt4q3krpbx>;HaYu%Uo)f|;w{Wv_^{{hsh9X|q%-qG(LyVpt zaiYKe`+c1rcGiEN$=Usn+X6SpiTDdAHwPEzU#|_GDvCHNtYziy;^>KZd2MGq4{;vR zUl0EG;lI!J`&lwBPOffN?(X1e;`jb|%0G|&`|;{-c2?m25T_C6{^RlgJofL$YuLJY zxPY;Cv$Ifi_ONmTC;Q{se_r7K{D?nZOOz8aT>lup-|x+@qu|XE#}eiI>x~x23LmRv zL_(58QoJwq&H<-+Iw!S(TM9Ov-7kM<4QKKH<@b`V=e+Hn}B1my?h% zi%4O_YJy8XT_LE)q%0bX9W7ovXAO5Qb=|j>EA-#g{l{ca2!yJ(5Um%%n*M{66$VzTgQ+~e`zrz8+m63f-$2Kk zvu8cDR?>-i%8X0|{rfAPc(LP{fU?o{W5MOo_RInN+0DB}ftP1={pMUmP?!hwjr}K9 zyU%qS`GxZFcs$2Zi5nU=lZg<8;J?4L5|L7&`ET`a)m~V2Z+N}hu!uWG$uGG%SwShvD3buj^B(dvCeLy79Gdj^I<&6AWF$;* z^p7veB%(YbMunbqE3PSzt;tvy&rsI_0nCKXYm81?9cB(WN&G#)z=;Ncu3TDA0x$B)NKTb zt&7C5i>$njFDq9|-ien4WCy@^2tP`g z(JH?_zk{7D4Y@Fm&1aEX8_oUBXFu_dTUjs5|8!r-8Z-7_(!ix@WB$qawzwI^Zr|Oe z^{mft9Yj#ENL)vR)~Rg1_C`^xNnD@CyC2QiILv+Ri{3izeu_PR5;>J8bt6`%)>?JF z{e0Iwh8^Wl!fhooJZS$tb2g2bpWy1yb8n>-mLxi6hK~aU-BdEpN|8eZm>t)S-!)BrV`edhZ@43A=t>ZbrUq!tvh8yy}h^)%JthsW2V; zqw14G_DtgQbf&KFx`YZ1;-p`U;`0?3X;A{Jx=fW4-a1SPqhgazNX#^R@hjtXU6foe zN>+1=sX$r>r=Q}>N8+9lv}Vn7Km4R4ay;iAKf(cXOx`FGR$J*)j0)aLPkA;f-IYkoAI(uN=l%nh;_kRxdVGXhl*AZL28u) zFMh}iz~`FW@4<%!7Pgy^x|l;YvNP>hZ7m+=NzwU)A`7@p@xJLYK>jl1p4xq)Y; z6jnP69UbXb{FPIo_~NhDu}Rrn=VJj|P+D5+3xd&Amo*>GIS+tzXsS$>+U(UqP+w z<DnN{@ASFoC4ZKuytu}*lYZWQgNlapSf6$?!+1WS3|^Z!L(fmz zc3QJ56(!t9cWM3ir1nV@m>7_W6qHb`@ATiBZ}vRc#0e|1+^y1>Z}H|d|NK=yD6n+5 zUp$S9{T=wWv$(3|(-iO9G+&@8z$#TB^--OoA1zY!`r=@$)<#{g&F}5@LWfCzEah{i zqH6r4zG1c~%?7PvC95Byp`N2u$MF(Z1hqJFF}MT*Uq-S;=G~vvJUMn2z}-qyTX2{x zNA$kC*{`!wJY^~K*nf2h0zBkeoSz5F<2}@Jkc}5vT!x78qkchFU?|JJ>w2%m5IRn zl=o)iW8-5B0sW~;qfhul4@=atS#$mOqsSOuYBhV<;ofMW)5tTut(eL0w9ovD5J!VQrYvf?lc5~9kWVx`oqhK0acp&z}(pN6LoUEhe{ zbpQPJQKn%;)^B4j#N_S9y4Ed&WTDzb%b!M_eP@w0)WXtf2d+ssw zgoh0+J#{V%U%L34>&vH0yvK?Zoj-5*ppd4lKo>vNpPc*gnt&b)BO_3Vqa;%{8zAMj z{Tjyneo~bPrGz;z*H<$sj4Rqq__@w``mL(WSD&J)dO{0(=jnV{B&-Jgx<_$T zLIujMrkM~leF2v*F+o@6Dct-Dp98Ne4pnE*m>=}DeI|YNobZcDa%pWfn=w7X5BC{m zf~0&7n)a#ccqSQiUMo{t<)rXES8ma{0ns^op(_1v*Y)l>I-iS!sm(zv&dUE{eLHgT z9b4bCo5WO0SR}VUz1GI2s4(w)Ce_3B7Xe%u{ZDOYS$&_4`}$0aY+c)!TfMrVK43Q z=Zm&+H4#+`en*z^5>;%qZU(wm`w04*nFD1`U5CY`)0BE^B=n0L>uy+YLw!z+2gK7D zsS^X(6e+rU`u1FD4_wPuTkW$jlib5qVE@OpA}$vdTrQCrLxv7ye@Km;Ed#>rhqjFt30K+s;-|aLka2V7(COT9$$?4ZS-Z?-1?maHkT&7bs z=_7IZGkI}z-^hEO34pIO@qn{;l}0Tq;grq^pif&XP}-dgf+BTyYpUuw0F6Tc3)6vd zPenn+`sSN2{dA?O<0=W8vsI%+C8NM<_e)y~JX=I(sL^ixw(92j_YR$gvdU(q1R4_V zls$}KZW22UxzoXm(}Pc!(VXoGZBLGyjU(`=$J!yX3cGVnM#tM~2}j_>H3uakzeflU zj8M)EpBM@v4QFkajZhjV~McEV|{@$uO#%fmS?>;8!53_XU* z*>b(w2`5lKHzzBNL@verPh16JG(!aY+aoMexZ_THwx{c2yL)j%lXZ2g?xA2Zx}Cdno`K^;h|T)y`ETS&M-V;4 zU(V-?e@)9V-fG-7_4E_1Ibk}WDr}^eg955Y+@r_U)1ol3BvzY{`yUbzJ-P7WgC50E zg=owsJcCz$4~^Dkv6`fx<(5bI(_xe$IRrT{k~5#*3m8`wXp$&;z3O zHbbrD=ep6W6h*|eg5X^{vklnR_V<4^SSb67)m7M5weNv-scv;a>-E*Sb4|L~{&wx? zrWSN6k(rLTlZdsDWjxT_#u0%g0lN_C;vwDZb7Q#s&7FWPGsGVNcwv#3fO%^b%HcMx z=RD#0i61soYZH&?#aj8&s76VieK&P1btl8PeiKC49wt~T;=KJB@kr;oKCNWJ{$R~4 z@mk<~Cjb>Lq<-#=S!n<3M-}?v)6Cf(NuPfPXX6$cq`cNHv)&tsA6J(!%6s@G$bTuk z_|BuNdX4gW!E#DRc}voZ>ONX4lWHsy196iMjnV3dHpjd3V0~&yD>` zAH05e$Md+svJZayV`IN=DdiM=Tcbh8*jIM`4ORaIo!8~gd%%?Vva033lFdT>c3Y^> zl)c_SdN>B8S^@dc5N{-xd8HQz%AaO!w;P#7X18^w+d|3lZP|=iE$LYd8#%i;uex zGR7@v4?Jk2hl!^dO0Bca7ZacL)lj^(|1`GpQav|^j&l2z+pJSVRP{!lNSSE_sRb=q zn&cE*>CYNYg>I5b_w7@fWbiw^pb)Omt1Y3>f0}zryp4AHb6mZz{MtQeme^gU?xj{o zU;qsHAzZwR2b)Sa&SE?0-Md7Fc~{Z*Lda|pT@UX&P~GH}c=H4w((e{KrbNnN$O5Y4R#Yeh0*S!EH`zinggsw#Z0upmzx@bLL&>yLiQ=IG%e&W3!C*s{8G3`!gcSw1wYV z&=TRIn5d{R4z)O#wR;)daz2phf$rtL4`1kjk_2(WY@wbmAjOIppn0OQ^Q05EUnTz} zxD)>wI*ZJn98E~W;rI6QN`$6Zmhxs&XJyWB{b7u_f<;y*jO>a0uGRgMCP;O`e zLZ|H~3eW(G6=n)GXtPzVmq)YYW4kl8FL7u@b^VTa+0&cMy^23deBqkJDtwXPceoif zbMH1QuP&cdU?b#1xCXy!j;P)&Di%Km&Tcn;tqh+n^ZsqWi3)Nq$D*-a{mmY@Teik!AKtcbpR;5wwB%ufUO*kFHIQVZ{-UMA= z;q10c!1sx42I8pO)r5p2R0%9nw8y!Ep4|&7bp4Ted$7T4N1s_Fs3Y;pi!UfWo86!u zf|ZmMSTcHBwP}4cH~y`HS|YtHRjs>5xSJTjE}J$u>%xF{x?jGzS(9<+YiED-&1o3- zGw!#F7EU2#GHr1#9elh>2t}yWh=P*y0X>_mz3vDMH$h4lMUPTSP# zE0g=YJm+NRx_sx!9d*+KqO$|$2N3Jmb~2p{m-PA0?Mew@q>Qq>0%8JoO3se`ugp=b zD7Z`|!`&*3+w@bRdzoBjAzL=MoApk!k5Xxt2jX=?U-v(XwWxRY!Y!oFnZfPnt_`B9 zH5}xw4drW#7e9NadiNvq|Kf@shOFRpY%~ceLkFv?XR!xYS{M?B3jihAjG&GvH${t_ zA)30QSrB$czWiDZEYBt_Y%0?w38e&qRGxt;@$fu40YC z%;>(9ROHy$#HgrwrGrB@7)DJ{t7@;6dRH6k#yR`@$&$pH*~Z6(pHH}yt!q@x57vhD z7c93qjqR})&?MVp_O3RupRO|!h@mJR8^+by8v}`>`NyYzY3u!|hl$cPzq1cOgZJAq4N> z2G>apDD}N1g7M?M0Ji%m8{}*{evdNW{)hJbrw*+k05E2eL3C~VzcgLMuRzIv1yLLi zw|V>D7{@$oj?AihUpgLSkV zB2>W(q|L3hc;(fZfv+JM89re^NT;z=c*IZ*0XS|m!^xXiIjIT3Xe2KXoj+)8BaPu) zxB0El#nzBGY0w&++Y=L9MYGdD(tqJSKuVGVQaP6?Hwz|&VGpQ#%Gtu6MW$ULZhvwB zxtV&WOg`HWlHzCI7<|7*F&6>wz~lGR_E5eRs7dysjmO^?Oz>#M%L5zc+@cLVKFYdn zPjd@YARz5{gV`d?6~Lk?E=qpK1RxnG{rK8fED1x=YVmT)5PV#4HWOdfex{^fXGj0U z$E!>vE{&>;#cVo*k>fN~zPb3zgyT`a?vS{?aK8kg+nHIX(r~lh=ZCo`?;ycTjeKdf zD&w}F-j`>G)P)zWXGc3*q8EETc>B^bjwe<_X+i+1KJZ?#Bf0$o0oVBgy{xoHSps_t z*b8)9Ba@}u;8MUQh{T0ORt;~{@6+T^XLRnC__p~~VH^Iy@Sho=54J9N6aRE1Hh|6J zQ9GU3++QWc+DG+Ytt5b1OGTKJ7T1G@#gh zZW8!-f93shdIk+13HhTuxMU*`?1T|a=D@5cyi|aMS;IDc=Jtq1yjY1d_AQPHuJ0Y! z?X@lo;$q+Dnrc-?X6o#c(=2);2{$KSX#%0BSZO3T=xQUJoX0Y<-=N*UX`!YWxzL9w zx<1nU2fZ5BArHdNf$~G>+xcsw-#zj1vP5agTsz!s<8bX`*t4RIMtTv|Elw&da8&(4 z?=<|UD*~{jmzLlnFKmUa(?EkWjp|VGY2dwrK6|F&MZ-0>0)ne_tGl(?>YV34<65K* zy}Aa{{d)6~B^dq9z~g!EG7;-kK9?_sqv&{4-_n_Q>d4SK%&n2(}<$R#@n!4-SERDNFiGKbZ= z=^2!!S9E$JFfv1v9x#!_+o~^KsS?(v$eQB%UtOGX#C z2p_VDv|sEi+s!prvr6ZE;8HeQ0gO~bHfp#Czy=%Msu_( z4TW$gTfAL>F0Pu^1BX<dECZ17k(` zda6W`H()_)%0J-55S830KQS&;!i~VaE_Cv-exfv>%m{Z}9%Z)l_Ui6<^>j+!`KlLM zGP(Tpma)e1^$_2Q}OMeb*1C zOuWUl&-Hy}Ie)xuzyK&a+VxpQp?|>} zp(fGG{#CI#&a}L^-(^St*!2b|ILg~WOEUw~>-}Jb-d#wb0s6Nu3h$>ny2U5idviA+ zuC>wii(LO0 zc}T{mPd}VkD@JdLyR9gAKG`gNX(*7zreF6?XgP*=0^iPDAOjFc;dCjKF193_9N&Hf zJPC)5^-beEPUA3oeKnJ+Emn9!_`fSNfEVQ~Ci?Z(P&BHUXNZr-@gf%z1d}paH@$GO zu>uYaip5Lo^Vx(080rz8bV-)apx*j0@Q>aniq8s$c-KgvHoDrzXBS{*++R zt`*bgM+uQrICq<(Y|9%JPC?|Vs~>pA2P@c1pX&GaXhT=F?z3qM6Ct3A{`~(0RbpCt zifmeHC;uB7@j{@HmfJQ@{~H<+L*Rg%oe};2f&(N!69P1{^CIi!f5AYopNxQ^_|=Oz z{08~|8M3jH2LPhZ!R=A{e~~fAk^x@XnAt7A_kRE&V8~M=vr3IszSlGRFCw#oIs$h6 ze|V!nsB937lOe#o9n3FjfLP-SP%0GWEZ6SGe;Gt|vj4C4&<`c$8X7uc!HuHil_`@T zTp`gL{o$P;TT!f%!Bef6RSIyQPnskU0FOYoKUwP&YX!*a39(7ix932@I8$Lz1d=e3 zeGD-kGUsu%t6-6ZE*O#OPvfm4_Z%iufL~i8^t4xp7FZ_;As|6g5VF5MdQJR&X^9|O zRkuzGhJt!UnK?n6$$GA9emXlhGr)!!IPfg^`MGX+x-XLvt5oY)Qn~y)2uH_-TZVXm z-@7GpE`G?Dz;majug~t9CksAc0_chjX4K-PnhgXy@zFh7F_6HX6ShJ@C!W$sDqs3@7NNq1Q?B@nqy}b9(6h|IqiIBxS?XIIfaH0^Se&bff zhsFp2`-yMyK$LvQW9O$y*C$R|F+yY}ylp`kVo{fXStLb^ogpiXwxW{Y;dsbb#e3xu zPsS|`0u@ZU*a}gc$XpWV+qG|TTVEkXajLG#q1Y*!=AIq|7}Q!1!c^G-iZF;Dq9$A0Gxd<6jQP+yoT z`D6G4yKW350?6;#r{v-OwiW4M0zkSFjy&GO*pMZ|J>C=c+@xaFulqKCJ4%2-<+gSK zFlzFOfv1clHe91pE1qUuF|;X{jFi@+A*!?(EAYh?5z<-jr*!eK9)7Xv_@#A5D3Gd! z^L7=P2nAl;LxMMu2X9a<-k`ZxMK>Wo#x>&m-mMZ1sbT)m#}+%|l=Aw?gH@Uww&RY| zgLM(PsFhaWCC`w)Kc>O!H+6bJKhHOPXKFypGYcdqF2uvPZcedRnkd;qd}xQRkY9|5 z8`jda9zK4Gir+_YM;VDy{br#3HU2TM2vzA;gJlmj%(8% za~CRit4B)&CoQo)LAdA?Jz!ZehDZh_Z#-BNw2v!;|7aItDOK;pBRHW_x~#9eXX~K$ z=B*ix`=!#}O%sPPJ<{$Z0t5HmfAU2*8MrGU zm_!QV1>XZK`ais&BDCTC#lqS*%Q)&aTgN!sVJSs1l%sf_fz#3{ulin_^yI!Z!S_@y znald_AM86|l5{Xzg2iEOL0ML*5MJ|0K?Ftpb;C*t7+u^Hpa2zMuJEGzj8av=t7oWU z;niJKDHHUdJJ`R~SsTl-r%FpNED#dslNWKn_$?fyV5~gT5>^TSM{|I_1w;Qx-re6G z?*jRJnrihoFTzJQbZOjmy)?UxQ7O+Vi37pHypL!fAB*1C{o{Rmi%G@T=VB381=k#= z;u_Q)`}NjM6LorqZ{bc+S9aVThHWJx1RX0BLp)+CmghYXn_hy)jGe1ozXlu}q-$J> zdelo&?EOB8sY=Y_o5owY0zM0a|U{vhj;`ovro!^^wrYp}96d1QOpHeeQi82OM1|Fhzck6V=?`{E6odiA=RBv2&>S}a8ErjX%S;jx@R}%QWI&E1PO8?fN!tJg30RxA5W02BodPzzTmD zTqNsWo3upKNn6|GzOGbl-uDxK7Q}|w4DAUyWW$+ZK^Ti?U7iqBjLN%)CJ%!zdLnER z2|M@apJLx4<+6CeeH(RJpwLnV;2+1%4;jB^IMfN_XE-i_3PUlNLnq8g=Yk*_10&%u z)HMz`)ZUn?%DoQu04*0o326$ylEF~DvC=g_Qvpf=84*2DVXN+b*!Jb4!f2VD2IkjK zX;y8&C%!bjsTZuN#Gt|1--3vU9Zksw<(9Eee(U*eOA?4VAku6rFmxIP8>(^<%=`C| zrf@514dokiT#dj>zkrxdQ;W>W@~T!9%Q8hy>8Bhb(u&|3qC0_`0M?7;WJ=|B?D5&WJ0G$Po%(u(x~Rz*6GTQ zxG@rXVPnjn<#W{6-+)FK{hs+7pgDR%iEMrahnqp3W@;0%5BE5&gU)7d&L9RoNlhfPn=vg-O4}`mbs>*#Keo;f{l#)CP!@fo!bg3jr67=SE%`z$%qL2->P> zSh>FXeqAK}6svg|sYGAKZ29g?@g+#w*->;{ot61t{21_SxBW;Tdj`B;B_M1S0htF8 z2|}zo@ILjck`CtoA=r6FY3zJ{>-oBYGMM>q-jQ%0K{pEeO*MZHQc`^9+Hj^ea!4^3 zi1MWXnb3-51;A`8u-IyT3H}R#=M~Vnn`~uGTa_&krI3Fa>7SuIwjZDyvkg_IVw~N; zCO4>T*%kI(juqg8gJdJ8`}#=gFQZ(XgcZ03;}`kR42huj^lax9pj8#V(v1@!Tsk+&1tW;ofd4+0jhz?Sv_E*aPd97o!T{)vAo4U;C*tu9ioiM`$@%W` zc%kaCMw+_Bvj#%*;l zj{EED$9rb=k08t$W_s7wv1t$z3 zJ$&`*L+l^(8wdG-*~CV)b*~#6$W92u5Hqd+K7G0#j2!w1gj?wl(k_Qy7fw7#tAAsA zhVY54M{_FBy&Sq6@!|9^Gi?m7VqXxQWdb1;{yUfbczi(4^_=HuKQYgCLW$5A_hPsX z04*>VV=#)Gu3jab{6V$3Y_?qpIb-?^Pr#W+_NR~;X4lXq-32|CjPcc9vM*T7;7fr7 znLmwD);UG&^=ykv5e$kD;xho?A!(oPvmjA8+NtsA7Fll5I*y9g~KnuX`OkncV zrLRIaoo_d6Hhf2)JKCM!+G6auzIqbAaDvc%|7Rc3Ulrg-hH%6_1<<_lLZook1UF6918chY-Pjy%mp_j0Zm)IKEcEq%e0Z z!Ptb&+ySEgEtbaskuU*0606z+NAB-^7e5f;_zPX$mwe1zmhT8Ii4TBqb|)p9sJaKK^h(hmRcaDsWrsQT?mA&ce~vd zJCP9)rhD5m_+pGM?WZNJK-L|5>!8!|W9bUus6k)>vw*C5;UYyF_=|RRh?7@!1qTUR zqlxLQbcCrI1Ji`H;Ttw2D!sB}z-RG}sn!8Gujo17_3RKPsdU^>zpf*6uiP$Y!RfeL zwUS&ts3J2eFZdJ@O4K1}z1e(@(_CX;d5oVCVH^rYI6dc;c+PVzz7$3bk+(jEcRa?O z@ejBJ!aIY;v{~sSP*M|#?mLMIszsJJaeaZ+Oy+hP6P}@y%;d+^=FJ`Ls9%j#TM33O zvjz2T$f3Mf1)UpQ?6y4SCX;N}croX_5U?5e^PA==2%13naZMK%dx_OkK&q)f zwucyyoNL0)ys2E%C;E;zu_?&_%_gZ0edg?R4nk(H_isE4129qtp$|~spXJ+&CM$#y ztLvjppLODp%-wZv-ngwj3B3fp%PxCVc#CZqPuawD@utCw^xnW*#pK(5>RRZ5MYu4H zT=zyf*EWb>xY?Es3;BW zjlJ5L$?i+unJSKfg)uM<^e*6eYRlw3QF<5Up{L0OX8sKeb%E1swP1c9y*!U)=Nx{z zi3s?EVp<@BWZ33GtPYAoi`v{77znRx#szo=mc-D7zwmd(q~{hPTaT>35bl{^F#>Na{(7)?m^{jZ9L#T?98w^T^=GWaNfjU#v^Go$Q%))o1^?PX#!3Y^8d2wuz6TN1OKyo-c>(Wt{dy zB}^zp0n4B#IPuhWOHD1NLslnf1YL)B8uLwR_t~{BX3U4KS9Fk2>1hZokxK!z`NK@w zKUy&XE=rWoNd6pRqkuQl83-PFDsO?k3fh0Thp}>>z9z}hpZHp31$%KCHoB&Hbx$*= zeZ3i2ZP{C=2+ctFu__f^HASc}0RNup6=oGzf?Z1^UrE^`vPHa`<#3%|a8+Vilyb!( z2sWbqz-$Q>dfoA?A~YA^V<3N$!DS|Gyji^w?I& z=c!4Z4;M9&f*0U~j8Cwu{c{+&OgoWEO*#V8{LE(SbrnIT*+W?^Q``~O84?K1gz&Q= zrY8bK0R<9qhyEZ_I=-Xy7*FKh&AX30yr07iUMu{*ujeWN1`du*O!u*-DG^ld4WY#0 zF)nI?bNSSn>9G^-F$L?-S2>R{?!`L_>2Jx>Cg807D%-u_2TXWYXFa_yB_7w@le0rVjwcv;Zi zR&WJ_4wA)Zmjpt7iV6JxXj2N%GXM~y#2=I>pde$e46a_+42_7HnrDb!7B7vmKj6Fi zeR3%572&(xrsV9FVZn5XZd5GdccW{GSPL_s7(|X;H&-}ugPfO>6*+d}8q4o#Vc1 zo#$_nc83@tRl~@Lxj0vT%W%TT?&@D1F31+r3$t1m`o#h-8V^A269i4g6Yl3SKt}9^ z2E0NlVGo5K6OcHI<8FPmZvEu`I;fOh#=fhmICh8_?Fi~E!^P%|iqd;o+-lok(ZHL4 zoR$rGj@%{XJ^Pgub_3FY0q*|fY>*}W5PJ+lXSPCEq-@^+ZJo{#^{v^mv5ptSrQmb* z+MZql)?eQp?bj)_0q0&*xqpMo4cTCekDUy%*HuL~2J$)o=;J;)k_+tteJo|-hXmOWIFXO`>3~v~-+@L`Xr1zK+y&QxQu}icvy%omZn92A8 zxpx&X4BX5Qz?EM4_OXx$v`hq_=pRrcjThaIyZ}*LbL&Kf>#IvE5X=XHGs*l{cfg`} z%zH5q+l;22>LPH+=@D^ML|Vx->(d|*6jni!o_IR}A?5vsgy8$0dax3~wc$6V-=MG` z!X8Mypg-97> z1)goZj;Ti8nkZ%9D!?j^NR6QJi^L+p4*eszg;+-kxM1$rAj{bNt_eSxg2~AeKyR=L zb{$h>#*9-a0@!bVEz<=nV0mzBp;!C&VW*i2WwnVj;ym?Gf2VAzYI=7$d(Z%da{B?&1oliC=TE zjTE74kUeJmv!5F^Af%jdAPf*zJ9~fTiVtSVtSR3S*F$qaQ!7lmEcd4Y@`1}eIj4fpKqCaqv? zMTO&B<45iEiAo~_kZCTPI-mDmsAw6H%s>`e&vO4c%lg1n6oFIB>xWa7IhyX%1fH02 z8ypM*cTE(^eR#b5t#WJVn2dBV1Lr1LTrBlWe)g5IrKV6DqgKW*qn|Xk`W{Qa zD!PbJMh*P-7%ib02t-Mx;x?7j<3tJzfzq}ts6sI@!9kI3M^n1{CnHfAS zJR zaMsil9sr|T^%@ik@1E%}u#>rX0~&B3Kg8{lkPY@b4*6|)vzWnY?+#p)x-Vxme9->_ zN}@2Z9K7az_wkYI9Kwz%<&<$?K@_8-GeA$Fu7&r_bOd$KF2ty<;ns)ud1hrV3iH>? z{tPs=_O`sm7A^91iN*~W=xl)=pxqO34b@Tcp1<$9`~tu429IF~h|Ac?;QiSlIhugL zIWOee{ZFYZEaXRTaTvbp!}*ydgH(v)5u<6o4ieh#a2(bFHi_}|xD27x{1>qOW86p& zIGjUufQ-cpPNY6?{FxU^K49&$;FF{6k2dw({M(?acMC^c3_q40>Br39$DXK!_3MER z3dQ4K)9yNGU=!8pnvYtRGhF-iVEw-JwUDjC#z`*g6VSFCN!Fv7b3SbK!S8#_`$(fK z5lMC^!4sY}`8}0v42WA2Q_BN1TI7OxFFd4&|=|8V(XL1Pt={uUt3&Kg3qJ`s=W40h>t~o%*R$~%+TO=B?6qt z#CScCilDE2bb*$1NF&Bg$R&zmHFb2<3earKMe?WTWH}ND8Ot5)7HAwLuqe{`2ze4- zZrC)&&0F~MH1Za*;5DFoKS!T>K4wXooBc1%8yhwSgUX=RwId)GWX~E1~wWG|NTETL-UYHWm&p@z5mP%Y2K0 zxX-jPu}n_O8UuP={k{rp)AvHSGE$*nt6^grdcfE;OojmHpZAfDg+M9n#;XK ze(znwDcl>17D1A-oJAC*O;D`~h znOhnw^xY%OkbNLJtPXO-F^oYLy5U1eSaX<1%8ch0;g9VlV~48@nAAUaC}{bHHiuJ zkHnJ>u}bA3aVvlLsm^niCC5s|yyBQe5M%OcXVWF?=~2bQ(j*IY zJ9-f=a1`L5{21)ltJ>8?LxvGSMjJr{N_QR|pQd~_G0-(@uY;+z`s*$G7Hb-td)y(tKZSy0{;FWOw=DRI+Q05lU-K`V_GbWM;SC{EKsBMm619R|Zq1W?ZB5 z)mSA^x)2*nDEi?O1ci?8YeG@$SHseg^TLSM5Zh|@02NtTjRC#*UqE-2m_?&HG(FvssXk^zowdDegbsZ z+Qk!qB#HohO%n0mWearel%YNUJetAF9NnQ^*J-gnlC5`}+NxYnSH*#17vx;S>{44m4^2H?1+tnIitI9YZRJ7tIXSp!9m-`ZAYQCNH57 zfIhI9qyO^KJQ@7~6ReINi3#1>4B;H$Q3)G&B2JXyHq(ny9 zWH(-7UV^ptD7PP4P8Rxxe7|Ao3|WbrNq;xoppBOKgq4Vd@wi2 z!&{*~N2kv&oi3NqpGHAeEVT_{NPA-Gj5ES(ma^w{o+~v#lzlOsK+o2sVdC5ZR*&gNDiy)a5dM z>hhQT`1bR-X#ESq-n->yZwggjnGxXmo;xQxOj%kqAUz90v88p_(25AY4llZkS#$^4 zfq3O*Zs4FKQcIFvCe+6m^~ng7XIrm}BQisKDcuSpmFpg9fm_(EHUOv&bs^42qX{^> ztoMDHbd7_7M~9J94K5}&Us`=ffSWeq+H~5p;hgLM!v97XjwWX4gtcLO(MCzvPcbgu zKkmMEY`8D1hVj${Oe&Wwt8=f}dM8h~b(*-HNMWvdw7UZ_B;fW3!MiDsjadRFK0~Sv zR=T0A+XH6XR`V#w{AWqQK`7sPK+r(47jkjT)?C64lf<`Kptdmf{~J`{!hAO9QMmE&nQy<9Lx))hWJ1UkVmyl{@|i3 z6j6h=M2Huzh4`cx9MXLm%w#>x5S~00YsL7`YzG!S%!_%Bwdf>zNDV)|kpeUpb6UdZ zDh1ju(jT^u(}l+g{L;zwgbcdUhd0F536PWsSk&@dyK4LA^I{Sto6PGwsVK-wc-jorN@bV?um+pxe zqI_Fe%d?b$E}M@!q7*mro*BMiFgI6JBJ5upgX_s&2RDR_xC%BdBFGC|1m(T3_j7sawJO#;Ts5fZR;vcG8tjV z9!a<4Uw(+QRj<4IOXKLKbCni^jq{aTFv@E$G`Qr(P>Vc5csqzsl>jo~ATTi8Q*JYW z9Hk{=!d)GXJ4VfQ6XkjjfS1Go_6?-b%RaNjA$e&MG&-?#c~7miFhMo$`++9Qjbtbc z2O{Y8E;v*|0lEoEvYR6~9OUk*V@Alh90>?>*Ch*M<7j{mp#y z{{PNA4|5M5PF?%jd#ztB2+!%HP3A$qfVemClK1>)Ai2U2?y2)oxFkXtk7hU_4e7qh zOk${BJDyFJo{7IMfLQF>xhnBs9iP3&-cBUa+TL+4t`9k#&3;7F3Mhf)fvep5V-o1w z-te)VhruSHL`->Vv-WCmi2jVLN7HLckIQJQTgyo4m(sCvQWGRe;gL zxQIBU{dXp)p1pglJc4z!RUDX)I%m@8_sK~_Hz*y0Ug*vQVH6sb#*9vd)DoroHXq=s zze3TKjyfk7_4ooF)>#ek!#)ATwsMVM&#-kqM5FQL0`>HOJVSNl?R$y{R z89)6-D;m6Z#T>Ae%tIiTUCU3mdL)0Y)Fiz_yf66-%j)K7ak(IiyKe-J_ZHHTL=dXTVRg=~c& zPqZ((M{->wh>16b4jp_2HF+M0H^NU;f@rX?(}X5IJ_nkub66mb%&EJ4mGqJ8@pp9w zSr8|`F8v6K9{}4t{>{*FXtyo_wz_nppY0n~Gvjj5Ho7j|{|g7b(3=N>KhA=10q$;` zccnx;6&=S?Ii{e+)5EOC_X{*2?Zo$;X&efPL${#{dw@V+39QPbCeQgVVAh06eZpK0 z_Pk&1GEUWfYnSl-oj`=l_f#SOfPgUMD)kG;9W$*jPo*2kRL(n8=KMlq|@V-SE*dZ=rAv z1(Js}Alw+l0OjOa(M61~hQ?w} zmeXG}zXPq|)o0y9-0@TSf@;mX*aywM_?|8&tk-b7)_(bV5zyT^MlwkRfFzJ5VkJ8p zJVG)zkgwfyM+T3K|1{3OAXN<%=%A7Sdo}jjAFH%~kss{6Sb8NQ;iZb z4&u~lKHq3O41m!NZ&4u@?vL3oXIK<=dw$%Nj!~JgHjM=EEOq23eshkf^djh_K zIO1%3GsEh)KiSi4D12i?@*jXa?F3MW3u-yCCF%_@rPW?dyJQt{^w^7m)IC9I25VE^ zr5u$ujo&y;I619$yzu`pYp+$H8k#7(y=7+tIw@rh&*wz)O}t{H*uO<{4J69H}w?$P9YlqAr>dlfMHqO zh1@v>a5QZp@^`!XGQe3x8xge`zeN4ZmTtO4%s}tcjQGj$LG()!Pg>m7NdU1JZlVm_ zcn`qlWX&%6_8OolnmE<6hq(NdxU*EN8zEFRKk;iqc!fiSj;}Lh!6?ODCyUYCdS8U0xFb5Ps9> zJ8(k73ruscrXE27I)AOnlQKe;5dSg2Ki3ZH(jF(?uAXJeGL!Z`yEItuX!s{xUy||krcwIE~ z^)o-XQhVXKPz_xQJo1xB*FcOomO#IR&r;7NJ7=$fEgle>pkEbN8 zjW?HEa}Q5ps^GP$Hn@Cl)x5EYYZX>(sz_)Fu;gC!C{r?PNa0cm5%187Egu&joF!bg6O2Sjqr68eQ?~4y^_;ve1@Y>C7wfj>V8wWd0zlij&jfk`i zwcSs$WMsQzehR0)<*=-qN10#kn>t(uGq?gUj?Y+Bkq*I%tf0Mk^swhYY>U%^zA~_HNa{aD{hnB_{ zZgpNL5lgzk=UFm?vukwgy?`4y7rKGyvw~IP25=}|BFiY=TVm?0-ThmO29ox7FVwjw z3^L4A0=KtVHiHD)lv#K_WJT`p!p*Q`${~Pr;ep7l zxxQ_O9k{oic>+hvto`-QV~f|jjU}KIj-f@jm3PF@abA1*Ty7A8ah$W}BB^)eBc29b zjB2ycqaV?3A-MMrRYc9ps-NKV1Vgg%@M9TD-31$NLH|#CKaL%j=F&lzrOl-W)l>MG zaI!2%_6VGHONNGsNlFUo-*EEDMNwxcTbJCy5v$rI6IUsz$OS&B45ZdXcZDu7h`PIk z&O{Tpnx&rzwQNMaSc!vRrcU?>#o-_Lh$wwz<`K)uPk=TCwRRk`Z1k=8(}9c&2`9u!SMj=fE}85&M@>!mv>P(%{ilz5w9gqLN$J&UxPK zbWoui8?zWrLio@}ubOK(Jnerr{o0qR9=R8KBMU?$RF_M{zH(6$m_8DwwruyTzK#>h ztp<2B4-zRELow;QD&gPxU}CpTs6j{cB;8WSNtsP*12aP_uN4^yQ(VfrP1e1unwMh8 z(uIErBoY{pu;R`rCq5D*3c-aPg((ejQ>axy`e)wa)l>AsOnl~h(B01`d5$Cu(Q>tU zJ(KizAMIrD8qWwlvosV^xOk~)y*-q2o0$(I7>7aG7?SK#q(fs4he{`@oIOfBo)$Nw zX)9KN^z+Iq1#`SUlH#=62xRwWa9A~ZXs&=e^l)1>_WL?>5T+AsdRxsRzqhvLI;^$3 zbb=zL?&N{ZrIwYTnY{###p=FTiS1;`{fm^%BN)7wN}|sHD5otKUps&1qkEJP zi`yxP#PMWDFA?f4p7OQO$VGFMyM0ySFsOPNBSEqPUD;5DnTmf$9J8Ifz%3dF(fosG zsU?l`W_4y!884%&K$(=0KO(u8%95-X>6zIlQB9$=W5&e0!hIH$Chnf`MRJ| z(^SmtYO?_bA?g4Rhbg%#f=?H0?CBLHpGzgEmn6~3X>96mJ4Iom^`g9RtL5G43&kyC zOo59U@l;iPj^K-TGP3SJwH+-a;~j~Gt&}-sgT-Ypy4M8idVNR27>*3GAkAi%;;%j0 zAE`H17sGdkqpm0X)En9PX<0 zet~1GxcR0Lc#6nAuqUG4+!`lQ@}UmLYGsCQ*i@h9m)S{TpQ7n9 z2w+z@pt;y~+mtYAjpU9iKlNjLc}QUF=5~3FN%_6R@l(b9TQToqersvsu}a1VY58{a zh!65Nh0d%|4LwsOxbwRE6txIQAFa1o-y=#7-rH49$sQ>&;r%UVl<$42a^%url0Kbi zQ%Bj!uZD4INk8=Q+s6;RMm^8pKCW-0fJG;@Xf9Qf>>(r?!EJG)!p)@}D$wP+*0v^F zOTCF?r{vvi<-4#{NiP*JPUfNP&dN94lB!rA=3p1~!+b*6Vf5H-7=)D&iIFsL1v7-| zLUH{DzIDhR|EYF7Nxn?KF4rrG!Iwm!3yM2V+EX!8e?ZKiBjD{BEQ1*Bs%>7AH!S6n zU0G~&0P!q~c2mG%GhXg1Ucwq8po5RSaEpFJH>~pu6VXb!{PL97wl039Rjw=Xq#u(7 z6kYxd1a=x_Qe3;DiA(>{w(v;^br?>hGpX6s_Zb86Z%M>h<|3ZEpxnnA8!)eayo7aG z8a(q^b#111sk(mQ0&F#w7N;_E=bb?G_ZsF~lMw$R!irqB5HQ7jJ26#qvb~hP1B;hO z=8h6ZQkNMPWxpPzK+)Ophj(LQvdg3=hcY*TrcNLgF00A+u=4`Ry{ad;uL}Rq$-M2j*Lq_`g>8k{E?dfUCSMx((UKY~^ zye3Tsk(CB!LeLRLPCppzzXDP;T%~|Moy_`2mSYL7IJB@@2c`NuM9(Q=eiMzd1^r2! zuxD^y+HEB-qhM96KqO?jGJW|YuUIv>Af+XDr$d+O4{JgDyRRPu={+Bc11lK7-y-w= zl+s;G5^Cu%B9c$aiiY-2Al?-zqF+dVdh;+uXtP$bzX*2z8e4>DCMa=MI=3L>uprsfU8{?F&@_Av&aGYX05LhoE=|5 z>yreH<}Oa}4epj!%X_8)rc8G)H9>*=xWW8r9|-XWP;0e|Gu|Xg#uz$)UGB0A(;rM0 zH?0Y=@2uERs`${)y*eD!Gnp^<0w;QzWhbeG3G|tylMwQ@3ArkWcqw16IPJrwLy-NH z^|T{x6Ho7Y8&W$$ma%6zA0&q7<@Q5PN4{oOxckzGCJ6JNG7|=1(N0<7InbBb<8IV} z`U1%y+@LyV=L9-HBC8nEFT%|7yVT%XV!}W`-{$)d^?)h!@QSj2=6?;k{$xi%2qk(h zzR|%VYH=Lv0KNqpA*Uzf-ZRm(c}fpCwd%cHC5jbTh2i zDIZL1^qGE~g`*p4TCgw2Lww?*59D0x<_CQje~afmQQ6Rxmw~O1 zVQRMf(Z}0uv#B$8l1+)kG-JqG<=)}~-pWi0TMKhY#7(~O0Ytj3d%jtLb35bX1&vFl zf+r`B5JQ_e<1Uz0nhwB~l!P>w2nIU;j-;OAHdueQVh=&j?GfGywx1UfIN4*6Q*#{8 z3G(+n#~-2~Nt#70!~Jhb8D_CVZIfw2E@i%?CYSC+A)fBp+Qp&=aCa}9R-#U2Y?P0t z&AEJyPwXo!pCQbr#Zn*tNKJ#V%1E0dLjkLsD?cc#o=gC#(Rt-`jkHeh)2@hGf>8a5 z3?J8ta!J#2!jFNtOBYSQVZUXkX1wk*9efgrJM4W1UX#I6!nau7MbPmpBAK?{ohX?C z5_YsHE?v29hD{4$GkkOG<}bmfAyj~Y2q9wv`NV?$s#8Un!G|J}@hG$=`D{cED_^4k zcyTUp_`_ysM?x5i`vOCniBOmcT-bxF?-=i$l(Bsn&H58u&Wo5iZy{wfqx(82wie}D zE)Y&tn3+|zrAQ7{l6+@rr;jj zX%R;+RRpxYEQ`PMc!`eY;~^yD&PS1md5+Y2yl`TPeZ+GOpPW7YGJ47RO^d^6cENPn zmJy6{)e_4#{8rPUj+KeJuZSSe<*m>)%rdW4V;ZRd@#@NIjh}z$Ru&W_y|Va;L`Jc4 zUT2HU&(Yx7jd&UmVa3^~1umG*RH0(9Tec04N>q32r1@enhnYkuJp;r zyu(>@GQ_LbRltYkl@!J2+1z&{P+2GAO$n0H5`6ZJo?kmLbzSo}%C{ViA*{kAQFN^| zQ_bP`jTQ*#%#$&X2OA)8>=Q)qoriUcgz*F(q4`?)x$?eD(0-*{ntZa0kHe8XuUTfP zXkxun55~0H#91#qdp_SzAyZ)iS#+AHnPXuwI3qWz=B+<49(nI$8JrU=Mw|JaG)Nu{ zsWngG%I*5HtOZa~{JFm;wgM9tiq9}2hp(g#bBzAl6(6U4$p?<&eltFefxiMOiM}ny z+?u+PdX%*OJm6i_WUTBKn=}Pr4~Y&G%^D9EUIo1F#A;Q+CzAzY^a034U;FcPX8b{e*Pje4mFeFLZ4I|s zn3Ej8v92|2Uu9dMDB>daO&ilFQ|MNGB4}P*4>!k-z^7K&p?GH#mZAIRHMAT`raFUq ze3NME3-5`n{cgeDuy~1KL^POL&%S3u$T96tkO#U37#h4s*9luK*dD=oT|h*V>IZJJ zEyt5l>8*?gqb<2_5rCv^;I-SpX^f3vJA%UWd!m zv&)rm^CachgH_J1kij|Bl;TWCpM?4R|nN*IXi7U*ZR$|mse+p{_CN``Z`txj?tx@)Q8G9ezqDeKxr#Z82gOCz1HbkRTKK=}6Nb7EnzwB(qT3{fdNu6g2s=)!<)%Gn`)46=|~oP@5Y zg&NuX1qsop43MsJXh662Q-PP^63DA5*~pLd!W_mJA9W_F1@0{4Q(I zjV_HK)7xB8NmhB5=uh6|sH=E4vn<*8YH&gwMlyNC@K{wmU1#QFgp@v4*@|~RQuDdG zuaEbJCvEv?XY1|0eL9XJPg4ny-@1NDZwat6{iUKdA4jqBf=^g0OwQAAXK6G`tku*n z*ox-}ORTneMIHcWRE+dlnyKAx-V~>GKV#YQabp#(Ps3M*qef04ByJ}#Y`x9gt~qbCq$WeQM)*?hZVmnWMxQ;`?0Utmia;4^?_*%ryV~(a`8j3`mY2d` zxDkK5lfk`Zf`3c1I~9Xw?#wz+B}IUNrw(CTEyf@fUdz<7^NenRPCOzvLyWKxTIZ>v z)CZ!HSKyejNDw6ZT?IiQbdNijl3TxK1}BOU7*J+JnP0cE7r4EGwNhaFYi752HYvg; z?wgiugk1b{+^bHRZmf?lsT(~eV;#7f8)@H|UCWuH6)xK)A8iDwMiB-2lNf^RVLJ%Pi!v}b%LFk$o^{f|| z_c<|3LxIW@ed-GD?;9e2a{~)EB=Kr-;;O{1O59uBNnIZA&YGcG7PzQ^m&bamj(KLAMqG1#JuoBrji^kI|ttK(R+E_%(-DJ zY5sCSOmd7SR2bEEb=k#!HWk^XmW1q zH!q&(QSYd@&@KaiaYW99yODQwGEKM0a1@LrRur5%E_xdcWJsouk>7gT5OhvnZAZ_u zVj@xne}Hm+6mUyt=viZzp%Pt*on?FG*C!@BJ_F9!S-B^3vGbsB|G4}MI4469u&z|V zvg!6HgO;|x!*}p=E#P%3_R4dyUuR;(c*jv#a9U75m8TB}V7aFck?5p$$ zFFtI`$-rO!{mf!E%T5`$$U{ zxin1XUrm4O-Nhl8l(W+i_{lbJl*+_U^PKu-q}QZRdyMcGTro(JJIivCqF%4ruzd~$ zt@|I=?W1wcv&r0^5uN8P<-Cz{qy~01P$S6ZlI_UwvCP^bN<3!j7AS+RIFA#P!r{F+s02?G!M`H0wae$ z7SiReQaJX9C5&Wa+{;}LfK$uqQ*PrU#CbaGbsbt@Fi>?E-FXGL$V0P%#Sw=Y~q%d@a_`NA7}l z3~_VWVeSth7Uj#CK@MHh*?9RguJ!>ooP&@|U|+XX7CFYaA1mj@u8_$*uyVN5cypO!4gfC9_zxqVZvWxAkmx>vVZ)iBC@m(V#cLe4>pNghX~UKp z59j_hSFck<>CSqn9u$CjhJ{6fX8MJS!s-m9ll@={#bedrHG@&XF^8G(0=tx`@U+{C z{iNg_*I_D;Kd5ndVnrEi6J}TsKfmN4gLnF>x-;PDr?rqMTAQw-@eYc^xUU0)L67AM z{}ov&{R(Xc1`osf?B=_aE9Wr2+W%R~5fQX? z?VHP<;mW^Upab!dR4@B7y}R~7fkP#RFZ+|iQhP&glf8uEaEAL`fq06w+Lhxww>qfN zByoTWkO`r2^V~g#Rk)Js{hEpV*m5i=fe4vH3|@>(e|?87=z=}_owxKm34-SK6|={H@s@q23evX@0|-=8Y{m60A%M4VqfWAsY%vA(bD z>zfwZoBAZQT?K>lifsfP%F<_@ua~(MFbK4^re4vhwA1y%z%H$iKa-M@sb=NSr@*MH zxzRTZ97{h$;`Wcteo|#YS9nHA3Z={#8EWm%muJM3eP}0m_uS8DzW*Vb*Jr0zTs!!9wYuZC~xU9GVrM|9gdpMYVnXnl|9(;2ZnjbQIW zip=S->NANeXL1SH^sQB-cL8Osi$A*{#6Dlg-iu9cWcm1_MR+%v|HU>(+@DgLWLJ(b z)A!)!r3;q&)wT7WBWzW6K)TmREOB5|L+iWDi-L9@mx4B44=Ng?&f%`r+KnuVPYu~p zb{;3GxMjbwdiYhi6bupd>Yu7DEO2^Qklm)~LFj+c-ZCV>eb5HjBd6ULUW8|~$!Z{J z)o2V6xD;k&j-BC}b^)a3)rq-+s-aeDFhkwFS# zMY;7hwTX}T_hhn!6V0h)2R|{@UT%h*m$QpqbHtZxS&QTXZE^xtWBcaLPU!2kZ+vL+ z-R$~&`uOWIL8U}wo~FgWv$*qWk4R~OO_pu7vT>E7S);^I+XTQu>dsu-hkHC)O|O+} z7lBNuwaJV>tDmJ3OzGG^r?DkL&bPittS_uv?KvFs+}Kc69z40*q1A`>g2+` zrB88(EWF3QmM;6%PFjb^XRuIb>tt2Y^z+c{vVr`*K}Kw2{^hf{im!wE9sAuE3O#!Y z_Vx2*f^nUNT{kbf530US_83&H>Q$8HG%_CGaijNwA_Abvl->-A)u`yCjCRv&XhvLgqKaPrst@ZAq+72ZL zmaPTw2kI0uH5u^vZI+2tZ|OF*ri$jTo`{~)kdleaej%0KrnDuI%~ZQbzQ=OF_-pqr zL4rMg&K`DcbZvlq!Ly73D58bBX1chF-HRJ`*y93VHb_k*2Y$qJ!-?~H-}B;Q_awD{ zNe6GiY@eYmsroj9iD3NMXAXfTJ?aXFp&H}AA_rnM_MVI$(I+0g^JYV_67*21I_Yc{ z7ti{PrZQuy8fo+mdcL1P#wbqA5fzhV`R`@z^vW}OV3I_0!bvJ0>2x~_-yK`7y)Pho zT%)fd92rGPDIVF7SuEA%v1Rh~Y$0Ywb9=VpgYk<3pU{f^!tR8t1qovvQXB*w<8_y= zji**z3Uu{yy&n)Onn=6;fZZ1V;Sl4xB-=a|+P!|nXXbvFDcy0M9)P|1G;{s|z9IX( zd03+b635-RpG75nINWq(1;0%$Itrc~E47fXPP>a@Qd*E_CkX^8VJ${&)%?BvyiAG%;0%l$ zv~6sDfX#6=oW;H(b|{{sB60o~SLedBZcnzow3n=YuXtfGG`_@U)9pb$KJb6;(Zb)S zlchl6uAaz$R*QvHfVQV@@p`5ot7AXAtpBgME2@-N7=j-$zi9~v;M~u{eH@hdJO5fcd2_{?ji0w3FzN!pLT1jOIYdS`2s0m+YtM%6B~*&b9n}j zL`O)KUbz>6c}Lm=s8#T$vPr{#Oj3eviPZ#21=S6i{J}-b{7W3OGfhOgYz| z7mH*R2trO)9WX(LE;EqZBFG!G4y4K`G_BSE(p3WeIYxdCvD_{>2kxhe?Lwhwl(&iO z7)Fp`k~<75v1x4k*{_4V4!rTeUx!sNi0TDI0zLeOK%yLO;>O4uhR$RfY(JH7vkMjC z4@0Oc6m54aCS|^5qBhAQZm;3Pwd<(DsP1@Q<$}r31o2$kAWtIiq1peX0}rg}9~B^Wqy^#r5Xx^tYl?`$|CRJqj|>nkaB9qn@n z$^t5>DMLsk0xhr9flI1@eUg#qsMQb*%d`iP`%@sWszBlK=sT$Wsz5zd`@PE-s?OT) zNVeHAz5(91l%iexh_IuPNeFuz1N+y}FoS`%Z6Z*Pi0-5uq(3)6>*@e}V-*Yn9ZBRB zwY6)9k!K4Z^a7Wpph=CNppDP0^@DoI6R4?_g(cO&dzScX4`4uNO+15LEQ~knz4_sZ z?WbH09^`mOTu+Ew$^c0AiHrO{iBrmU%0F+<*UZF}0|EGH^Rbo|TDBz) zWUW;y?OQP)e`6j;*!-DLTCBQHl|w8;Ikcj$Osj`^duTXEidOF+)hG5Sg~Kp80k188 zW~0oc&v%AL>W4VkGjKcH6f)P*W~&SYvF0``msj6moa^LzN zbe*0)kZLuTE6Uzfv|)cIC|u(RG2)5Ee4Dka2o9kDH-~FWIP|Zz`KJqVQ9Pn)C|dlI zxrJyH&+Q_qCxaZlr5?Ih=0Pg!#XZ;e9gajljrK*}z@L#~0oRW-nH2gB5d`a$#Acjh<>0j#bp@4`-?VW&xNuL-Kt!Kx414FAqNM(LZHe zf?ia3&XOKOP&`Y z8&?i0*X%|TLR64xWc~;kQ5!G!;dxdb(<9LUB_!bDDbVb${0rVrBPA`4dTSR%m*JKw zni%xyd(KpN5z$uGU9h9e>7yTRi2sAN@uZ_$?%7#*n=VSym$xTQYiU z0$NwJwKwm+pu*5R6P>BCkt;FxN>msgQVh_YvzM;IM>5t!%)%EJ7CK%U)eSw&;X*4- zGLGJ#%bgBX;^rNVfF*Ihn8haPz0~+Y(3qjK@CA&t{BgFp2iVq7OkG*MkMVEH?$A{jvEVjj zub$_&DLBpA#`<+<ro1k`CYWSf6g$TyuWd*^Jp7IK-Oey2Mis4~ z=^_^^t+q)PGG04Cgo$;)35N-NWt8YCrP5#w|5uB##yK#)txn!q0(Zx}TM9qLrsla9 zq2mr^_vXgjscFl65F_xQGngwT>&Jj7V)cz2(=!yQCC%Z|QL0sf%CK2o_G6(=QjFn=v0-FIJ61+ zmAT%tRw7QBWg|!v#S~|Nx$0{|MVI+t1ta$vs9s*ZFZCyI)Cn@u)~)wwswk{kbzfFa zB=_#UryFfNU?bNs#Qr8htITb{t0#;+Iy_1hKgL^;MV+GO)a}e?B%74E0npLPBQU-l z#rrJXzSi7c#XxHNLK2T_^#<>3)g0Zmtfg6a_!H7A6!jkP>t2nFvz08CJ3IHJVq3xX z!Ek5q@@56GqC=l9gL8qZd;57olQ6yPfFjWD(EUcM2ej7frhh2QWtOz2tho|Li@DS|fm^f8jnHoe7AaZ}G-i%B|* zyCIUj^bqH(E{3r*mNVA)jkWR= z2^to7*o-5VNqoj$uRS;GOc{H5`V6h-%PrANF|RVsASJ1V5j+07A#}!l_nulZ=)Qk3 zr0&4E=xPz$ZjThV!Q=Q#AT%#RPk3DQI{hPIWuRW;k1eAH#Jxf30e(ffQ-l;L($u3HjoK5rJsMSO2>Dh+q(=gw1&PWZ7a_+(YG zQp^tAYW`53Poc2jphcx@q+`o^gc<*^65zh7cinyXRN;I^cBOO$ZULpnpm`oLiO!(G z=tYk2i^e>+F8x6xMFK0%^h~pICH?EFULn1VV<}fMvR^3ndNI4wY46HQlRPb^ywTrk zTW~k`w#}W_H>{sh%tz(8Ja4wvL|ePqYKQiEj4b$duSME%DP*jAPTW_)#O97G5)gvorG9{5EJ547mF5OUWy-raAB4 z_VpOX{kgBbhGs_oeOb9n>?4_`(cvFBeSX%rkIX=D-tTUPlwj`E44~&9`8Q?|6;idN zp4tS1ob)MH3A=bLqqrXPMWtGo%*Jq`0-bgao~waUi#*)5b*YWR6yCi?rUU9d_N>vO zy=@O+@NAD!dhf*NOl{7uAd48=Qg?Tl_-3T2eM>%y`m@t8oxgb6@z>21M-=>}iK1Ln zAJX%?J9=hs-L!a-qZYIhqh;8)?%sV})#sjZ$0gAwQGVJF1@R|p_a5W}byC3a;#DG%cw<=Vi#P|K(3?y;%S2l{B93BmVoeO%uCdatg&4Ty-`D~BDa zqf=}90uP9iCoFOfeI@JiUiZFc*U*}hrqbM$c!9nn^y%`DqSuXHTC%n+Zu1{`8~Ow< z3Ji_J*D0?edbn|S{Z$XncEv<_`?P#x{HR@yuN(WV>Ra10uM#(ZolX7z?mJUzExp(n zYV+y?f__Ey-%at4p9FQY3vr>m0>Aul+3`9NU?$c4KU1&-Q zK}0#qdyn+@H=rN`y&3KG8}F2)m?LZGX@lbkNU-pJo@9y|bOyC+1GFG!FKdezUJ7a$ z#Qpx0wFel7Sb-Hd$d41&!Dtd_`LRlM^R=~td5mgDApSk!o0;`CHaMvqz}fjeKsxdF z2SU?AliJtZWqxfB@5Y<2J-Bxa;gfYpVVJ!nc)eW!6l(P4Z#38H67xOyPR%>YQ#L1?5Qbl(&atG{nuynt)7 z86Bbysg9Q_G5=sDQusi?s=t6Av0^!A!7LoYWM=QWKzEuNMt>v1f23rSxV;Hk{d+f{ z!OYc6t<(tB%9*I$V2< zad>A*2p4|PX`@iMZwPjGcpG+ay&q2EZh9t^r(1OtWDJrn{Z-~$M~DK~8b(-FgYs?+ zc3AQmx^9M8~QS$Aq zD(~<3t*FP7hVr1NTAH|3q?mmjXYvBBIB?$@b-u@)I*gmB8`NmyCM`PO>0cqV{NWvi-Dgk7IIFh5Q-A8iV{P1I4F9`90ybw}$Vf3?-hH(s zWce*h1RoWM-$QlSf428BrJQPZtKdJ@KkgH_GnxfBD1Kk>zZev9&dS2=M4Dd$_1=Hp zwD1TN8=B;ti>gxC|NI`fp?G*USxyIpF8{B83~t7QOK^I9mtzf1V5zixi$s@&A3C|Ld!R4wI!Dg%bHbHb5fD-txrNPLuT_cXJ*Z z4!{%%zS=hrj}N(fP0iUi{!gZfOKQn7{pdF`K6>CfVR2~HD|pwpM&@m)?Gr?2pon6{v}z~|fd z3b8M=rBX_D>S%RH`2kBEqHr#u7>jn+-RN8T6G*Ru&b?;}8G1%zJg;UGYLuY6c*TQp3GmxR zuy%|eIXJNE6+L+m8j673&nJZnSZbpnIeer-N)#hA{oaRi@Xo-QCvFdx?pbb-u9S^= zLHzXyK#q<8*BM(fs{j4i0?5Y}i6=ll_Ct(*GdEUMkqlCAApDmg{lcpy7Nmx<3q>3p zFk7Yy^a|GjSt|s@WgD*$Z=Sb=<^7euJyozx#a3JAVg(=i4S^?yUK%-CPDZ4S`(&T%9(x` z&w-3DW^jFTTF_-0jbLmH?mg1+VG2^a(d#!0&05+YJQ`y z&;wK)3mgzL{8wY&kmE+BUq9D3GQMHhll9eBv*#gtC5e>})K4kvM$a z@y!vQHk?r*+_YCopH`$7pcqQWvVOq0j&}_@#-x94=|p2FV&<2D9)l&v0ZZ<2!mKc^ z!)ald#%h=@(}IW1kl7D;8rEVAS6+q!$!KZnhLA+sDB0~_*nVgx= z2NL>9xEp+Lj1MG3lxL0Y<@o48al}QF+Lg%^l^_>~OzOas_kB}zdZV2=Lj0u(cz(9! zD#%AW3H}@_3OC>#b?I0lDyFjY&8^RGWPNQmBkOUrrtqGFx!?0VT&0TIde@4LpRRr2 zhAuxJ5NIa`;)ZK+vIHkVWR@U$rUGz9xvKg5a?flePDda0gcRO>2{Cg9 z^0F+NnGa$z<^^dzG6xsQSg|z{&4&pU6CXML+3Mo1!`m>YszP|0SCppb6pGq7c;isC zju&ADfc;2GhC!?h%LVg(HL+z&co9#647-a1bBwAU>so;7*5^5`>on{FZhLmDo(&2N z_m+ghs!v^GZ(rgQ#k7f%3>zW z{n*%8zxQVk1Oii?ee@nZvhs_N-Bc;X4aEhsoK`k6`9aRt!5%Y$Uv1m}wy-|~Tghw8 zQD`SC{T_-EEc>?{Ja_cHLVKTD0z~{R_pSG1MAIYZf()b`)0x$bd~8>Eiou|Tw2&-; z{i#bU9#3<|MkUFM1i7?1VnIazR+k#JAOyL_f5mrjxJ$sEG4B&khwK?uuxFHUzgs`S zn|_)JUkRKiLlrA`1I!j@&JlAX=-ZOP*RbPp12Pu zBSkZ)z3ri~EC!*Gt5feF7~em^t>cjQ#7${dLj2DO6+8gILxm=o za?Oe1g=x4~_vi?|y<<#~QC)4A3WY`wE+Om02ZYP1{37Sh9xG>MS) z-<|FjVnlhqZ(Ue~&g+sE^s>?-3%GRSo+K6;FrlI=1FW(kNzXzo=&6$&7f+ z>}fok_pvJ30`@~S^W44FJ*r=8^x1X2qVu7C<{a(+@B^NH1gNd=xzjdS_xqABS}cL; zts0?p>OQoUunEGv~I19ROzeJ`{$MBv7iLGOaVR8_PhV7jiH`hdHv6R4)-2Zph5iS z2LDB;{w*1TXW_~v?3Ymeqf|kig;IryNKxSLL;v*%rdRN6vyw|!)6=os*|LzDAHtal z7V1TA{Mm3i8jPL=>Kbps@ts@3x^;IbH=5&TnQq0Nv5@y>JRz*9-ueL%Cs`JAVOo zsw!A2<3O$o`2tk{Rgz~FOjt<;EuL>tr36Chz$_Fmdp|2%yM@q90$|b%Y$rNR9>tMe2zbAb_+TRazgmjP}(-HuXGlZ+xIygBX7>j~K zs|aBtoeE>1P5ub8&K==~5tJx~%{jHT3=m-hL>2f0F4Y(W<}DRgPnH48H-zk7C^r1` zb4n@rSr`hyOi!%9_SO$$O%~FcfQ9vJ#ymL!$v$SGI7DSEF~Y4#{t=Y}%x=)kd|mpt z5prXP@9ORdImj$CC{0CGo~J(R`mQt$+Vm1&-NuL|HZ@YA>!d5q;F6^MdeLcISjb$4 zMFb|>eW{v%l@LOu1XAZJ*rlgZ1RWQQ+mFg!K3}1_@Cx3yh1aX$M}}_qVT{}d?%IN_ zD~X7fISwyw6{-aW2@NTBK=`Al(r72SFG z?jy$(`045Lg<1rymL5pBVp5H?&(btDdYWf~kc+*%#eEuqUV z%Rjh9X51Fzw__2sT~KjV)HWj)TjKL-e*^_3+O{V>*x)*?`PXp8t)% z!&^Otpd>+R@wiMR`j|slc#i$r>xTI5fZk8s{|d8A`6sq=VAk69?o4dmmar2r#r;NR zi52Q_sGco=!k=@U{1LaoO_27NLU1y_?Ie)&eMs1`*j#$(ID;KX6{zaK(Kx>BQ#pzI zS5C6Q1`9E8J246r6=Z-O+fxiA~&<5^%VsXHe^&KUi!s6@9?o zy^b)CUc)rpMJaZDSDx3>sX{1{mcMeJUx6}nHYh<}VVmD_E!tx!VO`%Bpc;A1!2(*K z#~evnOlx_7Z%A?D|D-v*vpd8EYLTVK-^->*3Q8Bfv?frg0VWhm;R54^)}`(!@0#3 zmU=D;%iW!wnz(8zxMnjrD2#VaPN4APpLucf?iLz@Uk}0fK^=$xO#j@_)Wi z*stMJ!hU@VurGvS+q+qRBKQe`1CaYe5Gh&XR;jxG2VZ6)F@`evv!kyaG_do45Gzha zc`?ON1q0gI_Au4P!0`OP4*qFu(k(nvM*?vy3Z(l{B7qNxBB?vQ7KsK6|4S$s;{OkM zZyru{+rE#oTFOvarYOmrA#>(&8AIkNGldG7hmy!L7nw3-iX;*WWlS;^8jPicS{0>~ zS)s76+tB+wZ@+zfzklpM_CEG;^g28**7~f^=f1D|y3X^w&a1D-$Bpxks%(yKVLMtX zMF>no-8&;TxnKh2Ki6~UP^W5F<)w8P7eaKmKJg%09{Kwl?F`lmeLHfsJ5&3XhZuR< zv^EcGTe7@gTYI`@&l=IAa(h81H?*rWs!y4Oz=Z)VRXG4C_xGM%m#L%0C=Va)DmETP z%ptvsXO1#fWlO}vkVaa!<7?_i zcq=x0_$)Yd9UZU`xqcq?Ape8YLILp>NhGW)T~YvN=EF*K1v6A?S|JDmc&L8xKV=4k zD%7>`iwwcP=w)`(!s*K1j*YsOlRy%o3)QV3=`M}uvZ7=Q7WS}a0R)ypCU{7GjQTVN z%Z`S*y{%8r+0N8%R8Wypj$s$uz-0``(djeDnwMihzDgy12pfuaE3uMKLfg=OsK(15 zVJcZNKBvItUuW9J@$jKs0I5}#$lzIB} z>9;qCa>G4DGQTr)RpEI25q~7F0`j36I(Wl#AFtdz2FQ4J0C?{~ymjdZ^!_A%on zTmR#PP<)V3kyYj68;&5zep~YCKc15)`svNNC4YJW(wB6SBtP#I_?mk2=;? zS~!3Xp-PGk?N^!X>Vp+}4R9u%Tq>Sg}!I zJ_Z%mrw?dlT*$Uy?9O@z#gNm3y_D-=l8ptGrzf3+h2b6#cspX7h2XP&%L&2tp9eGA z0;`I|nFG21GdDxG5hOPD2;A{j*lia$k8b(v2UtUh1vsk|**_a}Z7HHhgfMz(+RdH5 z_Sn{+-6jR}|H-#8I7>iP46F@~ZA_qdOcAzgXxERRWI&Vl{6R%Fm?7cH5pekI40I8` zh9q)Nhd&yH@lnX~9&g1Zo{if>w1Y>r;>O&$9R#+V2JG^tHwwrd?d+jCDqt0ct<&*9 z5k#b=sCu0R81oo3cI6L)!%5Y#N0fmO>>hm6VnqfUu3$Q+n=Zzkr0-O;ee;ZgqAoZT z6?Kqk6A?uh%xtX6N}dJXny8*=Pe&ls*O$d7Ad#%h30~^BWmdXppp9DCl=$7j2H3x0l_GG-1=_-RCCz?gJ()+)ReKbnGH!RDOBVQPg7)&uk5%9|5kh?*41R2K zRmT(1W%NU<*nRNox4sHntJa&Iq-TjcqmC{xOY`WT6Tbvdsdfmfu*jp)Mb%6fot?n$ zJOA5@8G6Bg)q?XW2dCUl*)cXKFfy`zd$kJ4^xi>pQ@fUPX#&y@19*BS6AV$UJS#zU zI}rd~)*Oq-v@sBLl;0uuz6Zcbvm0pEJ`Z2V9DqkuWCIST?sIKQE7R1BQ{78ksCf_a zdG$D-7lwxz;QIFexoMBlP0LE&f`Gh7lx6!h%f-Lp*?cgR=UNM3Qux3YYI&mMfrM)` zdy(@HnRLVQ534{xv%iDEXNn{`oj-Rc7Tq0bzHR6{zsLaRxxfxN$c#cJ#(JUY@^mBgi&;Q} znj2`dFVS~{FJ84qJ2Xjs2Vr0tjIFzWzu_dkrWPpu{P~mAjVVR_l$)CdWH|3AL~t>^C!}zi5styLr$5}X6YG7z6IcUF(83L4iZUL<5aiq%y6Byrt>1OZ zl&eQg8yT>VEJtJ2#`JDtIAJMnw{W?~Cer~cH5`<`Y{N1lO%g7VID>5n0lCMD)Q zTemaWgAtPFX@U%Rlg7L-IUODVz0%)ML#ETyREB{gLorlBHH~u(2qzfcaU7Z%Cptyy zN+&>C;SMZ}gO`Vs+(yV=(1ef>zbDKXGMiA~!;QM1w{`r|#gBWb-RXFrZI|cvwqdM| zPEWe&feA@IXgz!G>-0I9-L`f{>PaM#qIf#}mY8w`&iLv^7vRuQ?YWLOfZ@q~*zRr+ z4utNfy$bZ0bEpXFS^UFaa8r#1O7xs|&2zgDr=i5YPMZ?uuos2)C%g@8NeXu)E&FgHTk(q z`HSl#xfTx8ZE@J=^c4h`vGHImar0neTLJbv8SXw{q5LW~?f$U;V0!RoYHMH)3xKlt zfwUFqnyUc;{ynUiWuu9OwM2UQ=M{vHphEdw>zQ^Zk{-6FJoz9Vvdheg{=D~TT$`m%BecU3wWbgNo@@hW>%a~f48~~}EGfd5s6V-F zE3BGmhW8Tia1G9q7PKV4jXRBY`{v=he$bM_sm+)B_=5?RF|oXDpHFp5AkcV07Niez za)f)N`PSc4jEnel^60}HyT}Ye3kqpal!V_bFLDXS`pVf(@+h51s5hA$n$z4Y1%@aW zmIGf;eohJYx4EP@KMDzf7IsLorrK7ln6Sw4XH1Azy2K(U={*H? z(J&Y-3m040R`|h$MVNGk8z?k=k6gmgKzSC}n*Qy8SAk6nF-{a33FM%^yc8-Evm4qB zOU=tUW^5I-{J(MR4p>hLp^8EH($D=wxB|9D*i4wA8k6Ta80Z)%MbKLn#Pv7O z-9D)|16uoP;CNnreNp59RvqOpi;oNfy68d2qv!hzkz{L}S{0O)QTTgv}VCT;%10Ds(d9rQ)zX6u{0UN#UGlC?kVtMtUX0yr72c@cz=M4084Itk4?I3j7S5HD7;+@ly z6vq-248(y)v#g`5gV5@ok&0o{JAxP9-5FRSIrt|y;YW312bTq%+0!tvfm07fK3(&TYr0K~NI=f-*vRy4RT@Y_+? z2oIO1pLW1-UT))V2Y}GZp{e1}Mc09f8wxOK;0JKZ_uW$XO)J(2KRiczmHZc&v%y7@ z_Z@Y&pgx|n88hWJn++l)gnD2>+nE)}P%40yXh~C%`wEtxq{!xe6_txTtdD-^EIlm- zM$g{yfJbUIU;YJsKtn%Z4hjHEtp(ZsZ;ILrW=2=DMpM6*$LtMR@9Tj&$p?}#8JDLz?C7Can1gs#@NykDpuX8^ls)9D6xz)~QyP3XJK zrwH4tW3UmmcsUZDPtKg$ zcVClxy>HM00Jqs868iZ6imt$sqPxU=Li=x6O&)UY4xcRs^8XIV@CqXu5YyeW_&}fX zU$E1R0Mu$-JMh#0is9%@6PSUduJtOQ|L;G+wcmx+DlbLy=x=b#9kHd4(~bV?T_7-a zL13KufBXYII)nj~tQyk9Ut_V&0sJmoD#>1z!;! z%BWKyKhv5mJNOF&lb|bK!sv`&WQH4vhBGX8%`FFFEWYqK(xj1HnIZLqA>r$7|w!b5wmxUe!wMF+x%Nlsj^}f7Ob)|7;+HL)(&?VH5~|E z!$X7v3hb!=*tMvPPnk*H2iRtzzpU=xHF)^%s;$f6Yf`17xSP00ZKFfjV`MJuiWBHHW}%q2E!?5aEK_7+(a}EAwhjP>J#weD0l@n)SoA16-rb{6Zu+Fx!P1sIXU#nk zsw7--7k8{JNJYp?5|+0Dv=ro^4dLPp>FUZ2#`Cuq-&BnLoc(m~Nk3~hOsmrxS8(ji zX@qM&1><5Q0TO`_+I0#nBA`dH;CHq`h_;YjL;V-CFiU4Us^`vE{Td1y(J9!w{LuPM z!2)OW$*i`q?DoLrF@PJBcgX`rN?} z3Z$({Y0Ht%vP2Il6ZJt3eBP4l?BnK}7fjZ>G5(cI5fD*Pv;T(5-+L2D$v7CQ{)Rt@ zumQ2RwTf_m#Q-jZwQ5GV{q&vxT6_7q;#nbYy6_j8B z&pds8zbAK1fo0S^P)WQ2NfSdr7j&(<0F4k|-#L1@ zZC~KCm*dci`P`{IHOAZFDHF(;B!CJooX;0QhYx^Bkb|9LZVQe&(SVH^kAeS+6l+t1 z9EYLV*wkguR^%5p0D34CTTwc8WE5rA2q^~TUq7^x{(=lqS2;1M7brtVB%VUev@mGa z<_DzM8NoA&Cc=|EX`_KoUq!9>sb_`^9o~kyfC0b$W>d zQ0i;XL}|$5p{#1|>%VsE3&7?Mh>iY*+U5k!%Nq4$M(KY+y6Fbx@wJNFijA9QU;;dh zA+;^qH%DPd(iZgZ46vaF@|33xW}(~%VtY+005P9~|KsCJK7(E0<%`-cha13d@Wl`4 zUgcP%H%A<+=}m(sz^I)7=(A2-2HYewQxbrtL37AO*V<%3QXrTF#;={_cROXDqGAz+ zVK*3KP7zNRpem}5oWP`>7eexG27uW(w221+_M`3%O!q9p?IAP#K16=ydYtSN(gBG6(4z4^xq_2O#*Z*TINNfVS;V|Pv~ols=}r|ihztV4z~O4N*4q6; ztN{cJk|2ySKB3hV1|?gwZYCOPZk>wI2a^|Q<1q};Sqnn;4Wu!jQnf@JXrn1$4^b%GP-(EY1HKP->|+c14D z>X{}}O+GYu@6&}O4A1If*ZC6I-zd$s1RU!dINFw@F_Z4+wN{b5P)KvGoipAXXt?3% zOw>CKUG$#|*q!oXlR62@>#G1H>%o}i@QUEbGfGgLN0PXP(rd_hr8-mAL!DG$$fAu- zV1NYPASXB}ZRMCuI78oEs_#C_y65G!dkBmL>oU~Zbr+eJf&KjWTPfeM?)D5ep^vHO z#h*0YKbD_M2;H{FG_QvV&7ZbOP|q%U_d*!wXWfi@=1pkM34Ddk>id{CHuk zUpxEam&DgcvqrcR0J2Gt1sZ9gnM{JV^{*!yez}9eR^O0Sr{b?us2W?VDa%Wo^1b(s ztY@-prjSm|(HSk5B!JQ;6mhO++_W!Mk;{W~!3579RZg+Fs5-FZK=RtmAJYJ%-cQ_a zIc!Z;fOK>{9x;Kpzi>dishq@;vKCbtn&IP)jV^eI)EdSjD#{7)Y0??7TK;<<@L2SY zQRznXd>w<|U-a9Sg*24jNI8&pXnq=XfwDz)D)$8dK<$#T(->yrSlvB-$@%4PKnf3J z7DojdH7nh^{PsN9R;znYKmmo+xpmHp2>nWb8ol?*2H@_94+(CTzn+sopr(z2Ko*La z)P(1cwbHmm&P>@e5~pUqLo2U)a^z3Tc*_=*>wMyuR9%C|{A-ZV?mIl+o`DZn51gUN z_C0LNU23udb$Z=rptGN4po%hz9oz zy*znP$ogKlua|NOkJV`5CCQ^_5t*Lh3N*={+gIvJo1Mq<)KX*4D_~vpi?@IR5(={M z6VX4R;C&AfpKqw@hv!UPGHG)_3TD|2fHP@J%E0+1D z*JZ_7_P@Lb7OjjF6K5v-UT~DgTy0X`^K4TY*7VwPrA|SUYbq< ziD8oo*+JiGM(7ZI-~L zFi@wD6khu79)2R3k}vmQtn+~7s%rHBxGdjxX75ER`C-MX{CN!{zLuxgADR0xriSg5 zZAYFBmLru=T~CjZTIN5Nlog_U=8L zZVRN)`ow0yk3hbQm!!zl@z<5^Nw$O3k+qubRzVDJqljz2z|ZT&${k77lF$# zfR|q_c1|bjE@1d9lx!K-`$#hqUtVPEjV(Z2WzuImxTl{bqt(si!N4-AVP33%W=FE8tS)NFY4 z6LQK-eN(Ny%VN_D01m&w0bq*0_EYld?tYZQU?bGELH$;(J(+=(Sg&&+m3V-!9b&Zb z^rnCvQdfV|Poqm{{n?{QSJ$VA5i& z13+|(s;aIJH+tb-;10$uP98d-tHfRvhgCkNe#S-#EhaUvgZ_*#AHMCV+;g_tbV=+yF3+LWPVg=r6bp zy087D?upFn-~TKBB@YMPf@U;x&aD$i@TXmC$MRyV+0ite`;Wc>mY6ZDVk})ZfjJ1> z9(~%=>3Ya^VewjA`fbLx^>mf350Y&$!6(xwYkUD$o&QNq2-~}>ThLp|eQ4}$q1D~9 zO3tjjEgs!AUiV%PfKxOB9F0Xf0hQ!9;-MHQ&JU;_N!E zz6P$Ze<6q=sGE+}W9O|94L9)Jc#P2HKF^mh3u_QUHF&cAu*!deykt=$L={@3kWn3n zxt^8p&aWy|C=DaY`aPJYX$jOuDR5N%0C+(NvrKG3BLNW@Z4LxaDUiSYh5D?>(U3C$ z#!&8+gK^*BdRz^%@~XFIU0q+^N&&wWAE7SDqun6at`7``48d|P&!gk(ELiH9BjR*s zq{APWJ}Xesl>pa*#0^&E$0C?b^OlhU0uC7ZXr5AMa1uQ4Mw zR2WX>1vi}?huJ;BVdf2p|Bnxzx_c=bFcGu+{%*{W4g}tA;bSU?)N=rU#_v#uR{>4h zg60SQ8hp26cCq&vYISE!KrJsNKtuY&c<*kH82<_K7w4+@5}W*48LUyz~V4Hu;0La}ex} zTmh=g`VQYn!c4#W#84`0Jerj!G-d>fmOdXc?+v~Mu%cjMnQV%mw++0!lmxnq4M7^*-}taaoRGwssJ{mO%00k5ZtxLOw~zBCj#q zTubEGQz~lTNFw#5`DuQUD9k7mdGPM=p6~C$T2q(va58k0eBbA=isnHcF6lz7$lHS4 z7)FwwNOXLm=8Qjm59DF6_*N^ueIJ4W36tXD3mO6u#D)}lRz{b%Qvhez%2EWXB=tD3 zRFTUF79Vv4+I3Afngh6g{Q{uy$D~pgES)4FQzetnJ2(mX)Wj%!t`8d5#lbJ@Y<1vm zdCRVr`*fE=*?Nl3Ev`O4Ydp2zu$Mg9WZZbDoQ`F)No9ROm|dFq3CWR5%K>2F3Q3f4 z=c}krq#Rn4-ljA_%QDA_YkpGTy}G5^{FZ-~*W=xvY!;(CZ_==Q*`M;o?^fQsaT~Wc z!2b+?f)*no^!8+?kQ+(4a0{tYFxT%q>^yF4?Rg{o*ua`kNeEsU3Dn=5;+_v2PKel^|*=u=$$6u37>PCljgFmX0P=WiHD*{OEnhZMX&9Z^=C# ztE#s=?bw^Ig=+TGgKuFzP@#2!yIXO-T+f(Uop$=I_z7}OMnQr5$cp9E%0}g&2~l7B z9a_}+s}wKD2+a&2HGf~la^fmCsbHwA!{hCYFPJ1;lC4}-2jgTuG06d{YEyL9y4pS4 z`5}fDQe<(zi=g+t72ao5?4naZI0 z$|cA1V;wXzI0={i;g|%otJ|_t=?boKshdr0=gISz1NX^b#RO6Ik@tU}eYmdEBL591 zVUg|}zbDD%zd$R%(hv}jECV-kivRp2DUc$x36c4G{_}rre3yXw_q|zv{LNG2QC5@B zUbOWu;Eap$X2LnCK$@vS{C@YKPhA-)=gjJyGJ9cYq;{vyMz*X1*)rTBa-|<8yMeo+ zht9Q&29AHhW`t%WdwBi5ZQyUVoV>9G$1#7zBQ zy0H7m?J8&1%~piw1Av2}&JgI-34#UvC3AWwc7~(-j@`o=Z#uV_r>$Ez9tC+lK2K3O z&y5h38ww8_jTU7{lA^Iln%PNKik)P?$^&EPBcUJE)cm5Gu-6(^R{(HdkB-bXhcACL*6sUR&6A(~{qm8`Ch>1Jh1- zJ&o3XYe5AhLBbh?%kX-vz*tUuee`r8=b{}0$Tsi!f9w!qgE6ca^jS9g%61So`JEq- z1+V!@K%G?q$ZtgMr-3_pyPT01C7{mF>ioaEAS@DOeU1<9fi_yZC90@2z@`&L=ICg1 z%iv*D2%(-=o&pjjs%KY3Ti`%D4xPz&fhHdSD8LO`*UJFVpS?`F#o~!cm6#cu+NX^h zO{om|q5m&ch}`eblft<;g=M0kt(Ux^!ABVj{(8`fz7r`qzC+6-Uqj5PFXTXcbpk9o z9*o*#0-+@w@oH86-z}NNPa4QZae$NoMLQV=pu2*ADHxC{z+!RHNagF%oQfkKd(sJA znr=W^0#KWcd-0mp^b|n32R#O>N+Ik&2iER;Jl#iu^~W2HYORRGmcltY?`T~^R91Zv zVR!#sW@h6P2)YeRt1^31(ggW)+y?S8=vhr7p%^QrONOH3<>Z|5G2E| zr4%{Pz)AWBpmiz1>+L&jM|5)GZ9#&sU(M#$NqK{9eZd$3$eVyMO1%pRy_iVUAvV-U z2rk#;4)dS8ay%fzr*e21WcnF~q2HIAS@@bSy&L@@p&J1HskTF~FgCqmAS+$tYD3(^ zie(D(e^cJv@vlp4Vued=V*TD|!G=?V#jbqH^lP$-JwVSB?%Y}j%MXrYGG^}s30ab+ zikg|=#K`A_PsUREZ4}ho3RZEk_}38O!k6iA%}XAjeDysnX%tNM;(z9J|5V89HybV= z5olgq>kcXZlq3C7Q1fi;r~mvg8gYAvnm2d)e}3?XI{ts-+nA6ayK$2aQ%i3SD43v4 z-3usbtz3KaFTj|O+SebXJl>u{qEX~4ibh=&8xU4Q9~~>-f;J!tfdsXz;5m)RK!uSE z%n9aYmdXA!3>s{L#LZBZo&pfDGF8yZ5*Ye51_O2pgG>t2^M{abLG?yIOv@YuBWmmJ z5m4^>lO~Tnqm26o2y9O#0Hyk{LFiV-wP4;rkPm0OL&!kE`$h2{-I zPXlitRicD8SDw0sw)Vm`T^h;(ev2fxm7hM|g>tjm%W^)S-==p3JZN~AUi-7QYbN;W zuV)88U;Uhw$aw>b`wHm6Ls_6au+vF`R+0grsGf(}+y0u+-!lq$e-d;?YlElH2)M9c z2C-!aj3BuIpSt4A^syq_2XzHen9z3fp3-Z7+=Av2=z& zf(9}o$5PYK?Jy3O0l7-1AYc8bC9v0TX*z?E?cDe)eWbirVC={Q)%8;t)ySx2RKULZ zre%pFuP7v&j~jBK3HRKj?ok(LnJ!-@AqP8yL&^(Qmh=bAbdWl$Ns`gr$DVZx#@3^r zx+%kuEA&ShdT-mvSpXto8F{6RZh8rD49%Yt!#5l|l9FJG5t3XzyQ(l~^#b~GueFu9 z)qz(rnzuRArL?+`ABKtAoYlq!BbQ!LSo2`_?UvVuc}iHA^(lRg(R|x0nvj><&UvF5 znTQS>rgY$^w*QX$PW{f-!CY`o}TN|r`lV^AZ%;Bm3p5*6g6uqY^tx^fK-+@sIZ6n zH=geRaAy8B&^)8bVKZf%e0pg0e6lsN107Bs2EA8aWY2#EgNd zl(l)xA#0;XZ_{D$@7dACd7`k8Gf>H%cUvO_nd_v|HLZ^Dsd0 zG|GauYU8fVa_pT1K;q8e*<38)A)R^D=7_iuEQXFKOBfGtv_8Vv2}mW)0AjGy_=nmunEWo&ZEva4`zz$FyYVz`UB%cZ_?12LCFPuba$vY61?z zgN`2V*`h(`AiLwaY2e?;*bc*2u3yBRN?U}-_);pm)@{G)ae3HJ>;H1GHF`w++GF|* z(-m85gIY3VuHd0V`{mEARSy@Qg1UDer!u;W-$WB~n*Gqu)OS<#xaGp95aV8O?YJq1 zH9UrzTW5WT?4BsxMaJTt1~C5ifMyu3UI}I!M~Z(Uz=t1C z-E}HLnZ>>~Ny60(nIzvh*^`GErw?yeUxvA+PPY4@7e57pB~`NYC&{bO!*0m=2(K0D zCN@ngH<@x~E-?SQ;&-aZS=ao=q#U8I0g8*J$9EWA3dx>5daFF58)b?f(hHay?`0K2~R*08KXtag0Y5M=qP&+|7LKIKy4BV|t|lF5<^q?N328lU_JEGFvT74;|4`T>PL=uj*{mz)i-s zj;u0tJqtHJWr^B!`YZN83%#p3^j$WLsx_^C*oVN-?FGVZ<}Q7zc>4fLzm{z z&^O1ojC;4yoTj6bO+fcvj=NqrXly(nbu(;0|pn*<9u zAskcL?(l4$WNRC2BOrQ`k!nPFsY{h7ap4pQ(7ohLZX>M{k+uX!uv!kqoj<>;sfDrT z$#<--K<6I~E|*&qn{b6IhyrADsx3WlLH()B&XzCG+=C=>Co|0T;GCDsm;~05=H=)% z>V3&DZ}l*wQplGNjrMkJ%>XLlDg5vHlLNi2Ad9H(p9kjW(j098=f;)3LJzq@XhfD> zOiu@f2ij9Zfc*CUnUS?N5po4`YA$G%ht`5P$CG3#&%sgSSH{*aANXXvM-q%~XmlHyU zWq=F{ZdL>WG(VrwbN$fa$+r(4ysGkeXAA3Xzw|2DynoI4$@LCWm-T!F!d8JEVo<$yPiBV7AXFvY-d4Tq77Lo*NM)o zBw;&6XQ;T27>2rupu`BSa)}rJEQa;$7<=VT85SvP00KGTx06s_)%_R}o~M4VYm0)V5HpMaM=Yf-auzp3S+-EnzMaNy3rL} zJ$#>T7ZAzK9eexW{$6QMJUm@aH2RRu<5W;n6~n0$1n(^~(4F=Kx~*PNlI4K2mV-0o z@*`b2c2Gj|8{|F9QL^j98-_0JRXA$`<13;@c5^{1uNY9jlhfd>WDBaFBC8X*WkoQ~ zyMFpHGb|_u5ZuS#JK3YMoCPs;7I>gi+d+^LrOO9Ypy9e`W;UTK$=>NK)ko%y{iS&n zwvA_YLUT)|E+-+G6VI*r#E*|TnH)1!BnmnSG&Tuko-rW|?E@!rN-0UqJ+D;weAYAF z-YCPDOo`*A${jNYG&lGvfNU>j*iJ71?ZACMp%^h8 zraG`AeWaHUHcI)*Oux!z?~ItSrW+k zia9Vn$4W2}k8Y&d@&lV_Mz*zc8Nix;F2zaKz;B|J(xTMIVq#702M8y{%cMog+H?>R zG++9#QVX`~bFOn6qe%((xElRJ$zM@8%gha&r1fMWtOgI)-2Pf3Lh<@A3Hym_z9ACi zD^OA02tN4Rm6VL)+{wo3ZF|$3_c#`pKgY%$(k0B3|8`%XIv0+m<&|NZ;y%V;9^J_I z1=ax@)KxFP%S4;g@tOz<$Y{heu}_iF7B@a2n}K>`!IAXDt<9;dnkV2kMa!OTV}cNy zJ&CQ##%Afmjr#nP#&rybF;D}KHyzva%YJ?$Z(vA*@*Mqn3n+~F*^z#%7Wi9J3k)|| z^S6vA6w>Sw4#vD%It1?zP$ozzE!A^8*-p($sI9xfBbynzrrg?vSsZHf$xiTZ$=R`=^axF^gUJy;#sitB zFd}~0fIgXVcAGI}g*0Qtz_a8FRJpSFN0RjI;Nrvinlj9@7Mxj+6yBODBna|E?{kQ} z^RhK~6-KV|B7?xCaeVkLHP_uT-uXyFjMMMXjr?BUVgw>S=o;sW#~J-H7mP$8zjD4S0spy^bKmcc;zdWyvk*61*a$9&Vakm z7s*s+_ZpB;JHoojC{52Adf1q^#mb(Iyq3x-hHvU+IpAwMT3q`{Kj&H#j7`m> z&6k6&yh^`w+~U)T@#3OfWoBH**3mqnvdCJHTaCg=*hT!)dnFvcq6*{Q|UmnG*DD(YVV8Pv|wJX4r?;Wh!dSeC}cuNI&ZwRZ0FN;vr62{!$^E zqtM=5_Or2mq)fWIA-nRQmwR*mEPMN+L`GA>&^^Ph*^W}7!}{tN|IVnQcHygT7304` z$HS){q(1xAeA^3m6=R%$pTpMd+VMPAQCVg7y#;M_I!RVEM@CE$TSy8S{nj%!FuvCQ zKnI|SqYKj?F%iu`$cYb?O31tAHj5Uq=NM4pthe3J1|SVi5Pq*CA#CSCl))c6pt$$? zpR-s^1J2@hX;B6g{#r9ha!ew5_WUQY%*q{5^8a?loiAN6sE%LVewZ}o5-9B2AsvyR z)Rmn}&i;8ZdpXX3ck|*)g)r%_0dEy_9dw6PAZhFt)JZbcmwPB2TA4T=e!hciw!yPm znp{%xOcu782BnSc|D85w&E1G}6NDaUlj2J-vUI#Z-S;cI+$4kCuWUQ{+su#PuHANJ zKk)4gx@&P!i6fM67K3PR7g_UgdBoISeym*M)jZ2Z)DLkrJkQaP^M&bU+BT)ckWwsB z7<+zm{X)($#RD-(i6(eIDteyL+YJ|qP%o6{Vziy@lvoPtoLtgdMzu$1035Vt$)64LI@SP0mX+dM=y&~ z<2!X@Tjt0jHhw=#Eo2;fVs7gziq5fZ{h?T+M3VD_{9|`2blE9TR^dM}v|DV4{{E|t zfOBzyo$LkSXzq=Rz|%ur8;vo;T=vdRz}Gy>!1iXrjc`^Z=`CNpaX2j%y3TJn^s&4= zK1aF~x(TT0alkV&5`;H)m zO@aiW1BlkX^2NB5vylE&Qg<_Mt_M$AIPjC+F~cHyNNv7_R?`XYe7^T&g&Qe_Hm1@ zP8uGDWH^EsTJJq)<1cS2iv`Tf@E~z|zc5t`O-yvZYn5?tcj`3lie;8?;i6Vt$PH2W z^B8s5GW?pMmr@vdtI~k~-v2;L&ypu(TTO^Q+k07d9K${+J~YJ%2Wf5alp`L`x~|^L z0JX!3b`ZJb)kR*oe256~MGRzG7T{?08(@rz)04I`@ZU51$tv7AkZXq;giYO}KIHTI zM0Saw$fFp>YTc#?nMPdaor&E~VaHB5?{WRQ3u7%OU0n#D&ib9Sb2jmvw%n^7D^9Zq zDI8N#$w%Q0Q%|WoH6BQJCUM80yibq1VTxOXj6zMB#O?`>{|!gfR`uf@N!+8*c$H|E zd?ESwhYoE0i1nU%`J615iWesP8!N-)0vY?`z;evAxUKkhoi;RXa>0$)b#chLrzd@O z?xtlZlw?L1>G|?{7N@@YmIn?KefO2*{37=h+*8*DX`Xn_mCg?xfK=Xhs7qx`gpioi z2x#9I7MhC zuOBMY_*BQ9H{rxV>nRA4g0|Za(6Izik?kjyId~ZMu{k+0xHU*?4Us9NCBA7YQlP7` zMU@6b*_)=HQ2KS5M?V!Z4JFSH;!2c4#4UPd$H=6nd@4MLi6d{14r{Wl2a=ZUGL7sn z!5;YFUW6PIMsP@CB@SN3G7)!8P!x#{A$hfv2Jfv&tCltnhJ;HrVSV}w3{Mj|qf}i! zKXVD8mL@R%zC&P@mNl($_jKthv*p6E!ft4dx~?yz5cne*R7xh`#yIAa#R{?H8_yJ| z&%Z23|4Lfh#d_TE2#fxm^DIuG8~#c$(lg!xct zB$}X=RC!{h_!U&bFFHQMw!Q%*t{6x|mNJ*9_0=YcIaEW9)O$fD$4~r9|Gut?Poe2- z0{$3<0Yws9gXeGP7UOm9$%}BgPiM0IF661?DkQ1qKsP3X5wtl=IhoeumbJHHLT^=Y zK+FVdsE6=3@E=Kq(PB@Zr60Uu-xRvLcN%{NL);onB)>|eZ0Oi^n>lHCi$D(c8)qY~ z_PxakeqWAz!^^*ZrokU;1B-Fv*rSU}H#(+j-!~mpx5wyF1~Jw|cF0`x^gk22<;3x+ zjxa&zvi9b9%KH&VJHG9|`ejB;qU%JR?{PdP7+d2%wIyjJzAcXBGI$J){nRF4$VNaX zOFMgCy2F&pyxu}en%%D=HyFHMXw;51$u0tTodS?;TNS^Y^RT_ubqj+xI5>8Fj~G+X z!d^Rm*6SP;^$7V!5v5e+$Hedfpd=~n#Ec=5tLYFyGeaMKMJ=36nyGbV$9Mb z_)~It`qj_by|{i(KiCNgvOyiG(w9Rn_wVTBYPGCQN833tUOS=toM&px$EnyW%^H|O z9*ONb%!i!}dvHWXeervH6H{mFIqL{1Y@{b>8n+x!6K(3hgjf)eba8+Uq`w*WmKTrR z#<2@qi;mp6`Ww1$z&x}1=Bx=XFN&DuYp|1Cr^vO%DS>|+f%puX#;~mmw?Z(e(yx2o zyMq!Z_X_lGhuQI#^?ZLGlb=;MQ`!x@IT-uqAUgy>IAUx(E_Q|C~S)qGQON6yOasl!QO^%yl`u{xEU(w*r=Ml7UJsH;4p|7v*Ox8)()HV}$OhnU$ z1>?`JVb*MK(LRwf{B-EwZw!@f8q_=12pS2PeJ*ta8YOSLs`8sK4K=YYhpgZquK)Q} zEECIqd0mYioX_!BVVqBJS?iDkOrKMzXMDK z#I-M2WoGb?W2Lp!efOz-A((1Hzt!*??h+}A*LoqYJ>BijXhn7X^JeCY5T-x26=SeU zd!T2B=t8E+??=}>ny1t)Y+!ILXYJU-QDg35DfAQd$y*K{qhzMYe?qs9IdQ<2zH+jP zv+EQg>+`|GB9Y@@4Fr>taWo&m>+&+V={*6!*P;DF+Ge^jLjkOkI^^KOD7XT&X$wLP zJ^(NDwI4zLd?`DX?Vrm8pG5&80o5m|f(HKV)GALNK zOg4kW31;qjZM6aS_rnmG*+Z-NZo4rlxu?f1veGzNwk+&x`MJB{)zq?CfzIxxpDd38 zuqP>K8}wtnCMj|*@}b2-6eo3gXCp-~-8J~y$>DLV;fW|{=}JN#t>$7e9E4SJm}7VZ zx_%CTU93S{CUoX>pXXWZOvDno7n(zcPttX9(1R}BL zDDb3Ftlbb!`dD9rVtH1#Q9q6wwX&F6vWHKACKN{W?vE=ya6kjYRt6XwA}uX5(fm$H zUh9}h4P9=)$EBY>@;alTqYPLjd!z6^s_m!iN98?Km72DZS2d!jk`p^%z!k$r*p5|P z9HDcD885$qjM?LcH9W*S!9l`P6#dBuhDm-oXLQF_;j~0ZrLG(X5W56$#q2-=xXgIm zZFN=ZgG*S7=TRsAq_ZFJB#A6jFLo6)(bH;fB>-t?4HC+0lsIXDd<@s_2MLQ}Jsx99 zycOL!V@D_{q|+PGK?Cl_8X^yEqU6DXMMUT(D9T#(vy zc%~tctmOML4@jS&HGdfVK-_)E913IM+R;=i$L4Sf?>8nx=?uq9n`!Z@*I0YtD(8fk z(p4~e+#!=n=ctXc5XQZp2&1D2kKxm`haPOmB;O~yb25LH$AIemfi-kx&A_!E2KkI5 z07!4kYC4MrUFmm_hCfw5pJ&(eaO}jK6t@pHa{KPQpY0Hlg2Ys~xYbNBD8u=cMfHP} z-63(`PPc3U1X+VQt`(SuiMI$SnJv?&XVuP8_50MkNgy5uOMs*tw;BTcm7h znQq2)Vq9>nb9t!!Pi*N9av&qj$P4sXnc&v^Xm&fwTP(>lG)R+ z=p@9w0Ad({BWhlT_3M_I=6Stcyj^8B+=(kr_76sUcj`@CKlAF6HhB9v1+U+M^{-mo zWkJ9Emz85ZpGEJR0h0&v=37Qa@gz$O%Q3+k=T&J!R|;FBNP=rlPYhWAp?qVfvt7U_ z1OL5`VX3uTiBN}Vg?Huk#o$F|RW7SQtoG9%hw`GJAWVdH`wyVWuI7F&c?C!YLUfb1 zJHygCR|F2KU^yN3;@2CH;Imck7>d&YBZvNllkJc#OEyX8wP%kf_L zj=iMj5ko)~pyUMl&EjS0C6~W6G9E?<{fE{IiV?_k#)?l0JgyQN+7*{3g%ixu7faKq8)hIxdp^e4_<@O3Hx9BqcIn( zuBiqAXol>q5#=C&c7s7$T~Phivxe%1CD=&afRpi<-&wl1BAt<$= z_5;KwY(_vUTN~p;zAI&hLJgjPBIQSw7t(zSMwA1Xx~O`Ow6lMIQ+ZJM{eg{}Sd&Wf zV@L|PgK9OzsNhn_m12k3ta%)7r*(IYdHg>4DE>%n7!-@zE58>Hrg>%9?x0G59#&mw ziYwT~2EvLNcET)7EurD%Edei{htUs}Xm|y>sm(ngJ2ENI%jQROHf`--f|02Jo)9rF z@T67tCqR^G;7G0m22eK!t_$#*mj#3NmKBMRumQr1s!wp)l zUe{AG@N)sJQx2vx)Ru?j<->D_KxULyV80v>)?Lh2c2gs65nS9 zG%l!gi$fQ1RzUkN%iGp6`XCjk&v`DzdTxqf>kyuLka}7|58K`4es!+5Y{Khnz;U{Z zBokDW6kJIm)10ji<}zPTGuuuN=r$u-^Xd1GY>M>9x*qL~f>l!c_VAVaJ3imtEhe7T z^K}tYhM~_{*F>h6QfbaUE6eB>UpU3PyZ*~FiJQ|RdhrjyR6Ap5o2~_b@t>jh;|CMv z*7))V9afAER&nXAvvAYLy*RBaeqK4Mkk&JW9?}*P5 zG2qLXhOTsOpa%+Hlv*9JPsvIbd{NfUYgJW}MlT=*O`x~TSfr!B52BgLmTu^Tnu^jC z1b83?_|aE^EF}WCYmZU4kN+h5oLp5O+c6whZ~cSaOdj^7m||!1%ZRrf50h=;$A-$H z7TTqTBX?{=1tp;Q$Arq&{BfgE#D`Oup`9PUtsMQX$#xWLe{2C_$k@uWaE9jCjRTCW z3A>Y3u|#oe8Ci4Lxnwx#%qi)46NpB}N`X>$=(<;#w!Q8Q@gilVpW=C+T571rWb+y$ zc6RP$koi_6??|TzezjwZg!o^*c4a#rn|RtYKLb4oJ4E82`UOvDX4!swU};Y|mtXWd zpq(FJb?053U$Hw+R2TK{MJjq0NRkZf_{dj(f%fC89n{0VDw!C1?S@rw`4j`u)X6+J`cWvC;eZyP<*did=$1E`v*bcB$Y3J@II3H z#c7hOy&sA^-Zw`Kzoz%duGO2$Yk$EUyCpUXu`zAi9#dTX)N+BFTb?w$IfB?AR+SuG zKC0F-q9PO7-P(k8)jK~>?UA#}ZhCg)dFGLL)eBxzMCJJ}A9A|uj!X9aDiqDG%xJqL z$ot9ag1E;d*<)toqF%`J5a?T4IH!=zk^pW{>Z?TI5g2Wg1O?&~X!^butU^2;|G^s5 zbg0Z{{8>g~NHGQRuyn#@A?Zun$;7AIY->v~g~pz#Pqf}>M@+6iqahU%YUUd#v$ z^oH2+>CohXEAdRC&4JGWC36JviZ?jRy6ic1NqH6j{-}E&F;J*ocDE$KsYC2wd$gf?YV?OD|S`)q%^x} z%pP-DBl5{gQz{zV)0`AQz>IacoJywY16v);Gs@x>8xk z@J1Pw*$?)@C=iLWI>%qNC62hcU(Nu68fkxs4%naU=kl2Jj?NR_LvAvfe&T8(5U(*b zCU^$UGLl+of*u88ED>0CeSmfHzl&x)S)sbL%5s>f)7Br$05DcGGpYPEzFl~hnkv+w zAlB=^mE8c~pri~^ezxRO(8xc3K{3nz!#1HGpCKbj_&Sb_rH-YGWi;ou|82k5FqE#$ zT~(Q%m-$*k20;_B`!u6V50eK`k!>l_OK=Ap*18?rI<=O^^U|4DLs{AaX02=L6RY4?_gF512UeeF zqx9CpUr^D;q70&AQ#tW?6X zae$(a)#!|QJx*;$cIFsJl=b}lDGD#q(~=(|`H1Z_1T}K*$88PDXaf51SKP#suZ)NL zX^I!Nr(RYp4U!x_3YQ&_aHl zl7=A2O+Fak4se5l<|$oq_!H(Yu?F?GxbE6+8QqaAw-Dw>^*iy(F3#h!r}~=u zz3)+dR}*J8w=(|2vT6RKXtl@TTPd$(eQaw7IOU}sNmBwoA4{~uz_9U;k7eWRABh0I zsVRJ-js$Bze{^Y@-1E(6Tr&1Oc13zBjr@Ydq(MU@%vsgvspjP2D|X`nQwMzkjFN;eYym~ z47smBhWv2qu8&(kZu?07k>cY;DAp5fv7MTCK0SV6+2S((P~@=aXk!dIhAuWbPmX5U z4fl;_eRdl=FvR#Zk0md4*Z~;p^hdVzdJOw>(Mx;rt4VpLdu4%TMX|54# zGv}|2cLS)9To0ge;;A#6+j!#-QPp0CX-eIpcDeC#ASFd>49R1<#|)1dAMbw548;!H zJVL;hf}ZXAPoW+%P9DpuWZRm&!cv9*Q-hbgTEqa?15O}*n{DxAJrTtQ%K&qc}yMfE@qqKtTV1FbW0SPl-rHq@?MiT(%a&?tn7Ch3Dt z3qg0Nx_iNQ{Qkw`GR~SLAE`cW5A}+g{_HQ~^&vbg#hi}d%n$ZTl zep#S7-s(HS^?$MV-r-pH?;kK1viD5($S5LvZ;2E_ifq{-Xz^LWaUw6G> zYms@3bfE%g=E-Lfm(JzvGnGZqUrtv4{*M+R|Jk5hD96IZ^1Ab<QE~r99E;rV>dk-wD_-RqZ_s)FV(1}^%NyTdFci+lXF+3b<|<;}01A>U z4D-iC;YZ5x_Zi8t1LJW@!B|?Sjsw~|>Lc&C#$s?ENL?Y!3hJn$AvEucpxeHNV;mg3D-vZ34-yJ+f z(b3@{=}S-teU$z16T%@LlALn0s9-PfnKsrhgPqMsQS(3_z*}Qf_!;z1PyGO7vq)u z!W-OH(XxZ%RaU+q=h^^+5PO%Wy_@}vsy(Xplgo!?;X%npp|iLqa?_vnygvi_w#&jL`GX;Ij6hehLgm2xi%Edt`T3r$_Cm601dx@<06 zqo4YRT(M<>sr--saIG{JwewM%e9@(v)qDH=+e35RzF`UJTv$ z7**aX(7-GCCa2E!M)P350%?iT#JE=EK{F8J8qBC3F&Nj_v4ar`ecPnkpg;T^<7A&u zc+!`*f4%%l^}V^^pF!X~0tRDcWE^jKq(eO)9>G9PeK|3Uvm9tBo}+{eajtf8cESwL z5iBD26~oSV6%(d>(lv4=Q=~4!&-C%z4yT= zy7(T;uO>!1N=qX}v;zJ^9EFGQsBTLcz{uk`db=i1mGE}cU}z`)$U3-2B!O3yX4C%C zIXP!1i_IHFjuvZ5rL*kjmPR$W)>g6v`d`^K>TG)H8sCEu1 zmkRhgQHm%Ic&Q&g2K`Z)@B$@4=yvFh+^(O(9$!zA!mp1T& zE?2RT-jXkWeJ2R%=*Nn~KI$HCrO8XavK{7&ZR3TW;+HkSk+F)nFD&@(Q z6K7J56t*W8o$1Wh5YFT)xnhr|l}N%*4;7xk>+ca-9CrWE)Ws3zUG(s$`fqw!B^6&+ zk~^!eEZcd<;rOmMb7R7n-on6tiJkN?>Yyz@^dG*&CI?oOs_IpdBhs{SsWs^`6xdUEwFq(YL z!pIOvL9_TO0#EyipWLXj>}W^N@Pq}m%a&5!qwqeT8IUz=J<4;IW01{K3krnR_9=Cq`1E3m7t7FITd zy?(qOE?U_sE}-?~y~!aye^ghd8ZzeI#&?`Yp;ApqU4|GHSf~qT-(RI5A~%aj<)yq% zI@^|O9k3TZ{z{xOQ4H5iyo#u4q|y`5*qFqT=F84U5RnP@(*-XQ+e8`&MRI+ zqmz%O<~6}M*(g8o?c=VbK{hEtaYdLprb3L(VE`Ou$~0Y;nJr z1zZgHRVnW8!jd$`HO>hT_u*P`60?NSBjd)1#c?`6!kWKv8Qv5$8K z{YF$xez9IM(WV{;tiROu?uw9{Wy(?xat`0DNSVE&suGqpaNTUih<00~Ey7JE`swwJ zUGsZKx}R0p(6+y>L2htg{#4t&PEvJ7Tio!d&SC50`v@-k*IOHYcA$}n<+tH7CXTa2 zyE*#2Tb1lYd%a*Ff91m=(Bj?fc|qc*iqobLUhGEyBkCW)ZTe1I>I*bJonDRvOd%X2 zg0U98HJfbLnOEC{{e@a4M=CE5X^8F#VqS;&Q^eF;-hUHY;G+^bM;ad`a3M@jb4%GT zVlMbYXPI%`sMV#$ji9j7ZL{_CDT}@cVb@9rUP;+;^7c+N`0Z5hj*%qveL2i6;ld@qKStU@4k;)szDb$ z$v~Q0n|8${rz{_ow_Xv$_tPZP}p(o#&QOf2E+pZq8W$Ge-*k)0Di_p9%W`VXzhD6IQreGyU z{^PH*F{RCb%9?sQ6susdagBUfpE#aXaAO>()oj?EYl7V7SWC`N%I05<+g#1INj(;` z@nASit5!{+UYV+^z$}oR_XQ0P<$Pqo!ABwG=Ou)9VZHQh=1nK0RzXuq(5W_5P~A(4 zXbs`s);T$qTj@)S<^aQKv+okuXwc?K^l(T2Zz=x%tt{+ z9hJD0N4N2sE}_&StBMGBSRBod@Den7>rA7ZMfXJTjNe$57;`xXol>cqMXPNbhp#IU z5!Qz)ifT#}X;Y(wCW7~Kgd^Vh3lWO$sJ-B^zYqqRb2LSU9B>GxYr?3{i_He#W}bU- z&QIkxmDPjaMO8~NyG1tPVlIsmP3TWS?bqTS!nk?t&>D5eQ8ifj{QcuLvAypuEaPxi zx$tB@_l%e)+tNtFBma_w!M(m4x~U9xG={nRnH#scWW^T4Bg}j{kBB=yc51?{a+BN5 zG8>NvOOOKXM4Fws%k#z8x=1J1rqc~MyA+1COMP!(Qhq{eM!CR6egQk$AfGKT0C#3} zGMbgFOhon?K6UgnKqX47_(0jxIkADq0a}KS#ey64%rDto+L&}ms8wRZ?|=UkEq1|_ z9v`)0tRmUm8Ca-&+cmomTQpOx^&u(QW;AQ7c+z{0Ohv5N_I{ovG!apF)adit>%~?a zYanLeSUs=cB6(GeMV62w9MeiT+T5v9v2ZI|%#{Y)$X)`F4}W}yrYlY*=YS3(?EZ({ z^}vlOn-bxc7!hnlb)njrU8Dx5Ms7f9^8_2`%*u{p+m#6I!dI!*ln~olsUC~*(oKEB ziOW+P4amt;g72T-+$?{y#27;n{U{*;IRRAWAb=Gb=V^)Wp5>}XxR6s#kC_Gnoewn`5ZkU7$_l}vrH|8n^V z)3Mhn*@bR~aW?LJjxq`?V(3H0nW4okBcezkzY#E*w-LP;A|%UM1Py#8D1##K!epi& zWY@jPxfVqs870nP6FK`#!=Eju=`W{p(x@m2>^AzRH4dLNr~OZTf7~((X#aLiXxh`t zkLw99%JqOaioz%iOpNrSH>xf7{IvAl`o;orqdP)&6f$!;uu!#NM1Qn(RW|@QY7YZ+ z{&08xuPe!umovj@4W`)#DrXN@tLml+>{CP1D>bMhVzB%+bn**PM0;<}(HXY|>{kRW1|hFxbgvCEyDE6h3idRb-Di3^G*VoA~6E z6a{0exw!$WPydGhH86y@;FBN#)*v5?8HnixM_lu+7t^mPq982I^-VoFprJ*(ct6Pj zvkh=ZM&p@H=Gq!;94H|d8Rt1V-G}En{zc~vRHafJpZ}Y-V*^-YkWgOI0wst9yuX)m z&&y+u0}fHU_&=$UPZB{sDca9m02#Rb>kq~g7jZfCo(-iJCF%q_&u2{(CSwGKskl_C zSWQZatd_r);&{XoJ57shY0U`0;O7E?Z-*gH4 zGkv{+YfkEJvn<1^T~*Zh>xc|_{Q}bw|`4ioFkYy z!T1Ip;*GM^`y)h~ZbXLjYrNX}Ki4007(mbI1*5rVCnZ48$+b|t`{%Ho&LbDlY^EA- zm;5Iu9E;)ug3S0#%ew|IC_nX3?-H@Xw4SFK6izpq4u8rZooJ@CC;!hyAr&eOdK9_= z0jx`O@`8YPX!36!&ejc?^2MWcngd~4;A;*_$(|o*jKB?|(r9NII27f*pG}~QiFK<7 zYH3&>X=_gSDK|x&+xypcD?=WJE5ouG1Sn*YfXS2=h)N5=Qdh6aQjJe17mL%63*|+$ zbh{Dl=KKjbt9q55RT6|fJ@8=vkw+k88$=rimS?<(OSRlh5G`U8xRQ}$g@%6&!T_b{ z!s)L6>2rk{L|+XZT%7USdAacCTfRM>%HP{Qmb(!|A$8a4q(4w_5qWjHsl9311Kc?| zvYTW&OpL1B8cb~Ji5-Z(A4EE@AVdsCg?geXmj&SB5uiSx=|Q%33Y@M>2xS9e#npD! zzg~iEm(Y#-M!ezzu4MR?fyaC|qcAWxDKu;Ldx3RU($(CAmW_>eP%3lR-XTU##@K31 zG6aPvwUQyp==%Cg@E*GbR+Ub+H*Y5g6SliEN!O=cEo1C|-8L_gPm_<$gEE28=Xb ze4GbC!3%vBNQ@CMh_oSF_T@+LzlHe~Eb}w$5VNdtmnGipJ<=cU&Pe%cyb#SPk8?R8IGo zSq(zW&qZfZ^)qFc@A8ciMzY37GKG)R;ouRJDk0QgP>lVu*U3Hn$b+p=KpJS*$qsu_ zINr0ikR|H6>J2PXRyNm34fa4gf{~J&S7=!`{t8sM**Az*W7IT98GRUgc`MFz9A|!Z zh=96wfSG^Q%S1>z0(-Qp2FfpszR3lW7rSGtt!-SwO0)Qu|Q z7A{m$WNv1&C#}9=hork%eD$nuW^*Ti42aG@sZmCj1JRuNtWa`_pW9LA25NeWVcWKt zI(Id_GOfTh2u^{3Ls(aX%@CE@X2spL_F35+SR)U9ujsz{Q9HE&bMIsrLi7Sd?1dlk z2D1ocT^ERPe?$&oSYs3`zToN|{I>-&sE zxg7)(d&&Wg1fh6nZ^Z#3CUB>QVoyNQ`A$~SlB#t0^EG&jg5=sR$Inyj(*G1O3da0p z4!qpo$3WqG?WPkkBQIcBnkpA#!xzghuzhq(^|Q`*6n@G!wE0+S z8=L@qvhxvP;sfx>d$bkoMNbq5p!ejJRgj@wr>!9j@{I*2O$Me;+*zZcqr|aqh*pN$I8c@hXtn z&JPZs+x9rIvrK>`VNBtf{IS^ok?IPTLIW9#EJG!htZx&Hxkq@g^PO}vwLm|zyrW$p z7F-C^a6IbI&+E-*MG*UpCsL^H*Die~Uxw!#B-MiwW)rjohY08io`oJaP+SWGc;aSw z9HLw^0T*vQ;J~bomSteLXiyhOz6&>k?6v^tp5Kh;Ab8*-Ud(x+6J{VwA*f7@kY1CI z_qH78C~zX#XT zl1B+uH$Ku7tbxXMUs%8tCfgZ#S%I-O8tpi+hhF8xNSTDVqd$h3jR)l|pn)dWR@C;_ zDeO6|MOHxzR0K8W0Y!ZC+Ilq*_TN@ty6d^TRrl`_BkLYwH32WX5d9Sv8P4(hiPwYb zHxb0Ct{c71sx|#60Z}z4Tn}*HYu5A{n7zvSSm^??XY{~ioawi0Q({#%;wVIO$yC-( zI@|ROXlpWu$g4(B$ov14T0#t*=R zcZFN%brw(&_+7vNy&(BQM7SyFpA8}u=SjcP+x>ye;JI|@4)Q8W#8`}KOMEHN6sS43 zp7^6wFl(>6jEh15dtf%R?HRL(*K_iFcqo5sBjOsT3B552ud%4b4n$-(UN5!y0<=rt zR0a_R9jpQ#{l|c&*%Y%}d;yJc#;hbcREDb8LHHA2{Ys4cqNq7#ptBGn2@fk*L6>r< zJ+_lG|6$PsEVgx!3Nz3E7{C8Ky9v?*M!umGQfJYgFK68z0=Va}ABd~tlnLK%2vxr% zbOIVcEuLABg{vUWDqEF^j`1bdCX47aqTfe=ryK<*&rNz>>&kw-9=e;@L84%}_o<+m zxIFlQEr5y-NygwRtJn2|bCYJAl2Xv6$B?qTvg~nS8-r`*n{%CH+>pu@=g&pRlD`i2 zQJJ%mm-AaNK{SperRO#%Lgxfgqpt(E-@UyPKmEGnpPlH`E@3=7_Amw31qvB#zqcI& zf_b-qC|l96O#T?*E7g{!CMfv(ZIw0MDZ`oZp*7Z=(I6Z>OR;yx8?0;%swzP{&Sz3z zIi{&foq0WYw>bRl+ih9MT>&c`rZMhQLjq&H86O(W0H~n}cb4bY;rFG0v&AFk@UWNiwIystk{1Oi z&^g|t@V1Rtol#P8RdggT$sxx+?tv2~$U8s~8!2+8d0Kf@R{ zv<$yC-8^5Fi_-3ih36*dNnQp$1M}K%KbQw}EzU#*IMYuR&9!R;P2WeAoxg`5wV&);EA5w3{!T!ROXh~T&KeF zk7y^U>C2ZR$0%d#qPhP4B|6^2Umh5FrQ~kz>8Oq_MDuijJrF;0?e~k9GLG>XdoD?S-+P!jP>4|~b){qmEI$@RQX2L1Q|PCO<@;=}27%e;aRGSL)dl5w1c3^Ss!X|mM$-e{nQhb};qKGdzh||1 zPo-Sy%D5DV_MbyR+#FCT@omVu<=IU+X^Lb`&wum%P zU?osxz9^q&ybn%hj~wv7?Gz~fy3#^uDan6)=_ zq^S60HYJSYkIJ)*5~B~(3|aHa+W131YN0RDnk=3| zV{DU-Fhl>w*4ODMeHXr|E2!s$<`vwv7`X0XW};UaXmzAp>Uu${iRY_?hzSE-aV1^C zIg`e>R>fW18@at?LwmF41(?{g<9Gr+&be$_FE2F~A@JWH!|1ZPu=HO(#s~^7ASj3zIeP0J5wQ7)aQyn0 z@`wq0U}~0N!d?FEl5?!YqfQ9Bl$>s*oZClJX{2n@-MDvuAWD-$VwYu&g&I4grGrIL z^8sP*KN?+vD*;s&+7^4Yi_Z5lfg(=xX5soMMBbv1OU zn!`GWs3cO$7u7al#GDF!N|aexUnKCKYzSSaz0}8wdK>#UJ2&s~V8=?x*yCTiaCfP0 z2KVC{31|a;t^<3Wpp_}j9}&!gTaQ+=-!aurQX};3-xx%@wQtTd|zU*39fcyfw`&{Gh+4KD@0~Z&~P^*`x$G@IZnMZu)io^98 z0w%pNQmu#uUik$*UpLU;km1uW$lmGr;@=TT&9eX+i>vpgqW-uT?ZPwi^F-d7_}_!4 zk5MoWB3oGp1wao%ZMvC=Al9%YbHQTw6ww>P$N3Iy{YxJXlPg!Uztj~f>Haf5q98;H zuMj<|kKkx}9d-zAL;?bYoQIA=ZgZ-LMIcfsn};0(3*tR+t>)$8B!K0dIX@! z70F6p$qe&Pn@)l0wg6JY-}}8F{k0yC1p9c0vs{sG1rBzOcD zufvkU4ce`E5K~@(F7)MCt2ghjAk^aPG1Ek-&eTjnVB&f2jYZoTg8&~0NDk_dFMqa2 z6_D20HLt6c4SoE|RVS6nCf~Q$vZM_lDWK$74-{ixiJ$F-9Q8Nz3R>yGWbPc?qdM#d znNc&1%(l6Gf?U|Kbx+Xf%LBNuP?97~yh%rH`OhX49+BOTTptYiG6=M;#xRN@MXYII z8{qalBDMqiYFD6LT?9+EL|7b%F5*6>6L>ME@@5k*F~Qt4Oy>Kh=1z~-{a;{#J)Ae@ z!e#5djag)Bce~O&5N2E|aWNpmMK-#?Uu`ra#6)zGo`twPt`L(&bkNO@P*ObBr|*_2 zU4DjJwbZ=}cV&PB9{=0Nl-kacE>~2d>_xN!<^`&8y?sUD*z@x`YDzSbg5qU0+Gk)yzXYQ*hmIjow^yKWmp7~)K_j9E>H!~pVOqBF`#2ZKp&pe* zGXOJxSKN2*W zn8pRt8W#mMl(oVZ?%&li>zJp+_-+fw)+K#6>=J)G-(PjN?J;)*+!#XCxi}{ad+b1z zCGd<7nn03#;BXTBx(i&61Z2Fve^6lBt?QEnRV7~DUsa18NuV)%^ptOW&awyQ7cZtR zNRUQI9h%|;T5XG(q{vLwFjWW>V*LwPVydP<@Mr_%jU{k#&Z+THff}bnp3;S(o*ZdR3P^w(j?e?em(1`F=z`J+2Kv{oBYZ zy3;T*eOZ!+sfWys-*SGKJBYyn(q_U$Cyq@S8+Vl1FwhhKq9mAFewQa?r0iPd2^wNgY@iu?vA_sv!fctZ3}(1?!O|!hQ3qAXL70j)wC)a+ zmiON&F)rMNHV*M-ElY`d?X*)y&gjP$m zSI!u6bmOj2b9FOQl%k!~0z=PaCUQI=_G3_)S;Kj>f33w<1Kd>f1LI}>y8q<>ne`BJ z#4B##Fe=1u-g%+VCl@BWp|ab}lamCamFs2rGm^AjfFb^xz% zQR;HvpbzOaETV95D(HDc74JcD_KczU#@C=9T1AR)EOv%K3!Se?{c?`*m^kVQWboX| zdXYYs#|rk#mvrv8c$ZH|=-gN8C2!{3*t8_-?1lcUah|BF%qC-X?_Z7ez-*P9dym2Y zeMh+x;i>H|@Nf7IYx~X4Smu{G=AW>apH+SiI6cG5JcEPFZZJEHLw0Sm?d)UXz3qT% zJ>Ea=Rqswfb*CQ{2s$gfR1+hrMrhZ*apr^8?(ChYGLgM_(hE3yPXb=y=6P2YqDFDX zvjbUwjQe40oeAmv0^{E>A2gdJwv%^YLMy*eVmQYfNHc%w24OBm@M++L($vb_#+!|2 zuly`DsD!OoX|818fH{pi8`Jsyf`3=(JwH%DWb7OICZH8INMXRbxG$hV-{GN4%l$qk-5rk-2S3M(bgbn6iG%h$q<$)987YTq5 zUG_K*?D*aTwbtM)4mZigv$+=oJqde4Q!B7V7F(X+=6(*S9)?Qdu%NB>{7s72__J}x z9lv{n%Cz=2@d9lKE~=UnqSP2JhF5ME>EtOt^aImb#vUUDqZ11s_G_2h@6>V=eg(F6 z2pepBxS2FE9o{k>(j&9|vuJlTexcE}83lNdcO*dU;jnisX~bGd_G{*A3XZ%UsN0d?ErVQoqUN$@|G9>cG1>U>*mstyk;Te zgT_KVmsbMuAfn}>(j?QuDa6YU51~5B-TtfLEWdB1`#rnQ=F8r)vVn+jkD_)o3DMde zSEFTF42)8{j?MD=g0{(Px!AuEmx@oL=SgD(5%MyCLjkAUjCg0r47#GvJ{xNZdJZLs zgAy+O8-N88fKjzN%EPN%+EYf7wuwkYlwno4K`>f>ug6=aTr%x81B%JT0&CO51S*UU zS^($y;7Kx>bStfWL}1<<k(o3 zQ;xv$w(W5uyssRU<1Jo{s(Gsu)z1&O0~7W4$Pi1zR6pgu%l)?&wEh(B9_KOMGFgwl zp47TGP7KOsX)U0&mZxo}v{9#FCqfFFVief(bkS{in<~>vA9=I3Y24dvd}96i{sX@luP5g(`Hp)}4^*Ygk_-q; z3>-?B(nz{=MJ+V%t6n}c=b59fXXs;`vj16f`tqPq0u#sCb3?Ih8o#0o?t2x@qN@|V zI$v>|6)}mrw{TBRS&QNVX%i22ICNSJM>gu_ftF!k`>)A^!T~8$*<_KT|E#du5dmDl zIBs{M8&DJgYVIl?=-=ie`8M{yX<6VF{PV{eDRk1*SF%Ye@m*o1*EL`m(gL~--6yeY z*Z!zTDEwhKFh7{|$B*UDj}@@O-T7ENNJ)x&UE~?1jh-o872=tG+Ng5_!^#j8x~UDzpRUV zqjBG-hr_8csl%iEKssd0>p}f<#}X!&)yb`HdxdsHAH@bOm9I+vdm%`9UG z7~Mt0^9ZcJ?N>%%AlH>K20H9Qm`dusd0<28i#YM;Xovu^*MxY`0o?F>73=|?uY#N- zPge(0mdcjFmO|iyVd;lmWC*$xXEY01^+b~g-Q0Klpw_d2T&(F=F7>l#5MOIQvRkp< zl(OnyVaIsTL3`^x*cP8h(mRi!7+A6=-8?|*dEgfb)_r~uH9GO=V#MD`ZG@8tuMp7eH96u&H2?QTUr5*ox?R( zVwvJWw0|{Q)*JEhzDLP!_7zC_{a=|bu~Tnx5f-pR5%n{xEx5lUvu=1py*96#$wgf8 z9;GR~XQnhzXJHe<42O;GkW3x|hkk+7Py@2h7l6^82rp}i-U1YANw6jADX3rvshv%t zg*fu3U=5=M&Xs%!1%UOv|IubRVn+Rx5OGL=F;y=p-xfOP&5}T@s|hW|b{CMuGZsNt z`f2w)`6wQ*2(WFC;G-8dm<*RDJ^*EhSiuZ^8Nt%2;Iw{dN3rfR#DWv)H@ab85(I%XHShWI#$TGuJWH>kJo_{+lyT)$ z>BHC)kUaDvXcvP!NIHhBW>04Ev-jx*1qKrprKx6^{KEVs-c+ssxwK!#GX^R|x-zy7 zd7nDG-%Qqg*ImmM1TR!%E~7>ERGmu-YH1EM64gzt{531{;^NF-{ULfTU^C*q-qSmmMN+Glv1Y__60py ze9i)5>;pB?0gf^G!7;HR5WQjIWx$PZS^*lBcfX*lOQsXJQ~U~FVSxRs?`{^qR`$bh zdaw!w22{kHl09fXv7O> zRk6PUb7w9uDH`%8KzcAT>-X2uoN<9|`QiI-oUPYzqDYU5LxZ($f?tYD)2CZTlp$2S zrqDq?o6<*2%?;YYMwe}b`=Ow|6#DW*FziMEZjl8XwE#ZEYyGdcVcr)F%u)o5i>~E1 zK2`^*2oFIdp@MMbv{0F12^(WppAYBAchh7 zrW~xxSa5SrzdWGFidiFfoA&ahZ+L;tuB78Qg%?w7(6`HuiLAc8nYL>K4WrW;9R$=~xPFVr0veaxr&7} zw&aPJG)K@Lc@UH5w-&s@oRg@1k%|hX$h`Tf!p!VQ;2{{TpGSNU(nb#N3md#JfyEm} z`S5&kKfbu-DFZug!g%ob5>|K`K`;EDV;bhE>&Ow~FCB;e4kLPoE>^_Tk6p6U7uv0x?~NNFx!Vu1l#gaYh6o3`oEOj0g9!4D6TyEl`Yd~>3cVWJr< z-Rpr9_~AO|=MO0QG0Lf18wIRQ9F+q3#&!1*+m0AAoI-ru3J`DJ>ZLd$Q1^%jv(;XJ zA-}~;&oz^FK5lQo!_&P}8FR*v8}aAo6#^yQ$Y8|zU^~udSK>zBvlCPh=TI6(7BDS@ z`)tfJQ(vl`WV7r>ygE%P#F{}6IxGLs4qnM-hA*b?%l97c` zn)#G|spV^;%UhT#7$v8({=rRypkb$QHSu}8RU0KU7;3PLcDH!`roM;CMewJrp9PF-F80XJW(4OC|d=c18 zD>($Fy?bX`M$-Wj+Baj_v~}tT!sy}cg9zGC8G(>i5ROlFT4Kda6VBEFv+3)b zWL`#FlG9A1?0lxov2`@-=HizwVt-A2VUhUU2GypmyrNSbuhT+)W3MxpD+lZR3Ho=* zU4cLm{gBB}?ruDKlUzC#^f#tUEj#p16@0+r$@ENq_fS`6oml4ZxF$bx%dzo{{oEFz z%;1!iA@x_2-KXZu$LGkUGwbG+h9u%NIu#p#4r}!MD!cnY!L4zkOCh0(@`q$J*NwwF z3{70IbsN!CSsV$%Z9vR3t=y%*vC@FJufa-Q*@{W#K6Tc@PU!R?Afgo+vAuS7gX3Rx zLsnE~ov-VLN)YV0+zdk1yX=_FrQl#z6U~4j3#6eE|4-@v?;hZZ;{Wm;`p@^U2RX_)a35%b@SOr-UYhI^&j6;<*d!svhbHvC>qiWD!Vm)id{1b+;c{+^6eFvIee zKi@I(_kaJ_?|~aEH@LxS$;5s_oMZpjpNLpEs-aCwhyUYyY}hGC#77}>m&4m|4Ad(H z0RASx5UD#|?4>@SrZ6R57!H7`AmXNkxYr?Mma{_U0ZcMp*M^UFrd$x4Ngz5_$?h*C zLhYaO9wB~pAsIf2gHTNV-h)l+H@ap&7SfA(6#S6m3-sWEj*|J2oVW85T8rh-<6_#p z0P~|P=N%q65ySyEP?EYHRBel(hlVkD!yolmPYh($+`Wd^3EM{j)cmi&acLF8a`>;B zqG$0C3#*#*kjtI|mWl6SjTS&WM`sWfVp-R{zSP;9Q-sv#-vc2Eam}RCYJyCPjqW(a zFj9k!8F|dROmCQGNKGen$un!gVRf0>W*ie{&iV)Bh)`@WfNY`$e?W4cByj0zf)5Iy zyA-cqEPceDAWTuwpr&VNrx~BKh$immH}z#~NP0i0 zepx)f%$>_y-L&lkpVYMSR%GqE@^W&0b$s;Mnhvls3--VfPznl|wV;Z?N6_gKn=T|5 z2#GO40yMqZZu7W|-cDv|pt3#LPa(32Ulx1@Cqpq}aS&>y##~7TSvPQv&FvtZ@7L#7)R#{z<~U41|^InUZ-(!?xz$ zq!@^L@WpeCu&EEpdV0-;D9($J7>K6Mgo+#NM}OW@d*rMcy&h6!z*(#A*&B&=88doG zXJiIKkcG8kD-IWh0%~JR7e;>uUI4FMfu5Xuyfz3tzd@28A-pla3C}PhrMQjI!32RET#*Ls;Nu67 zC-@?$09xBuk9P0>39`uhc3SrrkGUR=El!si8MagiCO?QHAEmB~JyfjQ5_OLLnVljY zt>uSQyTRACa};XYsx%+euxaJvBT$DwjK{0xs)vW=MC9ar zj}EboQMIj*f*jw90iPcu$x< z_iL;-PI&3VcH{SqTuUA7)0Wd1{lRLiV7BZMjR?gP6yQ;p9X&IBeswR=zDW-rSV3|{ zkt9XyVjR{6@rZ$?51~4sI=_g|?D53JYr^K{pw%rs(Yy_-X-nbLYn_ocQd6%o89iZAFr164xX7>s;}u)Z`L)bxC>FYV~4d7 z?D6F?$CLRP5hqJ8I=^EaTlLB7(t7HWw>T?}oZfdhJqwpS^wT(W)U8Kh@5I}U$BiR? zJ-vhR&h{-?zjpFZIOl&ta}@(nBtf_l{bjA}xsFaNqKdV@i0sZS+*vPZES&#Iz(#Q% z`{0Lbjk?}}^IQy?ST6wnPVJjsUHzJgm*5$u^u+qL#OpFjJAh0EB+3J5qA?(yD9F#} zNX>AghuF#*tYa_9^hliYN+O2m9ADg9! zxz?#0fp{H_8~Ch>w0z#1)~!&=UnHOFt|+hgdbeq5LIf$Rp~3DM^s(n{D1lN+?)d6OlCy8QB~hu^LyR9HzlrYWU0~ofBO6 zfDEkT+6*20L7p-=Bh7O+?!Te_`;6@2jP53FIWR1?-t@vHqAe~hMc_RxpdQ!&r|Kj0 z0m{{L4YIf%J27PPcRfEDatqy@j_+Em!&3fSQi9S9|O zx62+6n3+i!8W`Bn<$v-t`Yp_r721ZSF;rt6pvZx@Qm6Ix`|qjOo$bbDY9%`nb$4#h zm5t`>KWaBFv3{1~fV855f^WMB}l*Dt!M=@>b zYx@_!1rUIf*pJG!a4X|GU0?K$Exj7qc`PsFGUwbUd9gUwZt7>Tkn+sk;fzwIv+~2v z*E!G0$?bd93XFbFbx0(o$Y@@bB;fK-%H==*M5Ms-w&`%gA^1P<-QQo4!P_T^fmpr&`9pKu7&eW@$;8P&d8PmL zpMYTaG%_Ku-ufRu6o@7iBA}|+DG>i(ulWochAjezBpeCu{cpdC%@hux{iR$9>HkW; zmLw1&nEU@v=l|vId=N7IlqY_bX8iu!A5;cEzJ2?q`f1E`f?F$jyxE2o8<&G$v6*=& z>_YCi#Y$ad9g0yfKP7~HI$=jjz<&GnsU3X%ebM6UTPzW!M&o#@|9)5CCUIPZ`yX;H zF8_Z;g$MkFedhn`KXW+@1)hZUz!EltG5_mVBlkgfvUIZkZa2GP*9Ec=U~nMSv~=|z zJgJZ?yi(WpfvB4gXEQmomP|!EvQDT5Wup-U&-qn5n{O+;A0X6m0+>RG6Gp7Prf+^j z!{n|p5`ZM`G(PjF$vOgFrzGP5zqg+So|GO^yL3COKNw3!U-*mI^|XLM^qQtx2@h%l zmrU-r4#J9ml{LqJJnxDGu#~Bc^9l=6ZPLwml3IL4oNMD@lHUuIg+- zxR}up=)8>N{H)eaKXgP4jFEsRx=W@`T2=PDu`)lOfab9q=Gua9zrTvggy^PMkWEqJ z0Ub(0FqyPQ?YJee`>EOSw_8&9?g%_N0o_XBt5vJ9Ky_^y=B5)n0hwRkfKQmD^Q0B{Udm_Znf^1zu?bdKKY}}i@F~(=buqa|BUAi*|C)86do1li)HNGe3|~G3{fV%@&!ac$ z_>{ln_K{SLNSyosN%DGuvq}P=$?)eJ}MhM7g*r(0 zmR9sN4~v@Z4j?2_;c#Mbp{k2TWbxfbEAb*|SQ25@(hUK+3rOTDaD7bSPTHQ4YdZR+ zg&-T+w}GW)gUCo$q&d@!YduKMiHq9oLoVIN1jN$`*m#byWhQSv77SrO0b$#f>r#s} zx^YN$>z^Ig@uT3nn|)dZ=7p(|vz>BOr}! zv3hv`n*ec|_#L+g?6sRn!{QCmc>Nb+0aBUX{CR)uGwh?kfP$RSh(xIU(Q*MuMRVL{ z+B{hM1&|Fr0|Rat{#UG679$|t*_m;#DQE2m$BTPFEJ=NE@++m8<`^$8m_~XdPvpAC zj0!%#$6Me6^bm@gozd*39-huutqfh$G8i!Fr!;e4p>qyWa^gXzea| z3!aMFP3b2f<3oV7UJ`jb`JEq7NIiuSX6lxJacxbG=2YGD;R!xdiL91iKp=Y$$_K8~ zy)K158|mIqQf%CW3p1nmwux?U!~EWsGam1pr>+|>xxh2$MY?HtD+bLutsgL0=;<#m zsmB7#UeEl%T^yAU`br_%6BTqVI}e}HGQXtUr+m-5;@~! z^&sugj+-(|3CZ8wvWxWmaG>_ATF0X^QfwwaSozKlAcN0)P5}i>n_vOtP&AmukOhqG zB5zBM9!#&k9P>yUn_5GV(2TQ}%^(x(-sPq%Sb?~N_1i5Iyr0YSzLvjKZ+_R^F5va8 zCo-)pb*&oj@3Y-AxwE?*#JNa4+K`!X(SrD(C*MS46hn4`fD$rA0el1Z}}xnckh zfyANBy)%7=#?CgetJk6wj#qEoGwZFG0#ha1rh6w{Ck#>L+H*Er>4GAc?y`0?t*$+u zfI4~tm3Exr@WEx+CFR@mAoRSua*5>y<0dGolLZ7N=^7Oe%&c3!=E!j9=wbV+a=Q;> zM?<`z4V~1Hc4A z4t)NQ#wnelDv|ZY;;Tnkg#6`a)iMoUKOC;FM$;cR>t12@Oj_Y6Y{0Cthcj`EV%k88 zdei=g_iA*!_tsLHN<(UCO21Y^-(|<9hBJ}A!(BnpOTf|Z_zflGew`LIzt-%m7UFLf zDIw3ZN3Z=;iqfS$<5AKR4g&0xI!B70&(EAt2Z3&Yq?R16apWIJYAlsT>IbDW*ExE) zD%m=(1cb}C=AG7l^?OuZSuE&~?mX=1p-pjUC3%2<3QlS1n&tK$!yU#yGPd7kkEJn7 z8|B-4uC!3$pFj*8Rl-Pwi7W2lAO2poI-~s)oVBaN-op#&bP=e$_6aM-4@!Tn-Rhr@ z;EWa8e)a1;Q$b~{t>5eCj*~SdQpY)`M?6oj;k$W!p%I4~4x^1JBOma3HzoFa%Jb7o zokUuF$Mw5X@;u|G;c;@YuWsM&`o5ZQBo&Bu}&=Q{Sipro5Qk5Wt&zy3e% zeR(w1ZQC}YNSO*HnL@}=h75&mo+3latYkdG6awSU)jp678M$8pxD&6~&q*RAFGd+6m} zx7J59ki*v9-;D#KL7XySDj*$M#(7pv11+p&K+AQp`&N=!dPw?Io~zq?J#SHE8yh^y z%bv9m^Q{m0lV^OkJVb6dgQ6yhEMukg)qp)V*|WPStNXP%j0qg4e+MPZ*^oC=Bsqlo z#646JwatR7`q^eM{yTG=>342;{mbCF>0`H7q`dP&1>?K`yGpt;o+T)FRfQGT(Z@q5 zWBmP5AvLMzq3x;;i|K_hI5}tO zTIil2ZY}!#D>kwtq!hYIcgX}#yW+K%d+5D66~(>ze}W$KjN^rl`GTFKYu?@09eSbn zG~9x3Cl1m{SDxe=k1azuhUdq}k`arkAiGLX(OZeFkUF6tew@5xes#5?XNSE*yy}^I zzbzVfCxU9IJ(3HxdvJciJ~!Gm_sOj+GJ1b}vRa=2p}IRmoZ|JCN>PX5NdosVZ~DS& zzB>RVT4R)LZ1D$>iY2Ru;uTOL3}1@|<}deLIh9QH3!mqQPRzuL)w!Vo{^8v8oeHZg zY~k&5HWwRYNXYB-;6W%OP*3Oay){*+F+FxFE@`z}D!R*OK){{<n(ezi=S#J>q{s*=g@U)5#%ugqqZ$Wc zGOJ2S!pqi#_A9khA2RyGq>hZdp?3)_?0peFfZunkC`pF8!eK~CgnZOBw3=)DOJeVr zDGW+yQpBG2sC)qt|6o32lvw0vxq_0`_I9Ox1^!3kHO{B@G`OMmwn%N%^lC+lNWR)j zD*@LpWY)ib8LqkPr1DRiS6zL0LR0g`*xbbA$KUom1bf2ly%WtEtaB?+>CrlFK$+(y z<=!g4gSP|$fs=O-V(GsL&*HRzlzG;pA>@J8?XYV!WU7|;Yz;g+IdBS)3 zLf*m{ljDs3@)p?e{eFqr_N#%gTDdAg2aXV#OUUKE-ZIT zQom@(y0wIvFJ{yN0~j}LysGb@wUhr)^G=@?V}mHTCun4_#-iFYfZ_Q!Dx>p#>nDbz zhPLH;Q#>e*o^5g|`88aN>g??L5?w2^IdcLMR8ziLIHq^W?q~|tD#MZK*iu07MD4;* z8IZHm1MrP~zG*cFTt0b>Hl*@=iUcjYph)@ABa-u#%Tvq~yZ3=P4babfiQDm-Bz4CkX?CHXv+L?n zMByJfb4qjU`p8ggxer_rtU5E)T)zA7QwOUDONgLR-*&dwmymZ{%sdbvPTQt-4mMb!u`euvUsT?GYCjdOVsW_y76=x z^*tnkK8+k6#(x2F{?l&{RfIsbHG!2EI zF%QzP80eiOiL!dHM3&cNNy2dDvHX%&I)P~r8vTT9_H3^=gp4KdJ^lcCSM<}$5RG*f51($C+_#XX!h=oM3SgqS6e zS2^77kkIq{djW~lxcw6R#rTchsFn@7e17v|&C#{~Ar%{V;5IV%5Ql)I90P9Pk|OEy z)KeEi0X1v}3;Qvcj1k$~tiJJ6$d^~@`p1en@tTcLe3~H=OdqX#3qnHNflz1(nW9=Q zdL7M#LF|Y=>^3BO^9lY(sWSe3E?5H5zaZ2Z@gslrNNv zHu8KoGb2r)7*z@uC6^$_r=K{P;$hKWb4W{qT}*Yv>FLvsYX>cFFxi>X9{*K*HhJX5 z9<~ZJiNShrkj%D4%^cJ$XD3!Gp!Vy@H+KPId-h(;N=ED1rook0vA}@G9q_|TCJUs6 zkKm=_`3)`h>aUz(8gd;EE}g6mqWW$y0Hxg3r6V8%cWZ39l~Y)#^-j<&>Nq;JT42Ym z49GoV1EiPM@t)v8;k|F;#UbbLxXnRNVn_y+p&g4=lb3LVoaw&0>dC7C4iz+crIY>a zJ!P(wx+Qj*C1DT^I}SX?yP-kIlF!QhrXf@1I(RH==jHSdF|f@jdmRb-szwmbobugX zPDKDBw>KC<=uQ+Qw-^c$m`1=qI8jVJU)1Xf$`Q!(h<=~_?H=(;4KAt9m=Hqt%Y?~N z(Mjke@wSn3)QAKL*Wq^my_;(|@ox!^vqe3)vZ9cPek}0+nxM(VMqtlNOZ)JQ0Bo9B zHA}GMceA$oy4+J&krgOdksDr9302m3WE?#3vI4A7FhfDZkEux>>!Gfmt2160SF1j; z7&q2)(R%18shagWy=h@e=ut=(kXA4RIuLs3?Y@*M8HJ_6sTOB!C>)Yml-%7rJjL;n;57(TB@^TX zfTofwvicqcZ^3fKoy+7NVdQ(Lv~ux*@Y@; zMaxzS|Kr`;v62q%WpEi9rS|=G-}C!63>>B!Sm^_M5-J%6i$jo zO5%lx^vXe@ie#^;obe9EC?87i0sBeVea0X<0~)tW z0P!S}nY%qlV2q)V8($uGS({~(*O`5QfBO58xvUyGD82aRerWVw7SSSiSgLK;<2H(? zw0eWfcyv%Dp2~THx8CBg>?d@!x#oGUII z#h9@5zT#{5@g(_#@gtA~5$ccP8VJLSqkdZWB)k1XuC#`-K52>T2Jl0Nn5^E>m0)@}VWFIZaJa+N-2x%c_1J&xE6_467KQ%Efm%m9kcQ;r;CS;t! zS~Fr$=9$uSOA;xA4j`RW?!~-0?Uh_@%6*$iv2Dod<3Sf#K{S{U2Ys~4N&Ya%gf1Y% zT{UKLZ-pm?loqTDH-G`#zaGYTn*<&_RX+$jButWu`kce^T}LQ*_AkcyDQ7+Ag@ag4 zr}#d`i9|NEmEfaroD^AbGrKghTYpC)98tS))Np_OBkUZ(M^;!7JvAJ#XJX>0RHTEk zPHz-+)#WO=GBb0sxoTbQH9lmZK5U-N+8^3{2XxYoK@cxA5un*m!uY=E|APtx;a(BL3crwPh_| zli18RXWsJuc0a(>ah?A0Bz6|N&>48{_cJ<2cW}knq<7R(HydrAGzx3D_`GXzoO2vy zuTvS=NeR1vv9%HYf5YbM6#e8P5=1$>|BmMq`40l1|KGg^`>`&AJ=!GHMLPV7;B9Xz zZSkH2-Yb+&=A<+{mm%#kkSlD1wCl4<$eU05B?>h4ZdG5aN@>fyW`7jg=>SfZyGjJ^ zrdR*`+bfibp4C)S`g`94!-g6k|BqjD<497XHZ7?IpSGbPufD#%+f6Je^fJR>5SFR7USwJ3z;VeV*!A%*+IcJl}nOco6a97F6O|NNg>z_)%jfb zQjF>X@&^ijKIfQcBVu%E9Mx zV)R-136Yo)EQM_E@g8_uXI87h@2Pg^E)O3*#4p}pDXKpHLj(ngG$DOsV~%I1iNSzi zI0I;5Vc0LX6CdrK`fSX*oz1|Em;etAb&bIElsaRHet+G=C!uZN>|+v&VIwJ5?z){! znc*}@isHocFic^!WN*3DQD-7zYLapbzFtXc!)Q`}sog*Ce&1ghI!L1`iW60wQSQClQJ< zAO&`ltSv|vb0sQ@8g~1~q#t;p1JmjrY?l*+eOq=*Vp^&4VTmbEtT<|)NJj4We@%({ z6Jh6<8ZSw@nuq^+v3hW^&t~a*b!i=)?faa1LOdvK-f8nWl5COv#06g>e;{95JzC~1 zd-!2G&2FE~jdaGl)VoGYo{H`JeLv~?nN1l>O$PEV+TwunY(9cn4vfu+*8i8LraQ^*s_Hf1v9Zxu$)V@}WGO zk<9R)kK0P_@%q}_k^WFFPg z7Whyn{oA9i=j8s7naWd1;n_BT`~Nd7L@)|;5zUZKWL8up`pO)6nVFbdz%-2sN-RN- z$;18aj`3^jhkmLBK%)^*6nrQF21+36{t{U2H-m6D zPr6;x=ps{CuR|rdr4nk4HC~SS5ooV@ zV`F2@1W%xb=~e5<4#$10)Jh23ZE!YB%S zNgT4rg7_Lp91+EP&L;o9mel6>R9+E-J>?&8o}bTv=F2w|(LIib95X{tVh4l8#o1do zc_ZFaM#)BbFWl>0xYs3d zXT(_ei?VTUvnr~q#)?sOSZA+XRPT6j(IN~-LPVtCJ3QjGpxm$@Zq%D|aq@+nxdtM; zL_g2}$k}jOJ^BY-O}Cpya1Cb7ZdqXMeL@MS-7 zmn+lvHiK%UjoI_jhpTifq2ik8U^c4Tz9B+org~%QMAH8Nlu=um!J5?Am@?~2LJZ0U# zdEY@K|Me#~b!9*WiaZ-L3z0oo6Y^Ndx&BA=JbV&UWBQEs%EU=b<?Lj$3Z@gYDy4Tr7a2P^@6V2pr~4V~1uO|mnAS@@l z^Ws57FA!T@QDL;g^2f2ZP2TRf`9oE3m^4V-Q7dB!QkCU~NR!bM5cBT!MZwS}N&%@03lxXp?u#s@sbHIfZ zi#vE+T*;_peqFVN7Bw&$710nJfOxV8?=9nO>2q^)g&W#nNb<6N4G8fjaNDql^lU*> zrzE0a2}N7-Hf)^ur*K5e}guz2{wrRyD?qnf?v6poGjj2|YL6tso{hfZpV?y%fjkSk$mgs9#!R;` zR9AjOEG<SVg4L16d!Nqr3O6T<~P%?y>A14mq`W4Vymu>4m))7O7l(MU0%^ipgp0S(0Af8 zhiQo7{ELv1esQ=6+ue(ZVuNz#!DXt8$Ahs8Zf#O9=5L79hoB$fYpGAUO) z8w@CMo#Hd4%m6lSXYsY!@Z>DXm39TIsjeI(DMU7_IAe2HhM<8mFl)@_>QFIq-GsQ; z($5AMc_t~&ZYa$;OPD_1^jW@o6)gh_1?$>IOuK>vFa`EY>D^z2*(_98toEbN*1s>;$$h&Sr zjGD>DZG)hrCGRdwg^Q4yh`Go|3@*@q7|z`NUx@Kl-YBvG4J0kE?$xv4*}X1B!NtU_vfs z;p^Rm@i2Udrf-j585e@X#Ic_KAtjCD!vSY2S4gRYwTyK-6ZYe;3s?9w00$Rmd=+JI zZxW}DT|-b|IS`6}0?d5z+{6kF;d6)spg<*AS9wOcm${?CVFf3O!uuC1=3Q8R0Y0l) z;$Y5d3L3|1ov6rt#G|-9bZQ(zaHfnUMeKi|Q)`s^(XKC_?2;Q1+OM|plRtzBM?QmQ z?`Np^@aduwO^vIY^%e68XRf#8-C-0o;eiBU+&KySbl&Gnb^F^;`q6>|c#3MS#klYF*;9sOos9I2iJXi3>{1k83Np#{nzlF4 zer*+X-^Ke`2 z7V(4$ORtYiqk)^CwM!f6#`3iOQ=fsn7}?3nIzMySY_qwa7YiyUznp~#Z*tI zJwi`|F2wHWeqfu3FJ5L@&lryN(75^eWYO?>s_4qI+4byn3BWz^yx$T)MmYvgZEUKt zy}~8=#{noBUj=4xnO(8@O-OIMK8!gsvJKBL%`_N7(dq19I>q2cY&xF~HBcw$<2!Qf z{|E>ps0DFnv#!9k&UnyBba0h@_`{G)Osgzxw5<}(lcY-w?N|DyWdEpMZb0}@NlzZt zRUZ`)6gV2{MdR3U1}dS0vS8bZ4hZCC0w&$>(!npj$bfTC?;1h%@p*eWHF^GjCNiaH zYyd#s2(=-7Mem|D%uxB+0xbj7!zqVtb$imO)e#;>N1dh5T(A#1N0>?G;1nBmP~1;C z)~-Mu8!kg2r`WB_g3z61ys*%{Fe<=(@Mz>s8b?}Np2tUx0DOH=X3JxcKnkN2cu^z+ z#2DFC+pIxN$yf7x*3k-anm}-^W@QDFoqxt>eci|GHn+N9~9dn3)#_7||QimZUfRA`!?y%gHk!i;! zuq9$a7%zvyD{^m=#G(XU#QKS$pZ+rN=MAQ$WIva~#Jfv|UlESlEC3sGd-I{sKn)KU z;NLrGP0}YJpS>MUr#nV>m0MNtcq{utbS==bEuppX8m8?bIIXW&E)ZlbMq$!x2D_qU zX^Z9AMc~@yoCk>gz*7XzSd7PkhL?2J#!mLck!$%|**6 z6aliH(~*B{f)E|_@k^2-GMEU_3v)NDr?D?Y#9Xrisa*KYDh-eu zr^u-v{`*)?Poi)SofYq0F3)|o@vAsrncpjVND^WgIEK);z$8&yi@nOO-{rADVCZjI zdkYY=)!y1rLm)c3d3U8n*R-bW&TKfwP_M=S4BSnIbFg6!jNU7Jf{%LX-U zVHyiaU{<~fu{-1%97fcgfCHAO01+Ne^AXxBmo^*_Q*(=?0-qyBl=iS)MT5l)&)%?v zZ=p(l+EISn2HHo;e5NqeU6Ju*w;Z+Imj)YG_RE`D;R6PH*t|^kiFvKyud(>P>YT(h z!QER!-y%UPeygv|$9m4L22HNBvo6Hv$FvlDb+I0ASVzlh0L%L6e9F^}O)mb~Ls%CY zC@dq+Xu`tE)f;!sLp{=E9I+IyiD7;C%5imi!puU2Rjp$df_;g3T>lO3P zKG!GnX6pKB$U#a$yQ1b=MPv8nt`8U`341JDoz|YpqYASecr0L3c3yI~btC8JW86}q zckM<9XN*5uizm1LwidFNNV>bb!+yT3(aGf2)&%e}mvmrlET)6W%ANh1 z-rdAuz@^Ek#?DX6RZRdJ;ndF7gsh%z{_$+NB#Y0sKijUrA;7jDiZ7iTnZIHpJwh%T zE5G2bGd^q-EF7o6w!Tv#s1p44B*__iJeY^%JS{ww7LsYSvqGDAB23}~zx7Rh(_WgU zLDvOKM11)O5mUtO0p5sHD@jJM1UfU(ShpHr@C@f4a{-jqx+)(Q7#DLNEiftB<5}uS z?1`^i|EpP#K5z&YL{qs9F{`kFM2nV^g{>YN zws#vn>z;R!Vt8~2T}>m|)*M`o!ti%1K`86T_o{xK4aPnYljz3rLOj=E(c&mptMzhK ztTEa=$Rw)zqvOZg^@@l|=&_8!XB}t0L+FK`G8}c$udX*F_Kr<)!00~|qc z>pt-!1w{7{|2sylc1mS6uhelqz0|;K(5Gd7JHqGPg;4WH14d_*EA5TeTjq2gjR*Y} zn62)qt`gECl6H_)-eK4*ejeaS-j#eiwc!(mn$faE^0E}Q*}?oP8D5$fD>^BVMeNp% zFBa-Ai>ge)F$@18tB=L|)kJ|UOdR4w-Bf&j6gc3t#MD5GWX-a0lc444{Y-hEa;3vY z%on?qHCMT96Vkpp*;3x$c~!4IR?YXf?K_$`lErs%GGJ@NE%%ztZ}#71(H0{ORexp> z`W~8xiud41zPp0*cl9zi# zMC6#SN<6{vlARs6RU%bXyXqxtiYbjKfWj?pXQ}TouZ_$)y8t>(08S?;5N3<94!@dJ zDjybUj+$_c3bYnz485i(K7I2neb{cY_|L^WW(Lq&x_8`}^3I3aJ2uchiJC1fcg{Pt zibBdL3>MwcTmTswFaZY=nR(m8ZSot({;x3|b2c3Z>1MMC3`l@7p+x>e zhXXXK5eby`E@B6zSY zRzj-J!Q@65CTa&1{^ZypMK0!bxYRG#Zv894WyHJSmp-6}c`-GU zQT=d-+x3yXgF%2qdWF)h1+MaJvAho75G&yRb<}vh-><^d>ysjLp823vY=F-|)WVKk z0>S?7Y~{*GCP9*Gct&7k-pAkvE+1)PLNNJXowl=W%NEw0|he2R=|J(J)9~QUX!i^YLFBZN7q( z4E*GHhrzwXq%$Ohl5E-gx5y9kVOY&xGWW7;SN?G&qXkXx4gDls%SJ9Zk(m8B{gzH? zSqtQRUFhzjfGW@Z{GBrp&xP= zK!{(j1i!`zhpS!>sk0-0ug#~S5*jFaLk*c-K0Kg6#g|Zg1YFZS8Xz-KTf&eJ>y80P zuQwyo?=qYZoOC$+`tYtqtI_DmD5o-8Dq66*2TI@Rxv2J>oOonNx`Z0i*2Hg)TW|** zOpkd-Xy}D}osZzlYj7+hrgJP5Q{cCq4{FZs88R^<-Me?zMmaI8EZIF7f;zN~Sn7!h zQ9glV1uk@f9XE^K#-05p@g(oQ`Kkp?osb&!#sO9tlex-eE!1+xtFpk}h|&69wLTmT zU&b(J+@d2@8chG`pHmB6j+mxh5Y)cG+S-R`I@ z%8j_KH(#`(ApuL+IX48haA;G{fMVcv_hsWn!r+awyqQ$Ne$~PkRGQr0^O*YDKu1Lw zrYY^k$al&(^X~QbFfoeqdZ1a`hE}4u)WNk`fmu1bVAMoc_Q&Z~txsK{VseA4H%E*6 zrS(&ISIWDAWOfk|*sUMt8@!SS)HK0HkQ1;9jpw_2M5PP62Y=lLFe8D16Q|Dy+Lo$Y zR*w(0eF;>_vgf^_NUGiuxO|%_vYuM!Oz=kL@Mvq-~xp=#x#Z;xC2Z{!V>OQ>$_02bs z1eRxvibBLHR=Wy(nAe~I@C{LMCB;vqLyPI~pRvbOXz zwJg2tBs%SPa|IfD;@`*gcP@9=7}_3CP-&1H>~1pAA!tiR7@KnDr^ zfI1_;wzN-TLeSaW6e}vXyoGx+c^$M*z2)Q7sSILrjZYG{j+c}8H}3B`aL=KuH+RZd z$yEoJ_;~Qja<~|VV<63Kzm}|d;DbGXT`&W&Ovs~5E}ngV-NL`-Jbz1p{`1%WEYtt$ z12R8xiiwD5&rLN&xl8|9+W*z1{m*j#XF2~?k47^5^0mCv%!1LS@*#p>YD!v)h4NPe F{tJI7qU!(v diff --git a/docs/doc/figures/sqlflow-arch2.dot b/docs/doc/figures/sqlflow-arch2.dot deleted file mode 100644 index db16cea..0000000 --- a/docs/doc/figures/sqlflow-arch2.dot +++ /dev/null @@ -1,50 +0,0 @@ - -digraph G { - graph [compound=true]; - subgraph cluster_1 { - label = "SQLFlow Server"; - node [style=filled,color=grey,shape=box]; - parser [label="extendedSQL parser"] - verifier [label="syntax and semantics verifier"] - parser -> verifier; - verifier -> {codegen submitter} - - subgraph cluster_10 { - label = "run locally"; - node [style=filled, shape=box]; - graph[style=dotted]; - codegen -> executor [label="TF/Pytorch code"]; - } - - subgraph cluster_11 { - label = "submit to cluster"; - node [style=filled, shape=box]; - graph[style=dotted]; - submitter; - } - } - - subgraph cluster_2 { - label = "database" - style=filled; - graph[style=dotted]; - color=blue; - database [label="MySQL, Hive, SparkSQL, ODPS...",shape=box]; - } - - subgraph cluster_3 { - label = "computing cluster" - style=filled; - graph[style=dotted]; - color=blue; - compute_cluster [label="Computation Service",shape=box]; - } - - parser -> {python_client} [color=blue, dir="both", label="SQL | table or training info"] - parser -> database [dir="both", label="SQL | table"]; - verifier -> database [dir="both", label="SQL | table schema"]; - submitter -> compute_cluster [dir="both", label="job description | training info"] - - database -> compute_cluster [color=red, style=bold, dir="both", label="load data | save model"] - database -> executor [color=red, style=bold, dir="both", label="load data | save model"] -} diff --git a/docs/doc/figures/sqlflow-arch2.png b/docs/doc/figures/sqlflow-arch2.png deleted file mode 100644 index 92d017cd212691212a137c1ce6913bcb4131c1e1..0000000000000000000000000000000000000000 GIT binary patch literal 0 HcmV?d00001 literal 96987 zcmd>mRa{ls`!^y;36dfr4bouHNC`*a(A^;2UD6;30!m1CdFbv&L8QC8yQTYG#~FV! zbNOEXFZ|57!FBdtdp+@u=LE`1i(;ZbMn^zEz!Vn~kw-wds|fxCpxy&l$h#1y!GG@9 z%Zt87DC~W*fq)>0ATA=L`0mbj5=x>%>xI8SE?p2?&ijvV?|fy@${&~8)>gzH$kVfa zSF$)_e!4kVI9_u49{v3|a>-r2I~(cm$r;Gs+uR)X&XyUuaU>YXd@=THviM@Z+8yWe zl*#4vay5xT2>X>GKB^!BBDOaI5;?-1-~UU;>BAr`i@eepgz@LnpPzoqZT|1?f*;^g zd8a}@5vvIO_b249TaEvF#al2P>7y(SG-T^P%lahJdZ%`thk`x$`v!i0nvVz?c{VOh zo8!N4&0DY+Pf&K{N)!R%zn9aIuoVNpEb;v3S;*JG5_UIK2nheZEQg4iE7D<0^1mmB ze_+@lSc3Z&2jQQM_!?hJ~nvsDj%tun<6IB_nX{kRD_I*F8o0e$x#g zF_PgsR^%>BW%zTSIZtKNP@{O-y!LaFDJKKK%5O6T-pc*iwvVKDMr9$^((5>X{=x4( z3ZnoIM)X3?0q;ML0Dc=u>1frD^v~`5wfRJ^5rm>B@F1KN|9KhfID-E_OZ%WtWr|vh z9c8~l8qHVJ&kzl*G#$!}w<53Mr2qAxesiDNyY(rhc zVJ|O-QTL;*xu&L_y1m8^d9a=9wG=Na*2c5+0@ae{TE{ITaJ}v`F=uZG|HXRwu@CXT zo?4OR4y9sLSCjl+quB@pt86*@?zozqbM9p`4ntDf?WXz)4Lao}az0Lj=d0jU zoGcaUG&Pdge7nP8*oog~G+wN`v3>9K{Xa)E+X7JtIcRE_v)}hi98zKJW^St4NbdLd z$QbDN=YQbI-0$T(oknaUeK+s(%|qGOZMP=tn{7iJr?vLHA7nlqg52wIb-Mx77O!%A@XIa?n3AMf&Tgg*hOKWHV?CC@)?|m1d$w$e-$Uv5#pg zkZV~lG=Fd!RWm+6YQ&S~qNvREDp0Jq@8xdO_Bd9mJDqhq?Bg9+&>1V#;=NThWo^D_ z9m&zCD%7&;pt(^z-gJq}RmxSQFEr~6?MddFN;G)^6ZjduQ#sXpbKd8rY_`l)DNyhK z&o+tifzW^@+ZyZ@uO++153Ysqov6CqoUd6P_HdqhC_PR=T;w#N}tsLco_zwk6Ei1*tQe!#b^!1#dwZ2k(r^H=3wdg#NT!D6yk z!vCZy>Nnn~N_x|!i@9D`OJterUnF|U3bpE7?SpLIPFXj3oDF()j3p=+sIQPQUZsP; z)fTHhYNEs`YIf|py!nnoc0`#|zNPs%m5>qlWmd9lZYZf~L(C{ojz#r?zxK16EuGua zC`J|a7Jtl!I|1EX4w6Psbp7&;`;xV*#eU$cP}}wL?!CI)Dr&l_WsM0kne)1BC}O}e z<7IlW|A^b}tEp5RrRrT2tfT88mq4n(@t1wChN!#CJ*7P21iTve`XGc_2G;cl{-O7C zJ(GAG^No7qSLN9bTNNpH>V`$wMScGD;_vwcQR93W(!p^Zs6NR`@kG~LZYvVEOUe4i zy+)v>up$y8a0?brf(#N8rQ)vWb7CizkJhX=6 zRv6z!KJ@JGCu#Qjd zGs$jZgCoyR`}j^1z4p@0&mFavpN*Sjd`65>kM-(?+`g}|U%NOZuoP&L(>6f5<2h&a zgKa?kccE(+0%7+h4f?YB^PA6)&lO+7k}NAHj5s{6-jPWST;vw}nWsi_sHmg$Sr>o5 z2UB7^oCfD}o$6&!FI~J{D}hR+>*v9Z$i}R8d5a12q6YWLX^+#HwFfGB$~{?Bxi@_R zx7(rgn;DBpQC_1DN@SsL)gZecYKtz@x5gB?BhZ`tJFiI(Vsc5I6Y;xOheGM?PL@Ii zCY3cTR}|2da@5L<8h3qANY_)6o*ngFq+QrayDx`}UOk7gZKgR0M+H^x5ig+pQ^N#5 zpceb*zCJ5mTciqoHFz?v|3KmKydN4~J!b8XsG!DWbFD9<2vLffk?qk;YKt;e;KY74 zLn9fBrz9j{JeYj;L24^iuS=SDXL&AnPj}#g1&wW|0d_3S@@oZ?)1ZalepNDJ{Y&xv zq1)yysB1Jy{b(BB`Eo?2qX7$aW$deIL1ggPIC%u?rt6b0ET`sQ9pp{@K6Hif?wYv5 zH%42N@8Aa>vY$3&iq&0U_FsV(nIbq~ZM$Oa%lIPfTOzmpRc3jK*4QK&#EhDue86g5 zZ*6eToxMPTA!og-f$}9tA7wgQ3i1ujZG^1a7P$H_0<m3q0{E~&9@fuNnFJ-oM6?;#NVPv`{?$i42aUq>XLcK~9}$#nA6h?Y9xUr+ z;b!h8x&P{zeI!IFV&&Bu1)_GAtZ}=vD~MNX@j0wR{E_0{jQ^pp@!HE zP~o?Ahu%fzW|<`3r&zwA%@U3i5qspcJ6$7q-E3`8kCTb>GueKt?@FU0YcNNF*4o>4 zOw0NFbm0No2=8t+mAl4fnO{7I=BQ?e?Xzg57n7{VtBHR9gf(3R!uwoFkqI>+dQ0_` z42N1(Ihe^u`So&svEFwKrQ@DQ1EK}COFx|wCCnovv@OD`!$zu+#tQ3p z55c=0KZ%UbDuprjo>8t!5~8cH?(^e2_HKVCizyBx|C0JeSX*`0ZI5ryUxjOtdSR)X zvfHxXy!Tm)3Sj`5=h^q=aTR4b=+Xy5?O5v?(oq#?^2_jcnXZUOUj(D5QjxOcA4HL6 z-+U~Bw%dWOK=&nbGEHw{_y=b_CJWJ&PbV1DpE&=u_bX4oJPp5iz+jOpz1*#H zn{q(Azs9k4DuPZWya=u;GS6@)YG#cfasaV$$LF>l22cD4tTOpM1B=&ml_By3W zb1TK*Q5w7S3)3+2i<2p9k0UWkNmFNNg+hE<&~5==3_c%&W~OAEUW|@sjX?g#!BGwC zol{eJe6gK89ndp1Ru%(QEULPKFmzVdN1#HKVfh z=~>4{lM0(FF*+wF>t@0@{I=mrwV&vZuJHr{Wp9j1E80}j(f8W?K?7{QR-uy zre(?x%Rgy3dH1i%Oef&OV8_99Yv`i##Pe+7=EKdVbjK^ip!Rk)A&47s$9bXURHb%o zBuD&n{hzTM(IZc5Ca;t!DLDuV%FPu#r1hjs-6nPNiV*@Es(xf3iEqDY#dXtFWT~`T z7z-*i-A=Ho%^~~3EItBR}qJ$@5BtVv}Fi;&s7b?VYnjx z1%N)%c&7%zzHe2d(IrXK8watFcpMFQO3BvuzO$Y43&Qpm4a>+0QQtztG0GAbpZlBp;%bpaF76Jz`uB43@S@(? zKltq57s*~M0^w1@*Gm&N&Bw^*6KAiyH| zu8y80#yx&7u>UHe^oUmvk`;!MW_+(#6Ra&j&o;p)_9i9@sn%gb&vA3OqqF2W-Cw{` zuo91ag>}l9E8r&~k`UJC=(N}WbXy`m2=5ut%LK{(;50a_MDY1U4Nz~pX~xjMk^nea z_^9OB>CZ5Z?)>L{wTl6~`6-cu2lKBJ`g@5xsQ;&z9(h?h=I|Ji8~S#(-i5<-klK7G z^Ycz|3l^FJT~1QmOXEp^W~{xg4gxFzVgzW5GvS3{5InkCXSYC`TT#?>GdZm7dk-_; zaclG@{xQ4R@|aGoLyiY%&CmCa)1grel^_oja{VO!AAb-Q2q+B0#r%P_MC$$P)43@L z)&>LT8OIoA?RqZVdo&d+qa;rFGeq|02ma$Xw}O>1HOFX-jG?*og= zovJ>8K8$~Eqyq$Y+$*;uz86T;Z{)S&IjyZwp4fOdNdfSI+o)x`ldo3x)2i{@h>V?Q zIZQZTzb(XRGbhUxN3p_8Ndx@3xxTqNlwqxJyxiqw-F2EUBs3rV8VEnwEDcBUTO)Zf zNiM5#Cm_XNm^(E5h$dxl+b#)UHydtmTK%UxYj*%AQ!+*9LT3XhvX=uOm+~Xw96JVF zdII_nQ|F+yc8{M6B49*kt80CzF%E#Wq^X;3&0gcB#m&{Z)3j~dxry?-SMJJ+c}_c% zP9sVx6X5wXPl;b)XPn9ln?uK+z&n! z08%z(>q!WdR@h(MpqHAcoU-D@c>hg@(|nZ4E2+1wP>wO!T6|}9E|&05 zQZZom3A&EsH7|KqtVvE@;s6fX6gc;rL&Nx{k6*nj?n&Yuqb8^RL`_=QaH5_v__Tx# zFe$gl+ImpKrO%v|1(lPX)K1jUX4XBkEr9&kMuuVzgIOE7yqE{~nr?S9IM4r zn_2Q0G}20u(eH-kMAbl6w(L(8@?-+wIbPi0`fMrG={quE?5qUa?7OAUo}Q#0M{-^} zfUPlC*LYZnJakzc`LinDsfFWzO$fOgUY95vwi7jq|{^?YO zVKALt2v-8Y`|8jzhmFBjjD%uSw5NfX9)V-r4P%6tTZMJKRavKX+oi$uzm6A!SuY$R zFFg8V+z252Z$3TJ1)#Gw382JEi-{Ks<-#j14+M@*NjNNG>7}H%B2VVLyu4Vh9E-Z0 z%GpTaqOL^^^}5>Rzgr$}!4r51yIhOKi58sC$x1E-i1Dd_-B+`q7p-qTs#lm9s^fkA z_+gZ2BiP-Rh68yIAjW+;oSK6Vkk@cT$B&+l-}FpMN@o$UdzZ#^6sfp8XMDffsrXxE z5NVk%N{eu3S(Dvbhor}EwB2@A6lSa%j+NcE3aYmKiLLh2SElXzj!*B835ZWeD&rH# z2Uf_i7I>_W4ei;UtjP07HuO=_>E&N>?U*IyNsgqm9&8D0j`CKp>boVBKcz$9j5Fe z1P3?J{|i*@G^mQ?GpA-SzFSfMhP3$)#mMF?b}d{pT;0|vm&4VG<&KE!Mn!d%ht}tS zLTzAtTYqf0Od$uu=Z``mefRB@5#LtVAUFSH*#F6GxskLS1hRtP;12TB)nsi@uc6xM z+W#8nI4Qt#PaKNwF$xOixI?9eVobPedJ?&ZvVsdVYhDP{QqXc2qVgzM*8D1#Mdc43;xJsUZ;FQ&Y(^>0}zafMh<~Rvd|FR)@07HWzN{G+&t9K>VH}3~#lyw)bdbw%OIw*2_mi2*#%Ek{ zb!QqMSbG%HF#G6udVAEUqBu456I8NuRt7 zzT2xq{?a4B)Ssb-i>$VM zag1LQrSpqoS!jx1{6*G*graG0tHt~`pV|6!w#JKYQ+oxh=>8Cf{B z6gS|)`l}hw){pFfr_$ zxdqpPakmdeN-D}*(6OO2>)MT{WbEjDv|DP;Qm@5vUe=H^7``06epVlfRB?e{ixU*} zQ|U?3{$suX&DNT)H4MSl7H%sZ%s4Ez+UnrWO&z#U{=V~9h+nUiNZody;2=Ux6@Jjl zJ%{OveZ+sLz*G29QCz3rkv}=zB>5j?vW8a{6bhw!(P&@@)E!bHH zvz*>)dN9i`9L71>AkU8wTRmu1L^eiiQz{e^006ud?B-ylbt~-B1kw3q^UXmqiSVTif8VkJk6nBgywnDMw(QLm6c^m4ZFQd zMuGaVY>cwn2sg{v>@dVo*RQW$mfpAKZ!; zrhl!Z((z6&X{TaLT_?wEl9r%DBghLT#cHeHooAGDD@v?@$=Ug+)hUgS6}O>(P#7VWj1 zG|)$$y*YT-Z9<%uWNCXtrc8$a`R%)P3j3{+@|f5f%;+Ac>4AK=)A}}!sW8ejP9g2T zg}yJMAmhJ^E=K}KklYG!(&63y{CCeGfb1J5V7slWyA3;&wn6RwwQcOX3(c|*&7mc; z4Jl^A0Vn(E8{_&j8Y<=YBpz!pW6;vgHR2aPg(J5Z1?L7x@Nw=Xz*qug|3#$l^B`y5hKXe(6A@6pXM z!)DDl8fP5Oci=bC{Ncvw>Te=KFclw24p5(#oNh`%KZdf^;Mp6gREZcC78;pwx-u1Q zm(~!TO*)MlJHs@#Z^bveyAZ9IM$SICOd3Gynq6TsK=~?e|Im3ZPAb`g!&H9Oew}>0 z)G&E`bBc3M&1t9fUDlt&A13O3`|z9^{@0J#KA>Fn8Fk0%Hn^QSHN&cP2CS>Iha@UL z+sB=qHr-mnr8S9(Ekn*3hX`iXV)BJ>i;5_23S*~)!{)Oy{jz64j*EdtUD0t`@6udU zzPlcGbK}HoxxglP_m&I#pl6wysDI_Gc0|P1YwINB7Clt>7QcuQEHmo)jWywbJ~$WN zr#^T?ndfzTbJ9)~Iu4rfWi3Sg))63eSvb#mR55GSvfajuXj_zrPZZYGj%Zk&kE+I) z(#gTHqXzF1bFkH|k-l_hOnEw9lrljR6Z@IWnx!z?LC0Zd2^Hw39f(3P*@lOm))rg zN!_-e6S6s0_*w`hB$N#OC3oR6*5j?Yz(|Zt2lVi4eaTI zjhb$gbCdg7K&FT10T*zDI`qXQu-nlQ)w42KC+$Kq3~28Jmcn&4NZ`|Pz!_P9yyyhP z7_~g*f~!VlJc8FyVbi%;$pq9;qKGxhY{{P)p&a%;Y`n|ZW%n?HlftYJP8Ko)P$%vQDAB`L>T)fzo_S3;t3dXUv^m8fjaQ$nJu8tt#gNtFcW z;{}e~9mE=l;CnM*M5q9A=gR0ed-w9vw`6Q4VGXx8mnY|IDHzT4v&tdq-{Hp(KJ2Ju zRT{e;6;WLP;=n+a$zYhYJ)9Hq`7wK^XaRLDevLfOyqqY>3dgCG*REC+gHbPS>rR*9 z$!?udFku4UzW0Pt!V09kd|HMmcP%-=c&gD#tFyoVZ^k|q6;xIn>L=cUR(P_0=t5Cb zDLytwm|C&xfK#XhYVRcA8?A@3WSn!HlPm#sWb3qU{b6MAMzg7r+t&ExY{_L2YF?CB z2$Elx~BYEgbU>`^O z1XRpaYPf&D5(9eh2+`6^E$;;>k;Sn3l%bkbTrU_;(IQ2k#(1 z_RYOJKFio)A5M5qcOoL2miN0?mGP5Y4_XP6TF;FuSE4m00aLYoTH^`~9#k(aU^g`$ z!41rH=|Cql%5?aC*!wa(Rl2^3F zLa;nG_C!ZoBu`~mf*UtvP=?o5v15$K8r+pmD zGD@(vavFfFCf@J4IAI5-5=xbg1f1R9orfEs_e+ggh|rK2Bn7v!3?dp9{E3yTnIE?^ zBZBa6z@wMpbvu+vyqqV&AsQM-*O_l_A8)2itQ?caTByT}B z)M6ARwp9vZ*R;<*hS~?fEaOsl4C46w^ZOWPjfWEp^mdNxX-0joB4`T}73x__-;TNF z>xZ((bec%hy>UWyxc@t%7=SF2i<@dMfhrP(;3f-kG@8QkY9i%tSal}>4H1M;LxCzSVkDb zN<0>ji174NW9ZM}kj+hnWuNdQ!JdOHg8gz3X1b1A^}F5i5yJ&$I6L)qe564c!3|Fa z)y%9EkHTqNJi?Qjn~N=vAGv{3__;bwo-F7N<#1s@AKRj#Vgj@y9B!4|LMB!Z3SVUj zP88R{^iDSOTrRZ@yS>-xen*gEEZCK1R8s`R2_djF+(-oIB%?O(M-T%vvGPz7-cjp& z`E(7Xk59w6fDBqRZ(T)X?C1Q;7I0n)Jf@L5tZ}Wos#%_FZjP{>{3> zX1$4D3YWS5$|MXRlVtWjx%L68kU zd%BV5r;{~}aOG%|O-}lSjWjF`K^1l}FL+N5<;!IK-A;u4&&jDKB1l}^>4O>2#1ub# z{|S^m>ouIHO>zGfIEP=IXYYR2bbHzKNkc>FQA4`Ii&EL)&6>6D(fke#^O4Gen)Sqq zUn?Jq=K21*Q(SPTPkSDEd=Lym?M@MxQ!di}+@ce7HfV8z45OiRoy44Zt;VQwZEGC}OAZe7y|6=Md>v zE&hT`cVL$}*l9tq<}Cn~Oh#y8H1z=QksO+A1$-&M{}|i3Eib>?nGYB7e*-qnm`N5GxN-NA8^#v zBmv55jp*tp$7a2ApMH@B($Gm>eDITa>h6PA^JGh>ha6~DBc0n9TSK(@qg-pBE~g7yjguUsXMiA( zTt_*fKBcUwwwuVF?96{~>$FR(etMFMQIzAk@atl$XbL!$wC4{%;d@x9yy!+0y$$p+ znq-;fA3(&~ZY9(y2b5O*a973X5$Np80bQCl>^fm1uR#Q)ERq${PK&C9V}M?*F$h_h zPAdYn!%TEqE=}zZ7-zImM9DnLZ=d+Jx%dmZfI`qI99)atE|0h^6a~Xe_|URBZ?{lx z9K~^GRzH+a5+~1s*$$=teD#&HphrNA!hEzw{EUrn4PTu+r`Zn5C@4X07aLiYptB!M z6}r3ESlz=hh0mCs?6OK^K315t5xgZ&Q)FjuHFP&zK|T0Pmr0f=^YM?Ye&1fhNDM zCx?vXluqFKNP$c|nZr`Orbz1Oy~t(mrFg2&-i>^kqTaRK!AGhPx_uX^av9ASU!?cw zku6xfO#Qw9j2f`qIbEU8mWM&-Z)y{3eA|7J4=*d{sd@8m=Ytpv_fxb^4Y|(1tzsy* zH#!{&9AmUp4*u&+*rc8}0E98*2L!+y+gPY00O%{-y-vRCdnkT=Ui1*$H?-SfvY|}4< zg4~n9$?DF9yGI7(tFOl`XzhZxoYeBs>%Q?n<-;3!P~>jiL}o_z@MEBfz{`|p0SCS) z5H?|Ws6w))S=r@p^wRcuej!bh-Pwk>a3BeMPj>$$PDXSNT&7fWdqEEU=Re2U$CfSe zw@b-DV044Vy>8=a8K8s2x{!;a&_14I0^xTKJ+b=djOIqPuAF*|JR$)*6ANY`45I{& z?0PMhe*0xxx-9UxZL8&hKlw_(!?na}hKKArp4Zu`CO$InE5Cwa)MoD`f3mA#Ma1?Y zpyM;=jL6k?%&a)3ZCd=|BfiV=NI))f9Ta*Rf>tYlp0lydbJ|G>syoNf%YX)40xFC#({tSjRt9LHAB7-QItfko5m0e6`w zsd&iDMY7xOMnD?7d5O91G9rTye)Lr0UY*D`fMrm8qBaue#G*oWh8z|a=!Qb&)9TNS zk)H1WIGB1Ma24jI;qfN&XC}M0`ks>fIZ(nARU5jKfkgi$7oWS(?Pg8j_FyYS4`Li1 z^c*FUH0yI7P{YiC5~~@GnIAR#s8HVyhts&o%d!M323G;-1gY0ImwTSwaF(e_!gTBw z#A+tTh%wqq^HWxqPGcx6nE9d#b_2DLaZbqGm>kFuy*$S|Etgyu~Wl{DmFiekFu{QGZApzHd+_npe15QRN z06~q(csh;AFTWq`%{lAo|=e@Fn1LIg6!? zvK#j%rYhwr`&_M z;YCp0u5AP4-Fqyr@v7TJJ(Cu-6T!>*4s?G~ckm8EUKFld!+lWg=`Y_i74n zLiG;qo%Bi&huk{MCm3Guw9fJmz(@+E`&NoX?G4zk1M_9is{>ITR)bYa_VY1dMEi~m zu-O&xa{s!v9jLx?)v9yC)EjGt+s?<(kfRuCO=1Nqf%tt76r1BQXsW3NOZ@8G)|^#= zdPRNgAiUkxa8ZU$E?;SH>s!UBn@h*tz{}JRNqsAtJOc z%#9EkF;4oMhxoCnBvdV8Dv_zyKBwH;n#1VRgobrojm7h=$q|eFxAvxPD_Ni#Q1!|8 zI^O_sB2DO`Zd6I5aPM*#5&9cg@ED8JZy}qiU_ls7-b-`-FIE zv4`N_gyGmQjjjY-7*UAFdUNfiQxL!IK=xbNi{we0@VId^WbL`_eB!DZCygmEnQ<7} zrP~W>*PL9XIyWC{K7IvVQy5@}MK$ZQY+S%S^!$c>kkoF7FVnOqd9y9kHT1nKQ$7}>zS-9GHk;2C`9IfR5i#_G_lSx%0>E)K+ZKxtYwbwV5!@O1h%ZDoS^2DPsQV0| zXJsiK-Xpa#`P3KDR_Uwg>(_?h4IcJcX@z>R>;`BKAJHgX`qRIEWIeF=G_XWOLf6)| zgC^z-HR=<|Q5?eyx^hr!q-=4daUacbR{_SaOiVlKPs|IRnmj3D^xzpM z`=2mQW;gs~so#HXPaj94BsFDw?C}Ys2%2cVS4s_D*d&NS((qQkH?cor8M~h_qi(mhOPc@OKp`k1zs|mduFL8N>dR||y^*f% zRNroc#kKcrtq%iftH_%^#@cS#|MHUYqgU2r>~g3=iMnl(w8bt^+l=(`1DTKs+;jm;+G~|h?6>|hPcp+S*4@<_fHY;8( z?B8k=984L;Xh~Xq2rSMvxs`t|K(K>1CYl21(+hi%2NYp15OEdNQPLb9p(B-sd)s4` z=_giIr{m($N_YRH_{skC8gGp3GAW^2V&>)6_&_3ZEjNYYz8y@Z22-ntJB0vSDX7_G zCtTHkOQD?f4Gt14l>YJh^c@Vg)lU@e4zlmNh;Wxv?1^tK#>@dc6L{Fz-w+j@RoQaP zu~73aIf^6m1@Hxu5!pXQJ2$t%KMoW80;_$=Y`fL%d9Y@{S(skIQ?j_Yk(F}7vE94= zBW3IXK;cN5QhZVFxcq^NEreqQ7gqY@LGW%K|CZ(>J`9(i`Xq(6 zyb)*#d;rw13C^{_&;CTmR_J}`(*1&osK!Tc$OuS=9~roiM%(C=j=KyU=Mw)y%?RQl zeP}y!-aNM9PAB3hZBBN48@?)RPgST}r>oJ*M>{#z&!9c6CsBWh=Z0BCk@~>A8F?(n zZ@`uUiBd>#w3+3Y3fChI4~92?YPsPK<(%G3U@_v@2Wx1V)Pt?;TQi+P zJL&rKmNs4;7utTrn(LZ&!+jvYzf%#Q^{aQRdim_231ddRFh`!8j+Vy z1uO3vuzZc%M}KY%Y2{C}6~$@$iTwp{4((Lwirr~OV)vuniM`$n3m!Ui&tHtc5D{;w zWFdXKccWx&Z`-dO;iH?c$Lzr*rfdz;GAWD=gWKh=9)Tc%>>4&R!I;2-r9@4xStMVg z{PNO_`wp0zGRk-pcf4oKBCp$<|K0tF^3zc}rp&;-1BPBPifHfAZy!qrJ0hQEMHW1p z#MNLODLSrrGZ5Vz?hN7A!^Fa$Kpi4%;RimY1BTX;SCBD!p0D_V*V3P}vQzBTl$xeV z%^(-JGQ=g0_;Q1_h2zhECWm?ye2l$#!o!}+8qaA=LZ-oNS;S5#{mnH?1hR~esc?(W zH~93M9eLz}B#%Vt8f}utn)U{5*qTn<&83|v&=U}mB7ZsiDHke`hpiw*@@Gb}!jfy& z0|}MXStGk3vnbAUmQLFxtMXUf1pzPYXEA&6$yqI>EvT6m2?*6AGRmRX(k=jexj#-F z>{vX1pVIBp@0#AM2i+mu71ezXs4C<86OQar=AiThAe_FMI+Uy8@>Ytnj(N7*1zjc%#O8auICy z;#N4cdwv}iDTpGq-6#F2dVSG){z*exY|V~m_c=xEA%o2-rcn~_T0sTWPYU??wKP(A zs6OGW%73e;IrU03t~=q8|KWq^vxr(9`Np33(G>6{9iAo+8o#-8?TCH72+$H8pxYoW zPJoe=7ZdDNfgT9T!mwXRiH3M<4wpaDal`45`t(Iqa6*w#^X;l;Z_?*V?gV>#_g@^S zEUZ*0zw*v^qKK9U1^8-3iBAknoh{WPegz~7BmIr6gIH%1gSLkH!-C4>ihiy zNkpi){bk=f<0=Ya|420f6-GKJyN9dx*G^_zTXXepw0=$4_i8w^h7WjRhE6{!>~=In z=sDHvC_IH8atR9ZybV!OmmkwV3s*UmnPEm&DY*)?eR@E@D+BiHU}bKLzl1Iu@agXN z$-{h+maA8sYDHyN=BHmSrz4f$_JHZc($DyEbo#w-0B}YCUpRUU1Gkekm-`2K>zby z00b}t#0qnuTjWg;TgQ4_+0OcZv%C8hEQQ^!1hBehlFQGeA*6E8T>q(eMiMU63P2ag z({8BcJDd0AaNJY`-p?s=TofNIXZE(PX?_Kx%ROtSL2)BST~F99=j?jW%R6E4Mbc#ugb1#=(*v-@cI`amDvoo{YMpwfo>Z6Dd7Z?mVMi+a+W%e%*aQi7&eNR1LO#Im zJgZ!&8EEJ_*fIE)1E(68p0&`Y9EkZ`9XK!0t7Qbh#F3Q8qh&i3`4;3B0y?i;TBE>1~?}$f;d0Mp3$I#v2-U3>LUAMiw5fQU1xq z!l`zs^ZP2?(>3Nsxc$J+>3lY>A8NRYKsmg{pw!*_YG}KIw#TTKY4SOdLZV~z#CzmF zdlw0pmS6E6VItC#d$;JIrw$s7;*RSDGg|-$xKj`N7#A&1{yGECfY+5Shn`4Bp{Kga z%KP`f1C|H^LQQT6VoV|X9MOEbDyqh6h0`COV5xni1~OlN*joml<$fxi7^`tt=DW$t zMb^5aI8h`IzwY}9-v^oVt88d+0UZLKf*w(#4ErOc4pJ6Wz{iz;UxY?ED)#; ztM8Kq$yeYUemVqW%BBTV!IRYr!N-Eq28euWz@@njOatwix-ds3E24P7+0Obm0jM`x z?=rg+jHeQ$2L{O}GD0&ef~VoKoZxGuTv`VLLhu1Z;s^mv@R1{uJBV-@iYkPwx$gRN zSRNt`1)|->)(GPs1;(+A0K40XT3H(zXTnQbR=WzPojIu#yPyeDw^(+otJ9K2?Q5XF zp8!60s?@M+`*hB$ihxlW-z8tOW)MDL&b)8vt23?NFh!o{db|m)u;1p7m4rT)z8)#m zn&JQ+emyeW8BAi1av)x6jO1SR)K=FlJ_29@lc`gW`d7PfWm}^#ImT1H_b0%>wPvmEXGBVRd`Vz0jHIqDzft_jU!o^PSy0W_ zNu3FXRn}UgkfKQwYTQM4EHjyW!2HRyz*Yr^dKSoq|1%m>$g+S+w;@4+jED_p9veL_ z)q#O%njb1N*&a^SHHV;rPfJ^>&4BvW6@9pjPy%1GV%5-DAR{J6<9o!80ro=+lWs^X*T ztis!`^x#08qpb(;qy>Uf^FVlW2g zJ^S@k_Z;*J=Z>yG%1W?)vje^@6kCYY0*0=mocsJO@GT#;jN zhS_NTMyrqCcOc(vBS+q}Qw^kM%ZtNwn!fZJtlGlvZc<>i62RCuolhjIOkv_1!!;TU9w9OG^hxKw^&UVJmz5%Wz< z5+$i!77=Z&QIOT;@@xG{YRJ`4he4vL0Hn8Uu<@NDImU715t z2E+Yhrg1Wx;C_}k(u?}mhmbjn%}`v+gf?*V9Z5{e%mCNqIZvv z7l|6ZT*tU}9Mf7$ymD0wz#_qx5x?>7g2;Y4MhzB;ru2airs`y{12n~ILdjXIbyHW&pK9J#`bm25NuGw@ezkncja!B(X zK1w^p=Eb{R?2Gy6**7fyv$yFgatR_A5eG=))D%luVRzs?Ly#;*_M$$78J-9>6koUu zH04;Zya?_m`~^C~vvvRUCxE-lU5cbvn10$IaPCp#k3|v`8`O^?G`gp*0=pgrDF^db zEec!t0%mbbz^%Bs0tk-PF=I&fmIE>(5xiw9Q4npTz!tdU@f=vv9%6YNP-Q$tf>M*m zu)}WXdl-m=r&o^*HR;`he)a(OK1>msl-==X}Oz3>PzJfV0Q1t)&s zGX)BF0avbLZI>ocD((*V6J?%KP#g!7sSR!51~uAI^f^+NF}iRSAM70Uq8~gOoeiL@wQuoF+|03QBDki4;2Fa&Sm2v0&;iG`}42Z6T#|d!YxVzTT8<6WD2k zt_eT|(SR#U7w{>`ce)_wThsx$aAJ>+m%s4;@%EKbRjysMvIV7-1`(0&l8{niQ%Zw0 zN=k`ziKqx0K~g%Tk?ux9N))9L1SA!Z76C~C!Mk4L=y!kKA9oDL_|7=T@7-@a&wAFH zYtFe^L#2<&!IKnR+I%ZxPc}=p3}n`A4jxcw^ay5|iCzZZ zPF%m~!ND@k873bz3-Q8uZNj(AQ*Dv=nFd8^j41!2hO@>$HDv{)e)*)^VXXMI`kdh7XPH*7S zDO|JBU>qo_sY|~U61@#=fh!~6^)mNf_XDUBBnJ74kL5-WVpldUL96WkgrR`^@h8$6ipITFQD2woCr_ z{tj^hqVs^JLJATjSAbnIc7LN=(2~$q$P*ipSt5=wU)tGHH?7mIJ>;rJL!(dlAuy7n=fW$zR@Fhttp_QSP7m+Hd}R~_)NP&Vm=XUw zE0;iD$CnL&&}4=N6h7t`IM1VRY4tYnh z7wVMP5ZPPf?5ZQl(M^S^w~jrrL;(|ZJ8>*~_t|XS1f2K|2!Y_lX`wcU{;^DjC}%@8 z3*1+vknLSs21VM}F~?2NXf;B?IQ$N*gL;5eGrK$>xMskR=Gp{|K>b8vJ7D{p6g@KKRMagD7q= zsh8b=SQM%Khx>Qx zx!FAcy&~%N!xZK#Eh~Qu6Ld=F4NhXZ5b%6?H)Dqw>@iMySum!fut^!!pWn9Ihp0eWCmch? zAB8)G^h&aAPJ&+Oal=4oW1ceZLrAfY5Uqx704 z>khDq{n~GGt_XvPne`h$6%i{YCu12oTyrK&z1No!QGwUMaJvHm_yn~8_G#Ggk$ARX zFB4H2_Fxioyq;wbPWa>*lDP5l7C(Z9{V|T5q-hbq7@DMXCB@@WcYvHBQa_e^jkx+D zk;Jg18njTaJKGp@Z|NqUPGf}aouI1RD7@H%g^n6%X6i6ei?D>syDj-RnBIlJZnCyOkSB5 z47F0?wnHfSD&R#j@cMY*EIiR|gH3tcA7jTU2G;>t9j1Q~C3JFg` zdOZ7>W%Erk&~xtaXIa*{ui8V?vd9iv94Qv;D3;fi5w^C>S8m8in}JFgdbZU_gtZ0l zk==$};2#vKrAp{zTt1O%&sO0Je7)Hc-931V%Lwt4Q7K(>p97d6?KHBvsiZcu90Hx{ zUd*h#&zA-G3ZYlc1t|CJvTpXhY|mzMVsaV<$X*|>b|OwdwA3u2NYclbG!QOpTx~zG zTv*Yw{<+=!7Q;T4^OW5 z@fA5-0I9a8uTqOO@ZXGF<>m9**ry@eUC$=DrE~xRm*{)#od@z~tTV#? zQrU8Q-Vdm^*7Fibb)BuWu4ZxwJqW);q87+q^bY?$ zuN&yKg)EVmu=-A)nRKf*15K|BA;{Tnj@n%nsJI%9jfbPZ-+^)XR%^nI!<`UWIq<#P za6-DU0<=-B&w{pJYA`F3P=0xwvkrlDS3cN_Y$Ic>jwv~E7!tK79YE`xPqn!5G~nt~ zbOp$=JA-)wFH+#0tSXY%sClh$l{;p4V}5`B&Vx4m0|?PO(gKBi7$&JKr&t|wxgehD zuOJuQ%}og5X1=&}rF1wz|5RSoTf*Qy!dhF>b zd>p$|6r-^12`Z}~iQV0~rm3xgr@|*4p{vm_{7I2-fUba?B8X>pB1PIdFeo+|sP(f( z8<`KNR$TPQZ|o;@y@v{llhi}Wa=#J>lS@V~O|eCt?nRt?1VA78oq{81UO)6PQuPkq zscv&Fvr~%xu)pbwkcd~UdpC`7PUOR#3{TJ^N~BW+Wqem$k@N7A0;1HxtSP?9aO|tB zAua32R|Ptu^0b@nh@mhNVV1$85EEq0h>p81u#(rr8n?W9!F( zh90#sLq_%U=Dj6aJuQo|Uyal*;hG*vPvaymDqG^J>P>yJw^O`S;$GLzdC z_(>-CCndi-8oxPs!R{Dw?qevKEJn}`H%Xo#>5_*dyqIJ5$+0CR_^QL$R!L5enu-DY z#JF@2An8uC*wH;R)##YdJ6U8>7?))N1M@aXgM5oLGtQ-I!B=Zv#HT1~$h%D7O% z3RObWz(G#rO)>*ehh_$N3x`He6d#t?wzDQ?F7OH^@}j;ss!%O{Tit-;FmWym_2cnt zuL|s2oJFep;_VMERC24O#nHQYk$ph)y0U8Ghup1`--7!{_@)GN zhLoIthjW3F|4oJ0dK!I?Y*%In*mq;bdow3kUFZWD!KDI`%qoDlO3DG2T-<>Fv$ z*(Qf}hXKnlM>2gXp&+f=^QsCo+ZkaJx4_e*PR+A^(uZIf$dHvi$rq?@8gn=;J@g&S zy_!)7#4enkE3=P%>OcKuEA7UH2ok{fV;HdpSnWa>`fRN@TF(|nGex*tJ^MZY`?WzZb!f-; z9-rTqN2oByD#hb9;`6xb!#@8SPS4BFUv8d6#h{OjZ7O>A@GykmLV+dXb1}Bu%=`L^ zZpPMhC}MdcEWBjIVUB={ngcn`%~_wG=cnLM#Jm$vSbJK9Zl$oB41U5Hsw@=7alQgF znh!q4Hrbn3#9wI%%?BB2BgR~_y;zsL(fxSr6KZ}co>8H(#~8w)|KQaR(n}M?K^Nhf zp==eJCXoS1=LH|w5~fXKC$&ERD5-4Ij{GmdGQfw;_tU3uS~CR07cal#Onq&I^sOn&A#$EN))%x7V7oZoJPe<0g^zyA-)=< zh8?4^TLndhpufEhI;s&6?g-qEb~R39RfC++VY~2gSd~}_b~K@4<%-T#PL0;xBy`$$ zsEB8+z<(~_A{@BHS`n$N;a@6Pu1LbSwUY4H zO`nHN{`9o4t)9kMCIY!t(GO)_B0rXB5s$5N&T-jZ&LI% zWI5Xmbs=WEJNuDS0cRoCE>@e>!cRRkerwqae#E`!xkF&FLqAZF!|1d>dJpeD)O6WR z@}y4{CGV}^Vr*IHqY{x^N>b=GSPHk88}XaAU|)c2)E0U*LkRE=M5scCkBvhGx_l0u zJ>oqBUX14KPgSC3$G;)C<~(5DdY@iDZkTLh{Cm#adwHvqAXWBn-sM}*+`V)zGKK%%obRkr^*=+%Zqg&Th zeF6$!v8ml7d;n@EaIQ)n0mGRkzvv-!yGc6OW*BvW-AdJm?Id^XKuf%@fiYp_b*m;i2H#T`{ZKIVE;S9!iv1(wU^hsaiD z7lHljGk_G>W`%nq@~LnY<)Xg&rE_rxmCBneGK}5MB35Kx(4S`>FVEsOZX6Hu1Xtlh{WDM_ zJikwpB_u)dSf7<)lunVAlr$Ym>&>wn<^%+ltm2y?Zr}&n#x}0utkC^&$Y_7D^)Ff0x!awBQ*`dJu27}f!>z}^30qI zHD7VjF8YxMD>@CxNG}4_m3O-gWZUmf>jttFxts{z!ia=Q=CNW0StJJIL*zqZJI z87R;#@m`2B>0g~Yw?uh)GP)6Q$rS*&ma}3$?1N0&5v&6%N%s{4LJoZOsS^N$1Y$EQ zCJ5Ut(M>#e=X1jiB2_Q|33H|QBp%b1sSP00UzO!<#b*z!IUzD$s#nH>#mb-Erpx{o zUy@5adqE%a=s@lmakC^}^lhY*`2M~?!Wp)-0qO%53jd}vKj_ch$DKKaC6`{=qe%2q z>fuU8&o46Zs+JBRO)NeDf9J2kb((i*68>?E{6>d`O+kl)Yfb|CJa#@~4zAz2Q)n?B z!3+THA}fX8-5Wg$boF-XAXsO}n}pJYKKv}a=uO8tUSfFytnD;K0!#b9|8*V0{QA{3 z6e?cL`%~QsK-ROLzGaA*6g2IU zo>(43x|PV2qhEo}gi)Bm^*Ab!SP#6)VOa#Lc#>h>od5Y+kcUh|A<(Axn=Il0d<>L_ zV}MOKeGBgJeM|J`7M&l#La_pkkEI~;h>(#8_X?OjMBX}FfIbK=a5afy4;$l58qOex zSgSoa*qg17T>9@vj-V*Osv{rwQn4Re7S}|cBqA`4=FfkfIRnRbkBZpQp3A?m)lXdl ztbcwxT@@OpKMT8nhaePe#R}y7^Z!HBK`%d=r1$$@S$f#WwNHrExL-!I?@ z_8~Ywv|hnNQLCDOp%x`J7KazLM3^gN8_2pig6U$0m)GTy@t(Mi?Jx4U9`+UW`)?pm zcnN-IajMIZ9vF?k&_alXf)+bzrhje|t~3qHC}_UrDy}&p9EgtJI-e}2XZ*R>|A-as zmk`0vJ(r$G^ z9%*-@4fpr1A#O8(2kA<_a5v=>Ht=Bh06r3aL=IeDlW%`s2-xmJuC_kELL3m>+!4cN zw>(@XY}IoHanx_RTmshb!D&Ik^f7-f>5aaHtEHw=QHc+?6@`X{{Kfm|$u(3yAY((o z{z9-huN1I&ECS=hoWt`s-_pldXYu^^mX2TqaEYIC9wp9UAjUx%6@cM}TG;3Aescl*ndgJFRV&vU4jI16Z?()sVVn@LPg zXvf|stjzli0-;Ej4AWCtB+yxHP=Wli=i-Jkbis;v^vg3Vmy$YxH;pkmO zsy~~V9VC$2Pvjpp8aLdng8mdRXDt&yV%t47?*X?kud#!Ni&o%vZD6<#}HuxTx7Bc zxPPYLFq!_eCi%Hc=r#f|m#0LKgiygdnep zehI$*)l`PP-|kSqe@iwN;zzlr%$q;!@8`b#`dtNP!%kATEA=>P9cr`;X6Ape6N z!UfPK$^nQQM>r2`srQ}9>!P1YAi$DNg{2COsH4f~O8hbnk|ZLqz04!RGHBgX+&lyh zaVZ`d>oBmdY`=_F4OKfB+d!ipBVLa|>m##KK1rIr%T}p?Ch)i3tKdS6;Pp^OlrYBe zK}P%G{)UR?!N&*t-H-^HAnYQ<9|S!ChrmQN$-PoAb#?N(EV(;FMgva!h%l{8MQdRT zq^kGt?IqtCv51Qq9{Rq!@ddoK__wB@7U3y7%Z9vYb8%Qf@e&U=C3&zBKk({t=U`am z{5(6s{VK>14n#s1U8l9ucpnsqKy@n_0Rww{gXh4Lq2a&b`-AZwG%J9Sqx~hxrKb!G z(VwC(eL$xPq4pjs?W|0E0apDpV5sVWzwY<>TZwkh$^ZhRsNIaWN-2Zx!w3vyu>t#< z>SX5*j%N=J?tXq>l8V@5W^cEEtd}bBqDMK>I}8<@JB9etR@sdSUo@QCK(^)P$LmW! z@Ys?};O*WC+=*Mbi*!^wxJoL4(h#c`_KQh~0lup3lh%MRDX&1p!5JxTNmi}fClJNL zGQ7w2PwcfibEzEm#Vhq2s(eBM9+M0E#hxJ-8j829qnL z!8kOxof0$raJ?X$FfMc0FM+2|@`PC55>Q-zV9^-?NjuB4iQ9;n2L* zhbo{O%u82^+DF)i^b6(L-JN`05bo!#6Y9QUcJMmIptq46+3p=Ji!F=B=8X5|){43B zrZhPYjbanTl?giR1*Ewu2l?JxK23@-TJLXxzlKOK&ojM|)Z1F|)4qtzxMJM7b5K;I z5Fg%5VVR1gEIloj!N<~bQ*~6arY|22I=)93D_zj*ULM~sU6aiM?7c@iNLzH7@qMMW zW}IsvOKiYdjn$9iJ9%y>sgdR!?a{pfI%dceO57 z?kAjD8G2#W)!L-4T$>W`9qR{5j8R5$4a?JT^%bzsSt=&W=$B_bITeCz()Z7XA%+N6 z0IEmsK~sD z-F56rO}8`kl)azxB-HLrDD8ys4?e?MATd-VYPMVyET*yBYd%V(Jf`Z|xX0~-bU)q}$3Ne@OTygxU;I54r#Zk)tJ0dYA+X?3Iyt<;8W6+35{<1Lxp; zc5%V=IJ*`CL8ze{pQ~`P{EarHyUYqj$^@)nWSXK6 zX&~9E-%arYN-Pp$j5Ek7UVfQHH?_H$V{bwOO9b(})(rR_z}`W!a)!##HDY7qO1$*K zlW9RUtC*8B(1`B_aAs~U+u;7!*5Zt?>w^kx$upO4EoKUq9C=~q7dR@xH+MSZ6o1QG zk|`EXfT~2}xcmFX=37;FU2iN#uhB@}UsuXe&gv9VP3E986*%ugolAfrT)=*MQ9aBx zGKk|%8vW$T%kRfe^;_2gz-M@yoef-cvOb3+D&R?JU+3mC!4HA7>vroAA@%f9^C6kq z#XgRVM&8Aq#GaGIwytyF)_HQ?=OH6vpPYag(pf|-U*ku0*E3x%ae7QZ5cCmjvkC1L zT^UVf#8*E=hsuc3HGZ0YRXdHU>fkJ3zMVlh0ZbOvt7c3=q+dd8UNC#znZ|Ez6Q_JD z*dzM#40xc2Qri&>MX{GHJQO?QovA9zg z8O|Mlc;;dh^UpnAjGd_-^Esx#s-cX2ah(3LM%?5R;ysOI!P9a%C>+0&uK?84FyNQ*@oru+eF0xk6?nt&u&B-Ld|gV0VGS8*)GEZ);Nkd`YN2ZCUbl zn}zLnJLo20ciqT1D96PY39sD6Y1w1W)j?Tfn92L>oxT}OtRx?N%r;YmV6GJ8)8_W+ zyZUnJHHP+LirA-!b+|r{Vfv4Qjbe-*$JN})m$-uS0n(B18B^EX#ch)`@1TAF1H*8H z{S$5n{(BQp#JhgMiDlPi|D-&SbHE^rWEg2Q&s#X3%?!3XHJ9M&xjRqF_kUzpp^Ii4oIk6?XW%z21z`WyZF4t$Bp{yMq$hT zclW||aIQ<=UwpUkS$jC+5p?`WOGUzbr2FI9#_8QRp6!=j9XY-lQhJWw9W;_PUj}0K z=d%;-osaTQxlXMTUR?WNnig;`REp&T1G_CL2{9H4yDlp4AiHBLL~B(K`;J2TDSl_Q zoTwiD$n5txUP2Bja($|USKO$nEwc*xPFO@xZ{?6X0dnb=-==?{1=HwEp zkMHo_pW}@0U}@kRBU1dfPoEI?K$jP5MPUU4YVtw-a5|3}5)t;7Y#q4M^XREt398>8 zc=hIh|EXu+voM;%622iOgL?-JbDcA_t(O8FFZSV6Gc#(5Kjd#WpJ|J*PxHkh;hTA} zq0Wec4Kb^T#@DVjsy#D6^S=J@VE@8)YvFyS);t_iAKF+ z7`gl6&NJfV=Lmf~vP(gyq}C__)Y7o<3De7RiBK(ZI=9}y}d zoKz3_D`%X1m4cqLQfiQMbKAGSdOv2|Yi?aWL78K)-gRYUWJr5a>EV%HmGrHtV{P>0 z52e_(aUTj04$1Q9P9+Gg`NR;sa5iTDYBwDBrun3eiD}Fz@Kut}i`!D{EViUVvKXHj zwP@(|a{rqu`PFTH5kg(5K->UtTrGTzALp#B>)8>q=Gnkn{u*oIEPSepHC$$1_mKMi zwHiI4dKS*H{&YT@@pOA&h0SO5cAgHF)`H5YrS&dPP2k(7JKKA0l@*ukew-m4aasYi6;OV`GefYo}h*mO}U0*yQTkeHq&W>WmkrPnSIQdu=V zF?1G%#iR|Zz0eY)I(4+#>{x%~rHl-VtH$I0PCTlthYwnKbbo&K7+Lu2EQa0OJ$`{N zWGz~<2nzylA-Ga^<3Q>3Y#TxGt*)atLNV^vc&j^$Pi8{0 ze5fu2$qx^9=HPj~?iRVPT)x^U7_Mq6f7?JjJ47hMSN4LD9{R}-ExPtW54Ptz?Mt(} z=l0H0&IyvpGST`w?k%`?T&cCXA*pyVa-_w&XU+#-{h#}(8sofQ&nvIu4 zU6jfALh7!ydU8J11>u7wmRxyDx)NtDJxSp!mN9~R^oCP(%FUF+Myr!)RFuuE(pKBunQ_?&M?Qs1Skj$D7!ilQ`X-r)GVrc&yu_s6-WUtQ{Ss zUNDN&PZb18zZvHoYQ2FSvpF;Mo`(TfT1%AIu+eqfNw7=HiZxD<%A4v^MmGCN*PH6X zPF7RBSE>o#Gtg>Y4&|WJ74D%fAWwSo{^~{x zlcgtEa4ke9Jfa@mi|HYv^(rRt&^TVqbU)Df%Hfoa+>CgkdXnT;#$%@snP#fAihQ(; zLfqPnfz{KokhQwOV8 z^n8u<*^u#f58ef)R4*=&ThZACf1aK{S+7vt)4;m*B| zbYGKq^D8gKis1HYTUZ}W@y9Ni>OJ0g3Y<_XMl;eJB?aFS8@@luIn(RHJXnx|C3vhr zedf+<0q^F9hr#PG`{HwoNX^=GQL_ng=Q(xBXu5uvGdm=`e2jUaXD^LTxL>ppjRR8g z5QoXQF?QLces6_2v>ff}6q$Q3r-wQ7$^YJ#LZw8t34KS)MF}X2WfvwDf|R^_oU`#> z_MPq*n_Yj$KhKh?s0za89f56zUd=3}s$>Bu`Yyz~qF?IZ8kP42IOnY^S6_%`jGZA^Ql-F^ZwpYI2n z&ri6eZ>hXyIH}34vG|0oQ>5qF&4Z;D!NDp;ZHX_8`N|Q(C;hcNEHxq*-uB^jM)PVP zx=k1rI)A=qOsgx3`d}AIdX-!6C2y2~+_iKa2{7|mp$?wK6G5-4zB!b}+ws(jPhVbe zk%S99nm2pa5iNN#SbMX!RpRHd7)^tb>nLi;cR0kWhuNw24b0=6j>e|W?#LgCdHdPzr4wJR}5>e`RDS2!2jn@KUsLE z@m+~gE+hx~_YVgg42pr5FZ}m6fByshB`glD#oe1FKc561=Fgu@W_mwPmeXJ09{%SK zW)ljHtoe+Ff3J@Gb9f}&SM<$FnZI}Q>o?$@q+?RTwuAm$`uA_36;T?w8tHu}|Mv&5 zNaWSv5dVLBh-elJJ#I(&hy2RFPZCW7k2rg~Rm+$pKxz`456=T?7#yLY5Tt_ca40MB z?x6bhjF>%=aA|GH2jbVTZ;;eLBfJ}im{4-rWy>OEy$4t|bwTeoLSPZFgj)#E^E!;q zVZc-^Fpv56!d~<%3w4&?Z#At4n=)@BLcBB+R|uvUBXeU|wAr-hny?YSZGbl|kOo4M zK?e95zFDq%Zs*lk4UzD2%-NbQI(m$){FEbHE|d%U-TB{^z696To*Z_s+^>3_pQ^zxlD-Yzf9ozsQePf&s&aQE(OWC>6pE z@|JmkS5?4gsf$9_Afhv$o$v2T$h`#@$=_T5JhLI^jXsLF|1E-5w%Ba6rn36;bw05Xb6x5C{2Qgp%8TF1pyN} zuNSz^U{hS1&2WDYgr^b0z`Pvo5V9RE)ky$%y;JE#z@FZlgK`ZoO$(X%t~m z8iDY-YX%T3h!B|Y+x#P}ucZCVgEp&SBG>8Y9P^g?d~h#5J$DTN5f*(?JFoKhn1mBI8`8L%n?p$6NqUSZE9d9Rvd*_z8wJ|eAK zL7*nNfSkbmI!@N#t6=ik;8iardLNkyYDPSf0+K{-yb3U|h!;*3#itjZxHP-^fKKx=}UK?lgho`*ce-f?>`99IEV|yB+vX_mjyT8$`Da(4yfniihhM_KY=ZM!D2*qPe$B*#THQlRJ7o8 z)qI3KT_XmN)%aa4C7!D2v23q@*XjsKxF6>QBSjw6dE71-r~^JTU4&;Z4p##G>viM( zA9{dapIEsUFP#HO{#^x?m41AC{6+$V?y0me;OEa1_@sql0H$y7-UTT!*V`vJBq-yP1hG z%-BKi2$VHQ>_{=rjBBvM5|CNln=t$x1P|hJ1$R9uI60BoHi<~XCRwRZRTd}szqgK# zCxyVAO{F@>%ZmXHY9hpoz9fj9az*mIA^x+YYPXW><^65?3Pc;7O=ZDf;pg($okd%zJ=*Rbg?q%_WY#SbLfQyfo`)|k{ zMC!`(hIYv|4R~FNBE3m>rUUIK@_;Ot{b6@d48zpZJu3g${=aXq z*$Ylz4JBd>H*OLr2Q2}s6ow$%JDvNT#ST$VLA+Nb=N!sG4EieSCjjN$EvR%xVB6m( z6FwvCEg#-&RhA}f_;!}a#5q9_fW8B^;djbdxfPqjvPT{ugsc!{bn8cUIY?bZ-#6{m zk!17A-TNDQz(N4lMUqXPU^^35`XrF>I+r1qX1mPycyM#<7rKxa5IuE1z!)EZBymc9zv4AHqC!O_vSRW->s*Vh#{|^2LVg=oZ&Uz?| z5#0)V=1IlH98Nrc9K;2b?jADpJ@Q#LbUafOdFpKFKB2G({s1 zAB$P%nsrtZmsZm1HJhSY=~qP1mb(a6QdvRhEz-ZAHyuelgf=;A2Jnan9b&X~_<;dz znWsBABi?WP%~wry3LEKDB*>f;dsSLm_HWdjG=`{oXz|p6F`OP9ePJj$;hd?D^>C@_ zjeB_y+H<5^#nhnxUQ2NnCItWAze)_hO1{il8I6Y=dUIfR%kR0j$&a1Eeq8El$KZ5h zdrvhO+08)*@b=SzE00gda0fp78w}Cf$hM4$Gh|B+ZAMxHl+S3zU2oQzvs4lJEy1dC z@!HKFhSB8~$4;KPsVaHky3sc)S9%^>nI+~N$_ zl7gCH;;9XgL#+fsw63L~c>6-o$M4DW7{17o>2O5VqWr&u--`&=B z9yUoWYyS6+(B`$Nm))?9ZoYm@qkG$pNcWT)9!7~x8kHy}DSH&nr;rM-WU+FDiY$l9USj{u-XL`U4 z(>7&j-Q0&}R!?0HJ%s|l^{*NUVmTTI*>5kXbaP@;jwJs8>isA}Q|FmUlyF;Im3u0^ zUfb*-fDdzA`N3mZ7{YswrfO7HpYY$ifk9Ooxtv=aOZ5Vzrz9WL$4hwD#lw1scBh;` zHs2OD=x)w)6_!d%F&@b2!@!W+yyA>foa%)?dnCUqB6S{!LfV}%Lmor)Ye1?_K>9q0 zoEE%R%U(A+A6{K(CO6e7oPtU0d{nv95RQme``!P!2|eT{w)>Y}^^=HB`I0FE{izGO zO#Lie&pjaq6{#0@U9_;-=N! z;e3yuN9%~qaOqPsQ2vrtfKD+U%pS71#q#_A1|8x!SnV~L(Z!6o-$7Dhe~`H{@jMZl zpW3g){M1l>w_sfS+#5PCLI>W39k4s^R|f0R`(SxJ7cO&>AK?Z|b?Tb_y`H)}vSJmk zTTmn)*~r#wa|c9Su0%Lx%3vvq;p z9DB()`M;dO8&?)D`p5KzKEZwm5_xHpws3|y$TqtN z>~A9b5E)7O@N?Eb0h-ef;RiUgxSdtNBQ{;h%?K5M`Lr0*FL}3!hthv`JzXyu98r>j zbH{{L`N?EDQTgRT=yRG5!EU^({RL_pS9@Q?c3K}SSG$pL8hh;)#|AJR;=wPaUgv}4 z_qB%}pwBz87`Tl%E78n(6qUU`eZHSm{^7-z@|C*J2~|Vaxav7(cVKk=2$FRvCkpde zB-pwmkrLUgk6h$|nzJ4lTq;jHX-J5eY){lmg22-fMk}}G`gD;0PDFf2*&lz&g_9}wwn?Q3T#E5Nfj-thKFmeEH2LZt)f z$mf0>uKpOmtBXx71`(ZbNKAbFYCKR@kowL8mgtF%1t1cBA@70g>UGkCy~Ri0>(lN- z3BJKQvv2B8YZ+_&z@q{<7W)&_k|2Y*w}H4IDaWaxmVSWVzE*uRFH~kR0YbwIQe^&f zwqoqgmzqI=J5R-2JgyX>ennVCU)VvUwZ!?CSMW+tLeo0wQy~uAkLn5md*7Kdb%xfH z$ti?HB)E7vs(M=pp|309JePiX$QZ6GCJT>?a2IOrcO3jkWr|3IZ=La)0Lv+4Kns*| z&3>I%NT8-q7Ml)W>I$J2?0OYQHu1pgtw1NYnO+Xzi=bzqqD}Gcd zxeJ9LqumI1wsLa}a43cB>h7;16M+eX^8c);d_TDDTebb#cT)5ibo5MdTt0ZrYY&c8 zl&-*Vfa{U?6jVN;l9c4dCptQSbX49bQ35ta6$q__vZca*VhdN+dA*(O!j6Q2WJXI; zou;Z->$umCgYyP62xC2xx9k;ZpUl$1JvbhE_F}$xd=jR<0(24*G7-C^?@N)#Y%>43 zl{qAk7ME|-L*Z<|7GEo_Czr3H$bU}m+i4+(592(0&%wO@)M1V42sYse2qn7#qp(N( zUl7(%MZY3f?Q+?48odV`k|BxJZvRpsdbD}t`)4;gTG^)d@p`^4WQ2)lQMFou{^NEB z%PxHO{ho<6sI1Qc2_@^rXp}M-o9k1RMb;YnW`{9F@51`4@vsC7fbJT~G2ncCnKSzu zC|3fkub%RS&X!hU8>U!w?&caNXTz7I*q92`^fCFfAP$o+`W#+tgm~%H0s=(Kw1k;{ z4keGch*~r8m>q;0%kAZnvN7UfJ9d*UuJjDLloGPr ztFru3qoX)sns~A0kRaEISzQg*%{N=BORlx+nvz}^cS)u+zVEcN58qr6m~YOsUeete zr^?x#^>E9};mD{pA?l)Iy2ThIEzWS~cyTyXRYeT+xEL`<0#8WVHP2pN3x)tP!NGwH ze~t&{ucq>MEZ^BH95PQ4ZfDXBzZO7==ysWpDV=f21E6a)6tO6=!Om;bDuGbZUkJkO zp2tpbW!V`|Cv9^|qw}@j)*)mQ%$HKC%C0&}%L{v$6WAof@}5vs-zUC6 zW0QKCpvHOu{RN(aNe<*gEk}V};HXz`By)<+G_( z$ov#<%X2&MW1NArf=NuuD~D4rx~@f&j-?yJ4EptE&f36kovZlBnE0JBPBu*PbxH#$ z&;t*I*OW}=<8k2BEE%&bdVr}SJnpzz4n@(i?;e4gB>*qf)m@wIqodTrX}5d+Df9yA z*s_6wZi@Ew_%p;`q6|V?$+$!Bcu&02nW$jCA;qPo$l&^D|DP2QdY2R+-X3i&3q56& ztnPD&_m98W4mwM6S~lPt3da()*a~X&CS_?*GHyCeIDUGm;WwBMSQT|a*Tphq)-Pfn z-@|B4K}~fwRgErkq8HO)9>2Cvuz2=#^1Sc81f-{YbN;p?EnyzKNA|W25Dh2N&V5! z>Y~h-OqcQs;zYOBd3{X_G9%vm z9hx_u_QC1$Dsj*5_>~f;Eki!c`CKU*KZg;W^9?k9{I|_N0@YhgyF7n*}T61AjdXfe1}e)O29m2nKxlD?9ZdoQ4h zG)pmdOIYlpG%b8|#;Of-9DjuU2dE-xJ^Lb$qYUJg3E z3LvqEt^-BU?0Z_kOiC`ahTRM>Z70Vg@jEJl&Me-V9yV9x<4tS>pk`VkUeiyYb-Wf} zUzTWg16CJ~na6J*Wvb6c7Oo>TAZDF0O-yUntwC!{2zzg5b#*SVbX%*6z_fkQJY78| z*u+!nDLNuurq>|Vyv3lPH0z-d5Y;LPyj0h)xG%X5ldMarqcqwEr z^T7ov(gpeV+UfWdeeC)&OE6{D8N)g)+mX?zM_>D)9lI69@an#Bi_?(Z08XPxC^f@e z9+i)V&vfL{-Ybrr;w{Ei_G(9u2XuD6Td!~DM8Ub{Bw+ z{0}z9SFH|V(bk=Rgqx^BUV4k_a%Y)ibd+@Kd7JT?vWkF7RnT%iuQkE#vk6=usCM{x zts^A)VT_~y2JHl*#Pn0H-taW2Y`SUm{jw7pw>bPU@d86=KUw!X_q~RyM+fXj*fnd0 zcj+md3OOq8jiXFsl+|k--!s+tx!Y_mE?;EY0G{Nx)bkQwzVqgYj(Au<643kH*;C+b zUB924tf~F!RLdZl6MyJHm66jr>xqY4!19bEXnCJrv zL1+FYGVyFzdnYeO#e|tAY{z{r5x<}WGyOy%>qr+(Ci3ltz6=`!jdT?9{a3E}^T_I% zKDj45aG@kRO)rva#aLKfx`n07Dqy<#{snU0p;znckZC?IkxdWAw;+>?tp7Yc=hu4N z13XfaEJEacyWC`+5IyqQLVht-uD*_}1OVbLeZw8Q=7ov+COqPpDtR0U zADVeHih$xp+QW z-jbryJ2%LH<`7xU-ga$>!@%%XvjMHG-bOQL0G4vCM6|=aC610%haMjz?`SjG%Bklc z^w617pDj-fn~>QjR^2!@WGG0VOsniPhQJuj#j~@|NNOYIeGEIfN@GQ-Oy53QBC>sZ zTDRBW`=?PSRUNS&x!WDTdTMFPctU9(w*A!E%J}%)&k)!u%BG#dc8p=Xd&;K!+;O>e zK1?7PTb-NtslFY1%O5K|sSbK}rM#vGCJh{nQ>LD4v1ik#O#MtIo+^`^!25=7CLMp_ z7*%1~_<81;uJPTAx65K>sS;WlAL8^7#8R&|E~HlLJrXrGI8#AlaHt2=33Q?A;;Sn~ zW+|s;_ix_oGb`OT;Wv|-ty2`<>c}QslKUXx%-B`C67OSWu6gYJaV`piGh@WtBWLm} zf9)*_Sv0Q~g)4u-S1re{84m-bS@IN)#7P=N^vl&0c*t5GG6vDDSk=6+s~Xg?{}85| zoy#AD!+SG6nxntGl!sw7-s%do6#4m8o!8LU#kWWLx}^fG4kgN8CkR$TMmROkbUF>S z#9obOEONJC7V_k@2y&IRDhv~lJ3BHEyy@iEOl1^5ed@`)Mm(chG10~g^^3ooMT1JP z>$i;d%Oy8JH1wkT;}bv?65Tbo*`hy@;rgrgIt}6U+J7#p76uh_I>Ga9ap|G%f*v$(@f%=dU*~iLFt-h7?MF7&$vGoWOasz(%q?cHrCyl(1OD=Pb zrN>U%e0~)#qWO$1;bG92yqyKL_@#-w4D6AG+jEeZx$>rVrMA!l@vfB~eU6+%tChY) z0D-d7T#-XN+2(1&D|+;+IzaB9SGntb*N#jpjDR{jSyyEy;I&=|`K`|Aop8lrG`KAF!V)5qcWU20*VCj2Lj784b5~RD- zlF{V{lcxL;Frf={KZfLJeFEtdXyXmlBYr{Q%h1d;{m`j_^=lPoDZsb3HliBj#6!@z zQ_tV#QiCdkA*q5{5H2WwFQ(9Z5+&~)8E162y_7fHkZ`DqWQZX$0Nse^7)By-0L`6F z_p2ShdY);r2<57eAmi2`bF)v;Py$B7X ziU~S%1Y{)w2m3oBdIroRB)K!qGIu8DcNV zr6KjwteSoHR!i=~pP$qgdE7B2U0RAa9ga{UB%Y)dukjEVbQ-F?q(q*ts6Q@3BS64Y z87>B9GTnfXewow$0%gh^2oB#e7LK5`M)k14oYqVRTch^L>y|88YFRcH%I z!PhJ?;03^7r@jn{aKH}g4gd|R$)Lb*^mU2gH;W*-MuXZce~gI%p(GRhn;HmS3DLrc zd;Av{QlbNBSujkm%ap;pLH$4YkDF0FI2Hz&5c%4#+K8`nUj zCPc0N@kj;lA>{J
hua1e@+R67qid-A||Z{muXtG;OwdP21RXff{;02*Hal=*!Q zXyZ-6Zp;IFJKy4h5xc{13zWoj}!fcMtzE?07b>Z?%zVY=IU{=wn)B_s z_(**uT~XF|g5!V#jv0)H`sXx;99MGyoJBZ!z+V9sVJB<={Ol zB*9O;7ax-yF4zmD46c2`Q(*e&!BT+;(o^k&i|~u;-|Ffpzx}vL?*OoZWXLcFm2r~U zb>cu^h0~C#v^v?}_IM0ExisOrRV z5K>l@Wor;4q1Ga+AlV98B|tr6vgvYXkTp5$95bXNIR?D&1cw8;E)EPNNURZtrt!X0 zkxGeBkodP7@_p(Xll&e^$#dAjwM}P#3HLY=aMC zv>gs` z-~n)a)*F{#K2N$!M)0(mQ>fi)jzyV@65!mc(8SpdLk@xIUIGB^NTb}I{V8HMnM3HC z>s*#^pt5!JfjT|C`2pLfH0@BEcBGdSkei#_U#{d=aWIw;zb@So` z*4TxrAwHqz2FP}F)W_;*Yr3yk((gUfhms6k{;=M?cMABsr_V!s`uQnG@3bPghv@%Q>#4=ntC95YC-$cX`weG`TJIh54KVCkm;C^5kR{Mtu(2aCWVGTVJ-BUI zF8@veTx)BUpoIzSb(kPuk)p&{s`{6~MAs_WQwHO6Mjjp6C* zb?;194FKRjPBimuEwvR2buM`g9Tw0|hu-pJO8>rDn|8VWvxPnl1b=Gr9FSgn(NX74 zm?s(cR{>9>WDZd|@{KZm*?u3O0~%DD_SfC0F$ZL&^Ai>U?Y32g2_WvJu|^1V2P23D0~a%l|2W1ynoEC<_^{1zJ zeT!_V03`*~Cb8B|!8Xb}-e!&&(Xwhv*rYDc>{O2_{3VnjIuCeyxqf-wd9-b1`r0c( zIsMMtQSxcRPtTKBrT?8CSr9jPBCB*sp#LWojUC>FwfTh{;SA5t_-71k^Tv8eQ6mV6)=`_9w<0{w?M8+GzUH>VI)HHr@D|F^8cwSr4vZ`f%E+&IM1QVH$6spE}3%v-SD%L<# zp%W=_jV$Bcx!2vg^H;z=Ti?nTRXg%SO@g&2j(7OympILYBl{4!37~-@x8}n+cffb`1LVO*&}qw9lgbSw zqK8vNAJ%&4%-q~2_lvUc9&?{kpV?WZPu|U^FWFh$p2PY7{<846Ml$ZfO0J)!cCkE1@R~;Lo_g)zl@xxu$sGqB z3XDf9Zi32}{!4d&DS8M}SwoI_B0P4mI$>InVwgeALoeK$|6 zN<(Q#-P6co6pCux8!)x>*Et{LryO}Le?7~;9{T2^BkDCO;!v#!JCe^pKpI!G5dx*$ z^O@gx)F;^-b&&8jCQ6ebUMV$M7P$CW_T>6%{s79X!xW@~*=_r1`PVL&$v#butVP$B zVm7LYsd>T{1}e2|Gba+XOi60jl)vgV0ghLegrbkDazL{TcO_mNSStqr;FdbefF>&Y z-@KXby`~g*$H(@AM?Y#5vLa$%)#1h!Mt`EHQW53+3??^Ne77;2*-fgDy=(Ex1$={# z5L-CC+aB4#rAZwcFSm-Zwxe^yfAIK#X{j!QbD}75Jvz4OAMBe6U8V3hSy;?KOn*QZ z8%Yp^ob_(?CmrKIDG`dFwyz!>T>?S7T?Ct7OSDa|dU5OQfFF7BY9V(7EV}i`niNzQN!5^UR@*(>#Ku~^%bgrfy7*8DIOCZJU1!L(W0h292 zLpz52LH!^28jmSaqN_*j7>O=UULQoqu7M%J#d00eMwqx%YT7LKn=v=wg{W#~j?sz- zpWyYWh|Z8YIMh3CE@xRbgW!7wA5lQ6$<2PVn%8~_esrs-LCr1+@5l>D>qZbs*bi>f zj7Qd)qZ2xiJw6uI>klB**=Q`Kg{*>ofvuGw?A@_s~?sE0D2|C??;qUHx}qK4MVf z^e<}UDm=48<_v_o33$Qo9}=VUF8C$5gId<;xB1dcNJtSy-<#m{RbOQGRyhSjZ6dgO zOimqauYpbBFD*}9{QA^n33PdH0X(&=hF#}p2hP2}y9*B@((3}^Q?q7oa7|?sVGEcA zfe6snWF|@#$gC3g2SA|k&mNi?(Lca8Zu431Cpfvr%5)oyUsac;!I01FfIKW2#0rdq zAa11s_=IF28XEa7>(@r^#=@26e2Z*)jS<{ZCGcbG4kv;n(OF?&bca}bhc;G7LG>hm z9k{?*M!sn1)qshWvkFL}>TS>KgoV-wAhP4>?@%85B&*U!`%Af^xH|sN;-7)6>QP6! zBupkwybYga4TV#J<%it8CuI1;ee4bW;8`U~!s;;yHa0mPO#xTFWVdGz*;v?20940WoUTX4>ncIl>K;h`s6h|%5`kF^QkiWfd+0?dPP{nQ*7rA6u*IFA zH5y(6_G&I@Fmn*I1k$?gMd{fW{)S9j@9Z!V9j*Ym^95untd_6~=PL>yN;&{YXN_&h zm}K=RDVEI(=xY^Q%Qy}4!!F)l;roozLSO2MLX)HAc!iNPDC|hjQsvZ|hUbxihu)fw zPb(K9UWC(FqqnYPw zxQo{VjIFBEv?c_E2kRyTG7vp8^dm@kzC(1pDX6Pe_+bIUQ$Q)S?)>@3k~=*F$0IjX zNqym?9$I-Y`06;0>%@Koqb2FupvkrH;2Fi>Di1T!dgqC8O+ zO<_cx0HJu~3#!{f4i%VBSFizMk)nNlR6(sqT56+ZH(1kbg1`L1@`97U++qUCUSGG6 zn{t|ytt%^`mrnFxTl;hszmnDpImsdA{Gn)H-U=k$Js{H9`iQlDl!RYfCZpj2D;Wzv zoht+ltRx9o2uXN`I;?YCnl1h~vk;68$#_hGp2ZZ{^&f0w90zN<1Vve`xo{p0Wqs>` zf(bwf$Fh!jVwE@}l#&4TlxsEZwEuKpjnn7{=dcwR3pDaOcN&ieYXqk`nogqM3?GAQ z(GzfWZh!(GI<$z_g#Vt}|3fHG1J>PQXXwOFspVpWr)VAkOPz;+^$V#3f*rDlS${y# z7{hCIwW@UGTZj#n7thD@l^6rTgMKP9q?x-1ZlyI7_Uc*C-Kk;|G&ET9C@WR>fyRLz zBtk05Z$yM>xv9HASp$`uI1>gyf>y`}5MckVK@_b~ zGpTZ%^oZHZ7B5HGyP*its{(3OQB+D32TH0r(uLB{@{?)}Q67}CylVQ`W(ZBH%O6=% zWSHh$Ap;`&ttrl55kB5(8oGSDb;}<^PQ$!0=Qv>Y$g-7YFhXCsc9EXO8kG>XSJEkG z?UZ}o@LcdFxen4tx{84vywNg0X;AU{H;yJ4{d9$>1h+uc_m@#pDqU;n3Ibo`D)6CC zUYB<~@Hgv#)GHqX57-v9mzR?^DZ;OtU!KZkEfn^it;cRPpIr0I`-!HC{)Lm6U^~xi z2MyPykZ#n9rifna$3?a3CAMB=Km2AyGKDhlqb54XJR5GKzwS@3@JnG06RHiFS|qhj z-mQKB#dZkY0?h};P9OIafGXx~glLik*r$nRN1QfGNOi;JCvfpKi(l%;)wMR`-zUvw zEl1VUpNxBzQ)TesJ}DVZIjY71$Y()v<`sjomhr9&(YmJ!^n5UK$v|u?f4L7b?)B@q zff~$o4mo`zp@bl#OGARb3=l%ae)mYIcQH`S`~GZQnA+=E9L$Rl)`?!cP)e}{dI(y& zKf9zrCX-Z-{?>-nz1^xx<~2=f4IvDgIxVH$Yh{{cjf4!?!*JN)cn9WSIm71{$GYM4OwJJVW^_`o?VMbZ^2Yu6`0o(wOwCpZvd zfmA?!30JQPH$hOcTSDlWSzP8Kh+{>5AIJ;`Lo%WeGR~AwkU2k}=d40w!+D7ciKY}! zIeY%IzWB$w4U7hr{Eq^uPb=BgebkN+2b0`)AGwz8sIFfDbq$hpct*o7Cci66P#&?D zC^v+hz%(HCcAfVXZbPN(!jY!uUR`JA(@VqB*FH5W^fL+MoG3fu-oF!L5MzVf;_X|8 z*Q%nxx%^}b@Vobmhv7!Mi_-djVX%qH#GU|tn>a=nuhBCr?95io4|V5~^=<-V*EC~1 zL8v*b$$vKYWY6*&9!_&W(h46#JbnD)!RP_B=0Sh_OY!ZiWx+IO{8=Zm4Sa1e2MyNk ze7)yw`m~#k5?%X{{(#&@djE3s6R+de=&MRh+EgWcemzgUosFixUoHik@S{=mkQ!T< zb0>GjJAiYc?U0pr?gRL5(n;@6fU!vxM~Tu%UFAe@2GGIo<&y2)1G-Y73|gJ=X!(av zl8>Le)?^Qt@%)Y;75-3u_oxxXpX><+5U}9e$Ho>&_~S}GM5TUB;y>iBqlkVP!NVwp|$0LY0P zmR(MCHj%Zr)$~hEf_R;qLxeOM%F<@z_ei1yn7R_)&%7xVwt44(C43+*EatT{U*uUCl&f<%dZS5p>WIO6!YKFwL ziO`mn+tT{FB@ksT;RFc3F~Gck`Ds2aaaR8^;rHfp^d*q0^8F`Y$`fmO%kos$ z1lSYFIou?ks4(nsA|k#C<;#oT5vj!}-Bn?yVVb{Q9C0A%*d$&=urlp^Kt^+$kg`QN zQNHS#WGQDSotwh`C2AXUygc?52v>|j7Vc%+&{-Ocaxo_aR}G+d$COm< zR15M(fCVoNIlr7Ay;tp!BRYg!O!suM#P%}xmRuB8+{zcb{@@t7lyAJwuASh?i&yi= znyXz?2*|a(5{{lmMP#L?-?8N$5Lt5Q&{|d$*@@w5_FxJU3+Wr$BCU&xXARxjUYmKP zoJ;*b3%ter?~TBPCTiJQvbWWqDjc_Kb+|{BGLGugQE7GP{nNp{kdmH%TG_4REf|?C z>6VZ-84s)Or@IQVY_@MjGakR(`D_|yH2Npw{-FC8pxPvHzfJOz2t>`~>Ay*`Czz0_ zQuk|AO}{jqXy2H)fr`sRYk1dk$nXhg-qmAY|G=jg>ki_Ne3g}2qPp3!sWcH~o_xB? zr&iZbPAl-hkI0}RdvPh<>8crM!33lxv~zVDYOTYq$@6SYmmC&7ea9g8*TwTF@gxaX z@uR~~HH1vHb1%Ll9kw`J#cN6 zDJ#LrtNfgUx&7O}H;SymXh+PvPT=*899cs-awho=9KYxM_9RafArvwACAbm9If>cF zN7AHOY7q^{S{LTVb351h& z6|~=h0Bm<`^fI2=J~d5NjN)gapIWKp9@HM&%oT8`6EQ?jYAB5=X>ZKsMVm7QIqKmg zg7%N14&2!o^7JW^D_jM$`DgQK(ipwy7K_byG}>;{34tO9M# zCrk1GW6!Vr(~!?uSnj>PVmdwqw!`uq54=S0LBZXWK>^D#sZy#!hf#_LEy$k6>nBf5 z`0#LAuzA#r3O>zppI<&8eZR;yCb_#JD#x&7H1#&&T6X?vGV>qkF4h=Iar7xIEzyFUcOBzVMW zKb)96v&2q+R|WEI-B-(a@n&-fQ1Yt=1&6q9XE26nzeZm%*uy zQ(9*vV6r+2L3)c-!%dLi$Rf~@H!&BV9k3|-7hIP^pCpI(>V8>kKmi{ zbfu>P5Lb^Iz-gh&m~yb~_@R*FIx#=GHM{LLp)Y8&A6#Rzy4lO5>-eD=y z7do5nxXY=R)|%gJUw6{S*5lbfIuwo;q1Ewa`O{6+r)8G9I)f36Kg1YsroZsWPtZ0# zhe15~x9B1T>aU0qTKL;UzfQL{EWO3sJ~HN5?LDdR{8aAGesB1#eqtd<#v}L|pO<@C z=eWI9N|*cZt#F*R2;OX24wjGOEwtO+RN zd8V1m38{#rVMKzH;2KH=1{`s~rQguJ9pwtn%wI?nY^&uUZgHN-3rUycWHd|Laq=T)XBiBwH~ z`RKaTG8`>e;e_FAI)i{+Cg#)#3nEjsP2TjWiz8bL`EDLnvs2a3e4HLs)_>cBnyKS= z(jumhOxYgy28pRW^;mmk|MNk&6D?bPo42}jN-hCLiz*{p8Esu#-_N2uC9EuS~2Sy=havlHQ{?Ob@`;u z653K8y)U=+WKR*4^|HmnwVJ43c;L)Ayc)Yhp>W<@e*ZH``YYpf1RHTZ7Nu!;ZE*{YrkKvX)scrj1(;P<@ zwny;;5mUcS-!b~>GtV+PxC)u&B$>?ncJ3%P%Jh4+OlvaJ2S3Z?TIRhf3s~71ojo46 z6fTQ-L=rIQcjLSWs$&oBuYx;pXZ~4?@2FXdQT}5*IpiEU$K;TLWFHgvyVed9Zf@Yw zZXYRAj;tJEn;evvtIItz(3n(UD;(4J)Mmz5$6$rb)KgSF<@YGp%A>_cJv3R>E|fT2 zDK1_1yN~=ultzdoE$T zcIuqUDH1=>+eyWCwARM1SEG<7G6--Ra-XVyA>p;<{|aAYY%i6r`1G7n`dn>m|GgkQ zxA&gE{~|U#0M}enr5;xK&PIK|3N@s{dLq|NwAoa$u;WkYQ-}VxDs`QFooWBC9??a2 z)!W&1-jwkDS4$;VYgG3DCPxpBUxxyjgE_U_M409}^4-TQ5&zYk7zKL=n$K6=!m(G{ zvhA+hZU)H1PiZqyqIT%~8C!3jP@-oXaQ^(c~Fwf*5XZ&;>{r0RvjY5xK8bnQ~ih z>Zv~;W$9H{ebxMX3SlIMQO>cf|D3nt?;xFp_GGqs4K*2`=$-*J4Vkqk0OT(B@GenW zE&;zh{8*Dz^qWpgHrdt2$D~1@y4QB=iYC3qiPs&d1XlWkDNpylEI(lS!GTaxfBD%n zxcF`1V0!R#M_`$;MTJF41?lwO$W7Q8HK$`(tHhx_!0Q@&5W)rJxVZvCA;u70`b_SN zI4N|s7Ka@{^rc>Tyhfg!zyeT*%=<>bFb^cA^n7X1K^eJC(PP+!cOt!D}6877F}qQs}A#TobkKLY4`@CZA0znMr2ZBdtpiYMzEN` zPh@Fbt=^r*gmP)^31~jd|8yO>b~{ImTsb24NAW*j4}Ro7_^*@vl4J3*m`T=~>5V3X zZJ~jvABMp;8-7VKA7BP^qJ~)r$R-eJbk&N=DaL_+eVySOk791VZZ`ThUCewC*{e>D zPuFpLwsg?Ztoa73SYSj}>sRn;tSwq~W?L#Bnx$(l8V-wr_xZiF`^ZDA4X{85HPf(K zi`U$&f4*$rfHmp$gGQdCR^`;16?(ZT9f?Wq!=&ML=UtrIjY3_GV!=$+4K2JQH-HN& zd@_`dG9B40zy+CTmO;p|j}YPuT2BW!9>}(mH4LVsO%A0L1M7I=Ykpu=G@B*^W_6H2 z8$tpC=NjaktvK>^;%E3?dw~C|B+xS8%Vip_yj5P5sd~C3^E`hBO@?6mWDdZ4?Cyu4 zpkT#ivkDPe2UG=;4EaVK09B2#Tt19Oi49Rx+0L%x7@_zlWU2=ty4nQltPLAA$-jDU zDuGk@2^W>GzboL;Yt}jlw;vVi;Pen6kaF}MFyi>A&;tN7*cCC3)2Bt1!&rB_3+)0p zU5lIqWC$cA3xBG=8DQ3^!F~5!3l}n%|4lq2gv&c2<``7Jp+Q^~_a44uQeCj{CwN#) zkPfqJ&E+eD?5n1&-UxOvMD*m=Gm?L=_&vq7e)-oq~{6aY1c**pebpqsf4A zTdfeW;Ot`)P^~cN8Z*t^Yq?w_tz8%gJ+ejQ8eJ)34(x>5OqaWgRJf0ew7C-MixfKF zloGcdJCss<+@M}>?)9d)FUdr``H8OrE(Syp>K*p!dx;Wz4SgQw5SpGxmL$QN~EUBAhYeW}XFF_Bwl<6ZDSSpdNId%-Gg zuyPCdJ_Uy$e@qg~+m}a* z_gI7XTh}Xjg-Uy-=zr41-@6Z!M1}+u7~9%QgWck|OWi@nK)S0TD21w-bGSF?4myuR z{&F#8pSl8mxV{H**@IzGK4p&WA}JBhabm+}#c&`Vu};Qko0!9Tjs z#rO(UA_J5>(gKq8DbfNpo{CL7PAwg_dVTg{MQC%t$b`ybIIZp|!R1?=mT(uw)X zdm)v$rh*^CHg#Uqb}eP{=00x?kDKaxE1NHxs~Pk7+f@4i-xWYec8$-Y1|Rvm={uqz z#ImgE?t<1!Yw8`tbXh2t5)_K0sPsJ2d3loLuW|u#15X{+;D7$X7BrENv-+s7HpiiR zmwrdE-P{b&hdLVu?T0Vkj?#N@7wM#eQ;L+=T!}^;X)zyK=oGU z9q_`xp+Kx{RXu*CtYnL@0|gAf?|o~vynd5@chfBl1abj&dL;t{HR^}g-$`vk(=%26 zQ*$Ok8|I7KBn~QOazrLoy$_dVz|pq`5MmyP+xVs9lwK0|{mZ+bbe_tpvCM+8fbw@a z)c|JmB1;?pY|@q-z_-t1i$s{vasPml&;emN_Ck6fR+|0%H@BeW1e;$CiH6lzO#`iD zP6G+{5EZ?Dj2SaB*r%2=-wxtkygp#v9*}H4!YkjTDyJZbDlgJPF4ftycAwT0i!A`4 zWKW%7MQatJim|NHW?yP6WMhR0c+GqtS*p$2X;adUFN(QT9|*y>M{nI!3OU-8bAgtj zv+kCjeOd_b|j2AvuDOd!_qVK`jT?V|>BioCQ&+J%z z{ew^EWVE1ghrhlk+(ckBc)Ak#oLhZ>#E5s1C%phKkSn=3QDYYlDOBVd4Tv#0wZ zhOox}se^01rE^DCKc`$afqj}%(@?rqJDMCpkk|1&-89}}^LMpx0NR3ZyNk6%B#!bk0| zz*-3hwo`&b%2`CCAjPPh3F(41{C~a!Q3A1|KENr|yc~NpnqT(lKO3;PHBkQbK2}^+ zUxR6~hVhLpBR^B`r5~0_IadDCj&}jHi$=Az?=8AjEci9E|HqFADHeDGi|$3LNnT>> zm-5bCASb81y0UP7&mcMTmJZQAj2ZU>!N2GMr8p3YUDVrG76U>yn2-=9#lo^16C`xi z9{w900noexUH-YYhTIXr(a(*bbD_y2|4(7_&qq#$Xa~J3hI*)g%BE53g@LFC`fR>2 znM>3{>whm3^mip~=+B9pcvsB-fB*2`pP=N0h=eAS>56#TqS^cGug2}~`BfC=oyxZ? zNpbnpd9`od-6WGG-7@#fe{1tU2Ip8}uwBzTKT}OTq85)NK+I8Vipyx}-O$w=G`Grzeez+TY85i3D>t8U@ zo-xCOvhNS7y5}F0%%U615+0x%WKsqTWCfDNgC-Py{%>r>XUHi-CL{Bj4`#arU;+V< zMxRf73)aE{R<7_vByYSdc#HO*V73spa-^}WaRx|Geqf4E(E<|x{uiWHs>?Nx@3tz` zAF2BZRDuwXu2<>@yg_^hgE%r1cQ)?x=7giU|r4_?0;N@w)rC z9bhxpz<=t34)&6kYANG4Z_@%Ha{;i)y7U9qSpRRHI;d4i0xn@*5bc0jtAqpvgL6-MdWf{;MG+z~e1z`CtzOzGGHBueZB$K%&yrzc&f6oobV2}iM zQmB}E0PsTQ&~W7dRP&W!(xZt)LeK)CU1}RS!blBdq_!ww=i0VJYY2?A6sCEai zBPYKb%Uj(=2e%22V-Rni2n9cshFbIx=QgISEoP=bO1!@j1K_SxUI)C5DtdV!y`U>Y zJSzU|b9;0ccUNmGL-PCO2LfPaImRkYWF5fgOuD;S7qgCJ);?JI5hD+{e{`9-vW7hi zNYxDJa7s|jGn)vQ-1*4#1SrV}DK^s@wDl!wfykg3XBCetxpWA))dI}P01&Ng4<@t) zK*x<8aquH@=icA@fIL%Ra&)94eukjeY*L}~8P57uM!!P3sEG0e6sV=HWkTT4j=~5x`fv_xi}68!PX9!gd87f0 zA>Rz+IjWpFD|#@OKx>^+K2ToNe;l0vGwN5#8goA;?P{m+=TB3B7Z|eJHas*xM*sW) zQZdY(fd-32zy;3&b!5l981P4>C}x4c2p=|N-tz6|NLo2xqsALvbZugDjy@NB!7Hn~ zf!qk<(&?4u*;FBQD0vWMu&3KnO_0&oeFKDoTlVJDlpQ_1@f?CK{?QeKNUzNx&%Wp ztpLTFs?zE)(Bs8j2gdjTkajVrj~zS=i2WFP*}7fux;dhv4z-rAu0hUB&@aRTxn`HA zz>6}dB?~c?4nT^?`0aQV)Yo%4GAQc-M*7RtIT+D|ip6M;>pm_}I*lu@0ns@Z6wl<2 zg_ddv=d+p_w^bZgXo7-HeF|?~BjCFT)bXJu2z$$6O4qx<*mFg;mzTmwwAwt#I z_iV`q&ko^WdFGqTYw%Y8fLW6VlBY}qez+rW{L(GutPUX4ljZh?(_Jhf^pLnyAyz&i z%qJAmj*Gm$#C*cR6aguBU#ETh<>^+N)L6ck6`Lurbm+SpTs_C5SH9ThdjcB250j72 zf`0CTZ`cJ@e*snbJB<<*;+^OSnrcSykMMiJOt_wxrU`y38ng&AAvQvnf_DYnr;txF z+!)>n0^$;Z&MOHl4ojofufD{4y+jklCC`{>;8Rlm3ctW^g)1Wk?;w&--g>h45?vzW z^Ct*b=9@d+xJh+C`Z3tVQPgRx1GJCxI?{7L%6&f^I`yuKV+B(Hz)EY$%273Y=-80? z^*Ag0`a#{w?`xWn&SW3#+Gb7L1&?#x{0yl$<}FAzE`*eT&{uxSJ-JI`eQnIiGA9~S zL%lWN-~nbTm*8QrDu&+{K3e*r`c-g?oQUq zep~YVn=+KdV>TfW>p>=+EPde#PKbu9#drx3*B?fi85mJmvq{&>^tz;LCW_ZmdLNdM zx_$VWnK2NVV4(-J@M9??7o`_-cG;+Clj$QLZ$G3BTVppOmPb~k@&`delf4*Bj#L#6 zIFr9$8;F1SE^*^P8&~i4AdfKtMkHqSFPR@vMzIJVAcBm91r^*kk#s7RUif zuh5!zbBe#X)aEuV@y`MWy^B@BTTR=L6$)0qio>juWb4A8GN(8`UeNHl?4wo`b8*+D zd~O4G#ikB_dOu|y1T0y$W3eE@z@4v~6E!ddgF{{rwK1dJw-^~O&u)JjWOmGj&^1z~ zzZ2wSQchgkj0B3hrthE31hYfAOY66atUd3g&eDbpv@eqGZg<)>797;Awe^})a8P~v zZ8x-J(t{Rpw4pTCCIXr*Ey9QcxAXT;8SmXe!99>TXz(qC3xOaLvV*45_f37OMmoES z=dZe@#n_m5;o2k2&zL*H0>X)WX|Uq#^oA`5c5Bk?05z3$&$vI!e0lcM&uqhd4AIkTf@h%$6$MoA>*B`o4Js`svpV zSYrMNP{*H#QXxxf0s(*JBbkU_-NJ;iT(OjE7fZo``t(`Zfuj*pCaJ_hWu zK9aLG0^k@ln-z`0h|DYUOV5|a4smSfN$^CHc;kn%?}D$j7V5M)N{<8cbIVO698+Cy ztBkmNbH$~GYenRf1<}!J<5m3*OTzt1!o*0r?<=}eytv8V-bJn%ynr#v9m_?2R|`(S zd-vh#$TKP8pzcTB%T)lbXb2irrnDAJ6yEv#-H(!2<`LE+lrXc})&iNGoV1gsWJH;C zkd?zew2dTDfYK(Rcp@Hh7x&VjvyPC@Q9^AB7!BqD-TUDY=|bPB_+Kw!ata zGuf8zB1m6efPMP<=;z$}mQQti-pL&Mn!nVZL+8dWDMdyev8 zh3sldOs@&KJL)KRwo3?Q>5eUB$Mo9>1~?vI3-Z6uw=%Kla7=@1low=u=F)uTp)^I8 z``p_yte6}(ZK(6u$A?x2F$QIb7N0IHdC57BLNHkH5Z1Ulkf9<&Q#qCMRwd$4AkpG|&Y`NP@Q)|T8=9s8#3&A6*I zXU0qIt?+A#W9JVV{;R-6Zab6)!a9DY+Dwg-C6FrXSS_lG5jjg0Brsj@vB(Uk21IY|zL^WMk;UvEd8|tz53fBM6cWBQUn`N_98=#UvHem#H7E~l<5)mL zj<>wMY2qBkoXj4M3>Wj|TkP$s5`Oa|fCn|22%r3cA^c2qc^z-!(Vi}=GG=*!zb<~0 z;dZ+8E&2L!Cq1@?aLF=y30;ib>Em4frFG-WvxFV@skL*)C6u)K_2 zyRZ6(>yIuGu+zhl6W~@n^nTyEYF5(9q)WKdod(6Uj+FaE*T{;w`>B0CREaxi>gSqt zF$ogpS=G*e`(8LwaT1JtjZ=bYaG3gVzgIGG-1zds=)>X??T|U%M5(GopkKlxLhcv7 zPs1@o-gXCHU0_!h9=-fF-=Gk?+`+HqaojY($`D$#t^zmJ&BqF3AQ z;i|L4`aGNcF~z9i+q9a67qsw)sCa1)p~@sJ?qOdf#BDz1O?&!Hfqa&ZuQsr^Y-8xd zi+TZS!RI9RKQ7eEQ{1MVrH{DVU3P)0PTmrRO*%D?rs()G`N*&7cqv2tZQ?BIMUGB~ zyLiPMGIl!DE27J~ckSJWdPmB)?X%gxw79*Wu`fLNFwME^jesDQNFe{VM`BoBz_nvN zc^HBthtIk;$+@*cS-;W|(%E31(MK#Y)Kj9)4Sq$ln`H zR6-7?3PdWVZ&MS9Uw zD_`EH-B;I}yfW!{sz%}|j*UbDS>;mvgTHyJbe{IUFn@Ssw`U+r=h|7yfl4s+RQP%M zhg{tKtptT6MQ66l?*uwqQ3AVoUkSvpw=rdoBhEsLDcyyfwqhAfc6H$)FZ@)wJ_mLI zi8pRxs^2#-O-iD$;m>FMn#NgN3PgKQh9Qa?HROY%s`uS9_VJbY>T2urFdVj{HW9RK zWtIdC!Ol>Ik&(~MD_@bt%&y- zKk0<={9(M<48OxhW9K_FrnY|huwadT+v9)3Jb-wd8>92~-VLlE-~N!?H5wJW-4UMp zu$kS89faZE%PNv%=7@8SykGfjob40QJq|=YNzX8%(NV38Dql+^tk*Z|n{+%DvWc2v zc0{?K&9yW$E{tON<`CL3yu;Bsj0DAnb(jN*BC%d=k6pWJj`z8!XQh&zy|S0SH;MgX z4@d94`BW@JQZG1A)K z{J%EX5H-t3282`kA)vEy5=9ufGfuxjwzUi!3utmP5b)|P;H6eE<4>(p5%8KJnVft} zyCvDl8EZE8x`RD?L2$8@c>2XYx%I9&9SJsK@ORiOSJpOu+cPXi<|rHwC!2neDSJnlOip=)fDUyXfJvw#e4hz6~&Bm(UY`!-m5FFu-^*IoZPxqi0r)J?KTsum#zWtUXs}I)~oNhZ zcpm%^H#=`cAhXn1FYCVdhD@t}3D8M@B!)=~CZ%>Lm8` zVR2)LQsfAikaUIOxj!XJ)oe%XqpWt!as6DGZp%G-gquD@xo${Ve|#@u+Vg^X9pP1q zdkPsjV?cM5;o$V+cH|Gm>s7qH0oOjU?Z(KzXU5(cTFcQtlVV`nW34=}8^94{a@5>d9DZgv9@oh-F5MljfThRPF1;CQKViLpy^J9y%wy!K|RxulRE|KPzE{4sa!xcrXj0ZJb} zr@1!{p??kuF&4UE1-jG;VMV$RYz9Z<23`FyG~(AA^t25R{m2l7r!6F(_PWr6-19dU zk5gsQmRdGm-_HK9LC>5m&#)r)9Dj9LeK2oIbw9gGVI9okB+!834yYD66qKl*j z66cahixA5R9|Hg`uJ!)&hf9~Al<^)`ae5}Y^k(3nQx6%Sb3J%wNU@(wc3|FhH}Yp`T-+{et<emx8A>9#Mdpf zm7#&N!H-D$LAvQE6GL$X!@c=uYqIZ;J$)Ni=ugFy=7HpE? zW3AzzQqR@4x6F9T#Y3+6(Z(Y+p*-V!mV3KCF5DhS=k|uyH|hPSm;;?e6#ct*2QX zskM>%zL?D03vp=VE@?K%zbN-I6T@2?-(T_gVHgK>P-TtDb#pi)j;(%i9dd}{UANQC zHQh2_BSs%TabeMY_=4)bIeM@fOUHTWfwd+w$9; z;_|$zQ+m6Xv7!vy*CkhZ+kQAvRL1aCy~Q+31_z7!!HZ^!Uz+UAN5U%~4{yHP7CH_X zPCGZ*iyi)Zi9?RMO}77t{ancpA{w5%u~G2f9K_gADnuZOD`iVqW1{1SssTf zyb6}0(mOth(4fc2=IUye`K~%gv7amBRQ>S{{qG3M4SOS&E%i+b>l%fkuW;DGy(NMI zx{*Xf%GBx-az0^P$Hc?;-%k-p>J(iLSW`Lr_#02{(q6?Mm9o4p9t$IayT1xQaos$? zE7RQj8DjEfDO9wkxnZ@WSR_3QGpvz6nywu@cI0ybiKVwyBo zl~3Jl{$X2_EX`qQwUfPWCoYkd%tlQC&vqYk`plshDIlM+9ys-uJfdcYFOy`_JCkz| zh~d9y{MIDw$m!6G2Ky+N)N6zVe-3uL ze^=niGe_HY!B9nh6HbDAxTBxJEzq5krhpg9bc-w-dXBU?jqj!Ge2#(2mpYS**eOm;G@@HoX&cLAQ$c})U&S{esXP{qWejsOT z{ZjB)xAbV{i7lHBa}W@TPdcU z2gGXqmwWG;vJDn~VyOKFzl8qE2tnz|*jf%L&f`SRLanaXa%$!^wpVVHE8*S5Kwzo6GheH`I;+8q2ZjGPA{xsU$)^w0uKB43n5QOUr^EH#=PW>~S*)^%9lvHYbI|A&C3QQ}18 zEvRu#|IJTz+Jg{&&UTq znuSK`h4=$C5^N8ApAr}#!vubReX|14?7jHB6L%rxh(jQKT=1K{o~u~F#AR2$2@J>sW>C8yp>!k>Uv`*L}cdeC792(KB9 zw;m>dqp@CdCaBP7EVq90gu(g=E(M@U&WXLE_vks(pX|!qC4Vr-!#WH<_rGP_WJk<) zb(j+~3Az3Sw`n224VZH^GhaEsy|cS>v)%ah{Prs4&cGddM;^S;hwZ!b{?RowCipRW zNtXC!-t9z&wbVG>^Q!RA^{>0Vg%{3BeZk9IJ*3mKw;g?tbkNw!y+uy2#j2u(k$DYU zw_TIF2{uO^E;3TBD*UIP{C`Znby!th)b%ZhaOm!mP6ectZb4G%kQSt+8$`OhrKG!& z?(XhJy1U_9=YF2|eZTAS5A%4=K6|e<*PLVg#tMqrDgYDJZHtWi!Y#Xwm;)3z{&Nu+ zmGp3G{%C7HBgH)~Z+yoZ^IA0=RDzcMLT1*!Mah2wq`6l3> z)BsYh1qSDCu_8L6~

#nE32RufBV2PQ6P=%xuJYrona|7r}{91+v z2yPv?+S$;>yg9V;Bt{8t<9XIixdw_Ov|O^ejeWKfC}fdT?q}E;V4sj$cb=dEL-E>Q z6KO}x1V`=4^I`V$X4Q^$+e7mqPGSKWga9EH?B1dDy>Gp-^B!O@25X3qUk#0aB?doe9OG0$N{VLH%GBF`p$EOiqA#ebzSh zl?<32k9xuJ)yX6voF7z9#m4&=@FZI!9p%0c~98YbJ*MkJSJdYx;w(#l;v z%F*{=OJKxOWz`YVmvEW_j3smEH1|h9YdZFO0QNaRn9d<&Bv1vJ)Sn;nRKYhXN3qH< z2GtC(lPb?7V6bALOS*W!m{~|W<$8XYn+GTwqs*|RV)L+UwX^czIQPNG-*U0|(2b{& zNa^Anst$4hqwaYqJ)M1!Rdj4B*5j+(&x0tdm~X$@Yrg=0jE)m$X#>RVzIcu{`NA0w z+@lT)M#)3wji;vZW6MXBLN-B6{3bvqIAew1$FS^D=tg~RVH`Ar)WBxkIU-3j^9q9kKdG&9xm>&)=G zcVb7inS{$Du#f^X=JXf%B8*Bjue}QxlctZ~gFLPScKMI>z@J>zY%`@s0P9q&=EM26 zQ$xC3tt2%!_?51K`tnRv(RRyboaO!dS}mOZy0tj#BMt6Cu2bn5bP z_gU-W{k52MjAVrw`(-nHjB$nAPMdfk-}V>7j=D%yYPR~^jf)hwOJxf>J%h^+$1<(h zG~F-f5wfnQcVdG)SNLbAu2;H=lMm~R@4{G<4baZ(AT>`_*YmNJM<#)&vei#BEm!Jn z-^{S%j#^MsPaNEy@0m*FDHTw_K4MdkrUP=})mpE^Js&b&$Rl6Xb6?&%JlM9S;qseIaxnoZc6pq*LG?k zi+5Vh$aXj}&4x#j4Y7<%eID|>urjFFG%rsTrjI}D9`m)i*f=QUyqL_ham3dCda27i zxC+bb7Y0Gw$MnbCzaIjr3R4#I%}X{z$?bgvTW?%}G{*c-AgyU5ENLV%3<~4XK8)KN zn=#vTI*Nyw`-mA+k=Yfl1#<(>Z!urq?Z90$k^-=f+bakR^<3o0L@~WIi-kiEDXyc5 zK{zJR)l_!jvFkK+4I<3=VINUw#hrSbuDt9xIrGr~Yzwb7#D7~Ee=_^sytHFj9zPzb z$s6D23O-78E45kubZ|xqnmzz&Ed@DjzRf<5j~nFAtXmq;)t~3*PtQ{6?#WJU^wP2| zdDM|C)xRO^aysjQ24VR!x=~Vtmd$ej7;<-2S=BwQne1pd=!wHfo>@#e+VNl2@%8W@ z;={|jk55~fJS*=<0;$h`$;|4K7EIuB67-R%FnxjQE4>5q@KAZ1MB@HwLrSs~ zq&nPy3P`e8U^9kvnl)dO0CwQ5*%iQgun$eZiVA|HffQIXdXR+$$B3o7R=%Bfyih$Ru~pvKfv>6S z1-7eT2#3f|4B?8$Ax_nm&wv6@^DoK>ARHv$Ffze@m3qzXUBI+EWim-9z}U+p2$a+x zQ=&kAowxi8^eLc%;=c)!dwM5UTw6SyAC5Mmbr9e0?`GV7;WJ<=#p!6L7F><06qAa8g+2K# z&TbB3T9jjf>`!Hj0mQ+*=vk>qxe6rU_z?2PIDuoQ`Aj>9h16ui?p5By1!(Mg$Z(a+ z!$C}F(hDnF6NN6A^hBUCv#6zHMjBXTc$Xu~dst=h-n+mYALiXcxse7-|DZqx|1V=$ zSApTK9dU?hCBzO3hNoOrXDRoz9Xa_;xfq=Z%U9l%)}74cuP*t7h)fvmPb%7T2j?UA z0{bnH3yl9>e-YP3I3;C*gGxFo$wARGIf)#Izg{c>WLkL`kmt_@r-;O3h z>@Vf}DA^zm0vacR}QHf-~Y#;=}D zbLN$o)sb1a`^CoH{EV1hWE9{=>&kn6S)@6-X#)}+h&(}=q#FOXiZ{7O!=xy2y0D!R zKn$*1T&r(Q|3d80ym*fY;^1_FaBZ;z+>XLjtf!ss!Oz_$Ximj3);6PW{pEfOLLZT> z>_%bQgtzffB@Z+qVT5weM2v>?BUm!j*o#$Pc?TIXHl2JuoH8qNa76S9eoHhR5z0+= zrF|@>TLzI~51m8hP6*}JPJ<^aHrP%RRGZ&vq>NyK77|*ToRV}P{b8GLcc}VI4i(EY z#bFf=@8bOp*x2X5@27gH=0FcH^P;Wsn`HxU_Y0G|RNP;SQNsGZX1zZ!0+L^@aY^?* zbAS_d`a+=;jYT;s2U|y|WZ1la75R)U_r6VSz&Mn1`ujGg;u7U37py@jr^-p$ax1+7 zDKtOg5cs4D6bqh@fXRTzu$7wGAzkDw#31ZLHwYov@>1bk-2hs_SH=YL~81{$yu?dDgmNsq&vu3E@SXv%Kue%oM9I$f|= zUz2K7f<3~!MWWjHtcE>=koc<+T(zhma?b+|57PulZ@TFiztFZNhdg;b(6CnLNhnjn zdW-6U9Jg|SnGv*uGcyrVdrft$?mPSs95xfGorY|HcqJ1W|B-=05#;@*<#x#G02Gg` zvz&}1sz;dAY`x$uEq_&}DQLcKu2KxKb>Dp^!}^hnjM<4-Sy2ECKLgckk^FncT9m zur*21jXiZCr}TI00q)_YH$0_rIDs_j}6*N-tEa|T0s&_HXb`Jti|OfZWpm?6O`y5^ zuv(i?VOCXr-IT)gw%N&j-7F4!vhmpsqwTAj5w_f{Dz4toW;x)lsgG=3ML%?aneqf{ zyPnuHDBcY4t0ZJGV32M3A=^}9rSswY4P@Qc{<)Ug;g=F$L?rQ2nc9;G+8(A#J(Sr{ zprT|ni3M=_C)-j*BYa1#EWP;`-0`aK>6Uf(B>H)yd<51`DR6!S(!}j>d;VhW=5-H^ z=^=AvwzDxlw^xMX8$IY;K_wn~?zWGti1qXy3`OIvum$WQ{WUF)gY&eH_e=P|0}RE+ zzmk~b1(V(d5#%>G{kBhqi_uV4udnU$&bf9@k}2Z?kePayU!;t@nH`<@x&nU11yoOd z4n@1j`f_YtIBU-Buxo<@pOC$|`K%K5SwoE{f4wRF49SL~qrK@Oy1Eie*^6`8w)mz> z4ESp1&ED`Ao~B2O2hd4>YYo^hLp!Jc@(Z71shgaiBtVYCrNXTZL?T-Nn6g`K%Za7L z?c7WagYlAN0|0I|W59Md{oD=^W5$-{_*~SyZpL2!hpHQ=o*$6GD8ypT1~?_4J|ouS z{Q2slJP7dw={VpqXfR7Zt^j%UjlgM2*UrLhb#|Zy<*>(f-S=;40$YaT(Pz~`TZ3TiaO5yj$hi2-437g>%L_S z9%BI-?-~Gyld$hFi`$oPW1h+X+;q<%LS%1@iFZ#E;x-{WMBcg;x}z#XE|$K5coY6< zVqNFJr$rmz^M-ag|G*yQ>zK#0^*H~b>bPg^WBdpfH>_Vn$#RBNV{~+_KSd%y%-$bf z_niioeu#JC^H@)*>TLSj2iBZuH*jZU{AM8{*}DwWT=)C54lkY8!%Gn8$a6oX{M7cb zFdC@P8BgV8*2z*j<7pxG&?=f9_p%PF54xxhAC^A`62A{57R-K}Tq8|CA^h4X`gOi1 z88CDOKY%X+n;RaRyB>=hB?uDOXM;{>f#A|^*Mxf&ZmV7Vf{5I)dh0f5whEYk^j63U z^KbUx<-a>-r^H^^i8AJQZvd#R{CmNfC0_Tyk)cnEYI1yi5_!w0xmk^&mRkPC^vsJl z(u-#CH*S-yt)UT}+j)KlOG1QU_Ed*gVvJsa!~=oo@YG~}_$3pdixhLCmAjTCi1F2g zdnSg<*~zs1af|fd5459VkgQ#}S4|=pS-ZBK21F-dvOWZw_R0YSgZAn* z=QZtVA2U^{#n;kq>9jkC@aj%}hahFB4?xQahc~-*Mx);~PkwZwyt-J0#LFvsXs`y%B|%u%4qPa%>S0t$FHM@`BMmZzJKOB*Ki}|`D@2IAp(jx zf^1p4s}FP%I9R;BN4|&)nZjKH^9Re7{{S6j-2fv^9=m^RPGsjsVFef&2?csQe{)+X zjFq?bCV^^2Kk_7K&FeD;uMu#sI_8ahjFH|XILYWxlQtz(B2kqcG3Dc1FaZIzrvT~d zz)J5-p4}KquvexAZa(bmPG5ZjX>(gEWj^NoP9EuoI!6g|_Y2*@uKvaS$PfU-Z955G zU~GbJlw!|<&$lDlrgZ>%BnwU380T7!(HZdI-S7wZ=QIQ9ygJMAs8KO_=1P-SpsSF` z;VDs%S19vJU7pGv=f3sj5#|x$b!S$g+}`eD>B%ViB?=?_$2e`5-SG^8p5hA`Jp+Dz zhK}=)HavLpKdd($t?2EmJ;Dg2ILqfFO>t|@^xW;AUimfaz3Ad!rlGFMv5;Hbx27lc5=Fcw=pdXcM&LZZ3~{^@rE*wq37P4u>ONyw*8(FY->VFBSw zZWyi?CB|!?sV0i8lwU!6F(dfY6WA1X%ZT)`bV_D`aLRoQ@nub@?cmKjEoRB63~=@r z^bxP4E3D#w^cxlJuWeJAm#M|W5`BVnQIQ0go65v&h5sNFae`7q6)b@d12uz-&HfTV zhqViiSv+rp!NO#kJROXoeetT*}8nSzUoS(4UO0EZzoeNw0BYIMyH; z{yt4jpsDyMr@bt}?aA!k&T~B|h7PeV`cSH!St1 z0_ls#a8bQV9-~iYg^}K^mBejzkzG+wHtK?wLuohZB9r=_qkKqRxs_!t!doAD%2S?i zx5HhwIXabO@tb4}Eh~Gwzo)78U7MMzQev*Cwbz_uJp{guMH{114D)ScG#5V9^{}|n z9yYtqeHx)5IAx|hsEa#k+_=X$wjx(TTnqQX=(ZR%pn)6Z|C3QrG$S}*Z>n+}WI^eS zAsF^o3ru2$4?y(RPZ0x$!+ae zb9R_1JRkRUygzScuGD~Vai3w11*!Vh)6MhR_a6@y72ghv29=tdiS0Z%1<38xW`{Fb zNxP6Avwe!Fdg~Sj>POiG;)w#+daaROie?GYTB`)^&O4PIiQI?}_z@x3*eYxhgVsgAU#q<|9HqP6eTXdC41K6#_~^J zXEx3`7y{2b!g;M0n<~wh=0tWzYNFI1>Kf))8e5mks1$p7H`5?jJk3kr^CE^gTDVt5 zO2zPT`Ra7^IgAdRD5p-d7&%%ErbXk}E9|7i(ib43&MUxQk9v&YjZp0qhK^m#pT$Bz zw_9N?!gXJWn2NTiYgen~rY%bF`0bq|O<>Afo+_DnT5C~GiZ(dx6GAw>e}u>HGCS>N zcO|1B87b>)@7v7cQcs3&^@$5ocwKVTj^Dls`8knT(8cxDBmH=PmvEm>`8`cts+FcgdO*jH&8lRDk=K)95>%>w1 zb}fj76f!)D`ayqafE}TTWiOm+pW;uwo}CY5)O<)&ozY@GpEc6bCQBhy#iY5}ZAKikwbK46_YV(yK<-Xk*H7MT~>X*yZss3xZQ#^BGpd)kr!^Qd4LW zL>P9Nfr-6;Xg)VJy}e5wMwL?M^FgH z94g$;!soUz{|^h`;H)lV>%5?SP5mNdyDd|GT0nl#fSDhYKhEcnP2t};lfP>i24w=8 zai~4&KhA;65u23xoWDA%=WX8|UtV$ArnwLH!w12Sdyq200&O$5chFSTfI&Ax;aKr| z7OE6Jh*TixURcryAxyG&&aHqB2kXio72a^bbj4uw9>ektLy~~I*GKSn7f=_(LtJJ- zA#6V_=+o|8XeeW4^~1c35+Kp@(&OW{`%X|r;pUHNEHdPDR3<>On|>;yXWn(f_9d(` z^|RIlBs0=I6^DYEzMuZwh%BnT<~u=WS(1%BCJ)6XJUUg|?hn4LtXsI7DB-Sv_pnr8 zsj&9B)9pNkbW@U{@cHhd5m@xRlqAL+=cLNA4*MlHiN8Slm-|dZcH|S^x*K(S>Zm`? z9RYflnlZTq_m^dfs|qZnD@s#rB+6Yfi!p~MA0v>orijL7ju_TQHA+O?YLwWQ_s zF%l;0oAKYIm1%-FPG$NLRdFWa?*A@HYH^+isrm&`hGyyOy@QAlC4~?Kohv2E~7T zjma(u9yi?1kI%?8DdSOqLlSEZjgO!i-X^)6=%1O`jo%DUS$0NasF3133M7w`<(i|m zwo=%AKdu}tvba6zWA`uXkS6*yVfS7ju@pr#UgCUj<41T#a!jQ>r3TldF0sRZ~quWA+T)H=M{bWc9fO!sy;qkK{t`B~gzl0?g6W zkqqYThPS=n=<()PECtB$)qRu<9v(+BjHOak)##1uz3sL9s3P{ob>eM}^RLj(CYy}+ ziCe2%Qg222Z7xo!i2@z_J3Nl_e3|`8KKij8Gt*_6A7CDlb${(Fwnn1f zY@4Aw*TUR0RsT^kTAoYph@v@M>m_yY@VA0V+2mpN*6kS@!+R`X5owsaXkn31|U#g$k&wc0EO0MR9U1cnKiD2rC=#J+VM!D+9c5Q3Bo)Yxq zb!KZ5bqM2nJ6)dy53vFZn}Vim@Cz~=oofc4VlgWSa?Cp(!b)#(O1>{@5Nwv8rM?qZ zzZ{u6n5b$x@7WoDE1%4=n#$@Uh%CL2&qXCa`f+^X|KZhUW(!HUV=zzub=RlaaLeCR z!&CwHM9wr-2St^`NM!gGiERpV| z6z7^X-PO(LVX!;cgx+drwVd1g)Gyg=2 zt{t(ALC`t&Ri8VbVQ<7qPYu0~LYFI*<-i8#d9{rHy>j1?s?>_hrETd)gP@JNZWDtb z3-FO+WToRa_Lit(Lz0P-Tynw-ztLmz{MMGBaM$8po@sdc&Hmp< z9HiOwXnwqE78GG|f14}@$Mc`Kl?PLfCWVl+Bb^eyS;fC(ynLP(HbrI?q_QKLjGgws zBlo?f$RdkZqHUGr%xtkHli2(+pM(!nA$Ip`u-lD)3u68Cq@;ojJ&=Kw7KzEkn~~GA zpzQ7w8)omP*Eyea8;z4=0)%q0B7~<_V+~NT+Z_|!7x2oxItLZ}Mm04%kENR4DZbIo$5XWA(R@bF%~(K*k*``lFRUx%D(UwDB(QF4vv zyIxac5W;{_kx=TNgg{tNroMvki0}25V%3|p8Ouh;;G|ncy|YPLqtM8X0a_`&Q;K9L ziuM)1N%+yf)0oAJSf-*;`mNY}>yssj6VN+V(N-_}0u@Bi!lDbjFIxA-|8KI8o(ut6 zp|WWb*MqWa7dlulS)%Yn)ye}_RTd`}j`_tG9Jq%Xu)yQ2lVIidxeLw{&kIV#sY(zc zDW)^&7D2AAwNa*neMKQu({j7b?fMV%{D7^fP%m}7`gOk)DrTNMP&(pfK4ldhDmjC6 z??cdO;nvBkQcQl`B%Zzr2e}7*??)Y-9S^JfiL@u2X(!WLDWwd^#XGN^3`#XR$DMsG-w|LFvH#9y_e`ynT>)LO%Ug&sq@-zgk&gd2hr2 zYW40VDQzYT)81RGrt5R$_mXMH%<1aUW?IA_*gfK7I_^^dSDREM=^rT63>U-I zgmw(&>4W))nI`CO@5f9YS-x(C4&36(^cxI?;GiiFRJ>bpnwBQgw|P{3OlX1+KPD<( zzHFb5^ERW=1wHZeP0EUt>F7H6)5~%*r0w8uW+};^T&^BgJgu%C&b>D1&quz+AAH3& znI%jN^0ZJ7`Ch-GzF))tR3ka%|$tx zmjj;_p}VSoIoYl7{PBQYt?S?8e?T@TK*S~)ba7;I`ACpucI-X-WT3Z&DN=JWZsfQK z{$!_$G9pFDlETvtp3G_0>mUg?oqEkf^?pb4pchAoCV$ij%ulC>Ucy;)t5)9uCEsMK z$DK)X58ZSraLrQqTtNW}8rr;P6cJ?kBikW7(nheN9giZnb=$*k2}m_zKby&m4dxVN z4Z=)7>)GAEtnAWph10GUDY#V^<3FJlm^UvIgm7`W%lY1OhL#wh8y0hX`6;a-uM0>j zRnUlxlKMTy1|a0K0L__dsG`UNaD}GRWH^l|SO2ns}nI1n?lTlUbrr1!@@@`nx~RG9w&{E3RN+)TUTh`xExx;vNVn&{6xXd z3}4=0oqR~85^ALJMhxInawoZH8HFwkNU|Uz3hBD=g?O@9dKa7#=+`!Y!G`Nr{12`Xd}42_Iztv-^|t{YWl*Kw5Y8Npk+*Ni5cA-gYX;vKXofC~Rj@8l2IscXXH?r8Zk5>y8K(6(n!pEVpx#2)E`-u#!UXPiSg#F0 zVo?5f`Y0R9nyjRj$~6wGpu!C;Kb`kM^QRes#*Jm1Z5 zfeE;u1z--Tuqw&%9wa>gO8E=~qJe$Vj5L+Uwcq+mdnhVX6d%R#Y78DXJ|1c*jx z-NU1Dn?SH;4CR?bvN^8#;@7;w*BYfw`u@jmf~ncqave0PHV1Cq%%CzZ=>CH{P#)=f zaEg*iWm~v9Bm{4fsvpE~aNY9kzB>!btQ8PKp|KJB?;W@>o~vEgXJET1m!eDIU{wQ` zsiT~v_sa8Lq!>Dq%VR1Q{7eBooS^m|@|x@p575_}XpZ^P#yVgl2h7eo;uO@nzbLTYV?*VxZ$m z2L&{3a$Q1YJy?`(IZ!?f4A47ypZdVVNqhwbq69j&c3jVdCU^2~1(w>z-Dv#klcWlb zaoQG~lp@q=|Gyqw9g8r0JxVY0{tI+ZBikWF}53TV%A8STkR_V z6@^Y90>8g;_wF{CI#EAJ;i?nVEBA}^NY{5j=0tR1@68TYkSFYZrMtfXKt%@tD?iRI zXLRUJGNE6gZ9e@q4)gfBpMHzQ4=@9=j_$(KEK(W_O*8a*g(tbV`pIqBdUfyR=RP1M zNxDh@_}v@b!RJab?zo8fdmyolFuGdp-?u+9zM${Rx-s~suYm72=4Re51MSGSEr!xv z5I(viJ=OJM?7kV9sgy6zbmt$XLU`*>k59z6UfUnN$us{9AkzJgxC!`vBg|4?0=5R1 z#Gm0ebOQC{Jf(a(JFCzc=g$1GCDq7+Tm#EaOv5(yEhtj8zmST;bhI_mW{8HtxPvss zxU2JS(E$-clsJ06mZg_CA&?q9jGM;B&TZt_8h|WSU3)LzkAW@?ja<@Azt{MlZGk}Y zLuh9<&1ElNkNCcJW*i>Z4|CyJd9c4kKFh-ogZ3{F-{iD<*I+0c)$T zYU0Wr|Kav6rWYmlC<&wIS+4I7K|BJ|L00c70T&j)j7xlBD+qRc(gvXA3c+~0rD_36 zS}qVO%L{VrVhPLM)reTqgpBa6KLQ$?h(s^p+6?vk*Y^fIV&Q%mSUby9=fke~Uyo!F zDZcycQh;18IY@(nocUsR`?$}GcY-WP8QFZ-h?~Ld@ba?~y;_-H1h`GA#w3VKg7U(N zgR|XjZ^a%0`8wYQ`13VwGE5@QNat!oMUF#F!V&?7oVNoyPvhZqn{S&g#!tTHJIG8Q z?nvMKAR9FhWnu6?`w0se{eeQ$iGz-Z+JKU5zZ7lJ#%Q0d&@JzCB<3A zEeR^4#u9c4(g$O-*KR4JF#EEa=Qw$zT1^b`-h7Tqe(8PuiR!G(77EJ5; zt!*wR>2gdmtLYifr{|S5?Bz%;Ix%Jk{_sZX_;oAIy=14jOHlFEX2gJ51q(~GcIGY0 z+pOzMMJ9OUc#s0eS~|wx-sWU~oEXa`3XW}U0KhFg8GffymsAS1-o%FZv|Lr z>}C@h<+sm4*b8^sz?gIB9!#jz>`&_rkY`il0B~a{3#Vc^zWbqiVo}~2SiIMp-wG_k zH*0kyf3&%L5B=0sUpU46dA|4j>N9#I_TmJ!UL86(fn3qZcTS|~*ay%r%7u}I-q#CRlBGHA`%W~P6|z5>c| zE)aUGxiwicSEwQd3K|l`)x(p2u+4b?6uX5G=Vo>;BCA6O^$b107W;zm5rdD+1e!N{ z5X4kcR;&HRd53f+9>q?eh0a}VVA-G*(-J=u?~kLuMKzoOEX!y1@sV_*q)ilg!fW=# z>o|3z#E8(+xJdt0Fd}a#NAVf4+RKIan+2gZ6IPmcFMFoV$SL9o+CWP|AF`aj=PvIspwq^Q+!3qrR;aOR6@S5n=>(W61%rNI5}E%YpZ>1BA>1#JF$6v zF80ULle+NdNY`yIPOzG`OnxJ31`jZ^euPlHY%}vVMmF$covf0B#?%)r^XW-AjqsJx zXJ%iyHtHo%Nw_${BSNI7?oYC8aNvv;HZDYM@sU*`KmKCdX?iqqMHT+NuCK@Mv^F#! zTZ-S(l(00gnPD0(`#OGl$Cto5DbY&qBrDDRErk-bG`z?+u^R&3tPAQ>6Zgl4(VDWZ~2ArjC@_9x6V6XS@bFT0Kpi>jLW3(bp8k+$J@B zRewQnD~+YDFQwr*@5QL7XwG{u9~;9ScZnt(^>oBevxqV zjC$Qc{(RHp6SD${+}gc4ADc|p8c9C(JA7$ywegF<4Y&Jubc-U+Z^xGPBA0|=L~4X= zB4^~o_V2YP+YU-=7R6ETnDypiWl1KFAY-nJ(6wWKRUK4yCikHtg)oOs!3x40_T6mz zkYg#mFfB>zaHB~il;hie;OnIuC5YAW(eikNHRz}D!& zBp^yu1m$|TkNbmdxbpku(7I04k%-U;oU5Zpqy=Z14vH0VxEw=jF?k&|{?qD$c85o7 zFT)90=@tLZN?R~$$b*v*V)G2(qU<$D&-vuUDOvl|*xLuQ7)xZ&&AoCJQ@`yi9!|-X z=1=2b1<@#xlf8W{_1{c5m%{Qj%F-m51E|;i>vaz~F0-%o&P=2_HJAwmf}&k zI<-Cl4kLmU=6P=cBJ(L)TY*-OX;0ZBSm5Ys5Hp%m>zkBF1cTWd4;+bpXS9pq!%EfUqi!t0 z!(3``kd;8&7x2;uDcpjvK6JOe&O0$t7+%g$iLJ3@Whjg&&+&VFYQf?&-sqL1jFr?C z!u_;9en#a`hOnG-x%m?nBE?OONQ#PSE0}I^ju=)WMnc9N!hx6>o%2fL6J!xew9@44 z7m^Hec^ReEcp^6tZ)%e9CKLP0s|@09V;WMk-N@nE1g@O-@CHa=^dRNWo-;elUbUE@ z&zs?wm3@q3P!OgIraXf zO%6Mi@?|mGp@6>14C`?wYOYG`D~i>OpMKlf)AMq5d3)D75q-4CEM!kc&9fGgte7*Y z`l$5NHk53>@IP(p7^7jYwcy%3cD6{aT$6{VWLe`^L;6&Bc@EIl-CO_uz1S-{e?oZ~h_iSZzB{ zp&G=-F)_a*x!!yt;#Z26CzEprxk-MV!~(fYPDdwyqEM}Y?3lokZE4Z|@C)~hZVC)}XFM$sc|XUTyC671-K*IUZ$%`PA8prihg%7a-0 z+bhZt`fESJ=%X3+9zHSs1h%dx&!-8u3d|3)LwrLmEZyHSQ-8p^0^KVwoPL4b)|bA7 zy_EfBuQx+O>FvBzC0+Ok{iNRg6P;*PM5Lfj8+{2UDI!#hYla!Q8Dj5p(z>x}6!iqK z&cuzU)?ZV|g#OTSiM}xCn~^MJU~V1w&4~BM<3#wptKJlEMh2$hr^asI99=>kLFFo3g;PG) zVi^mvZd)?tr5NaX>(_D+veHm0=z<+TCgM-)U*o3&KpM8y2H!6lY?dmWa~DvD7z3?< z7gSVAQMm{p!kBm%wASeRTKu5mIeh>4!@hze-ST>^>|X;#uyAz;+U;Iev;6HasxP#iBfv~mM@&ReC-7a;TcJ7FPLW_ikxqG zx@0jJQ`yo!6Z2ygO}y4PI6o7(g2uuD z5c`z_G4mT{q7a1G?I}hwS_d$VSST!Qp07#*E%2sLp1ntp7VfXUXy&+Hrh1;yvY?x% zUO4>6g@}4p5aG^)qqk_YD^qw=T1iqYy%*XLgqT6eyhvTJ-lipWTZ}PsUyOc$xeL*5 zG5Bh6V11r3@BS{@47_5Qh9>h^A{PZC|UyvNRmL3n$WK zbSecB-lmACMTS zph>A&C2@^1GC-&lIy@{?ZKhgCNfuV-;DI(`+&7L4eIIZS`z6Ui`ELXn1iI!L(7__^ z9tUu1oC*Kqx) zj5L`95Lwa%RJd{YsrMUpqQR4XZ)sU=@c$MYO+c*l&aO#p2R~Q-c7SpXgmN);!*JDK z-iks=uML2id;plcf59`#UgV{mFY9y!O6Ro`Ek2->HwR2_MqC5uK+qB?yHK~%kp8Lc zAaK?4`7vC%*GR|1`I*H8AEvP!sM}o|3RIhlU!Qd0cbIQ^49*4g>8hhJfT8CkP6G_6I-t zn`k0K+-!C-xDx54@48DzXg0|T@g_?d?;sItU&dj;CywP)A_Ye`&8}3=JZ;d}O$J(P z*ZaM~OYpz(fBYHI8p*c?_HR! zkWtW86^J!NTRYx?_O7O4yeT0o42j>}2LjgaEhb6VJYdv!4&!-;56e$@)^o=MW%u$a z)h+7tK#g{vArT?|N_)VB5Ge~tW-U5c>bQ!#>%rwjf*%L&KY0EB{V^&6 zYvSD>|0^OuC=zORVmp&3@c?wxpfcQcu-p9KzxNvU6;%XR-oFgS*{(<)FbGAyrCKbd zKLhCzb0CW(=ua3h^ZlOy0vJq5?3=rNy|jNBMBO6CNWeXbt{kcH3@+y^j-t$G>e#IR ziO>KK2=p1DhA!X762wDpr;WABrIZSiu>2+pU?Skk1#~+2XTKyF{@+XQIKz&Ka&9H? zhl{)jl24~YEdY4<831OlSl1yn`47hOKX1E9ay8lNPJ_RjplQ~nx~>CqktKWy{1^+s zi#D$p|7q|4az_G$!oe7JiVDA;H)$>!ehTbOSa{_JVbXajRZ*?ak1NlFU?MM;jt{vx zJi2yesZz4Q@#;q0xgUF98AnFpU6B1=S=|-!;PRBR0IK94(D`)810c31dxvd+jy@_+Z#vP1^k>Oe|HVdtjUE|JjY#3ID~py% zbvYr}w>dQ8||eE#!*)|JOChI z8c#B5a(p$p&WHHl8&5T!tP@MOh3wq3_0H_wAK&8;*iQM4Rd)h9Y8A$W&{i!M zFy1ZzjD8xi_idTX>(0}vKtyD0J6Z;EL~6}7HrLa%W?dG6%eQDrO!lZ zd^`0T)`B)7*+1y>GavcnRINLwT}CZib%`*i*!)}Qz50eisK*`jzRN(5r3(PywUeQS zFfh&d3&I8F@YHl+lJ)!u55SzwyswU@I;wJ*X8`IMT%0skum&4ESNq9~ngMF^_dkK= zV3#ci9eB=b`l!Z*RAR%FJWi;E%rDcl&lg&#{~45y^kylKs+R6;{ zuyxLPLBH-AGNMU7D$7`Yej5mV%63FxffG&*a^4}T%#dZ(O=w*8677Af{-h8c#s8T? zSx9cCHQLs;!4uCyYnbEr?IS(C)@$Ry`!(qCbZdMM8g}=))*uYiFQUbRc;SKKxaM!& zhl1uQxIe<1dhP!+d*xFShiMBZdUw{#KXdU%3wDxy?xsrbiczmaOJ!|=@qZo=@=ch>J-{(@4jAbO+h-)s zA`ier&NA7j^@ftF8{9$)&!E-%9n{E$V}ZtK`1TkswZlC7vy{Qi`oNeL_7#+a-Qz@{ zfA;5b9lwPYB^PYbd?e3MOH~fAiI!PF%~?F)XC~RUiaRaOkEQ%4tuOUT2$nzWNf7id zfOYZ^41ExT`zxnO@8p3gEd$jlxMo&ASb^&aS_6nIfPH$7bp#P@&@rgJk>g?s6v2m} zIz&7a7918`C`7!(_=SECeFy)tlYX=#TFiWPBt2ra>yGXuDTL}`Da2Uu^ksWM$PoSD zb3HI@iZ&o_b|U*@$kl?-GT&=x=ue*UxD;JAxRv2J4(i?bwM{MiDP;@b zR!8035acuAUi=~2OMP}2NP?C%i)^J3|BLOpDBY?;NfPrx2|WSICZ zPD%iD<`p`5Hv@@9GmgNBm+O?Zx|3fwkl3;h90-$td)(~?665rNW@>!xSkw(k#*g2` z*eFr757g!jAhV#O20Z^9 zr*dx4`4P3{4;nX#FyMGeFLd>cQq^{*Ap{BDXMX_^>fn5n=a=QfRXR)0SPp2T@&ZzV zbyOd1GILDu^I~w!AgImyzQuoys~w93brxG6$aqir4GuvW0mK<=50gF4OF~jW4RV>j z4NfipS-HsCI#Ej%;<%i4^t)^<;b<*UXe?Q2cFvPm9%vJt)&rxVrsNnwPBs+uR;j=r z`DPAWqK?P5%U{Sd6_^>1;1k@`#9EIhx4xBmt|pHy z(yjX;xU=`{B^=QS^#qfffKyqh5O3;VkB0ll)Z*+=0;QuAq>;JrqlO<|cK>pX#tI3i zhdLhfj59;I{=eehI;yHKY8O_d8>EqLkP-x?Lt45Um6DR~@JI<#(%p@MbO|UZC?QBU z0@96ierux`zxTcWedAsS!@)T0v-e(e&9&ByXFfAH3AY=auxjgOco|qnf!0Kn%1}!d zV~`M$6JVyY0`LMt@q``9bdTO=4`v*KnkmOXkr#eDFbqNnLwwDZlt-ZUPc_wysmyP` zaQEH+jLE;MG%88FUmPv;C|ao-D8BY+D=B(Ig{VIj7YG;AA(6WpHQAH1#k=Oae@+jw zcE(aENcPyl-N%j*()5@$0O>6|1h1aYg7zrgjTJ^8$Q*k_bL3a^RzcyL+GxO5I}ai= z&VH`HCi=tuNIE@VM)GAM^o)aqq3tS%QwXc%_)^LAJpv*67^tUnrH=XzD0Xhfj6kKO zn2%>)SCqFWy+Tx{nF?pl>z#*TZ+djF+sq7}h**3rrH(*@E!$l^i(ZwqE1Vd?Q zvj@|!!}JYVwFi1Z7}cP*Rw`Jz^8tQ2L&&uJ=^nRNL4#$+uEU_?318L1Sahf79x}3H z`7e~}$kAg$k^R}W!Vv~a5|^6&B3>J%I@_+Pw|Eni{nZGUwfma;DGb( zsz=D?)XHso%uE+-WFU@mC?!qjo2i0EwD_13VLp4Q^+LSPGiRSD zmv&bAMfVHTNt&IH9RZy97_}@;yr3>(oVb^60JJNF8*}ZK!`)&KSBA@Kz2-2qT(NPe+67fSA6i;x(sQR2D1UUlaXVz%A zs54>#GDsG=R4$J0^gQq1SvP`rhl}tY)7i?8t9y}w1+B~*%l1j9!X>%i|fepA* z8ldP>T@3l0^p^2+h#u~}gG4-)FAO=rfIAYI|Ixa_&byaqq|U8vXi5pf=W{ECrs z7coQ7^-(I=-8U1QJpn2yr&f{(Bb$5Spkmq{EEGS-4LfobJ*i+fb!7nuBaYZY|5G=~ z_Yq;3ruN=XsRcmXYK*aQ&lZRep&pcIsVt)#GR)74v(2QaXF-uIV`0xmjM^sX zcv9!qj5#|#r!wihEagJrX*z8@?`92bya%KXtU7(Bk~5OlvV^oyn`Z~$lwgGdU?|aQ1?>$b59tT(WF$llwIf5 z6LWXu)#{bF&U;U+&aT(k!(68f9F4R>hPi66gF9bUs^ufM(JY6ymw+TUw&7&c}0#$R-qbvw<0sTmrD6iJyiW2k#{ z;8MS*9ar9cu!FOdeU~AGCv*Qw;>3-cWFQ6xD3&NWhP!pezQ*cU6nrB#&n0%XMAoIz zhEGkuqnZDfEw$@r#0$;9``$6Q3egl%b6n(fZ|@mLI?OWOyb{98959o!@~-G3T$8K1 z>qmZq1^kcpt}(;0dvDVnj=0F}%3DkNP|(W)Q~)7I+@~@vEXol``FQFg#RUF7scN($sj6qLwRR{|yF-|40@ldTt9zhm03~gEDFYs+2{r&$C0Hn|Sh-QC2_G!> zH0Nrdfq1*4%Zo>C5(OuA+x1(};&b_FtC2q-tk zUwtN>@}BM;?OHs&o9*qdBN71Ns6}FHuCck7W`*7TZDcP5#7DTrpznopC~E`9wjKU^ zMq#IfE+n|Wf700ifPlG-^zrwm^<5>9rTJ!g%9wLTX?QI~NEptT#>RU(o;7IEm=6I>mR*{TQ%VV9_=;lK{i+Cj-(>=VqxHUr<}|GL(i`BRw4mVFT?o_- z#fQuBHJET1@4%q+L2*E+r|_~yE;L+ydFbdVVk)$QUJQ7uu>gC@_9P>b;g>MFa6VaM z;!iS(fnJY)O~S?1tWZq`Q-6uB4RPg94gakbl(?U|#PJm&KF7a;T#zeH07OzZCtFDN zpWD@52e^KI`xK@B=$0;k*)GSj6gXG-y-Rx12mtK#I4}dsZ`Jiv8x1`wmCEoW4CxzE zvz?B9$UDp+blKoOH?STI4`DF@>>d)`RVdrC;U5dtXx{<36%w;tntSP*DBok}N;I$j zoE*S@L+<)+HI(;AXLD?7&_-s*S)MDD_(i71b;+}(SJba^Id}WLHds!F4nO+5Sr=Z1 z3L`)VH0lU+5dnPnWXWtbAVwVO=K*M;Bxrg*{opBD zhrASndhT#n@&s-2q*A?xG=wk29m^*(h5#u(>9aLnxzpFkgZmsS$Q>)1r!c)xf|3P> z!x;a0yHJTSK(f^Qi6HRp6uy(!%$;?2Rp%gqDg}uEl$~sw8m>W$UOieJK%#{JQnXLec*S-wm4O7y&UieU%iceZ~CUYC3-zTJoFB$FNT=PmFOA+G2QCh5EX3=urQ{2BNQKFxXGeJ=Uy%*)vS{W~K) z?Bx%Y6@aw=yrRYz{^yoINBCU~3goiA{(8Y9{d;wQUnr%3`0V{&tNuqCj)XvkC!7Qa zNdA2kz+rt#Z^x-p_<2_RUWz^XAQI;^w5w)`D1Esb{XMzbKo;aL}GD zc?P(9x=@+TqmTib>sR97?@P;|;)SD$4R14m2KFIHQImv<_rO+Tp~Gd}v$^u&&$D{$ zATymWECdg<+5J5?&@V7P=r2UOhM+kZ@EQKQ8B+J$g_ryyr`{#^`9;N-|4d>8Aa)f{ zbGt45U;V>|y@iJMKfiIo66iERr|$drK)~43B+uYk$vmWcz;fT5)4w1tTQZuO%VkA3K{rd@&A1qpi}|-9I%i- zWk$6F@>`sEumImOaE?z{Bzo1VYGnom4-bj`ako? z2nSSCq6W1$gX1zv@LLVRFA*RVtbJC-7-g0KLV5Y?OQ7X1_oIv?#@LlBeA+ViMAe`OG(g`5cXLtgN z#!=A3(Nu3%ClXMhHY3yES$sc-aR0z<8+G1N4hP9k8jeIzM$$sK^vxHm3dBwMi3C7P z$4pkhto>e9f~xsF1kv8C2JJvcq1L`cH|FdCEMbffka2?XXIKM#$Af*qDs81}Rq77~ zyr$zp@@h;0Q2YTNX6tsWlH)_rShGksR~^B;O+}7%pl}&cuRDb6i#d>!b-vm6m%1y) zr+{DE66`|&4}t${Ex&t;(z$Va^;^@S+vg|VKS*_5r~O}lZImc)<9u#dl9j4CR(~qc zwL=DEZlw!~bgW-W5EQ&%05(uD{awO%;I><)zO-#JM^?5?wrefWfnrM4AXkt+bq2JK znn=y+1_(zVuf3pf9)5t-VpRVac)AcWYEkS_02b_P@j3g=ZXbpv_y|NSofuyir>xvZo92A&ydksXzs{lBMMNg}F@{ZKb8?Vda79wS>^JUE@lCVR7 zQ6UWBMQ;bJ@Iw$hmxIJN6)K>@={*1>g#opOuB+_`8yIUJSpqA&Thb@x13IF4ux0N8 zM$!=wu@*TTp8f@4n*UZO95Vh#48pmQCWNoIt-73>?uxeK!4xe9Le13or(C#W<0z* zP$kLfe46I$z4JBDUmojJ@~EijRatNs2>b7W%GW?4+Ht6YlN3h_`9Xr~czl?s2gsTo zCLY;MLcj*|V-V4%IC@iR_#$=V@T3-0(=hMEl+v`5;q4X+Gj#X5SgCEuhPHQ!`x6HA z)Y?P=B2@+=jwNTD^)?I*MGje@R#L|x97)`GZ~T?(N9ijtPX7?FuHJu}c_A``hK>-z z8WIdLZk5Iyf}peK^gNkKv@&}45Qw{5fDnG*$DVhfg0u@ToL4F#ZV*#B1m3WJPUiOg z<+On7u=Mr?%b-2z^sIy&h|V8)_K4fR*4VSQCJ4prRa`cx&P_!Nz~H;S{QQAGRltMi z{e2#b1yI!v*Ssw!)&&VXM5gsX({pipdJKf<*T5!R0f?*w?=#QAw!pgmz`XOjJ3|j0 ze7wEr3^HO2aX&q03y3=tBIZA)ZoNWw3d+G+;vb6oyigEIIbcNJvA^ltu*0$_-lG`g z^+K!orjutc*c;8EoI!kZ z^8#b#7nx6_UTQ~1JMx^l*yu^hQe*?t-|<74GV|hZiJ!-a9#;+-@b5^?y0sfF4Ml2& zMASjeFT+3<)~NrX`7KHXIOHSGPOK~+sWdKDFkL1qyJSp#`I9$PmB%ftb;fn>AFR;Y zwXoS?pQ<$i6>8>b`g&xo=(CAN-hL1oc|gc>a)@!g3QFHDWd@OkeFmT#rO`cr{qeUK z+I|TE6NAD9XoY-Oh{04{Z%V*bv{GhXp}7fo>Ez>TqzWs#LX$sWg!h^?$z0zTi4i)M zmYQi5w}%Q~s6h3{h1;}VLOsvEnNK0&ke0R*U z-G+{Nhw#}}_?tCQ0VgwV#Im-}?l!zJ`KFhr0SBf)-xCC?bmmznfRm4yaK-EPxyPta z+fuzNQhheI*OiTa1|e6yQOR>Z`^hcLF^D7m15)Aj?ur{i^)B?Q`;1AwoIK!wz6XRA;V@=d>)2mCh0N81iFzeO^9z+zX?Xm~K+sw=F?9wqtm)BC1qsv_#iyJvR- za47V)e|tHwb*iWPfsD~Ry<9trpW0Uu#gJuMeu%c3l--`iX{LF0@Xg-!#0&ChyF(rJE#6@z^nO2Z%AQZh5{!S4 z)@Ufmi}Ws4XeK?3JT#YiETlc2_dBKl#PV$ {^9Agy4$~ztcxQytxxHf}1=f3hk zFyIfsZGP+I*NA}IGTe-G9=C4z(fUFt?JfWummnoN0^kW`A!viQnB#U17xMt&*2(az zG*Expl<1)?MnwK25sKd1l81HMD-3T*Q4shBRc!msOJq=dK4HJ5iKQ0btHO5QO!N~= z1fVNh;a4$a*^LcE&Af&kg7B7Ax$H|?@LY!>r}J~-P!#G-Q%-j!AOp&Rk(c)2xHL%P zex3D&^yBdYdIInzIxlmpRoQukw+dotUrf4g6XDB{zFw;PhK=!&Av&{c6(t1gJ*IkvG5SN8A=7?}fNaH32V8Ck|cN9?T{GkI< z<;NJNZFR2lX+^1DmS#!Z6eBt38v&MLFJK|WJcHMf{`B<0&k=~;ODj%>jFmS^5*BwmEW02l zONYFa*vO~ILN@j&puHdVp{v`t+uqAgvAo?$O;UP6-nK6@NA2Wus1t-nj3Be0t!e!#nhhq}1Sx?fQPOr3rHWjB)8y5qt_I9#2X-STXU zaFc?sez-wDZ&~_@15tg{wFbBtVtX+)wj8|tVgyMQ5nfnX;ddUE!QI+^12%Ji>n+iz zIQQs@8sn9TCXC>p4u6*w%csr~&o(=6p*X{yPlbGp2X>Fr;)!j(-*n#q=Uj&SEp&Tt za#;vN)50_Q+(xK?J9~+?1uKhP!1yuYQ1t5aKby!Gn#~cV5+tt(Mq@^LZgg=fRIh7q z8CJ8(71NMfU25kEnt0iSBO2@x_6!GSuhj9 z<^eAUfH%kP%) z-jC0=Ij2j!sh(6@PN}^~xQcHhUI7Y+X|2GtgRYotuFf5n)Zs^V@0YI3`wQ&JCcLGuOE z1V3x#r1KYsmq1y8-jT%0|Nu<)k3%5pY6V`1N)}$2FbFCwqW()jF{}o zxA(rf1LEZ?R#XRJvpPm6WAFDlcN1Tc2Tzo*z2B$WRamtlv;;kcbbZ(KJ9 z-1k*CKCK6CX$+~m7S*rhs81=E#fS;kPG@{wlQR*J-EDVoS@6^qO z(B>>`UR{Uz8MV;C)N6ZpJI|qKUuIv=yC*5NQBlIfe{g@$Q{gP0@5!!4{Og*98Q)2q zU9$U89z3iE-m%uYulwu<`YNlOgo~@|B%f(`HK{B1>*Q^`DfVX?R%kG@)PMBoQ4#4F z%dkn~N=fLpmrFkn*?_KdF;zP2wNG6WtU2MY)t4(;)ywi{6mA|QtC%-t=E_^=1o-4Mc-7b%(|$vZZK2g|vd8F&MZlNaU!r=;Qpr@V+fA{3^^3*^ zR}Jk4c5YzWASZm|_pMSzlQCs5tYwpI>@G0)y7AZ|C}>c5SGh>b@y)c@N^+6H$l%Ix z)Lzks(#ZPgQYWXAnTpg@Nxy1k;}0`_g{r!kGPbGRp+>aB_53y2*SrlUGNtz=CZgkt zYD$xyrJmWZb-Ub7IBXm=TREh6sJAM?&#YTMzfydgb&P6k5MSkqpyHf(=EfE%*q*w% zNFXsJ8&{h+giQJL3LG+}k_p_=%zFLE7iWtjgV~no$)UZ32eKf2*D<8?i*QA8kt~?KS%U2e<*^u@jbXSL zojktf8=?%VD{~COhccV!(M{Fr`&la3knn!}M_CJ3icly!wH!At*JHr7`!G**$Z^bw zs#>|qCz9^+D(L!rzbg$V>SBUzoCBmG>$^^5+VOR8c|QkmF^Z=mWeCV>DD^@K`$P1v z+|`hNKLmW88VY_&0u%jdQN1YXw`dYrfA0MGlStSy4y7}tQ6Cs)P zvh=VP7gkqQwK_3|{ka;t_XlbbH1(UEH{kysbfGeM%TUc22eH3j6BGlgm2x;>|07(< zt8nVK4b;D&{QdohIFw2sn>$5&sliCNO-T^`{1GMw?&8l_D&K`{{)#&^K~nV3Bhs9 zagxOUeu{__3U+?sin~3}r4B&Tft_^)#w=-t>Hm9ffV2nbMA8e3S3ZwKGjcdBtM$b{ z#{28Bs6cA?nIHbS%~KKZCU7y{hiSY|3}Ttzpu)cYZ*l|Is6Tp?b^G>hd=-W&(e#IK z7#J9>@sIIm7v5dImB=#uLo~I5sEp>zPcB|fg#6+GyA*frjZ>F5d46j$^w$V-n2nu! z%!0~HWLhbaM>&r>xl#YR%Kt(a^kn7a zL;ohx?8>*gcH(uUlL!H>DQ1^59|rMt0Rq_xNs==(4G``F6xV@p8OuE5T8Xu z?Jb+yp9M`RZnU?$hmUfwK5Rf7i~WhMhxK@QyCmh7iR__^&IfXj-DP}Fi=gA(7Iwl? zQlURKBR0ich9{~j}X)^m_JnLmSLB3M)?f^08;gIe7wr9 zVeMEZ&_rm#5;4@EnwN*{sQBWS?_;6zZ91?E0B{9%v%+qz^ySeL-&t}*>DNnNo@#2^ zUHsZI2|6h{>N8oqxV*Jm`8!R+gcK(Y*MQ}TC(V8qLb_jG2g6&iDHR;=u~v`w&VNbK zm56((`g;|qu-SgOYx<-CKlMWpLm~!Q;`a5-qe^Ne2be4&Xy@iq0Wt;U8pWIoBPg%} z)x1P8-pkGLAzsZKLHSBC8rB`#^bvKPEacT<;1P`po{vk@mqi5(?k1(R($=}zKh&+S`2-~jz=k}Q?IJ7U=lC$|FL z>N>!Yu*g5kRlYn*5n;MdZu3vWmzOm0xV(5i0$glBa)^#JkBuCy2-d^E@UXeOy!>2| z8;S=B=R?2OdbPaKKJu2S&g~<6g`Wxx4_Bm_B%f@~S~E$=%RNuH4e9b0SK)*4YBq$*^<8d* zwq>?RAPNmiDIdtGzpVfM2w6=B+9eiLmRD9-yc(_Bgh5-5wa*M`=r*U8x}V{GYq`iF z0s$L0Xg6C<6V9!Q7So>Uktlfnc7!U>ujJ4WUWFOFiELtbzH*k;NSO&l`a}ELqW`zq zkdJwFzq3x?tFrYKXZ=d?X7Zq&p)d%_kCzjgu5+TJ!aXL;9}&6jx@n~P#0OxRpNy7J z?Mhv0X(WT~{)DMbJ_)Jtq}A5U8*ly%tG%5*(>Oz$!3K}*<@=%7jAH4gfZ#pMvFi(} zTG2os`lSkCi=;ckkSMrGc{<1M(3+N;1s<;qbgcHubs2b%jEnE)Q1}W`Mo_{MG1b~? zJ=_+$)4HvFP>y_GIjVDOB(~nqx-?=K;CE&#bh7cX!|pV%g2FqjmA%F&gz8ppYOOJ&-6+eEi_86is$zM$mr~lmTqZgBkF6Z^ z?<4^Z%LIetOv>GH2h;tjC+je$rr5i1>MTE1qpPbc$XDR=ZSfL&8pkP6%n(#WiXUwF z?@?fk{7MUFaxl;4i_7r))xsjs99Ur7ykWQV?1wP4IiRbw1e$rq$04hgK+C8*r=p^Q zR>A%eZ+dd;r#geKf#^iS9<9em9_hb?^^O;AS$d|Et7U9&XX+E^LHw%nwr9Daa5)+q zM-ufYGnmowqH^aynW-?hz>cM{eqf*!0>wJ-JMfo_N7@mp-{IKK>A7|Hd6Z`aB{ zN3}$7TMmvg2tV{2sj)Lb(ZM(GuX)4;?POw7zXM?yBmL` zY}9-)H$ZhilXLefptxKKN@ednHWEmqxzJU*1^T>M(xysNL-}@kHV*;4)h%5guX9WX zcfVlhrl!3LFkN?%-#e(by|MwlYzp@Hxt*(H&tAqmye6qXi71_oHmqI&PvW z7w75n&&S&Ws*;~yz6*|3oOr06TtK~zP`3+>n_?0V>o(@)EuM7edxO_$sOA@Ysi##1I_^Py*Pd61Sq|^a}YPJ!oq%AV^rNn z52H1+Foxxe8#Or6)v41B4*OGTa{-Afz-g^A%m=(EJwL$W!f5mueqib@FM0(Q30{lX zbD2lKTQ!tIcoH=5kkBHXdRqIj#ufWCN#REkv$m>k?>|G6F=rryAqMvm2 zS|{e+4_5v8wqIU**1mQ{g582FpylzFu)0u-3dvywS17l?3R1Igg03^vnQ?{NrHu8| z%Qv%Yj}+!Lb<>6?XjEM{k-FH_KjVWc$2a>8J|{GuXEHwye>e=efJT3NBWObshBgK;NamXL&vB%Djf7Dl(2M z7eI|wDaaxo+bC^P)KGYh3v2zhrql6YC5^X0hqJvaIoJbVe<<>4Gst67B0wSbNMO2-7y zt(%_j`tbKqI3Y@TK)_y*Sv5b;tUf$0y-0eygD7H6s(mpileM(6Lg&jTnse`jTEMJ<0uN-N zyN`Qm;nc~+CnYZqmh?h+0(Ksk60AfCi0wzGl{B~ks0WqH*fxg;$a9UlKdT4ikG?RYA(Q)oaI_;o{YDrb z$}0mhp{eXx`ia7Yh2>@eLuCe>`c+&+1_VZVR8>@UFg5wiV2Q)xet;}@CA9g0)Z{op z?#@?<7X-CeKU27f3)Ym90|8#x%Ml7N1hi(|J;$+Boh0s|osWcv-Far+95*q@Ean+7 za>ufy&brIM>|`4NU-ccGCqSdD0iiloi~2k++Ijjo+tHy(1kfB;!N)~DNZ z9T0p5ONVDvqbx%9YG_QZ+5AtjFkx#|b*`>ecTIZst-h{qmg&&Z{PfzLPa<@oq6Krl zZb#o?)Z0TRMP-vg1RqEG8FYZ9@R{E%k3Oi8%&wOPo9I8=5&QWGocz zoYmOP_PPf56JI!6uwwikR#%J1D&~n?W?hF!6pn1%1rjn>WnRlIV~aU7y;7ioq%0Bw z1(vucmg71M1y2h+cEH|<=zF@XAzZ0&xZNgv_EKbyGycL1?~i@jYLzHFwL0O_@-4eI z4KG{hDBppcWYxKYwX153EqUjD6qcgy6YE^N)Qelg!5<9eM-gTyrU^mJq$Ba1kF3DP zQ@k)RWLAmTez%i&zJ3ts%_dLwey>eSVRp74cdf}&uZO)$t6*Emb zKU;ru_SgnqBKPBwK>7yfhEeCb#RPhP8u>AsvBpkt@1D6yG%6<~IT+j)NgSh?eH%bTM5&slDoB;h5--eg+ON4;Oi zs`y$a(bH)A2kcsx6HIr!@emI@ObG!5`OVg)E-xx)h(8YUf8;Z(oYcw%3Lw-5#U z!5p(0sU>rVskSD>a=XrB=VExr1mkML&mbi=oS~P=dhBys6^>r}s#=}2U{*o*_<^CM z6;>v6*~DX+_c9&D<#ZR5EV2a@j`?(1*!#rQ(%hV}vC0z{#zp{aOcJh8$R);OO;6te z+`~Y+22u+;XRGCzSU8?NfpiGDtX0UqQL$c6RkTPr(kmWqK87Y4_7>7$;9Meh=V0e= z8(infPZw!8rf&YBivRY?4H6hF+F}*oMDJ^vt4u?lUo(5VGoKi8?Tu242Kpxm$l>YyX(^)1x(1H! zUHHM${$S$?4m?ra+83b%wk^yNEr0CEgzWxS+gcw(nx80>u2Jur+Z>}ircA?xb(QuV zP3c}sd2elpR88hUo?~cJz=hmB*?w(R;)@~eAm`&Vkwl!7J&QCP)GNWy)hJ=>t}Upy z74*%H$WC3qP-S6F^W~9pfi`3Ii$kOUoMAcZk;)Ii9VkD-)mg9iDYnR5Wpo%R`Qgye zWN&h)P~)+`O2tG1b{v3u+a`+>ya?>6y#}kvf}NEi07&1l^y#xOJ|m?frTOpIUZ$}S zJw7x%XMU2yP1^!m*zayoAktMbO5`oNBhWeRiTsQMq=!H{_ppowdU7S4TAzC zz60wp3z)yBB5}yr`E`CVDY@$DIXNGpP%JDMIN>%ECOxmS$`HloqPd*pJlV~PY ztDVKKmTFH4k!fW_e9}S^VzXIPV0Ab3Nf4?0spU~665ccFN^g%bVmq0zykz)VE@^S9 z9EkqaDEg4^C^dGs5iu_3(tOf}!kZ=qto+a%*rRxE`6o1oh(*k!-VwrfGBqgX46z^J zsx#lbtZ797A{^4W)32=SGyK7vC zu+#bJmA0vPVD1rjz9%M(y$QWVksJ9?1TRG@j;nfiV`(I-=7kMR^^Nl4){n(9!??Se zw!O|Nu8rooK2=Y~tY?Mf!UBIqm&hOs)Ulod0rm_5K*CnBd_bgzr<_W>)3lTEAj@i4 z0&W?>X5<#v1nvTZKdHQXigl{BzT=pzA7yu6d7gHf&;#st-fr}JLfpHH)BCMAe#I0m z)WG4Wp4FJ_YzE)oYrBy0Ai{lbWKWUJaI&#e#rvuo)4?J^&7uHFKxPN(ov%Z+u}w$J z=cV=TXo9@`^)AC@S=iYGsfz**wOQ%kZGO+&6}TK!?0i}L?>}M@_Pg<<{I%l=9bb2T4YWrJ_5d)WvN%Ci&MB5Gs)J?B8uVdUB zX^yN?&j*gpFF~Myz&G7@u@BmzfK~DXrr=J`9jw~DQi7Wtpbtj#V7|+w_VMXrVlk## z-P--`?ZrEZQh8P@CsnBI?uHJ+CIQKvLo6Eix^KF52%Rw2n0Hcjm9t9uMof0^jh6HY zG5xYA!KMw~zi_8{p6@jT!u38sp3;2Dc_k$ERJnQxRUlC+?vzh>e`BdK*J}DxAbr=s%8g6g4;)KIR~rTK?+o$rnH&S+Hw zdUF6hRtdlcUjPT#_2a?nMw6BR475S1^94YlP@xH0MQGH-hT+gi4ujfvinpf#&^LT? z_ygQE)fw=7)gVT(-)B`W8fP_4s)6xA$adBE#5;DG^^Xd_fEMFT+D+c4Rsbh#Gj<99 z#$>A=wP;lQt8c`WG_6-UZVru2$`T5JW|BolQy_d6xCVf{XKTqnFeW{^5rt0PSOc$x z0R3U2L&uFL`L`{Jy9QP8ZSnABw+)Js*z#uD%Y zK;kM0QIr933L;Aaj+?R-T_bJm{0gjzK^F)-wyTuByavL7PWo&s<;%?mC0>UBl1vI; zYd719a15FycC3c+ufG9s_)im8_j4dP-rlPf#u)%!aSrF3uttqQ4UE&X51RU@^8-Ww zU~GNV@^$x0Ny)CkO_t>H0$of0h{f#Gk<60Hldihzl&*dx@;7PAZ>d*?hK0r!ysl(s z`D(GTBK-~oTNr_^6ZCnnyk6rbff}DkPUrf|Jm(j<0+91Qjln&Zh zm%7+EFmq336mj=J{_g+snh%5LCa+U!G z9!+cjd&v4As(3aaP@g5dIiwvpM}&U)30!7~CA-2FcijbX4e$yff{?2HfyBy}E9-94 z&Hgr8d3$y#{gJx>zMNQBwNE7{x(p(o6)WRLnZl<_ifbT#fD9KC2BHDHHRNUbnj>!Q zktxn)L*83xEBiZrgBt~TRy0Fp8k&pN-e0|n5P-{GEeLOd#$7uov9J1Wo7%;4ngvy5 zJ^A-S1>3^+>tN?)GG0WhM?83)o9QqLqItdqH+&2%P7mW!kAv_KaQzD*bm$kseZ7^j z>L+9M770h~PBO34XI|K#jDy-3-?0O{{Qm8Mo}O=0nDZ9XM|IDi;x>dT@~aF_0I2I& zbM)ZS4FD(R^5~q%b(v2vL(EWK!4Dpg%{9`H*5)*N%4+*shl1Q8#CrN!dljlDo6=}j zctH0({Icu2I=tSVRJyTmLqmwUUK>j~U+!ec{UN~gDI}2?ko^Hbn+Vmz5@{0c&c$AO zE+_Niv8K*`;SvBiP4&4>OXz~Va$CS{QQ19 z`0Bnno*6R8Z2IZ<1S8{}uhm2i43wN%OyuXzwD|L-e$v#aSL*5SHwK~h7pL>QO(@Ue z&qMaI783hm_o^~`vbu|AgY2Wt(|gNoeyExeCcM@^+q>^np}d!A7N*w`W9iNXf?jVv zE;pTz>6Pp4i5FB0{SNp6X`n#!x1=F_jv)8|Q9?5>4h^A>RiT@2DYn7ft$(67*DS{K ztF;rT4GGr=nFLW=6$Y%vMK{LXS9z!RUg*37v`3Tav#95u{cyXkgiLjG-4wdjGX?ui zo^)Jyn7$XxqU;J0u`b^iuinGBafhR>?xe%Q&7RD8cy}nNE6I$_Em1z+X#GLi-Xom+ zQC6)#d2DHDf|xnY^yY)1dH~T4*E)Sr-_I1C?|`yf1=T$uN~xj^LSEEzve-0)St`rZ zm%>2RXI?gK@Hs!*tDidkLa&%=1=-`iNMkNav$D38v*3<=PD5XZM#+xwG{j9GM$h z;HV-Ix+{vx!xSZsH+nNu4|iV0^JUn=zUXPLId6TOXa1TKXtHp5q=A)!NAqT#U5{zr z_b5b1U#+kp3&lSa#Ly8J)IXi)i=$2y=8>QDqos+3=y)pJnfCie%;101iM$llrXazj zs)q~aWE=LF@F3>;_gV@hs_)O znQ(V8Z+7X2Vy(fhu9QT^ZO}ylO>w2-|)TrKgZE&}V?3Dy2Cj?;- zjpB#0EkJlSw)`{TG_!J;??jm#a?l-(YnuB?(UevDMYkH0xyPxnMuF9;>9m>rjBi(z zM>4$*zHn=5Y`IP>`6A~l@*1_jmLP2Kp`wA{eu0{5iMhoCl@|(T+Kyu>I$r&7IXx;? z%IV4no3C{fj+HxJBA%%ewd0%PWqPuE7REh!zkRG?)NeNM+Pz{idG}@G8mmIrxKzpvG25lLUxG2BbN1qHV_`z~XC(7o`00GRIuSzAc{o3??z><^q7RArF5 zzuelOx1@Xl?YPCoVTNbrdRH{)*)Ite+>g$x(r(NpOO19E#C!^%g zSR4`tiQ0X#N9X9QcykPG6(_l?IKQt;--er*DQpJ5$_$^HOdBys&RPS`A-XkREHlEz zk#e&=UajBxV&HMfA^{z0{PYX1OpdO_eJ06=OKepzYP{s#br4%P!v z#9jE`lLx-9mIpMSuY|?y9{><1(BOw6Ag}({^OaKICS(O)gV5iPYYMS->_Xwt<0JG% z&IiM8+`V<2bP47Ho_jQ2{z16M$6`_>$txgP*wa*flBD~HFs1KezW{L;{UVlve5RPc z;bkC}LjR5n;M*ww*yND*_tF4e>aNVp{rhDiIDoO<%!WMh+cE`f4YmVbyVv1$KGWYL zq`eP3syAq2PXDY%K_H7AJr;}LKdKB_1Hi#seWvusD1p>I90*u&C^+#SwQmM$|Nr(? b4}7%BKOiZi7=&H{|70W-?iGs}J^TLvsqpK+ diff --git a/docs/doc/mysql-setup.html b/docs/doc/mysql-setup.html deleted file mode 100644 index c2e7415..0000000 --- a/docs/doc/mysql-setup.html +++ /dev/null @@ -1,343 +0,0 @@ - - - - - - - - - Run MySQL Server and Client in Docker Containers - Sqlflow - - - - - - - - - - - - - - - - - -

- - - diff --git a/docs/doc/mysql-setup.md b/docs/doc/mysql-setup.md deleted file mode 100644 index 49c3d09..0000000 --- a/docs/doc/mysql-setup.md +++ /dev/null @@ -1,138 +0,0 @@ -# Run MySQL Server and Client in Docker Containers - -The document explains how to setup MySQL in our development environment. - -## Run MySQL Server in a Docker Container - -```bash -docker run --rm \ - -v /tmp/test1:/var/lib/mysql \ - --name mysql01 \ - -e MYSQL_ROOT_PASSWORD=root \ - -e MYSQL_ROOT_HOST='%' \ - -p 3306:3306 \ - -d \ - mysql/mysql-server:8.0 -``` - -- the `-v` option ensures that the database is saved on the host. The default directory where MySQL saves the database is `/var/lib/mysql`. This directory can be configured in `/etc/mysql/my.cnf`, as explained in [this post](https://www.mkyong.com/mysql/where-does-mysql-stored-the-data-in-my-harddisk/). By overlaying the directory `/tmp/test1` on the host to `/var/lib/mysql`, we "cheat" MySQL to save databases on the host. So, we can kill the container and restart it, and the database is still there. - - Please be aware that the directory on the host must be empty the first time we run the above command; otherwise, MySQL would fail to initialize. I figured out this problem after several failures using `docker logs`. - -- the `-e` option sets the root password of MySQL to "root". Feel free to set it to any password you like. - -- the second `-e` options sets `MYSQL_ROOT_HOST` to a [wildcard](https://github.com/docker-library/mysql/issues/241#issuecomment-263011059) so to allow clients connecting to the server via TCP/IP as the user "root". This trick works with MySQL 5.7 and 8.0, but not the most recent under-development version. - -- the `--name` option names the container to `mysql01`, which can be used to refer to this container. - -- the `-p` option maps the port 3306, on which the MySQL server listens, to the same port on the host, so that clients could connect to the server via TCP/IP. - -## Run MySQL Client in the Server Container - -```bash -docker exec -it mysql01 mysql -uroot -p -``` - -This command executes the command `mysql`, which is the command line tool of MySQL, in the container named `mysql01`. - -- The command line flags of `mysql` include `-u`, which specifies the username of MySQL, and `-p`, which makes MySQL prompts for the password. For this example, we should type the password "root", which was set in the previous command. - -- Please wait for a few seconds after the starting of the MySQL server container before we execute the client; otherwise, the startup of the client might fail. - -- Once we get into the MySQL client, we can type SQL commands, e.g., - - ```sql - show databases; - create database yi; - ``` - -## Run Client in a Different Container on the Same Host - -```bash -docker run --rm -it \ - -v /tmp/test1:/var/lib/mysql \ - mysql/mysql-server:8.0 \ - mysql -uroot -p -``` - -- The `-v` option maps the database directory on the host to the client container. This mapping is necessary because, by default, the client talks to the server via Unix socket `/var/lib/mysql/mysql.sock`, which is `/tmp/test1/mysql.sock` on the host. - -## Run Client in a Container on a Remote Host - -```bash -docker run \ - --rm -it \ - mysql/mysql-server:8.0 \ - mysql -h 192.168.1.3 -P 3306 -uroot -p -``` - -- the `-h` option tells the client where the server is running on. In this example, the given IP is the one of the host where I ran the MySQL server container. - -Please be aware that the above command works only if the server allows remote connections. - -## Run Python Client in a Container - -To write a Python client, we need to install the Python package `mysql-connector-python`. - -```Dockerfile -FROM python:2.7 -RUN pip install mysql-connector-python -``` - -Please be aware that some [documents](https://www.w3schools.com/python/python_mysql_getstarted.asp) says that we need to install `mysql-connector`. I tried; but the `mysql.connector.connect` call failed with the error `mysql.connector.errors.NotSupportedError: Authentication plugin 'caching_sha2_password' is not supported`. - -Build the Docker image: - -```bash -docker build -t sqlflow . -``` - -Run the image: - -```bash -docker run --rm -it sqlflow bash -``` - -and we can start Python and run the following Python code snippet - -``` ->>> import mysql.connector ->>> db = mysql.connector.connect(user='root', passwd='root', host='192.168.1.3') ->>> print(db) - -``` - -## Run a Go Client - -In order to connect to a database, you need to import the database's driver first. - -``` -export GOPATH=$HOME/go -go get -u github.com/go-sql-driver/mysql -``` - -`go run` the following file - -```go -package main - -import ( - "database/sql" - "github.com/go-sql-driver/mysql" - "log" -) - -func main() { - testConfig := &mysql.Config{ - User: "root", - Passwd: "root", - Net: "tcp", - Addr: "localhost:3306", - } - db, e := sql.Open("mysql", testConfig.FormatDSN()) - if e != nil { - log.Fatal(e) - } - defer db.Close() -} -``` diff --git a/docs/doc/pipe.html b/docs/doc/pipe.html deleted file mode 100644 index 26204a4..0000000 --- a/docs/doc/pipe.html +++ /dev/null @@ -1,284 +0,0 @@ - - - - - - - - - Piping Responses - Sqlflow - - - - - - - - - - - - - - - - - -
-
- Sqlflow - -
- - -
-
-

This site uses Just the Docs, a documentation theme for Jekyll.

-
-
-
-
-
- -
-
- - Search -
-
-
- - - - -
-
-
- - - -
-

Piping Responses

- -

Streaming Responses

- -

As described in the overall design, a SQLFlow job could be a standard or an extended SQL statemnt, where an extended SQL statement will be translated into a Python program. Therefore, each job might generate up to the following data streams:

- -
    -
  1. standard output, where each element is a line of text,
    2. -
  2. standard error, where each element is a line of text,
    3. -
  3. data rows, where the first element consists of fields name/types, and each of the rest is a row of data,
    4. -
  4. status, where the element could be pending, failed, and succeeded.
    5. -
- -

To create good user experience, we need to pipe these responses from SQLFlow jobs to Jupyter Notebook in real-time.

- -

Stages in the Pipe

- -

The pipe that streams outputs from SQLFlow jobs to the Jupyter Notebook consists of the following stages:

- -
Web browser 
- ↑
- | HTTP
- ↓
-Jupyter Notebook server
- ↑
- | ZeroMQ streams: Shell, IOPub, stdin, Controls, Heartbeat
- ↓
-iPython kernel
- ↑
- | IPython magic command framework
- ↓
-SQLFlow magic command for Jupyter
- ↑
- | gRPC
- ↓
-SQLFlow server
- ↑
- | Go channels
- ↓
-SQLFlow job manager (Go functions)
- ↑
- | IPC with Go's standard library
- ↓ 
-SQLFlow jobs
-
- -

In the above figure, from the SQLFlow magic command to the bottom layer are our work.

- -

Streaming

- -

We have two alternative ideas: multiple streams and a multiplexing stream. -We decided to use a multiplexing stream because we had a unsuccessful trial with the multiple streams idea: we make the job writes to various Go channels and forward each Go channel to a streaming gRPC call, as the following:

- -

Multiple streams

- -

The above figure shows that there are multiple streams between the Jupyter Notebook server and Jupyter kernels. According to the document, there are five: Shell, IOPub, stdin, Control, and Heartbeat. These streams are ZeroMQ streams. We don’t use ZeroMQ, but we can take the idea of having multiple parallel streams in the pipe.

- -
service SQLFlow {
-    rpc File(string sql) returns (int id) {}
-
-    rpc ReadStdout(int id) returns (stream string) {}
-    rpc ReadStderr(int id) returns (stream string) {}
-    rpc ReadData(int id) returns (stream Row) {}
-    rpc ReadStatus(int id) returns (stream int) {}
-}
-
- -

However, we realized that if the user doesn’t call any one of the SQLFlow.Read... call, there would be no forwarding from the Go channel to Jupyter, thus the job would block forever at writing.

- -

A Multiplexing Stream

- -

Another idea is multiplexing all streams into one. For example, we can have only one ZeroMQ stream, where each element is a polymorphic type – could be a text string or a data row.

- -
service SQLFlow {
-    rpc Run(string sql) returns (stream Response) {}
-}
-
-// Only one of the following fields should be set.
-message Response {
-    oneof record {
-        repeated string head = 1;             // Column names.
-        repeated google.protobuf.Any row = 2; // Cells in a row.
-        string log = 3;                       // A line from stderr or stdout.
-    }
-}
-
- - - - -
-
-
-
- - - diff --git a/docs/doc/pipe.md b/docs/doc/pipe.md deleted file mode 100644 index b54e497..0000000 --- a/docs/doc/pipe.md +++ /dev/null @@ -1,91 +0,0 @@ -# Piping Responses - - -## Streaming Responses - -As described in the [overall design](doc/design.md), a SQLFlow job could be a standard or an extended SQL statemnt, where an extended SQL statement will be translated into a Python program. Therefore, each job might generate up to the following data streams: - -1. standard output, where each element is a line of text, -1. standard error, where each element is a line of text, -1. data rows, where the first element consists of fields name/types, and each of the rest is a row of data, -1. status, where the element could be *pending*, *failed*, and *succeeded*. - -To create good user experience, we need to pipe these responses from SQLFlow jobs to Jupyter Notebook in real-time. - - -## Stages in the Pipe - -The pipe that streams outputs from SQLFlow jobs to the Jupyter Notebook consists of the following stages: - -``` -Web browser - ↑ - | HTTP - ↓ -Jupyter Notebook server - ↑ - | ZeroMQ streams: Shell, IOPub, stdin, Controls, Heartbeat - ↓ -iPython kernel - ↑ - | IPython magic command framework - ↓ -SQLFlow magic command for Jupyter - ↑ - | gRPC - ↓ -SQLFlow server - ↑ - | Go channels - ↓ -SQLFlow job manager (Go functions) - ↑ - | IPC with Go's standard library - ↓ -SQLFlow jobs -``` - -In the above figure, from the SQLFlow magic command to the bottom layer are our work. - - -## Streaming - -We have two alternative ideas: multiple streams and a multiplexing stream. -We decided to use a multiplexing stream because we had a unsuccessful trial with the multiple streams idea: we make the job writes to various Go channels and forward each Go channel to a streaming gRPC call, as the following: - -### Multiple streams - -The above figure shows that there are multiple streams between the Jupyter Notebook server and Jupyter kernels. According to the [document](https://jupyter-client.readthedocs.io/en/stable/messaging.html), there are five: Shell, IOPub, stdin, Control, and Heartbeat. These streams are ZeroMQ streams. We don't use ZeroMQ, but we can take the idea of having multiple parallel streams in the pipe. - - -```protobuf -service SQLFlow { - rpc File(string sql) returns (int id) {} - - rpc ReadStdout(int id) returns (stream string) {} - rpc ReadStderr(int id) returns (stream string) {} - rpc ReadData(int id) returns (stream Row) {} - rpc ReadStatus(int id) returns (stream int) {} -} -``` - -However, we realized that if the user doesn't call any one of the `SQLFlow.Read...` call, there would be no forwarding from the Go channel to Jupyter, thus the job would block forever at writing. - -## A Multiplexing Stream - -Another idea is multiplexing all streams into one. For example, we can have only one ZeroMQ stream, where each element is a polymorphic type -- could be a text string or a data row. - -```protobuf -service SQLFlow { - rpc Run(string sql) returns (stream Response) {} -} - -// Only one of the following fields should be set. -message Response { - oneof record { - repeated string head = 1; // Column names. - repeated google.protobuf.Any row = 2; // Cells in a row. - string log = 3; // A line from stderr or stdout. - } -} -``` diff --git a/docs/doc/quickstart.html b/docs/doc/quickstart.html deleted file mode 100644 index cb2b39a..0000000 --- a/docs/doc/quickstart.html +++ /dev/null @@ -1,294 +0,0 @@ - - - - - - - - - Quick start - Sqlflow - - - - - - - - - - - - - - - - - -
-
- Sqlflow - -
- - -
-
-

This site uses Just the Docs, a documentation theme for Jekyll.

-
-
-
-
-
- -
-
- - Search -
-
-
- - - - -
-
-
- - - -
-

Quick start

- -

SQLFlow is currently under active development. For those who are interested in trying -it out, we have provided several demos. Play around with it. Any bug report and -issue are welcomed. :)

- -

Setup

- -
    -
  1. Install Docker.
    2. -
  2. Set up a MySQL server following example/datasets/README.md.
    3. -
  3. Pull the latest SQLFlow Docker image: docker pull sqlflow/sqlflow:latest.
    4. -
- -

Demo 1: Jupyter Notebook

- -
    -
  1. Start a Docker container that runs sqlflowserver and Jupyter Notebook - 
    docker run --rm -it -p 8888:8888 sqlflow/sqlflow:latest \
-bash -c "sqlflowserver --db_user root --db_password root --db_address host.docker.internal:3306 &
-SQLFLOW_SERVER=localhost:50051 jupyter notebook --ip=0.0.0.0 --port=8888 --allow-root"
-
    - -

    If you are running MySQL on the localhost and you are using Docker for Mac, please -be aware the option --db_address host.docker.internal:3306 where -host.docker.internal translates to the host ip address as recommended here.

    - -

    If you are running MySQL on remote, please be aware that MySQL only allows connections from localhost -by default. Fix can be found here.

    -
    2. -
  2. -

    Open a Web browser and direct to localhost:8888 and input the token. Then you -can create notebooks. In a cell, you should be able to type

    - - 
    %%sqlflow
-select 1
-
    -
    3. -
  3. Explore more examples at example.ipynb
    4. -
- -

Demo 2: Command Line Prompt

- -
    -
  1. Start a Docker container that runs SQLFlow command line prompt.
    2. -
- -
docker run -it --rm --net=host sqlflow/sqlflow:latest demo \
---db_user root --db_password root --db_address host.docker.internal:3306
-
- -

You should be able to see the following prompt.

- -
sqlflow>
-
- -

Example

- -
    -
  • Select data - 
    sqlflow> select * from iris.train limit 2;
------------------------------
-+--------------+-------------+--------------+-------------+-------+
-| SEPAL LENGTH | SEPAL WIDTH | PETAL LENGTH | PETAL WIDTH | CLASS |
-+--------------+-------------+--------------+-------------+-------+
-|          6.4 |         2.8 |          5.6 |         2.2 |     2 |
-|            5 |         2.3 |          3.3 |           1 |     1 |
-+--------------+-------------+--------------+-------------+-------+
-
    -
    • -
  • Train a Tensorflow DNNClassifier - 
    sqlflow> SELECT *
-FROM iris.train
-TRAIN DNNClassifier
-WITH n_classes = 3, hidden_units = [10, 20]
-COLUMN sepal_length, sepal_width, petal_length, petal_width
-LABEL class
-INTO sqlflow_models.my_dnn_model;
------------------------------
-...
-Training set accuracy: 0.96721
-Done training
-
    -
    • -
  • Prediction using a trained model - 
    sqlflow> SELECT *
-FROM iris.test
-predict iris.predict.class
-USING sqlflow_models.my_dnn_model;
-
    -
    • -
  • Checkout prediction result - 
    sqlflow> select * from iris.predict limit 10;
-
    -
    • -
- - - - -
-
-
-
- - - diff --git a/docs/doc/quickstart.md b/docs/doc/quickstart.md deleted file mode 100644 index 6a3b67f..0000000 --- a/docs/doc/quickstart.md +++ /dev/null @@ -1,91 +0,0 @@ -# Quick start - -SQLFlow is currently under active development. For those who are interested in trying -it out, we have provided several demos. Play around with it. Any bug report and -issue are welcomed. :) - -## Setup - -1. Install [Docker](https://docs.docker.com/install/). -1. Set up a MySQL server following [example/datasets/README.md](/example/datasets/README.md). -1. Pull the latest SQLFlow Docker image: `docker pull sqlflow/sqlflow:latest`. - -## Demo 1: Jupyter Notebook - -1. Start a Docker container that runs sqlflowserver and Jupyter Notebook - ``` - docker run --rm -it -p 8888:8888 sqlflow/sqlflow:latest \ - bash -c "sqlflowserver --db_user root --db_password root --db_address host.docker.internal:3306 & - SQLFLOW_SERVER=localhost:50051 jupyter notebook --ip=0.0.0.0 --port=8888 --allow-root" - ``` - - If you are running MySQL on the localhost and you are using Docker for Mac, please - be aware the option `--db_address host.docker.internal:3306` where - `host.docker.internal` translates to the host ip address as recommended [here](https://docs.docker.com/docker-for-mac/networking/). - - If you are running MySQL on remote, please be aware that MySQL only allows connections from localhost - by default. Fix can be found [here](https://stackoverflow.com/questions/14779104/how-to-allow-remote-connection-to-mysql). - -1. Open a Web browser and direct to `localhost:8888` and input the token. Then you -can create notebooks. In a cell, you should be able to type - - ``` - %%sqlflow - select 1 - ``` - -1. Explore more examples at [example.ipynb](/example/jupyter/example.ipynb) - -## Demo 2: Command Line Prompt - -1. Start a Docker container that runs SQLFlow command line prompt. - -``` -docker run -it --rm --net=host sqlflow/sqlflow:latest demo \ ---db_user root --db_password root --db_address host.docker.internal:3306 -``` - -You should be able to see the following prompt. - -``` -sqlflow> -``` - -### Example - -- Select data -```sql -sqlflow> select * from iris.train limit 2; ------------------------------ -+--------------+-------------+--------------+-------------+-------+ -| SEPAL LENGTH | SEPAL WIDTH | PETAL LENGTH | PETAL WIDTH | CLASS | -+--------------+-------------+--------------+-------------+-------+ -| 6.4 | 2.8 | 5.6 | 2.2 | 2 | -| 5 | 2.3 | 3.3 | 1 | 1 | -+--------------+-------------+--------------+-------------+-------+ -``` -- Train a Tensorflow [DNNClassifier](https://www.tensorflow.org/api_docs/python/tf/estimator/DNNClassifier) -```sql -sqlflow> SELECT * -FROM iris.train -TRAIN DNNClassifier -WITH n_classes = 3, hidden_units = [10, 20] -COLUMN sepal_length, sepal_width, petal_length, petal_width -LABEL class -INTO sqlflow_models.my_dnn_model; ------------------------------ -... -Training set accuracy: 0.96721 -Done training -``` -- Prediction using a trained model -```sql -sqlflow> SELECT * -FROM iris.test -predict iris.predict.class -USING sqlflow_models.my_dnn_model; -``` -- Checkout prediction result -```sql -sqlflow> select * from iris.predict limit 10; -``` \ No newline at end of file diff --git a/docs/doc/sql_parser.html b/docs/doc/sql_parser.html deleted file mode 100644 index a416243..0000000 --- a/docs/doc/sql_parser.html +++ /dev/null @@ -1,265 +0,0 @@ - - - - - - - - - Extended SQL Parser Design - Sqlflow - - - - - - - - - - - - - - - - - -
-
- Sqlflow - -
- - -
-
-

This site uses Just the Docs, a documentation theme for Jekyll.

-
-
-
-
-
- -
-
- - Search -
-
-
- - - - -
-
-
- - - -
-

Extended SQL Parser Design

- -

This documentation explains the technical decision made in building a SQL -parser in Go. It is used to parsed the extended SELECT syntax of SQL that -integrates TensorFlow Estimators.

- - - -

Lexer and Parser Generator

- -

In 2001, when I was in graduate school, I defined an extended SQL syntax for -querying distributed relational databases, as part of the course project of -Distributed RDBMS by Prof. Li-zhu Zhou. I wrote the parser using bison (a -modern version of yacc) and flex (a modern version of -lex). yacc and lex generate C code; -bison and flex generate C++ code. However, this time, I’d use Go.

- -

I surveyed goyacc, a -standard Go tool. The usage is very similar to that of yacc and bison. -However, the Go toolchain doesn’t provide a tool like lex/flex.

- -

Google revealed golex, which is out of -maintenance.

- -

The Medium -post -recommends Ragel, which is a C++ -program and could generate Go lexer; however, it lacks documents.

- -

Handwritten Lexer and Parser

- -

Some documents, including this -one -recommends handwriting lexers. However, it doesn’t explain how to write the -parser.

- -

GoAcademy always provides high-quality tech blog posts. This -one is from the -author of InfluxDB. However, I -stopped at where it explains wrapping a SQL statement as a string by an -io.Reader, because it is obvious that we should keep the string as a string so -that that token strings could refer to the same memory storage of the SQL -statement.

- -

Following a link in the above GoAcademy post, I found Rob Pike’s excellent talk -on how to write a lexer in Go in 2011. Many works after that change Rob’s -implementation somehow but always lead to longer and less comprehensible -codebases.

- -

The Choice

- -

Therefore, I wrote the lexer and parser both following Rob Pike’s idea. After -few days work, I realized that:

- -
    -
  1. I should borrow the idea from Rob to represent SQL statements as strings, but -not io.Reader as other work do,
    2. -
  2. but no need to use channels and goroutines at all, and
    3. -
  3. it is technically intractable to write a SQL lexer/parser manually.
    4. -
- -

So, I switched to write a lexer manually, and to generate the parser using -goyacc. During my work, I referred to this -example and -the official yacc -manual for details -about operator association and precedence.

- - - - -
-
-
-
- - - diff --git a/docs/doc/sql_parser.md b/docs/doc/sql_parser.md deleted file mode 100644 index d45f7ee..0000000 --- a/docs/doc/sql_parser.md +++ /dev/null @@ -1,65 +0,0 @@ -# Extended SQL Parser Design - -This documentation explains the technical decision made in building a SQL -parser in Go. It is used to parsed the extended SELECT syntax of SQL that -integrates TensorFlow Estimators. - -## Related Work - -### Lexer and Parser Generator - -In 2001, when I was in graduate school, I defined an extended SQL syntax for -querying distributed relational databases, as part of the course project of -Distributed RDBMS by Prof. Li-zhu Zhou. I wrote the parser using [`bison` (a -modern version of `yacc`) and `flex` (a modern version of -`lex`)](http://dinosaur.compilertools.net/). `yacc` and `lex` generate C code; -`bison` and `flex` generate C++ code. However, this time, I'd use Go. - -I surveyed [`goyacc`](https://godoc.org/golang.org/x/tools/cmd/goyacc), a -standard Go tool. The usage is very similar to that of `yacc` and `bison`. -However, the Go toolchain doesn't provide a tool like `lex`/`flex`. - -Google revealed [`golex`](https://github.com/cznic/golex), which is out of -maintenance. - -The [Medium -post](https://medium.com/@mhamrah/lexing-with-ragel-and-parsing-with-yacc-using-go-81e50475f88f) -recommends [Ragel](http://www.colm.net/open-source/ragel/), which is a C++ -program and could generate Go lexer; however, it lacks documents. - -### Handwritten Lexer and Parser - -Some documents, including [this -one](https://hackthology.com/writing-a-lexer-in-go-with-lexmachine.html) -recommends handwriting lexers. However, it doesn't explain how to write the -parser. - -GoAcademy always provides high-quality tech blog posts. [This -one](https://blog.gopheracademy.com/advent-2014/parsers-lexers/) is from the -author of [InfluxDB](https://github.com/influxdata/influxdb). However, I -stopped at where it explains wrapping a SQL statement as a string by an -`io.Reader`, because it is obvious that we should keep the string as a string so -that that token strings could refer to the same memory storage of the SQL -statement. - -Following a link in the above GoAcademy post, I found Rob Pike's excellent talk -on how to write a lexer in Go in 2011. Many works after that change Rob's -implementation somehow but always lead to longer and less comprehensible -codebases. - -## The Choice - -Therefore, I wrote the lexer and parser both following Rob Pike's idea. After -few days work, I realized that: - -1. I should borrow the idea from Rob to represent SQL statements as strings, but - not `io.Reader` as other work do, -1. but no need to use channels and goroutines at all, and -1. it is technically intractable to write a SQL lexer/parser manually. - -So, I switched to write a lexer manually, and to generate the parser using -goyacc. During my work, I referred to [this -example](https://github.com/golang-samples/yacc/blob/master/simple/calc.y) and -the official [`yacc` -manual](https://www.epaperpress.com/lexandyacc/download/yacc.pdf) for details -about operator association and precedence. diff --git a/docs/doc/submitter.html b/docs/doc/submitter.html deleted file mode 100644 index ddcb2ff..0000000 --- a/docs/doc/submitter.html +++ /dev/null @@ -1,288 +0,0 @@ - - - - - - - - - Submitter - Sqlflow - - - - - - - - - - - - - - - - - -
-
- Sqlflow - -
- - -
-
-

This site uses Just the Docs, a documentation theme for Jekyll.

-
-
-
-
-
- -
-
- - Search -
-
-
- - - - -
-
-
- - - -
-

Submitter

- -

A submitter is a pluggable module in SQLFlow that is used to submit an ML job to a third party computation service.

- -

Workflow

- -

When a user types in an extended SQL statement, SQLFlow first parses and semantically verifies the statement. Then SQLFlow either runs the ML job locally or submits the ML job to a third party computation service.

- -

- -

In the latter case, SQLFlow produces a job description (TrainDescription or PredictDescription) and hands it over to the submitter. For a training SQL, SQLFlow produces TrainDescription; for prediction SQL, SQLFlow produces PredDescription. The concrete definition of the description looks like the following

- -
type ColumnType struct {
-    Name             string // e.g. sepal_length
-    DatabaseTypeName string // e.g. FLOAT
-}
-
-// SELECT *
-// FROM iris.train
-// TRAIN DNNClassifier
-// WITH
-//   n_classes = 3,
-//   hidden_units = [10, 20]
-// COLUMN sepal_length, sepal_width, petal_length, petal_width
-// LABEL class
-// INTO sqlflow_models.my_dnn_model;
-
-type TrainDescription struct {
-    StandardSelect string       // e.g. SELECT * FROM iris.train
-    Estimator      string       // e.g. DNNClassifier
-    Attrs          map[string]string // e.g. {{"n_classes", "3"}, {"hidden_units", "[10, 20]"}}
-    X              []ColumnType // e.g. {{"sepal_length", "FLOAT"}, ...}
-    Y              ColumnType   // e.g. {"class", "INT"}
-    ModelName      string       // e.g. my_dnn_model
-}
-
-// SELECT *
-// FROM iris.test
-// PREDICT iris.predict.class
-// USING sqlflow_models.my_dnn_model;
-type PredDescription struct {
-    StandardSelect string // e.g. SELECT * FROM iris.test
-    TableName      string // e.g. iris.predict
-    ModelName      string // e.g. my_dnn_model
-}
-
- -

Submitter Interface

- -

The submitter interface should provide two functions Train and Predict. The detailed definition can be the following

- -
type Submitter interface {
-    // Train executes a ML training job and streams job's response through writer.
-    // A typical Train function should include
-    // - Loading the training data
-    // - Initializing the model
-    // - model.train
-    // - Saving the trained model to a persistent storage
-    Train(desc TrainDescription, writer PipeWriter) error
-    // Predict executes a ML predicting job and streams job's response through writer
-    // A typical Predict function should include
-    // - Loading the model from a persistent storage
-    // - Loading the prediction data
-    // - model.predict
-    // - Writing the prediction result to a table
-    Predict(desc PredictDescription, writer PipeWriter) error
-}
-
- -

Register a submitter

- -

A new submitter can be added as

- -
import (
-    ".../my_submitter"
-    ".../sqlflow/sql"
-)
-
-func main() {
-    // ...
-    sql.Register(my_submitter.NewSubmitter())
-    // ...
-    for {
-    	sql := recv()
-    	sql.Run(sql)
-    }
-}
-
- -

where sql.Register will put my_submitter instance to package level registry. During sql.Run, it will check whether there is a submitter registered. If there is, sql.Run will run either submitter.Train or submitter.Predict.

- - - - -
-
-
-
- - - diff --git a/docs/doc/syntax.html b/docs/doc/syntax.html deleted file mode 100644 index 950339d..0000000 --- a/docs/doc/syntax.html +++ /dev/null @@ -1,358 +0,0 @@ - - - - - - - - - SQLFlow: Design Doc - Sqlflow - - - - - - - - - - - - - - - - - -
-
- Sqlflow - -
- - -
-
-

This site uses Just the Docs, a documentation theme for Jekyll.

-
-
-
-
-
- -
-
- - Search -
-
-
- - - - -
-
-
- - - -
-

SQLFlow: Design Doc

- -

What is SQLFlow

- -

SQLFlow is a bridge that connects a SQL engine, for example, MySQL, Hive, SparkSQL, Oracle, or SQL Server, and TensorFlow and other machine learning toolkits. SQLFlow extends the SQL syntax to enable model training and inference.

- - - -

We could write simple machine learning prediction (or scoring) algorithms in SQL using operators like DOT_PRODUCT. However, this requires copy-n-pasting model parameters from the training program into SQL statements.

- -

Some proprietary SQL engines provide extensions to support machine learning.

- -

Microsoft SQL Server

- -

Microsoft SQL Server has the machine learning service that runs machine learning programs in R or Python as an external script:

- -
CREATE PROCEDURE generate_linear_model
-AS
-BEGIN
-    EXEC sp_execute_external_script
-    @language = N'R'
-    , @script = N'lrmodel <- rxLinMod(formula = distance ~ speed, data = CarsData);
-        trained_model <- data.frame(payload = as.raw(serialize(lrmodel, connection=NULL)));'
-    , @input_data_1 = N'SELECT [speed], [distance] FROM CarSpeed'
-    , @input_data_1_name = N'CarsData'
-    , @output_data_1_name = N'trained_model'
-    WITH RESULT SETS ((model varbinary(max)));
-END;
-
- -

A challenge to the users is that they need to know not only SQL but also R or Python, and they must be capable of writing machine learning programs in R or Python.

- -

Teradata SQL for DL

- -

Teradata also provides a RESTful service, which is callable from the extended SQL SELECT syntax.

- -
SELECT * FROM deep_learning_scorer(
-  ON (SELECT * FROM cc_data LIMIT 100)
-  URL('http://localhost:8000/api/v1/request')
-  ModelName('cc')
-  ModelVersion('1')
-  RequestType('predict')
-  columns('v1', 'v2', ..., 'amount')
-)
-
- -

The above syntax couples the deployment of the service (the URL in the above SQL statement) with the algorithm.

- -

Google BigQuery

- -

Google BigQuery enables machine learning in SQL by introducing the CREATE MODEL statement.

- -
CREATE MODEL dataset.model_name
-  OPTIONS(model_type='linear_reg', input_label_cols=['input_label'])
-AS SELECT * FROM input_table;
-
- -

Currently, BigQuery only supports two simple models: linear regression and logistic regression.

- -

Design Goal

- -

None of the above meets our requirement.

- -

First of all, we want to build an open source software. Also, we want it to be extensible:

- -
    -
  • -

    We want it extensible to many SQL engines, instead of targeting any one of them. Therefore, we don’t want to build our syntax extension on top of user-defined functions (UDFs); otherwise, we’d have to implement them for each SQL engine.

    -
    • -
  • -

    We want the system extensible to support sophisticated machine learning models and toolkits, including TensorFlow for deep learning and xgboost for trees.

    -
    • -
- -

Another challenge is that we want SQLFlow to be flexible enough to configure and run cutting-edge algorithms, including specifying feature crosses. At the same time, we want SQLFlow easy to learn – at least, no Python or R code embedded in the SQL statements, and integrate hyperparameter estimation.

- -

We understand that a key to address the above challenges is the syntax of the SQL extension. To craft a highly-effective and easy-to-learn syntax, we need user feedback and fast iteration. Therefore, we’d start from a prototype that supports only MySQL and TensorFlow. We plan to support more SQL engines and machine learning toolkits later.

- -

Design Decisions

- -

As the beginning of the iteration, we propose an extension to the SQL SELECT statement. We are not going a new statement way like that BigQuery provides CREATE MODEL, because we want to maintain a loose couple between our system and the underlying SQL engine, and we cannot create the new data type for the SQL engine, like CREATE MODEL requires.

- -

We highly appreciate the work of TensorFlow Estimator, a high-level API for deep learning. The basic idea behind Estimator is to implement each deep learning model, and related training/testing/evaluating algorithms as a Python class derived from tf.estimator.Estimator. As we want to keep our SQL syntax simple, we would make the system extensible by calling estimators contributed by machine learning experts and written in Python.

- -

The SQL syntax must allow users to set Estimator attributes (parameters of the Python class’ constructor, and those of train, evaluate, or predict). Users can choose to use default values. We have a plan to integrate our hyperparameter estimation research into the system to optimize the default values.

- -

Though estimators derived from tf.estimator.Estimator run algorithms as TensorFlow graphs; SQLFlow doesn’t restrict that the underlying machine learning toolkit has to be TensorFlow. Indeed, as long as an estimator provides methods of train, evaluate, and predict, SQLFlow doesn’t care if it calls TensorFlow or xgboost. Precisely, what SQLFlow expect is an interface like the following:

- -
class AnEstimatorClass:
-  __init__(self, **kwargs)
-  train(self, **kwargs)
-  evaluate(self, **kwargs)
-  predict(self, **kwargs)
-
- -

We also want to reuse the feature columns API from Estimator, which allows users to columns of tables in a SQL engine to features to the model.

- -

Extended SQL Syntax

- -

Again, just as the beginning of the iteration, we propose the syntax for training as

- -
SELECT * FROM kaggle_credit_fraud_training_data
-LIMIT 1000
-TRAIN DNNClassifier       /* a pre-defined TensorFlow estimator, tf.estimator.DNNClassifier */
-WITH layers=[100, 200],   /* a parameter of the Estimator class constructor */
-     train.batch_size = 8 /* a parameter of the Estimator.train method */
-COLUMN *,                 /* all columns as raw features */
-       cross(v1, v9, v28) /* plus a derived (crossed) column */
-LABEL class
-INTO sqlflow_models.my_model_table;      /* saves trained model parameters and features into a table */
-
- -

We see the redundancy of * in two clauses: SELECT and COLUMN. The following alternative can avoid the redundancy, but cannot specify the label.

- -
SELECT *                  /* raw features or the label? */
-       corss(v1, v9, v28) /* derived featuers */
-FROM kaggle_credit_fraud_training_data
-
- -

Please be aware that we save the trained models into tables, instead of a variable maintained by the underlying SQL engine. To invent a new variable type to hold trained models, we’d make our system tightly integrated with the SQL engine, and harms the extensibility to other engines.

- -

The result table should include the following information:

- -
    -
  1. The estimator name, e.g., DNNClassifier in this case.
    2. -
  2. Estimator attributes, e.g., layer and train.batch_size.
    3. -
  3. The feature mapping, e.g., * and cross(v1, v9, v28).
    4. -
- -

Similarly, to infer the class (fraud or regular), we could

- -
SELECT * FROM kaggle_credit_fraud_development_data
-PREDICT kaggle_credit_fraud_development_data.class
-USING sqlflow_models.my_model_table;
-
- -

System Architecture

- -

A Conceptual Overview

- -

In the prototype, we use the following architecture:

- -
SQL statement -> our SQL parser --standard SQL-> MySQL
-                                \-extended SQL-> code generator -> execution engine
-
- -

In the prototype, the code generator generates a Python program that trains or predicts. In either case,

- -
    -
  1. it retrieves the data from MySQL via MySQL Connector Python API,
    2. -
  2. optionally, retrieves the model from MySQL,
    3. -
  3. trains the model or predicts using the trained model by calling the user specified TensorFlow estimator,
    4. -
  4. and writes the trained model or prediction results into a table.
    5. -
- -

Working with Jupyter Notebook and Kubernetes

- -

The following figures shows the system components and their runtime environment. The left part shows how to run the system on a PC/laptop, the right part shows how to run it on a Kubernetes cluster.

- -

- - - - -
-
-
-
- - - diff --git a/docs/doc/syntax.md b/docs/doc/syntax.md deleted file mode 100644 index 26694f0..0000000 --- a/docs/doc/syntax.md +++ /dev/null @@ -1,160 +0,0 @@ -# SQLFlow: Design Doc - -## What is SQLFlow - -SQLFlow is a bridge that connects a SQL engine, for example, MySQL, Hive, SparkSQL, Oracle, or SQL Server, and TensorFlow and other machine learning toolkits. SQLFlow extends the SQL syntax to enable model training and inference. - -## Related Work - -We could write simple machine learning prediction (or scoring) algorithms in SQL using operators like [`DOT_PRODUCT`](https://thenewstack.io/sql-fans-can-now-develop-ml-applications/). However, this requires copy-n-pasting model parameters from the training program into SQL statements. - -Some proprietary SQL engines provide extensions to support machine learning. - -### Microsoft SQL Server - -Microsoft SQL Server has the [machine learning service](https://docs.microsoft.com/en-us/sql/advanced-analytics/tutorials/rtsql-create-a-predictive-model-r?view=sql-server-2017) that runs machine learning programs in R or Python as an external script: - -```sql -CREATE PROCEDURE generate_linear_model -AS -BEGIN - EXEC sp_execute_external_script - @language = N'R' - , @script = N'lrmodel <- rxLinMod(formula = distance ~ speed, data = CarsData); - trained_model <- data.frame(payload = as.raw(serialize(lrmodel, connection=NULL)));' - , @input_data_1 = N'SELECT [speed], [distance] FROM CarSpeed' - , @input_data_1_name = N'CarsData' - , @output_data_1_name = N'trained_model' - WITH RESULT SETS ((model varbinary(max))); -END; -``` - -A challenge to the users is that they need to know not only SQL but also R or Python, and they must be capable of writing machine learning programs in R or Python. - -### Teradata SQL for DL - -Teradata also provides a [RESTful service](https://www.linkedin.com/pulse/sql-deep-learning-sql-dl-omri-shiv), which is callable from the extended SQL SELECT syntax. - -```sql -SELECT * FROM deep_learning_scorer( - ON (SELECT * FROM cc_data LIMIT 100) - URL('http://localhost:8000/api/v1/request') - ModelName('cc') - ModelVersion('1') - RequestType('predict') - columns('v1', 'v2', ..., 'amount') -) -``` - -The above syntax couples the deployment of the service (the URL in the above SQL statement) with the algorithm. - -### Google BigQuery - -Google [BigQuery](https://cloud.google.com/bigquery/docs/bigqueryml-intro) enables machine learning in SQL by introducing the `CREATE MODEL` statement. - -```sql -CREATE MODEL dataset.model_name - OPTIONS(model_type='linear_reg', input_label_cols=['input_label']) -AS SELECT * FROM input_table; -``` - -Currently, BigQuery only supports two simple models: linear regression and logistic regression. - -## Design Goal - -None of the above meets our requirement. - -First of all, we want to build an open source software. Also, we want it to be extensible: - -- We want it extensible to many SQL engines, instead of targeting any one of them. Therefore, we don't want to build our syntax extension on top of user-defined functions (UDFs); otherwise, we'd have to implement them for each SQL engine. - -- We want the system extensible to support sophisticated machine learning models and toolkits, including TensorFlow for deep learning and [xgboost](https://github.com/dmlc/xgboost) for trees. - -Another challenge is that we want SQLFlow to be flexible enough to configure and run cutting-edge algorithms, including specifying [feature crosses](https://www.tensorflow.org/api_docs/python/tf/feature_column/crossed_column). At the same time, we want SQLFlow easy to learn -- at least, no Python or R code embedded in the SQL statements, and integrate hyperparameter estimation. - -We understand that a key to address the above challenges is the syntax of the SQL extension. To craft a highly-effective and easy-to-learn syntax, we need user feedback and fast iteration. Therefore, we'd start from a prototype that supports only MySQL and TensorFlow. We plan to support more SQL engines and machine learning toolkits later. - -## Design Decisions - -As the beginning of the iteration, we propose an extension to the SQL SELECT statement. We are not going a new statement way like that BigQuery provides `CREATE MODEL`, because we want to maintain a loose couple between our system and the underlying SQL engine, and we cannot create the new data type for the SQL engine, like `CREATE MODEL` requires. - -We highly appreciate the work of [TensorFlow Estimator](https://www.tensorflow.org/guide/estimators), a high-level API for deep learning. The basic idea behind Estimator is to implement each deep learning model, and related training/testing/evaluating algorithms as a Python class derived from `tf.estimator.Estimator`. As we want to keep our SQL syntax simple, we would make the system extensible by calling estimators contributed by machine learning experts and written in Python. - -The SQL syntax must allow users to set Estimator attributes (parameters of the Python class' constructor, and those of `train`, `evaluate`, or `predict`). Users can choose to use default values. We have a plan to integrate our hyperparameter estimation research into the system to optimize the default values. - -Though estimators derived from `tf.estimator.Estimator` run algorithms as TensorFlow graphs; SQLFlow doesn't restrict that the underlying machine learning toolkit has to be TensorFlow. Indeed, as long as an estimator provides methods of `train`, `evaluate`, and `predict`, SQLFlow doesn't care if it calls TensorFlow or xgboost. Precisely, what SQLFlow expect is an interface like the following: - -```python -class AnEstimatorClass: - __init__(self, **kwargs) - train(self, **kwargs) - evaluate(self, **kwargs) - predict(self, **kwargs) -``` - -We also want to reuse the [feature columns API](https://www.tensorflow.org/guide/feature_columns) from Estimator, which allows users to columns of tables in a SQL engine to features to the model. - - -## Extended SQL Syntax - -Again, just as the beginning of the iteration, we propose the syntax for training as - -```sql -SELECT * FROM kaggle_credit_fraud_training_data -LIMIT 1000 -TRAIN DNNClassifier /* a pre-defined TensorFlow estimator, tf.estimator.DNNClassifier */ -WITH layers=[100, 200], /* a parameter of the Estimator class constructor */ - train.batch_size = 8 /* a parameter of the Estimator.train method */ -COLUMN *, /* all columns as raw features */ - cross(v1, v9, v28) /* plus a derived (crossed) column */ -LABEL class -INTO sqlflow_models.my_model_table; /* saves trained model parameters and features into a table */ -``` - -We see the redundancy of `*` in two clauses: `SELECT` and `COLUMN`. The following alternative can avoid the redundancy, but cannot specify the label. - -```sql -SELECT * /* raw features or the label? */ - corss(v1, v9, v28) /* derived featuers */ -FROM kaggle_credit_fraud_training_data -``` - -Please be aware that we save the trained models into tables, instead of a variable maintained by the underlying SQL engine. To invent a new variable type to hold trained models, we'd make our system tightly integrated with the SQL engine, and harms the extensibility to other engines. - -The result table should include the following information: - -1. The estimator name, e.g., `DNNClassifier` in this case. -1. Estimator attributes, e.g., `layer` and `train.batch_size`. -1. The feature mapping, e.g., `*` and `cross(v1, v9, v28)`. - -Similarly, to infer the class (fraud or regular), we could - -```sql -SELECT * FROM kaggle_credit_fraud_development_data -PREDICT kaggle_credit_fraud_development_data.class -USING sqlflow_models.my_model_table; -``` - -## System Architecture - -### A Conceptual Overview - -In the prototype, we use the following architecture: - -``` -SQL statement -> our SQL parser --standard SQL-> MySQL - \-extended SQL-> code generator -> execution engine -``` - -In the prototype, the code generator generates a Python program that trains or predicts. In either case, - -1. it retrieves the data from MySQL via [MySQL Connector Python API](https://dev.mysql.com/downloads/connector/python/), -1. optionally, retrieves the model from MySQL, -1. trains the model or predicts using the trained model by calling the user specified TensorFlow estimator, -1. and writes the trained model or prediction results into a table. - -### Working with Jupyter Notebook and Kubernetes - -The following figures shows the system components and their runtime environment. The left part shows how to run the system on a PC/laptop, the right part shows how to run it on a Kubernetes cluster. - -![](doc/figures/sqlflow-arch.png) diff --git a/docs/doc/walkthrough.html b/docs/doc/walkthrough.html deleted file mode 100644 index 8e32a6a..0000000 --- a/docs/doc/walkthrough.html +++ /dev/null @@ -1,272 +0,0 @@ - - - - - - - - - SQLFlow: Code Walkthrough - Sqlflow - - - - - - - - - - - - - - - - - -
-
- Sqlflow - -
- - -
-
-

This site uses Just the Docs, a documentation theme for Jekyll.

-
-
-
-
-
- -
-
- - Search -
-
-
- - - - -
-
-
- - - -
-

SQLFlow: Code Walkthrough

- -

User Experience

- -

SQLFlow allows users to write SQL programs with extended syntax in Jupyter Notebook or a command-line tool.

- -

The following SQL statements train a TensorFlow model named DNNClassifier, which is a Python class derived from tf.estimator.Estimator:

- -

-SELECT * FROM a_table TRAIN DNNClassifier WITH learning_rate=0.01 INTO sqlflow_models.my_model;
-
- -

And the following statement uses the trained model for prediction.

- -
SELECT * FROM b_table PREDICT b_table.predicted_label USING sqlflow_models.my_model;
-
- -

Please be aware that the part in the above statements before the extended keyword TRAIN and PREDICT is a standard SQL statement. This feature simplifies the implementation of the SQLFlow system.

- -

System Implementation

- -

If a SQL statement is of the standard syntax, SQLFlow throws it to the SQL engine and redirects the output to the user; otherwise, SQLFlow translates the statement of extended syntax into a Python program. Currently, it generates a program that throws the standard-syntax part of SELECT to MySQL, reads the results in the train-loop of a TensorFlow program. We will talk about how to extend SQLFlow to connect more SQL engines like Oracle, Hive, and SparkSQL, and generates more types of machine learning programs that calls distributed TensorFlow, PyTorch, and xgboost later. Before that, let us explain the system components.

- -

SQLFlow as a gRPC Server

- -

SQLFlow is a gRPC server, which can connect with multiple clients. A typical client is pysqlflow, the SQLFlow plugin for Jupyter Notebook server. Another once is a text-based client /cmd/sqlflowserver/main.go.

- -
Jupyter Notebook          ---(SQL statements)-->       SQLFlow gRPC server
-(SQLFlow magic command)   <--(a stream of messages)--
-
- -

The protobuf definition of the gRPC service is at /server/proto/sqlflow.proto. The return of the method SQLFlow.Run is a stream of Reponses, where each represents either a table header, a row, or a log message. The header and rows are usually from a standard SQL statement, for example, SELECT or DESCRIBE, and the log messages are usually from the running of a generated Python program.

- -

SQLFlow in the gRPC Server

- -

Once the SQLFlow server receives a batch of SQL statements via a gRPC call, it runs the following steps for each statement:

- -
    -
  1. the parser to generate parsing result,
    2. -
  2. the verifier to verify the semantics given the parsing result,
    3. -
  3. the code generator to generate a Python program, or the submitter, from the parsing result,
    4. -
  4. the executor that runs the submitter locally.
    5. -
- -

Step 3. and 4. are only for a SQL statement of extended syntax; otherwise, SQLFlow server proxies the standard-syntax statement to the SQL engine.

- -

The executor calls Go’s standard package that captures the stdout and stderr from the submitter process and passing the result back to the gRPC client. Therefore, it is the responsibility of the submitter to print log messages to its stderr and stdout.

- -

Minimal Viable Product

- -

In the minimal viable product (MVP) of SQLFlow, the code generator generates a Python program consists of two parts:

- -
    -
  1. throw the standard SELECT part in the extended-syntax statement to MySQL via ODBC, and
    2. -
  2. a loop that reads outputs from the run of the SELECT statement and trains the model (or, using a trained model to predict).
    3. -
- -

The training part calls TensorFlow to update the parameters of the model specified in the TRAIN clause.

- -

Extensibility

- -

By writing more code generators, we could extend SQLFlow to support more SQL engines, e.g., Hive and Oracle, and use machine learning toolkits, e.g., PyTorch and xgboost, in addition to TensorFlow, on various computing platforms. You are welcome to add more code generators such as

- -
    -
  • codegen_distributed_tf.go to generate a submitter program similar to the MVP but runs a distributed TensorFlow training job.
    • -
  • codegen_kubernetes_tf.go to launch a distributed TensorFlow job on a Kubernetes cluster, other than running locally, in the same container as where SQLFlow gRPC server resides.
    • -
  • codegen_gcloud_pytorch.go to launch a submitter that calls PyTorch instead of TensorFlow for training on the Google Cloud.
    • -
- -

Job Management

- -

The current design of the gRPC interface assumes that the connection between the client, e.g., the Jupyter Notebook, and the SQLFlow server keeps alive during the running of the training program. This assumption is reasonable because even if the user closes her/his Web browser and disconnect to the Jupyter Notebook server, the connection between Jupyter to SQLFlow server might keep alive. However, this might not be robust enough if the Jupyter Notebook server runs on a user’s laptop and gets killed. In such a case, the gRPC server cannot stream the messages back to the client and would cause the failure of the submitter.

- -

A solution is to change the gRPC interface of SQLFlow server to have a method that files a job and returns immediately, and another method to get a batch of recent messages given a job ID. We will make a design for that soon.

- - - - -
-
-
-
- - - diff --git a/docs/doc/walkthrough.md b/docs/doc/walkthrough.md deleted file mode 100644 index 66bf11d..0000000 --- a/docs/doc/walkthrough.md +++ /dev/null @@ -1,71 +0,0 @@ -# SQLFlow: Code Walkthrough - -## User Experience - -SQLFlow allows users to write SQL programs with extended syntax in Jupyter Notebook or a command-line tool. - -The following SQL statements train a TensorFlow model named `DNNClassifier`, which is a Python class derived from [`tf.estimator.Estimator`](https://www.tensorflow.org/api_docs/python/tf/estimator/Estimator): - -```sql - -SELECT * FROM a_table TRAIN DNNClassifier WITH learning_rate=0.01 INTO sqlflow_models.my_model; -``` - -And the following statement uses the trained model for prediction. - -```sql -SELECT * FROM b_table PREDICT b_table.predicted_label USING sqlflow_models.my_model; -``` - -Please be aware that the part in the above statements before the extended keyword TRAIN and PREDICT is a standard SQL statement. This feature simplifies the implementation of the SQLFlow system. - -## System Implementation - -If a SQL statement is of the standard syntax, SQLFlow throws it to the SQL engine and redirects the output to the user; otherwise, SQLFlow translates the statement of extended syntax into a Python program. Currently, it generates a program that throws the standard-syntax part of SELECT to MySQL, reads the results in the train-loop of a TensorFlow program. We will talk about how to extend SQLFlow to connect more SQL engines like Oracle, Hive, and SparkSQL, and generates more types of machine learning programs that calls distributed TensorFlow, PyTorch, and xgboost later. Before that, let us explain the system components. - -### SQLFlow as a gRPC Server - -SQLFlow is a gRPC server, which can connect with multiple clients. A typical client is [pysqlflow](https://github.com/sql-machine-learning/pysqlflow), the SQLFlow plugin for Jupyter Notebook server. Another once is a text-based client [/cmd/sqlflowserver/main.go](/cmd/sqlflowserver/main.go). - -``` -Jupyter Notebook ---(SQL statements)--> SQLFlow gRPC server -(SQLFlow magic command) <--(a stream of messages)-- -``` - -The protobuf definition of the gRPC service is at [/server/proto/sqlflow.proto](/server/proto/sqlflow.proto). The return of the method `SQLFlow.Run` is a stream of `Reponse`s, where each represents either a table header, a row, or a log message. The header and rows are usually from a standard SQL statement, for example, SELECT or DESCRIBE, and the log messages are usually from the running of a generated Python program. - -### SQLFlow in the gRPC Server - -Once the SQLFlow server receives a batch of SQL statements via a gRPC call, it runs the following steps for each statement: - -1. the [parser](/sql/sql.y) to generate parsing result, -2. the [verifier](/sql/verifier.go) to verify the semantics given the parsing result, -3. the [code generator](/sql/codegen.go) to generate a Python program, or the *submitter*, from the parsing result, -4. the [executor](/sql/executor.go) that runs the submitter locally. - -Step 3. and 4. are only for a SQL statement of extended syntax; otherwise, SQLFlow server proxies the standard-syntax statement to the SQL engine. - -The executor calls Go's [standard package](https://godoc.org/os/exec) that captures the stdout and stderr from the submitter process and passing the result back to the gRPC client. Therefore, it is the responsibility of the submitter to print log messages to its stderr and stdout. - -### Minimal Viable Product - -In the minimal viable product (MVP) of SQLFlow, the code generator generates a Python program consists of two parts: - -1. throw the standard SELECT part in the extended-syntax statement to MySQL via ODBC, and -1. a loop that reads outputs from the run of the SELECT statement and trains the model (or, using a trained model to predict). - -The training part calls TensorFlow to update the parameters of the model specified in the TRAIN clause. - -### Extensibility - -By writing more code generators, we could extend SQLFlow to support more SQL engines, e.g., Hive and Oracle, and use machine learning toolkits, e.g., PyTorch and xgboost, in addition to TensorFlow, on various computing platforms. You are welcome to add more code generators such as - -- `codegen_distributed_tf.go` to generate a submitter program similar to the MVP but runs a distributed TensorFlow training job. -- `codegen_kubernetes_tf.go` to launch a distributed TensorFlow job on a Kubernetes cluster, other than running locally, in the same container as where SQLFlow gRPC server resides. -- `codegen_gcloud_pytorch.go` to launch a submitter that calls PyTorch instead of TensorFlow for training on the Google Cloud. - -### Job Management - -The current design of the gRPC interface assumes that the connection between the client, e.g., the Jupyter Notebook, and the SQLFlow server keeps alive during the running of the training program. This assumption is reasonable because even if the user closes her/his Web browser and disconnect to the Jupyter Notebook server, the connection between Jupyter to SQLFlow server might keep alive. However, this might not be robust enough if the Jupyter Notebook server runs on a user's laptop and gets killed. In such a case, the gRPC server cannot stream the messages back to the client and would cause the failure of the submitter. - -A solution is to change the gRPC interface of SQLFlow server to have a method that files a job and returns immediately, and another method to get a batch of recent messages given a job ID. We will make a design for that soon. diff --git a/docs/feed.xml b/docs/feed.xml index cb6ba12..7dc3366 100644 --- a/docs/feed.xml +++ b/docs/feed.xml @@ -1 +1 @@ -Jekyll2019-04-12T14:54:10+08:00/feed.xmlSqlflow \ No newline at end of file +Jekyll2019-04-15T16:08:01+08:00/feed.xmlSqlflow \ No newline at end of file diff --git a/docs/index.html b/docs/index.html index e8f999b..09497b1 100644 --- a/docs/index.html +++ b/docs/index.html @@ -685,18 +685,28 @@ border-top: 1px dashed #aaa; } - section.feature h3, section.feature p{ + section.feature{ text-align: left; padding-left: 6px; padding-right: 6px; } - section.feature:nth-of-type(2n) h3, section.feature:nth-of-type(2n) p { + section.feature img { + width: 550px; + margin: auto 20px; + } + + section.feature:nth-of-type(2n){ text-align: right; padding-right: 6px; padding-right: 6px; } + section.feature .gif-holder { + width: 550px; + height: 275px; + } + @media only screen and (min-width: 40.063em) { #home h1 { font-size: 2.7rem; @@ -713,7 +723,7 @@ } .main-nav .logo img { - height: 50px; + height: 30px; } .main-page .intro { @@ -723,7 +733,7 @@ .main-page .intro h1 { margin: 50px auto; - color: #ffb214; + color: #1BA2FF; } .main-page .intro p { @@ -746,13 +756,13 @@
@@ -795,16 +805,25 @@

Extends SQL to support AI. Extract knowledge from Data.

Easy to Learn

Manipulate data and running AI with SQL

+ + gif +

Work with Many Database Management Systems

MySQL, Hive, Alibaba ODPS, Oracle and you name it!

+ + gif +

Support Many Machine Learning Toolkits

TensorFlow, PyTorch, xgboost, and more!

+ +
+
diff --git a/docs/pages/about.html b/docs/pages/about.html index 74db6ec..e2cfaa4 100644 --- a/docs/pages/about.html +++ b/docs/pages/about.html @@ -41,7 +41,7 @@
  • - SQLFlow: Code Walkthrough + sqlfs
    • @@ -53,7 +53,7 @@
  • - Build TensorFlow from Source Code using Docker + Build TensorFlow from Source Code using Docker
    • @@ -63,7 +63,7 @@
  • - Canonical Development Environment + Canonical Development Environment
    • @@ -73,7 +73,7 @@
  • - Closing the producer goroutine from the consumer + Closing the producer goroutine from the consumer
    • @@ -83,7 +83,7 @@
  • - Compatibility with Various SQL Engines + Compatibility with Various SQL Engines
    • @@ -93,7 +93,7 @@
  • - Run MySQL Server and Client in Docker Containers + Run MySQL Server and Client in Docker Containers
    • @@ -103,7 +103,7 @@
  • - Piping Responses + Piping Responses
    • @@ -113,7 +113,7 @@
  • - Quick start + Quick start
    • @@ -123,7 +123,7 @@
  • - Extended SQL Parser Design + Extended SQL Parser Design
    • @@ -133,7 +133,7 @@
  • - Submitter + Submitter
    • @@ -143,7 +143,97 @@
  • - SQLFlow: Design Doc + SQLFlow: Design Doc + +
    • + + + + + +
  • + + SQLFlow: Code Walkthrough + +
    • + + + + + +
  • + + SQLFlow + +
    • + + + + + +
  • + + SQLFlow Demo + +
    • + + + + + +
  • + + The Credit Card Fraud Detection Example + +
    • + + + + + +
  • + + The MySQL Server Container for Testing + +
    • + + + + + +
  • + + Credit Card Fraud Detection + +
    • + + + + + +
  • + + sqlflowserver + +
    • + + + + + +
  • + + SQLFlow + +
    • + + + + + +
  • + + Python Code Template
    • diff --git a/docs/sitemap.xml b/docs/sitemap.xml index a18747f..48595c5 100644 --- a/docs/sitemap.xml +++ b/docs/sitemap.xml @@ -7,72 +7,75 @@ / -/doc/build-tensorflow.html +/sqlflow/doc/build-tensorflow.html -/doc/build.html +/sqlflow/doc/build.html -/doc/close_producer_from_consumer.html +/sqlflow/doc/close_producer_from_consumer.html -/doc/database_abstraction_layer.html +/sqlflow/doc/database_abstraction_layer.html -/doc/mysql-setup.html +/sqlflow/doc/mysql-setup.html -/doc/pipe.html +/sqlflow/doc/pipe.html -/doc/quickstart.html +/sqlflow/doc/quickstart.html -/doc/sql_parser.html +/sqlflow/doc/sql_parser.html -/doc/submitter.html +/sqlflow/doc/submitter.html -/doc/syntax.html +/sqlflow/doc/syntax.html -/doc/walkthrough.html +/sqlflow/doc/walkthrough.html -/gohive/ -2019-04-12T10:17:18+08:00 +/sqlflow/ -/gohive/service-rpc/gen-go/tcliservice/ -2019-04-12T10:17:18+08:00 +/sqlflow/cmd/demo/ -/gohive/vendor/git.apache.org/thrift.git/lib/go/thrift/ -2019-04-12T10:17:18+08:00 +/sqlflow/example/creditcard/ -/sqlflow/cmd/demo/ -2019-04-12T10:17:18+08:00 +/sqlflow/example/datasets/ -/sqlflow/cmd/sqlflowserver/ -2019-04-12T10:17:18+08:00 +/sqlflow/example/fraud_detection/ -/sqlflow/ -2019-04-12T10:17:18+08:00 +/sqlflow/server/ -/sqlflow/server/ +/sqlflow/sql/ + + +/sqlflow/sql/python/ + + +/sqlflow/sqlfs/ + + +/gohive/ 2019-04-12T10:17:18+08:00 -/sqlflow/sql/ +/gohive/service-rpc/gen-go/tcliservice/ 2019-04-12T10:17:18+08:00 -/sqlflow/sqlfs/ +/gohive/vendor/git.apache.org/thrift.git/lib/go/thrift/ 2019-04-12T10:17:18+08:00 diff --git a/docs/sqlflow/cmd/demo/index.html b/docs/sqlflow/cmd/demo/index.html index 7237249..180eedd 100644 --- a/docs/sqlflow/cmd/demo/index.html +++ b/docs/sqlflow/cmd/demo/index.html @@ -1,13 +1,365 @@ - + + + - - - - + + + + + SQLFlow Demo - Sqlflow + + + + + + + + + + + + + + -

    Redirecting to https://github.com/sql-machine-learning/sqlflow

    + +
    +
    + Sqlflow + +
    + + +
    +
    +

    This site uses Just the Docs, a documentation theme for Jekyll.

    +
    +
    +
    +
    +
    + +
    +
    + + Search +
    +
    +
    + + + + +
    +
    +
    + + + +
    +

    SQLFlow Demo

    + +

    Before you start, you need to set up test MySQL server as described here

    + +

    Start SQLFlow as

    +
    go get -d ./... && go run main.go -logdir="/path/to/logs/" -loglevel="info"
+
    + +

    Regular SQL Statements

    + +
    sqlflow> SHOW DATABASES;
++--------------------+
+| Database           |
++--------------------+
+| churn              |
+| information_schema |
+| iris               |
+| mysql              |
+| performance_schema |
+| sqlflow_models     |
+| sqlfs              |
+| sys                |
++--------------------+
+sqlflow> SELECT * FROM iris.train LIMIT 1;
++--------------+-------------+--------------+-------------+-------+
+| sepal_length | sepal_width | petal_length | petal_width | class |
++--------------+-------------+--------------+-------------+-------+
+|          6.4 |         2.8 |          5.6 |         2.2 |     2 |
++--------------+-------------+--------------+-------------+-------+
+
    + +

    Train

    + +
    sqlflow> SELECT *
+FROM iris.train
+TRAIN DNNClassifier
+WITH
+  n_classes = 3,
+  hidden_units = [10, 20]
+COLUMN sepal_length, sepal_width, petal_length, petal_width
+LABEL class
+INTO sqlflow_models.my_dnn_model;
+-----------------------------
+2018/12/16 15:03:54 tensorflowCmd: run in Docker container
+Job success
+
    + +

    Predict

    + +
    sqlflow> SELECT *
+FROM iris.test
+PREDICT iris.predict.class
+USING sqlflow_models.my_dnn_model;
+-----------------------------
+2018/12/16 15:05:58 tensorflowCmd: run in Docker container
+Job success
+
    + +

    The prediction is stored at iris.predict, you can look them up at

    + +
    sqlflow> SELECT * FROM iris.predict LIMIT 10;
+-----------------------------
++--------------+-------------+--------------+-------------+-------+
+| sepal_length | sepal_width | petal_length | petal_width | class |
++--------------+-------------+--------------+-------------+-------+
+|          6.4 |         2.8 |          5.6 |         2.2 |     2 |
+|            5 |         2.3 |          3.3 |           1 |     1 |
+|          4.9 |         2.5 |          4.5 |         1.7 |     2 |
+|          4.9 |         3.1 |          1.5 |         0.1 |     0 |
+|          5.7 |         3.8 |          1.7 |         0.3 |     0 |
+|          4.4 |         3.2 |          1.3 |         0.2 |     0 |
+|          5.4 |         3.4 |          1.5 |         0.4 |     0 |
+|          6.9 |         3.1 |          5.1 |         2.3 |     2 |
+|          6.7 |         3.1 |          4.4 |         1.4 |     1 |
+|          5.1 |         3.7 |          1.5 |         0.4 |     0 |
++--------------+-------------+--------------+-------------+-------+
+
    + + + + +
    +
    +
    +
    + diff --git a/docs/sqlflow/cmd/sqlflowserver/index.html b/docs/sqlflow/cmd/sqlflowserver/index.html deleted file mode 100644 index 7237249..0000000 --- a/docs/sqlflow/cmd/sqlflowserver/index.html +++ /dev/null @@ -1,13 +0,0 @@ - - - - - - - - -

    Redirecting to https://github.com/sql-machine-learning/sqlflow

    - - diff --git a/docs/sqlflow/index.html b/docs/sqlflow/index.html index 7237249..d2d773b 100644 --- a/docs/sqlflow/index.html +++ b/docs/sqlflow/index.html @@ -1,13 +1,316 @@ - + + + - - - - + + + + + SQLFlow - Sqlflow + + + + + + + + + + + + + + -

    Redirecting to https://github.com/sql-machine-learning/sqlflow

    + +
    +
    + Sqlflow + +
    + + +
    +
    +

    This site uses Just the Docs, a documentation theme for Jekyll.

    +
    +
    +
    +
    +
    + +
    +
    + + Search +
    +
    +
    + + + + +
    +
    +
    + + + +
    +

    SQLFlow

    + +

    Build Status GoDoc Release License

    + +

    SQLFlow enriches SQL systems, e.g., MySQL, Hive, SparkSQL, with the capability of machine learning, using TensorFlow. We slightly extend the syntax of the SELECT statement of SQL to support model training and prediction.

    + +

    How to Use

    + + + +

    How to Contribute

    + + + +

    Questions and Feedback

    + +

    Your feedback is our motivation to move on. Please let us know your questions, concerns, and issues by filing Github Issues.

    + +

    License

    + +

    Apache License 2.0

    + + + + +
    +
    +
    +
    + diff --git a/docs/sqlflow/server/index.html b/docs/sqlflow/server/index.html index 7237249..35f1093 100644 --- a/docs/sqlflow/server/index.html +++ b/docs/sqlflow/server/index.html @@ -1,13 +1,303 @@ - + + + - - - - + + + + + sqlflowserver - Sqlflow + + + + + + + + + + + + + + -

    Redirecting to https://github.com/sql-machine-learning/sqlflow

    + +
    +
    + Sqlflow + +
    + + +
    +
    +

    This site uses Just the Docs, a documentation theme for Jekyll.

    +
    +
    +
    +
    +
    + +
    +
    + + Search +
    +
    +
    + + + + +
    +
    +
    + + + +
    +

    sqlflowserver

    +

    The gRPC proxy server of SQL engines

    + +

    Prerequisite

    + +
      +
    • Golang
      • +
    • Docker
      • +
    • GRPC
      • +
    + +

    Running

    + +
    go generate && cd server && go run main.go
+
    + + + + +
    +
    +
    +
    + diff --git a/docs/sqlflow/sql/index.html b/docs/sqlflow/sql/index.html index 7237249..23a2448 100644 --- a/docs/sqlflow/sql/index.html +++ b/docs/sqlflow/sql/index.html @@ -1,13 +1,300 @@ - + + + - - - - + + + + + SQLFlow - Sqlflow + + + + + + + + + + + + + + -

    Redirecting to https://github.com/sql-machine-learning/sqlflow

    + +
    +
    + Sqlflow + +
    + + +
    +
    +

    This site uses Just the Docs, a documentation theme for Jekyll.

    +
    +
    +
    +
    +
    + +
    +
    + + Search +
    +
    +
    + + + + +
    +
    +
    + + + +
    +

    SQLFlow

    + +

    Build, Run, Test

    + +

    Before running the unit tests, we need to build and run a Docker container that +hosts a MySQL database following this guide.

    + +

    To build the parser using goyacc and run all unit tests, use the following +command:

    + +
    go get -d ./... && goyacc -p sql -o parser.go sql.y && go test -v
+
    + + + + +
    +
    +
    +
    + diff --git a/docs/sqlflow/sqlfs/index.html b/docs/sqlflow/sqlfs/index.html index 7237249..07eb596 100644 --- a/docs/sqlflow/sqlfs/index.html +++ b/docs/sqlflow/sqlfs/index.html @@ -1,13 +1,338 @@ - + + + - - - - + + + + + sqlfs - Sqlflow + + + + + + + + + + + + + + -

    Redirecting to https://github.com/sql-machine-learning/sqlflow

    + +
    +
    + Sqlflow + +
    + + +
    +
    +

    This site uses Just the Docs, a documentation theme for Jekyll.

    +
    +
    +
    +
    +
    + +
    +
    + + Search +
    +
    +
    + + + + +
    +
    +
    + + + +
    +

    sqlfs

    + +

    The package sqlfs provides an io.ReadCloser and an io.WriteCloser that treats a SQL database a filesystem, where each table in the database is like a file. The schema of the table is very simple – it has only one column of BLOB type. All the rows consist the storage.

    + +

    sqlfs provides the following features.

    + +

    Create a file

    + +

    To create a table named “hello” in a database “mydb” for writing, we can call Create.

    + +
    f, e := sqlfs.Create(db, "mydb.hello")
+f.Write([]byte("hello world!\n"))
+f.Close()
+
    + +

    where db comes from a call to database/sql.Open.

    + +

    Append to a file

    + +
    f, e := sqlfs.Append(db, "mydb.hello")
+f.Write([]byte("hello world!\n")
+f.Close()
+
    + +

    Read from a file

    + +
    f, e := sqlfs.Open(db, "mydb.hello")
+buf := make([]byte, 1024)
+f.Read(buf)
+f.Close()
+
    + +

    Remove a file

    + +
    DropTable(db, "mydb.hello")
+
    + +

    Check if a file exists

    + +
    HasTable(db, "mydb.hello")
+
    + +

    Other I/O operations

    + +

    Feel free to use standard packages io, ioutil, etc with sqlfs. For example, we can call io.Copy to copy everything from the standard input to a table.

    + +
    f, e := sqlfs.Create(db, "mydb.hello")
+io.Copy(f, os.Stdin)
+f.Close()
+
    + + + + +
    +
    +
    +
    + diff --git a/index.md b/index.md index c25f3a9..e2c9c53 100644 --- a/index.md +++ b/index.md @@ -9,8 +9,10 @@ intro: "Currently support MySQL and TensorFlow. Expanding to Apache Hive, Alibab features: - title: 'Easy to Learn' content: 'Manipulate data and running AI with SQL' + gif: 'assets/train.gif' - title: 'Work with Many Database Management Systems' content: 'MySQL, Hive, Alibaba ODPS, Oracle and you name it!' + gif: 'assets/select.gif' - title: 'Support Many Machine Learning Toolkits' content: 'TensorFlow, PyTorch, xgboost, and more!' --- \ No newline at end of file diff --git a/sqlflow/.gitignore b/sqlflow/.gitignore new file mode 100644 index 0000000..1e4d768 --- /dev/null +++ b/sqlflow/.gitignore @@ -0,0 +1,11 @@ +*~ +*.test +/Gopkg.lock +/Gopkg.toml +/vendor/ +*.swp +.idea/ +*.vim +*.log +/logs +sql/my_dnn_model diff --git a/sqlflow/.pre-commit-config.yaml b/sqlflow/.pre-commit-config.yaml new file mode 100644 index 0000000..f76ae34 --- /dev/null +++ b/sqlflow/.pre-commit-config.yaml @@ -0,0 +1,9 @@ +- repo: git://github.com/dnephin/pre-commit-golang + sha: HEAD + hooks: + - id: go-fmt + - id: go-vet + - id: go-lint + - id: validate-toml + - id: no-go-testing + - id: no-go-testing diff --git a/sqlflow/.travis.yml b/sqlflow/.travis.yml new file mode 100644 index 0000000..89fe87c --- /dev/null +++ b/sqlflow/.travis.yml @@ -0,0 +1,37 @@ +language: go + +branches: + only: + - gh-pages + - /.*/ + +service: + - docker + +# Update to newer version of Docker to allow +# Dockerfile being outside of the context directory +# https://github.com/docker/compose/issues/4926#issuecomment-370501452 +addons: + apt: + packages: + - docker-ce + +install: + # Note(tony): TravisCI's native `go get ${gobuild_args} ./...` failed with + # `/usr/bin/ld: final link failed: Bad value`, the cause is the system linker + # being not up to date: https://github.com/golang/go/issues/15038 + # So I decided to skip install, and go get inside sqlflow:dev image + # Ref build: https://travis-ci.com/sql-machine-learning/sqlflow/builds/107870583 + - echo "skip install" + +script: + - ls -lah + - docker build -t sqlflow:dev -f Dockerfile.dev . + - docker run --rm -v $GOPATH:/go -w /go/src/github.com/sql-machine-learning/sqlflow sqlflow:dev bash scripts/test.sh + +deploy: + skip_cleanup: true + provider: script + script: bash scripts/deploy.sh + on: + branch: develop diff --git a/sqlflow/Dockerfile b/sqlflow/Dockerfile new file mode 100644 index 0000000..eec46da --- /dev/null +++ b/sqlflow/Dockerfile @@ -0,0 +1,24 @@ +FROM ubuntu:16.04 + +RUN apt-get update +RUN apt-get install -y python3-pip +RUN pip3 install --upgrade pip +RUN pip3 install tensorflow mysql-connector-python jupyter sqlflow +# Fix jupyter server "connecting to kernel" problem +# https://github.com/jupyter/notebook/issues/2664#issuecomment-468954423 +RUN pip3 install tornado==4.5.3 + +# Load sqlflow Jupyter magic command automatically. c.f. https://stackoverflow.com/a/32683001. +ENV IPYTHON_STARTUP /root/.ipython/profile_default/startup/ +RUN mkdir -p $IPYTHON_STARTUP +RUN echo 'get_ipython().magic(u"%reload_ext sqlflow.magic")' >> $IPYTHON_STARTUP/00-first.py +RUN echo 'get_ipython().magic(u"%autoreload 2")' >> $IPYTHON_STARTUP/00-first.py + +ADD demo /usr/bin/demo +ADD sqlflowserver /usr/bin/sqlflowserver + +# Make python3 callable through python +RUN ln -s /usr/bin/python3 /usr/bin/python +RUN chmod +x /usr/bin/python + +CMD ["/usr/bin/demo"] diff --git a/sqlflow/Dockerfile.dev b/sqlflow/Dockerfile.dev new file mode 100644 index 0000000..ccf1776 --- /dev/null +++ b/sqlflow/Dockerfile.dev @@ -0,0 +1,45 @@ +FROM ubuntu:16.04 + +RUN apt-get update +RUN apt-get install -y git +RUN apt-get install -y wget curl +RUN apt-get install -y python3-pip + +RUN pip3 install --upgrade pip +RUN pip3 install grpcio-tools +RUN pip3 install tensorflow +RUN pip3 install mysql-connector-python + +RUN wget --quiet https://dl.google.com/go/go1.11.5.linux-amd64.tar.gz +RUN tar -C /usr/local -xzf go1.11.5.linux-amd64.tar.gz +RUN rm go1.11.5.linux-amd64.tar.gz +ENV PATH $PATH:/usr/local/go/bin + +RUN wget --quiet https://github.com/protocolbuffers/protobuf/releases/download/v3.6.1/protoc-3.6.1-linux-x86_64.zip +RUN apt-get install -y unzip +RUN unzip -qq protoc-3.6.1-linux-x86_64.zip -d /usr/local +RUN rm protoc-3.6.1-linux-x86_64.zip + +RUN mkdir -p /go +ENV GOPATH /go +RUN go get github.com/golang/protobuf/protoc-gen-go +RUN mv /go/bin/protoc-gen-go /usr/local/bin/ + +RUN apt-get install -y sqlite3 libsqlite3-dev + +# install mysql without a password prompt +RUN ["/bin/bash", "-c", "debconf-set-selections <<< 'mysql-server mysql-server/root_password password root'"] +RUN ["/bin/bash", "-c", "debconf-set-selections <<< 'mysql-server mysql-server/root_password_again password root'"] +RUN apt-get -y install mysql-server +VOLUME /var/lib/mysql + +# Make python3 callable through python +RUN ln -s /usr/bin/python3 /usr/bin/python +RUN chmod +x /usr/bin/python + +RUN echo "#!/bin/bash" > /build.sh +RUN echo "go generate -v ./..." >> /build.sh +RUN echo "go install -v ./..." >> /build.sh +RUN chmod +x /build.sh + +CMD ["/build.sh"] diff --git a/sqlflow/LICENSE b/sqlflow/LICENSE new file mode 100644 index 0000000..261eeb9 --- /dev/null +++ b/sqlflow/LICENSE @@ -0,0 +1,201 @@ + Apache License + Version 2.0, January 2004 + http://www.apache.org/licenses/ + + TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION + + 1. Definitions. + + "License" shall mean the terms and conditions for use, reproduction, + and distribution as defined by Sections 1 through 9 of this document. + + "Licensor" shall mean the copyright owner or entity authorized by + the copyright owner that is granting the License. + + "Legal Entity" shall mean the union of the acting entity and all + other entities that control, are controlled by, or are under common + control with that entity. For the purposes of this definition, + "control" means (i) the power, direct or indirect, to cause the + direction or management of such entity, whether by contract or + otherwise, or (ii) ownership of fifty percent (50%) or more of the + outstanding shares, or (iii) beneficial ownership of such entity. + + "You" (or "Your") shall mean an individual or Legal Entity + exercising permissions granted by this License. + + "Source" form shall mean the preferred form for making modifications, + including but not limited to software source code, documentation + source, and configuration files. + + "Object" form shall mean any form resulting from mechanical + transformation or translation of a Source form, including but + not limited to compiled object code, generated documentation, + and conversions to other media types. + + "Work" shall mean the work of authorship, whether in Source or + Object form, made available under the License, as indicated by a + copyright notice that is included in or attached to the work + (an example is provided in the Appendix below). + + "Derivative Works" shall mean any work, whether in Source or Object + form, that is based on (or derived from) the Work and for which the + editorial revisions, annotations, elaborations, or other modifications + represent, as a whole, an original work of authorship. For the purposes + of this License, Derivative Works shall not include works that remain + separable from, or merely link (or bind by name) to the interfaces of, + the Work and Derivative Works thereof. + + "Contribution" shall mean any work of authorship, including + the original version of the Work and any modifications or additions + to that Work or Derivative Works thereof, that is intentionally + submitted to Licensor for inclusion in the Work by the copyright owner + or by an individual or Legal Entity authorized to submit on behalf of + the copyright owner. For the purposes of this definition, "submitted" + means any form of electronic, verbal, or written communication sent + to the Licensor or its representatives, including but not limited to + communication on electronic mailing lists, source code control systems, + and issue tracking systems that are managed by, or on behalf of, the + Licensor for the purpose of discussing and improving the Work, but + excluding communication that is conspicuously marked or otherwise + designated in writing by the copyright owner as "Not a Contribution." + + "Contributor" shall mean Licensor and any individual or Legal Entity + on behalf of whom a Contribution has been received by Licensor and + subsequently incorporated within the Work. + + 2. Grant of Copyright License. Subject to the terms and conditions of + this License, each Contributor hereby grants to You a perpetual, + worldwide, non-exclusive, no-charge, royalty-free, irrevocable + copyright license to reproduce, prepare Derivative Works of, + publicly display, publicly perform, sublicense, and distribute the + Work and such Derivative Works in Source or Object form. + + 3. Grant of Patent License. Subject to the terms and conditions of + this License, each Contributor hereby grants to You a perpetual, + worldwide, non-exclusive, no-charge, royalty-free, irrevocable + (except as stated in this section) patent license to make, have made, + use, offer to sell, sell, import, and otherwise transfer the Work, + where such license applies only to those patent claims licensable + by such Contributor that are necessarily infringed by their + Contribution(s) alone or by combination of their Contribution(s) + with the Work to which such Contribution(s) was submitted. If You + institute patent litigation against any entity (including a + cross-claim or counterclaim in a lawsuit) alleging that the Work + or a Contribution incorporated within the Work constitutes direct + or contributory patent infringement, then any patent licenses + granted to You under this License for that Work shall terminate + as of the date such litigation is filed. + + 4. Redistribution. You may reproduce and distribute copies of the + Work or Derivative Works thereof in any medium, with or without + modifications, and in Source or Object form, provided that You + meet the following conditions: + + (a) You must give any other recipients of the Work or + Derivative Works a copy of this License; and + + (b) You must cause any modified files to carry prominent notices + stating that You changed the files; and + + (c) You must retain, in the Source form of any Derivative Works + that You distribute, all copyright, patent, trademark, and + attribution notices from the Source form of the Work, + excluding those notices that do not pertain to any part of + the Derivative Works; and + + (d) If the Work includes a "NOTICE" text file as part of its + distribution, then any Derivative Works that You distribute must + include a readable copy of the attribution notices contained + within such NOTICE file, excluding those notices that do not + pertain to any part of the Derivative Works, in at least one + of the following places: within a NOTICE text file distributed + as part of the Derivative Works; within the Source form or + documentation, if provided along with the Derivative Works; or, + within a display generated by the Derivative Works, if and + wherever such third-party notices normally appear. The contents + of the NOTICE file are for informational purposes only and + do not modify the License. You may add Your own attribution + notices within Derivative Works that You distribute, alongside + or as an addendum to the NOTICE text from the Work, provided + that such additional attribution notices cannot be construed + as modifying the License. + + You may add Your own copyright statement to Your modifications and + may provide additional or different license terms and conditions + for use, reproduction, or distribution of Your modifications, or + for any such Derivative Works as a whole, provided Your use, + reproduction, and distribution of the Work otherwise complies with + the conditions stated in this License. + + 5. Submission of Contributions. Unless You explicitly state otherwise, + any Contribution intentionally submitted for inclusion in the Work + by You to the Licensor shall be under the terms and conditions of + this License, without any additional terms or conditions. + Notwithstanding the above, nothing herein shall supersede or modify + the terms of any separate license agreement you may have executed + with Licensor regarding such Contributions. + + 6. Trademarks. This License does not grant permission to use the trade + names, trademarks, service marks, or product names of the Licensor, + except as required for reasonable and customary use in describing the + origin of the Work and reproducing the content of the NOTICE file. + + 7. Disclaimer of Warranty. Unless required by applicable law or + agreed to in writing, Licensor provides the Work (and each + Contributor provides its Contributions) on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or + implied, including, without limitation, any warranties or conditions + of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A + PARTICULAR PURPOSE. You are solely responsible for determining the + appropriateness of using or redistributing the Work and assume any + risks associated with Your exercise of permissions under this License. + + 8. Limitation of Liability. In no event and under no legal theory, + whether in tort (including negligence), contract, or otherwise, + unless required by applicable law (such as deliberate and grossly + negligent acts) or agreed to in writing, shall any Contributor be + liable to You for damages, including any direct, indirect, special, + incidental, or consequential damages of any character arising as a + result of this License or out of the use or inability to use the + Work (including but not limited to damages for loss of goodwill, + work stoppage, computer failure or malfunction, or any and all + other commercial damages or losses), even if such Contributor + has been advised of the possibility of such damages. + + 9. Accepting Warranty or Additional Liability. While redistributing + the Work or Derivative Works thereof, You may choose to offer, + and charge a fee for, acceptance of support, warranty, indemnity, + or other liability obligations and/or rights consistent with this + License. However, in accepting such obligations, You may act only + on Your own behalf and on Your sole responsibility, not on behalf + of any other Contributor, and only if You agree to indemnify, + defend, and hold each Contributor harmless for any liability + incurred by, or claims asserted against, such Contributor by reason + of your accepting any such warranty or additional liability. + + END OF TERMS AND CONDITIONS + + APPENDIX: How to apply the Apache License to your work. + + To apply the Apache License to your work, attach the following + boilerplate notice, with the fields enclosed by brackets "[]" + replaced with your own identifying information. (Don't include + the brackets!) The text should be enclosed in the appropriate + comment syntax for the file format. We also recommend that a + file or class name and description of purpose be included on the + same "printed page" as the copyright notice for easier + identification within third-party archives. + + Copyright [yyyy] [name of copyright owner] + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. diff --git a/sqlflow/README.md b/sqlflow/README.md new file mode 100644 index 0000000..5d6407c --- /dev/null +++ b/sqlflow/README.md @@ -0,0 +1,24 @@ +# SQLFlow + +[![Build Status](https://travis-ci.org/sql-machine-learning/sqlflow.svg?branch=develop)](https://travis-ci.org/sql-machine-learning/sqlflow) [![GoDoc](https://godoc.org/github.com/sql-machine-learning/sqlflow?status.svg)](https://godoc.org/github.com/sql-machine-learning/sqlflow) [![Release](https://img.shields.io/github/release/sql-machine-learning/sqlflow.svg)](https://github.com/sql-machine-learning/sqlflow/releases) [![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE) + +SQLFlow enriches SQL systems, e.g., MySQL, Hive, SparkSQL, with the capability of machine learning, using TensorFlow. We slightly extend the syntax of the SELECT statement of SQL to support model training and prediction. + +## How to Use + +- [Quick start](doc/quickstart.md) +- [Extended SQL syntax](doc/syntax.md) + +## How to Contribute + +- [Build from source code](doc/build.md). +- [The walkthrough of the source code](doc/walkthrough.md) +- [The choice of parser generator](doc/sql_parser.md) + +## Questions and Feedback + +Your feedback is our motivation to move on. Please let us know your questions, concerns, and issues by filing Github Issues. + +## License + +[Apache License 2.0](https://github.com/sql-machine-learning/sqlflow/LICENSE) diff --git a/sqlflow/cmd/demo/README.md b/sqlflow/cmd/demo/README.md new file mode 100644 index 0000000..02fd487 --- /dev/null +++ b/sqlflow/cmd/demo/README.md @@ -0,0 +1,82 @@ +# SQLFlow Demo + +Before you start, you need to set up test MySQL server as described [here](../../example/datasets) + +Start SQLFlow as +```golang +go get -d ./... && go run main.go -logdir="/path/to/logs/" -loglevel="info" +``` + +#### Regular SQL Statements + +```sql +sqlflow> SHOW DATABASES; ++--------------------+ +| Database | ++--------------------+ +| churn | +| information_schema | +| iris | +| mysql | +| performance_schema | +| sqlflow_models | +| sqlfs | +| sys | ++--------------------+ +sqlflow> SELECT * FROM iris.train LIMIT 1; ++--------------+-------------+--------------+-------------+-------+ +| sepal_length | sepal_width | petal_length | petal_width | class | ++--------------+-------------+--------------+-------------+-------+ +| 6.4 | 2.8 | 5.6 | 2.2 | 2 | ++--------------+-------------+--------------+-------------+-------+ +``` + +#### Train + +```sql +sqlflow> SELECT * +FROM iris.train +TRAIN DNNClassifier +WITH + n_classes = 3, + hidden_units = [10, 20] +COLUMN sepal_length, sepal_width, petal_length, petal_width +LABEL class +INTO sqlflow_models.my_dnn_model; +----------------------------- +2018/12/16 15:03:54 tensorflowCmd: run in Docker container +Job success +``` + +#### Predict + +```sql +sqlflow> SELECT * +FROM iris.test +PREDICT iris.predict.class +USING sqlflow_models.my_dnn_model; +----------------------------- +2018/12/16 15:05:58 tensorflowCmd: run in Docker container +Job success +``` + +The prediction is stored at `iris.predict`, you can look them up at + +```sql +sqlflow> SELECT * FROM iris.predict LIMIT 10; +----------------------------- ++--------------+-------------+--------------+-------------+-------+ +| sepal_length | sepal_width | petal_length | petal_width | class | ++--------------+-------------+--------------+-------------+-------+ +| 6.4 | 2.8 | 5.6 | 2.2 | 2 | +| 5 | 2.3 | 3.3 | 1 | 1 | +| 4.9 | 2.5 | 4.5 | 1.7 | 2 | +| 4.9 | 3.1 | 1.5 | 0.1 | 0 | +| 5.7 | 3.8 | 1.7 | 0.3 | 0 | +| 4.4 | 3.2 | 1.3 | 0.2 | 0 | +| 5.4 | 3.4 | 1.5 | 0.4 | 0 | +| 6.9 | 3.1 | 5.1 | 2.3 | 2 | +| 6.7 | 3.1 | 4.4 | 1.4 | 1 | +| 5.1 | 3.7 | 1.5 | 0.4 | 0 | ++--------------+-------------+--------------+-------------+-------+ +``` diff --git a/sqlflow/cmd/demo/demo.go b/sqlflow/cmd/demo/demo.go new file mode 100644 index 0000000..3474351 --- /dev/null +++ b/sqlflow/cmd/demo/demo.go @@ -0,0 +1,121 @@ +package main + +import ( + "bufio" + "flag" + "fmt" + "log" + "os" + "strings" + + "github.com/go-sql-driver/mysql" + "github.com/olekukonko/tablewriter" + "github.com/sql-machine-learning/sqlflow/sql" +) + +const tablePageSize = 1000 + +// readStmt reads a SQL statement from the scanner. A statement could +// have multiple lines and ends at a semicolon at theend of the last +// line. +func readStmt() string { + scn := bufio.NewScanner(os.Stdin) + + stmt := "" + for scn.Scan() { + stmt += scn.Text() + "\n" + if strings.HasSuffix(strings.TrimSpace(scn.Text()), ";") { + break + } + } + if err := scn.Err(); err != nil { + return "" + } + return stmt +} + +func header(head map[string]interface{}) ([]string, error) { + cn, ok := head["columnNames"] + if !ok { + return nil, fmt.Errorf("can't find field columnNames in head") + } + cols, ok := cn.([]string) + if !ok { + return nil, fmt.Errorf("invalid header type") + } + return cols, nil +} + +func render(rsp interface{}, table *tablewriter.Table) bool { + isTable := false + switch s := rsp.(type) { + case map[string]interface{}: // table header + cols, e := header(s) + if e == nil { + table.SetHeader(cols) + } + isTable = true + case []interface{}: // row + row := make([]string, len(s)) + for i, v := range s { + row[i] = fmt.Sprint(v) + } + table.Append(row) + isTable = true + case error: + fmt.Printf("ERROR: %v\n", s) + default: + fmt.Println(s) + } + return isTable +} + +func main() { + user := flag.String("db_user", "", "database user name") + pswd := flag.String("db_password", "", "database user password") + addr := flag.String("db_address", "", "database address, such as: localhost:3306") + flag.Parse() + + cfg := &mysql.Config{ + User: *user, + Passwd: *pswd, + Net: "tcp", + Addr: *addr, + AllowNativePasswords: true, + } + log.Println("Connecting to db with:") + log.Printf("%+#v\n", *cfg) + db, err := sql.Open("mysql", cfg.FormatDSN()) + if err != nil { + log.Fatalf("failed to open database: %v", err) + } + defer db.Close() + if err := db.Ping(); err != nil { + log.Fatalf("failed to ping database: %v", err) + } + + for { + fmt.Print("sqlflow> ") + slct := readStmt() + fmt.Println("") + + isTable, tableRendered := false, false + table := tablewriter.NewWriter(os.Stdout) + + stream := sql.Run(slct, db) + for rsp := range stream.ReadAll() { + isTable = render(rsp, table) + + // pagination. avoid exceed memory + if isTable && table.NumLines() == tablePageSize { + table.Render() + tableRendered = true + table.ClearRows() + } + } + if isTable && (table.NumLines() > 0 || !tableRendered) { + table.Render() + } + fmt.Println("") + } +} diff --git a/sqlflow/cmd/demo/index.html b/sqlflow/cmd/demo/index.html deleted file mode 100644 index 7237249..0000000 --- a/sqlflow/cmd/demo/index.html +++ /dev/null @@ -1,13 +0,0 @@ - - - - - - - - -

    Redirecting to https://github.com/sql-machine-learning/sqlflow

    - - diff --git a/sqlflow/cmd/demo/slides/201812/automl.sql b/sqlflow/cmd/demo/slides/201812/automl.sql new file mode 100644 index 0000000..44a4005 --- /dev/null +++ b/sqlflow/cmd/demo/slides/201812/automl.sql @@ -0,0 +1,7 @@ +sqlflow> SELECT * +FROM iris.train +TRAIN Classifier +COLUMN sepal_length, sepal_width, petal_length, petal_width +LABEL class +INTO sqlflow_models.my_dnn_model; + diff --git a/sqlflow/cmd/demo/slides/201812/pred.sql b/sqlflow/cmd/demo/slides/201812/pred.sql new file mode 100644 index 0000000..a44d1b4 --- /dev/null +++ b/sqlflow/cmd/demo/slides/201812/pred.sql @@ -0,0 +1,22 @@ +sqlflow> SELECT * +FROM iris.test +PREDICT iris.predict.class +USING sqlflow_models.my_dnn_model; +----------------------------- +2018/12/16 15:05:58 tensorflowCmd: run in Docker container +Job success +sqlflow> SELECT * FROM iris.predict LIMIT 10; ++--------------+-------------+--------------+-------------+-------+ +| sepal_length | sepal_width | petal_length | petal_width | class | ++--------------+-------------+--------------+-------------+-------+ +| 6.4 | 2.8 | 5.6 | 2.2 | 2 | +| 5 | 2.3 | 3.3 | 1 | 1 | +| 4.9 | 2.5 | 4.5 | 1.7 | 2 | +| 4.9 | 3.1 | 1.5 | 0.1 | 0 | +| 5.7 | 3.8 | 1.7 | 0.3 | 0 | +| 4.4 | 3.2 | 1.3 | 0.2 | 0 | +| 5.4 | 3.4 | 1.5 | 0.4 | 0 | +| 6.9 | 3.1 | 5.1 | 2.3 | 2 | +| 6.7 | 3.1 | 4.4 | 1.4 | 1 | +| 5.1 | 3.7 | 1.5 | 0.4 | 0 | ++--------------+-------------+--------------+-------------+-------+ diff --git a/sqlflow/cmd/demo/slides/201812/preprocess.sql b/sqlflow/cmd/demo/slides/201812/preprocess.sql new file mode 100644 index 0000000..d32532d --- /dev/null +++ b/sqlflow/cmd/demo/slides/201812/preprocess.sql @@ -0,0 +1,10 @@ +sqlflow> SELECT * +FROM iris.train +TRAIN DNNClassifier +WITH + n_classes = 3, + hidden_units = [10, 20] +COLUMN sepal_length, sepal_width, bucket(petal_length, [0., 1., 2.]), norm(petal_width) +LABEL class +INTO sqlflow_models.my_dnn_model; + diff --git a/sqlflow/cmd/demo/slides/201812/regular.sql b/sqlflow/cmd/demo/slides/201812/regular.sql new file mode 100644 index 0000000..2321767 --- /dev/null +++ b/sqlflow/cmd/demo/slides/201812/regular.sql @@ -0,0 +1,20 @@ +sqlflow> SHOW DATABASES; ++--------------------+ +| Database | ++--------------------+ +| churn | +| information_schema | +| iris | +| mysql | +| performance_schema | +| sqlflow_models | +| sqlfs | +| sys | ++--------------------+ +sqlflow> SELECT * FROM iris.train LIMIT 1; ++--------------+-------------+--------------+-------------+-------+ +| sepal_length | sepal_width | petal_length | petal_width | class | ++--------------+-------------+--------------+-------------+-------+ +| 6.4 | 2.8 | 5.6 | 2.2 | 2 | ++--------------+-------------+--------------+-------------+-------+ + diff --git a/sqlflow/cmd/demo/slides/201812/sqlflow.slide b/sqlflow/cmd/demo/slides/201812/sqlflow.slide new file mode 100644 index 0000000..944beef --- /dev/null +++ b/sqlflow/cmd/demo/slides/201812/sqlflow.slide @@ -0,0 +1,40 @@ +SQLFlow + +* Overview +SQLFlow is a bridge that connects a SQL engine + +- MySQL, ODPS, Hive, SparkSQL, Oracle, SQL Server, ... + +with a Machine Learning toolkits + +- TensorFlow, PyTorch, XGboost, ... + +It extends the SQL syntax to enable model training and inference. + +* Syntax: Regular SQL +.code regular.sql + +* Syntax: Train +.code train.sql + +* Syntax: Prediction +.code pred.sql + +* Demo + +* Next Step: More DB +- ODPS +- Upon request + +* Next Step: More Models +- Currently only supports TF Estimator: `DNNClassifier`, `DNNRegressor` +- More TF models to come: `BoostedTreesClassifier`, `BoostedTreesRegressor` +- Upon request + +* Next Step: Data Preprocessing +Introduce proprocessing pipeline such as `norm`, `bucket`, `hash`. Also upon request. +.code preprocess.sql + +* Next Step: AutoML +.code automl.sql + diff --git a/sqlflow/cmd/demo/slides/201812/train.sql b/sqlflow/cmd/demo/slides/201812/train.sql new file mode 100644 index 0000000..89fab9d --- /dev/null +++ b/sqlflow/cmd/demo/slides/201812/train.sql @@ -0,0 +1,12 @@ +sqlflow> SELECT * +FROM iris.train +TRAIN DNNClassifier +WITH + n_classes = 3, + hidden_units = [10, 20] +COLUMN sepal_length, sepal_width, petal_length, petal_width +LABEL class +INTO sqlflow_models.my_dnn_model; +----------------------------- +2018/12/16 15:03:54 tensorflowCmd: run in Docker container +Job success diff --git a/sqlflow/cmd/sqlflowserver/index.html b/sqlflow/cmd/sqlflowserver/index.html deleted file mode 100644 index 7237249..0000000 --- a/sqlflow/cmd/sqlflowserver/index.html +++ /dev/null @@ -1,13 +0,0 @@ - - - - - - - - -

    Redirecting to https://github.com/sql-machine-learning/sqlflow

    - - diff --git a/sqlflow/cmd/sqlflowserver/main.go b/sqlflow/cmd/sqlflowserver/main.go new file mode 100644 index 0000000..fff731f --- /dev/null +++ b/sqlflow/cmd/sqlflowserver/main.go @@ -0,0 +1,59 @@ +// To run this program: +// go generate .. && go run main.go +// +package main + +import ( + "flag" + "log" + "net" + + "github.com/go-sql-driver/mysql" + "google.golang.org/grpc" + "google.golang.org/grpc/reflection" + + "github.com/sql-machine-learning/sqlflow/server" + pb "github.com/sql-machine-learning/sqlflow/server/proto" + sf "github.com/sql-machine-learning/sqlflow/sql" +) + +const ( + port = ":50051" +) + +func main() { + user := flag.String("db_user", "root", "database user name") + pswd := flag.String("db_password", "root", "database user password") + addr := flag.String("db_address", "", "database address, such as: localhost:3306") + flag.Parse() + + cfg := &mysql.Config{ + User: *user, + Passwd: *pswd, + Net: "tcp", + Addr: *addr, + AllowNativePasswords: true, + } + db, err := sf.Open("mysql", cfg.FormatDSN()) + if err != nil { + log.Fatalf("failed to open database: %v", err) + } + defer db.Close() + if err := db.Ping(); err != nil { + log.Fatalf("failed to ping database: %v", err) + } + + lis, err := net.Listen("tcp", port) + if err != nil { + log.Fatalf("failed to listen: %v", err) + } + + s := grpc.NewServer() + pb.RegisterSQLFlowServer(s, server.NewServer(sf.Run, db)) + // Register reflection service on gRPC server. + reflection.Register(s) + log.Println("Server Started at", port) + if err := s.Serve(lis); err != nil { + log.Fatalf("failed to serve: %v", err) + } +} diff --git a/sqlflow/cmd/sqlflowserver/main_test.go b/sqlflow/cmd/sqlflowserver/main_test.go new file mode 100644 index 0000000..536b14e --- /dev/null +++ b/sqlflow/cmd/sqlflowserver/main_test.go @@ -0,0 +1,40 @@ +package main + +import ( + "context" + "log" + "testing" + "time" + + "google.golang.org/grpc" + + pb "github.com/sql-machine-learning/sqlflow/server/proto" + "github.com/stretchr/testify/assert" +) + +func TestRun(t *testing.T) { + a := assert.New(t) + tests := []string{ + "show databases;", + "select * from iris.train limit 2;", + } + + go main() + // FIXME(weiguo): We may need to expand sleep time + time.Sleep(2 * time.Second) + + conn, err := grpc.Dial("localhost"+port, grpc.WithInsecure()) + a.NoError(err) + defer conn.Close() + cli := pb.NewSQLFlowClient(conn) + + ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second) + defer cancel() + + for _, tc := range tests { + _, err := cli.Run(ctx, &pb.Request{Sql: tc}) + if err != nil { + log.Fatalf("Check if the server started successfully. %v", err) + } + } +} diff --git a/doc/build-tensorflow.md b/sqlflow/doc/build-tensorflow.md similarity index 100% rename from doc/build-tensorflow.md rename to sqlflow/doc/build-tensorflow.md diff --git a/doc/build.md b/sqlflow/doc/build.md similarity index 100% rename from doc/build.md rename to sqlflow/doc/build.md diff --git a/doc/close_producer_from_consumer.md b/sqlflow/doc/close_producer_from_consumer.md similarity index 100% rename from doc/close_producer_from_consumer.md rename to sqlflow/doc/close_producer_from_consumer.md diff --git a/doc/database_abstraction_layer.md b/sqlflow/doc/database_abstraction_layer.md similarity index 100% rename from doc/database_abstraction_layer.md rename to sqlflow/doc/database_abstraction_layer.md diff --git a/doc/figures/sqlflow-arch.graffle b/sqlflow/doc/figures/sqlflow-arch.graffle similarity index 100% rename from doc/figures/sqlflow-arch.graffle rename to sqlflow/doc/figures/sqlflow-arch.graffle diff --git a/doc/figures/sqlflow-arch.png b/sqlflow/doc/figures/sqlflow-arch.png similarity index 100% rename from doc/figures/sqlflow-arch.png rename to sqlflow/doc/figures/sqlflow-arch.png diff --git a/doc/figures/sqlflow-arch2.dot b/sqlflow/doc/figures/sqlflow-arch2.dot similarity index 100% rename from doc/figures/sqlflow-arch2.dot rename to sqlflow/doc/figures/sqlflow-arch2.dot diff --git a/doc/figures/sqlflow-arch2.png b/sqlflow/doc/figures/sqlflow-arch2.png similarity index 100% rename from doc/figures/sqlflow-arch2.png rename to sqlflow/doc/figures/sqlflow-arch2.png diff --git a/doc/mysql-setup.md b/sqlflow/doc/mysql-setup.md similarity index 100% rename from doc/mysql-setup.md rename to sqlflow/doc/mysql-setup.md diff --git a/doc/pipe.md b/sqlflow/doc/pipe.md similarity index 100% rename from doc/pipe.md rename to sqlflow/doc/pipe.md diff --git a/doc/quickstart.md b/sqlflow/doc/quickstart.md similarity index 88% rename from doc/quickstart.md rename to sqlflow/doc/quickstart.md index 6a3b67f..75a00a6 100644 --- a/doc/quickstart.md +++ b/sqlflow/doc/quickstart.md @@ -12,15 +12,16 @@ issue are welcomed. :) ## Demo 1: Jupyter Notebook -1. Start a Docker container that runs sqlflowserver and Jupyter Notebook +1. Start a Docker container that runs sqlflowserver and Jupyter Notebook. If you are + using Docker for Linux, please change `host.docker.internal:3306` to `localhost:3306`. + ``` docker run --rm -it -p 8888:8888 sqlflow/sqlflow:latest \ bash -c "sqlflowserver --db_user root --db_password root --db_address host.docker.internal:3306 & SQLFLOW_SERVER=localhost:50051 jupyter notebook --ip=0.0.0.0 --port=8888 --allow-root" ``` - If you are running MySQL on the localhost and you are using Docker for Mac, please - be aware the option `--db_address host.docker.internal:3306` where + If you are using Docker for Mac, please be aware the option `--db_address host.docker.internal:3306` where `host.docker.internal` translates to the host ip address as recommended [here](https://docs.docker.com/docker-for-mac/networking/). If you are running MySQL on remote, please be aware that MySQL only allows connections from localhost @@ -38,7 +39,8 @@ can create notebooks. In a cell, you should be able to type ## Demo 2: Command Line Prompt -1. Start a Docker container that runs SQLFlow command line prompt. +Start a Docker container that runs SQLFlow command line prompt. If you are using +Docker for Linux, please change `host.docker.internal:3306` to `localhost:3306`. ``` docker run -it --rm --net=host sqlflow/sqlflow:latest demo \ diff --git a/doc/sql_parser.md b/sqlflow/doc/sql_parser.md similarity index 100% rename from doc/sql_parser.md rename to sqlflow/doc/sql_parser.md diff --git a/docs/doc/submitter.md b/sqlflow/doc/submitter.md similarity index 99% rename from docs/doc/submitter.md rename to sqlflow/doc/submitter.md index da6001f..53b0030 100644 --- a/docs/doc/submitter.md +++ b/sqlflow/doc/submitter.md @@ -25,7 +25,6 @@ type ColumnType struct { // COLUMN sepal_length, sepal_width, petal_length, petal_width // LABEL class // INTO sqlflow_models.my_dnn_model; -{% raw %} type TrainDescription struct { StandardSelect string // e.g. SELECT * FROM iris.train Estimator string // e.g. DNNClassifier @@ -34,7 +33,7 @@ type TrainDescription struct { Y ColumnType // e.g. {"class", "INT"} ModelName string // e.g. my_dnn_model } -{% endraw %} + // SELECT * // FROM iris.test // PREDICT iris.predict.class diff --git a/doc/syntax.md b/sqlflow/doc/syntax.md similarity index 100% rename from doc/syntax.md rename to sqlflow/doc/syntax.md diff --git a/doc/walkthrough.md b/sqlflow/doc/walkthrough.md similarity index 100% rename from doc/walkthrough.md rename to sqlflow/doc/walkthrough.md diff --git a/sqlflow/example/creditcard/README.md b/sqlflow/example/creditcard/README.md new file mode 100644 index 0000000..623c61c --- /dev/null +++ b/sqlflow/example/creditcard/README.md @@ -0,0 +1,149 @@ +# The Credit Card Fraud Detection Example + +The `downsample/creditcard.csv` file in this directory comes from [Kaggle](https://www.kaggle.com/mlg-ulb/creditcardfraud). + +## Down-sample + +To make it an example that can be checked into a git repo, I randomly picked up 2000 out from the 284,808 lines from the original dataset using the following command: + +```bash +gshuf -n 5000 /tmp/creditcard.csv > ./downsample/creditcard.csv +``` + +where `/tmp/creditcard.csv` is the one downloaded from Kaggle. + +If you want to down-sample by yourself on macOS, you might need to install `gshuf`. + +```bash +brew install coreutils +``` + +## Import into MySQL + +To develop/test/debug the TensorFlow program automatically generated by SQLFlow, which is supposed to be able to read data from MySQL, we'd import the `downsample/creditcard.csv` into MySQL. + +### Run MySQL Server + +For more about running MySQL in Docker containers, please refer to [Run MySQL Server and Client in Docker Containers](../../doc/mysql-setup.md). + +The following command starts the MySQL server: + +```bash +docker run --rm \ + -v $HOME/work/creditcarddb:/var/lib/mysql \ + --name mysql01 \ + -e MYSQL_ROOT_PASSWORD=root \ + -e MYSQL_ROOT_HOST='%' \ + -p 3306:3306 \ + -d \ + mysql/mysql-server:8.0 \ + mysqld --secure-file-priv="" +``` + +Please be aware that the above command doesn't use the default entrypoint of the Docker image; instead, it explicitly start `mysqld` in the container so that it could provide the command line option `--secure-file-priv=""`, which is required by `mysqlimport`. + +Run the following command to start a MySQL client: + +```bash +docker run --rm -it \ + -v $HOME/work/creditcarddb:/var/lib/mysql \ + mysql/mysql-server:8.0 \ + mysql -uroot -p +``` + +Remember to type the password `root` when prompted. Then input the following SQL program in the client to create the table `creditcard.creditcard`: + +```sql +CREATE DATABASE creditcard; +USE creditcard; +CREATE TABLE IF NOT EXISTS creditcard ( + time INT, + v1 float, + v2 float, + v3 float, + v4 float, + v5 float, + v6 float, + v7 float, + v8 float, + v9 float, + v10 float, + v11 float, + v12 float, + v13 float, + v14 float, + v15 float, + v16 float, + v17 float, + v18 float, + v19 float, + v20 float, + v21 float, + v22 float, + v23 float, + v24 float, + v25 float, + v26 float, + v27 float, + v28 float, + amount float, + class varchar(255)); +``` + +Before we could import the CSV file, we must follow the security policy of MySQL, which is wired and hard to understand, and copy the CSV file into the database directory: + +```bash +cp downsample/creditcard.csv ~/work/creditcarddb/creditcard/ +``` + +The following command runs `mysqlimport` in the `mysql/mysql` Docker image to import the data: + +```bash +docker run --rm -it \ + -v $HOME/work/creditcarddb:/var/lib/mysql \ + -v $HOME/work/creditcard:/var/lib/mysql-files \ + mysql:8.0 \ + mysqlimport \ + --ignore-lines=1 \ + --fields-terminated-by=, \ + --socket /var/lib/mysql/mysql.sock \ + -u root -p \ + creditcard creditcard.csv +``` + +Please be aware that `mysqlimport` doesn't exist in the `mysql/mysql-server` image. + +A message likes the following indicates the success of importing: + +``` +creditcard.creditcard: Records: 1999 Deleted: 0 Skipped: 0 Warnings: 0 +``` + +We can verify the imported result by running a SELECT statement like the following + +```bash +docker run --rm -it \ + -v $HOME/work/creditcarddb:/var/lib/mysql \ + mysql/mysql-server:8.0 \ + mysql -uroot -p \ + -e 'use creditcard; select time, v1, v28, amount, class from creditcard limit 10;' +``` + +which should print something like the following + +``` ++--------+-----------+------------+--------+-------+ +| time | v1 | v28 | amount | class | ++--------+-----------+------------+--------+-------+ +| 123590 | -0.341154 | 0.130741 | 69 | "0" | +| 135797 | -0.118853 | 0.11599 | 11.79 | "0" | +| 161249 | -1.04129 | -0.362025 | 7.38 | "0" | +| 72471 | 0.603151 | 0.0551408 | 352.04 | "0" | +| 120182 | 2.10678 | -0.0566284 | 0.89 | "0" | +| 160476 | 2.37223 | -0.0724277 | 6 | "0" | +| 135023 | 1.87987 | 0.0214597 | 20.05 | "0" | +| 117329 | -0.375213 | -0.130152 | 159 | "0" | +| 148637 | 1.83178 | -0.0378951 | 38.37 | "0" | +| 48293 | 1.26094 | 0.0275234 | 12 | "0" | ++--------+-----------+------------+--------+-------+ +``` diff --git a/sqlflow/example/creditcard/initialize_dataset_and_database.bash b/sqlflow/example/creditcard/initialize_dataset_and_database.bash new file mode 100755 index 0000000..560e618 --- /dev/null +++ b/sqlflow/example/creditcard/initialize_dataset_and_database.bash @@ -0,0 +1,127 @@ +#!/bin/bash +################################################################################ + +if [[ $(which docker) == "" ]]; then + echo "Cannot find Docker. " + echo "Please install it referring https://docs.docker.com/install" + exit +fi + +docker version > /dev/null +if [[ $? != "0" ]]; then + echo "The command docker version doesn't work as expected." + echo "Please check the reason and retry." + exit +fi + +if [[ $(which kaggle) == "" ]]; then + echo "Cannot find the command line tool kaggle. Try to install it ..."; + pip install kaggle --upgrade + if [[ $? != "0" ]]; then + echo "Failed to install kaggle." + echo "Please check https://github.com/Kaggle/kaggle-api" + exit + fi +fi + +if [[ ! -f $HOME/.kaggle/kaggle.json ]]; then + echo "Please sign up and get your API credentials from kaggle.com" + echo "Please check https://github.com/Kaggle/kaggle-api" + exit +fi + +export SQLFLOW_DB_DIR=$HOME/.sqlflow/example/db/creditcard +export SQLFLOW_DATABASE=creditcard + +echo "Cleanup legacy resource ..." +docker stop mysql-creditcard +docker rm mysql-creditcard +rm -rf $SQLFLOW_DB_DIR +mkdir -p $SQLFLOW_DB_DIR + +echo "Starting the MySQL server in a Docker container ..." +docker run --rm \ + -v $SQLFLOW_DB_DIR:/var/lib/mysql \ + --name mysql-creditcard \ + -e MYSQL_ROOT_PASSWORD=root \ + -e MYSQL_ROOT_HOST='%' \ + -p 3306:3306 \ + -d \ + mysql/mysql-server:8.0 \ + mysqld --secure-file-priv="" + +echo "Downloading the dataset ..." +kaggle datasets download -p /tmp --unzip \ + -d mlg-ulb/creditcardfraud + +if [[ ! -S $SQLFLOW_DB_DIR/mysql.sock ]]; then + echo "Cannot find " $SQLFLOW_DB_DIR/mysql.sock + echo "Waiting for 10 more seconds ..." + sleep 10 + if [[ ! -S $SQLFLOW_DB_DIR/mysql.sock ]]; then + echo "Seems that something went wrong." + exit + fi +fi + +echo "Creating the database and table creditcard.creditcard ..." +docker run --rm -it \ + -v $SQLFLOW_DB_DIR:/var/lib/mysql \ + mysql/mysql-server:8.0 \ + mysql --user root --password=root \ + -e 'CREATE DATABASE creditcard; +USE creditcard; +CREATE TABLE IF NOT EXISTS creditcard ( + time INT, + v1 float, + v2 float, + v3 float, + v4 float, + v5 float, + v6 float, + v7 float, + v8 float, + v9 float, + v10 float, + v11 float, + v12 float, + v13 float, + v14 float, + v15 float, + v16 float, + v17 float, + v18 float, + v19 float, + v20 float, + v21 float, + v22 float, + v23 float, + v24 float, + v25 float, + v26 float, + v27 float, + v28 float, + amount float, + class varchar(255));' + +if [[ $? != "0" ]]; then + echo "Failed to create table" + exit +fi + +if [[ ! -f /tmp/creditcard.csv ]]; then + echo "Failed to download the dataset" + exit +fi + +echo "Importing the CSV file ..." +mv /tmp/creditcard.csv $SQLFLOW_DB_DIR/creditcard +docker run --rm -it \ + -v $SQLFLOW_DB_DIR:/var/lib/mysql \ + mysql:8.0 \ + mysqlimport \ + --ignore-lines=1 \ + --fields-terminated-by=, \ + --socket /var/lib/mysql/mysql.sock \ + -u root --password=root \ + creditcard creditcard.csv diff --git a/sqlflow/example/datasets/Dockerfile b/sqlflow/example/datasets/Dockerfile new file mode 100644 index 0000000..7c7136d --- /dev/null +++ b/sqlflow/example/datasets/Dockerfile @@ -0,0 +1,4 @@ +FROM mysql/mysql-server:8.0 + +COPY popularize_churn.sql /popularize_churn.sql +COPY popularize_iris.sql /popularize_iris.sql diff --git a/sqlflow/example/datasets/README.md b/sqlflow/example/datasets/README.md new file mode 100644 index 0000000..182575a --- /dev/null +++ b/sqlflow/example/datasets/README.md @@ -0,0 +1,75 @@ +# The MySQL Server Container for Testing + +This directory contains a Dockerfile that builds a Docker image derived the MySQL Server 8.0 image, and includes SQL programs that popularize the following datasets: + +1. [Churn from Kaggle](https://www.kaggle.com/blastchar/telco-customer-churn) +1. [Irises classfication from TensorFlow](https://www.tensorflow.org/guide/premade_estimators#classifying_irises_an_overview) + +We can run a Docker container of it for unit testing. + +## Build + +```bash +cd example/datasets +docker build -t sqlflow:data . +``` + +## Run + +```bash +docker run --rm -d --name sqlflowdata \ + -p 3306:3306 \ + -e MYSQL_ROOT_PASSWORD=root \ + -e MYSQL_ROOT_HOST=% \ + sqlflow:data +``` + +## Popularize Datasets + +We need to manually popularize the databases and tables: + +```bash +docker exec -it sqlflowdata bash +``` + +To popularize the Churn dataset into `churn`: + +```bash +cat /popularize_churn.sql | mysql -uroot -proot +``` + +To popularize the Irises dataset into `iris`: + +```bash +cat /popularize_iris.sql | mysql -uroot -proot +``` + +To prepare database for storing machine learning models: + +```bash +echo "CREATE DATABASE IF NOT EXISTS sqlflow_models;" | mysql -uroot -proot +``` + +## Query + +In the container, run + +```bash +echo "select count(*) from churn.test;" | mysql -uroot -proot +``` + +should print the number of rows as the following + +``` +count(*) +10 +``` + +## Trouble shooting: + +1. It usually takes about 15 seconds to bring up the MySQL Server. If you try to connect it +before that, you may see the following error + +``` +ERROR 1045 (28000): Access denied for user 'root'@'localhost' (using password: YES) +``` diff --git a/sqlflow/example/datasets/data_to_sql.awk b/sqlflow/example/datasets/data_to_sql.awk new file mode 100644 index 0000000..36a9af1 --- /dev/null +++ b/sqlflow/example/datasets/data_to_sql.awk @@ -0,0 +1,20 @@ +# To get the INSERT statements in popularize_table.sql, please apply +# this script to the dataset downloaded from +# https://www.kaggle.com/blastchar/telco-customer-churn. Then, +# manually remove the few lines that contains " ". Before applying +# this AWK program, please convert the downloaded CSV file from DOS +# format into UNIX format, and remove the first line of metadata. +BEGIN { + FS = ","; +} + +{ + printf("INSERT INTO %s VALUES(", table); + for (i = 1; i <= NF; i++) { + printf("\"%s\"", $i); + if (i < NF) { + printf(","); + } + } + printf(");\n"); +} diff --git a/sqlflow/example/datasets/entrypoint.sh b/sqlflow/example/datasets/entrypoint.sh new file mode 100755 index 0000000..85c75ac --- /dev/null +++ b/sqlflow/example/datasets/entrypoint.sh @@ -0,0 +1,8 @@ +#!/bin/bash + +# Call the mysql/mysql-server:8.0 default entrypoint to start mysqld. +/entrypoint.sh mysqld + +# Create database, table, and popularize the table. +cat /create_table.sql | mysql -uroot -proot +cat /popularize_table.sql | mysql -uroot -proot diff --git a/sqlflow/example/datasets/popularize_churn.sql b/sqlflow/example/datasets/popularize_churn.sql new file mode 100644 index 0000000..5a66a0e --- /dev/null +++ b/sqlflow/example/datasets/popularize_churn.sql @@ -0,0 +1,144 @@ +CREATE DATABASE IF NOT EXISTS churn; +DROP TABLE IF EXISTS churn.train; +CREATE TABLE churn.train ( + customerID VARCHAR(255), + gender VARCHAR(255), + SeniorCitizen TINYINT, + Partner VARCHAR(255), + Dependents VARCHAR(255), + tenure INT, + PhoneService VARCHAR(255), + MultipleLines VARCHAR(255), + InternetService VARCHAR(255), + OnlineSecurity VARCHAR(255), + OnlineBackup VARCHAR(255), + DeviceProtection VARCHAR(255), + TechSupport VARCHAR(255), + StreamingTV VARCHAR(255), + StreamingMovies VARCHAR(255), + Contract VARCHAR(255), + PaperlessBilling VARCHAR(255), + PaymentMethod VARCHAR(255), + MonthlyCharges FLOAT, + TotalCharges FLOAT, + Churn VARCHAR(255), + PRIMARY KEY ( customerID ) +); +INSERT INTO churn.train VALUES("7590-VHVEG","Female","0","Yes","No","1","No","No phone service","DSL","No","Yes","No","No","No","No","Month-to-month","Yes","Electronic check","29.85","29.85","No"); +INSERT INTO churn.train VALUES("5575-GNVDE","Male","0","No","No","34","Yes","No","DSL","Yes","No","Yes","No","No","No","One year","No","Mailed check","56.95","1889.5","No"); +INSERT INTO churn.train VALUES("3668-QPYBK","Male","0","No","No","2","Yes","No","DSL","Yes","Yes","No","No","No","No","Month-to-month","Yes","Mailed check","53.85","108.15","Yes"); +INSERT INTO churn.train VALUES("7795-CFOCW","Male","0","No","No","45","No","No phone service","DSL","Yes","No","Yes","Yes","No","No","One year","No","Bank transfer (automatic)","42.3","1840.75","No"); +INSERT INTO churn.train VALUES("9237-HQITU","Female","0","No","No","2","Yes","No","Fiber optic","No","No","No","No","No","No","Month-to-month","Yes","Electronic check","70.7","151.65","Yes"); +INSERT INTO churn.train VALUES("9305-CDSKC","Female","0","No","No","8","Yes","Yes","Fiber optic","No","No","Yes","No","Yes","Yes","Month-to-month","Yes","Electronic check","99.65","820.5","Yes"); +INSERT INTO churn.train VALUES("1452-KIOVK","Male","0","No","Yes","22","Yes","Yes","Fiber optic","No","Yes","No","No","Yes","No","Month-to-month","Yes","Credit card (automatic)","89.1","1949.4","No"); +INSERT INTO churn.train VALUES("6713-OKOMC","Female","0","No","No","10","No","No phone service","DSL","Yes","No","No","No","No","No","Month-to-month","No","Mailed check","29.75","301.9","No"); +INSERT INTO churn.train VALUES("7892-POOKP","Female","0","Yes","No","28","Yes","Yes","Fiber optic","No","No","Yes","Yes","Yes","Yes","Month-to-month","Yes","Electronic check","104.8","3046.05","Yes"); +INSERT INTO churn.train VALUES("6388-TABGU","Male","0","No","Yes","62","Yes","No","DSL","Yes","Yes","No","No","No","No","One year","No","Bank transfer (automatic)","56.15","3487.95","No"); +INSERT INTO churn.train VALUES("9763-GRSKD","Male","0","Yes","Yes","13","Yes","No","DSL","Yes","No","No","No","No","No","Month-to-month","Yes","Mailed check","49.95","587.45","No"); +INSERT INTO churn.train VALUES("7469-LKBCI","Male","0","No","No","16","Yes","No","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","Two year","No","Credit card (automatic)","18.95","326.8","No"); +INSERT INTO churn.train VALUES("8091-TTVAX","Male","0","Yes","No","58","Yes","Yes","Fiber optic","No","No","Yes","No","Yes","Yes","One year","No","Credit card (automatic)","100.35","5681.1","No"); +INSERT INTO churn.train VALUES("0280-XJGEX","Male","0","No","No","49","Yes","Yes","Fiber optic","No","Yes","Yes","No","Yes","Yes","Month-to-month","Yes","Bank transfer (automatic)","103.7","5036.3","Yes"); +INSERT INTO churn.train VALUES("5129-JLPIS","Male","0","No","No","25","Yes","No","Fiber optic","Yes","No","Yes","Yes","Yes","Yes","Month-to-month","Yes","Electronic check","105.5","2686.05","No"); +INSERT INTO churn.train VALUES("3655-SNQYZ","Female","0","Yes","Yes","69","Yes","Yes","Fiber optic","Yes","Yes","Yes","Yes","Yes","Yes","Two year","No","Credit card (automatic)","113.25","7895.15","No"); +INSERT INTO churn.train VALUES("8191-XWSZG","Female","0","No","No","52","Yes","No","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","One year","No","Mailed check","20.65","1022.95","No"); +INSERT INTO churn.train VALUES("9959-WOFKT","Male","0","No","Yes","71","Yes","Yes","Fiber optic","Yes","No","Yes","No","Yes","Yes","Two year","No","Bank transfer (automatic)","106.7","7382.25","No"); +INSERT INTO churn.train VALUES("4190-MFLUW","Female","0","Yes","Yes","10","Yes","No","DSL","No","No","Yes","Yes","No","No","Month-to-month","No","Credit card (automatic)","55.2","528.35","Yes"); +INSERT INTO churn.train VALUES("4183-MYFRB","Female","0","No","No","21","Yes","No","Fiber optic","No","Yes","Yes","No","No","Yes","Month-to-month","Yes","Electronic check","90.05","1862.9","No"); +INSERT INTO churn.train VALUES("8779-QRDMV","Male","1","No","No","1","No","No phone service","DSL","No","No","Yes","No","No","Yes","Month-to-month","Yes","Electronic check","39.65","39.65","Yes"); +INSERT INTO churn.train VALUES("1680-VDCWW","Male","0","Yes","No","12","Yes","No","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","One year","No","Bank transfer (automatic)","19.8","202.25","No"); +INSERT INTO churn.train VALUES("1066-JKSGK","Male","0","No","No","1","Yes","No","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","Month-to-month","No","Mailed check","20.15","20.15","Yes"); +INSERT INTO churn.train VALUES("3638-WEABW","Female","0","Yes","No","58","Yes","Yes","DSL","No","Yes","No","Yes","No","No","Two year","Yes","Credit card (automatic)","59.9","3505.1","No"); +INSERT INTO churn.train VALUES("6322-HRPFA","Male","0","Yes","Yes","49","Yes","No","DSL","Yes","Yes","No","Yes","No","No","Month-to-month","No","Credit card (automatic)","59.6","2970.3","No"); +INSERT INTO churn.train VALUES("6865-JZNKO","Female","0","No","No","30","Yes","No","DSL","Yes","Yes","No","No","No","No","Month-to-month","Yes","Bank transfer (automatic)","55.3","1530.6","No"); +INSERT INTO churn.train VALUES("6467-CHFZW","Male","0","Yes","Yes","47","Yes","Yes","Fiber optic","No","Yes","No","No","Yes","Yes","Month-to-month","Yes","Electronic check","99.35","4749.15","Yes"); +INSERT INTO churn.train VALUES("8665-UTDHZ","Male","0","Yes","Yes","1","No","No phone service","DSL","No","Yes","No","No","No","No","Month-to-month","No","Electronic check","30.2","30.2","Yes"); +INSERT INTO churn.train VALUES("5248-YGIJN","Male","0","Yes","No","72","Yes","Yes","DSL","Yes","Yes","Yes","Yes","Yes","Yes","Two year","Yes","Credit card (automatic)","90.25","6369.45","No"); +INSERT INTO churn.train VALUES("8773-HHUOZ","Female","0","No","Yes","17","Yes","No","DSL","No","No","No","No","Yes","Yes","Month-to-month","Yes","Mailed check","64.7","1093.1","Yes"); +INSERT INTO churn.train VALUES("3841-NFECX","Female","1","Yes","No","71","Yes","Yes","Fiber optic","Yes","Yes","Yes","Yes","No","No","Two year","Yes","Credit card (automatic)","96.35","6766.95","No"); +INSERT INTO churn.train VALUES("4929-XIHVW","Male","1","Yes","No","2","Yes","No","Fiber optic","No","No","Yes","No","Yes","Yes","Month-to-month","Yes","Credit card (automatic)","95.5","181.65","No"); +INSERT INTO churn.train VALUES("6827-IEAUQ","Female","0","Yes","Yes","27","Yes","No","DSL","Yes","Yes","Yes","Yes","No","No","One year","No","Mailed check","66.15","1874.45","No"); +INSERT INTO churn.train VALUES("7310-EGVHZ","Male","0","No","No","1","Yes","No","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","Month-to-month","No","Bank transfer (automatic)","20.2","20.2","No"); +INSERT INTO churn.train VALUES("3413-BMNZE","Male","1","No","No","1","Yes","No","DSL","No","No","No","No","No","No","Month-to-month","No","Bank transfer (automatic)","45.25","45.25","No"); +INSERT INTO churn.train VALUES("6234-RAAPL","Female","0","Yes","Yes","72","Yes","Yes","Fiber optic","Yes","Yes","No","Yes","Yes","No","Two year","No","Bank transfer (automatic)","99.9","7251.7","No"); +INSERT INTO churn.train VALUES("6047-YHPVI","Male","0","No","No","5","Yes","No","Fiber optic","No","No","No","No","No","No","Month-to-month","Yes","Electronic check","69.7","316.9","Yes"); +INSERT INTO churn.train VALUES("6572-ADKRS","Female","0","No","No","46","Yes","No","Fiber optic","No","No","Yes","No","No","No","Month-to-month","Yes","Credit card (automatic)","74.8","3548.3","No"); +INSERT INTO churn.train VALUES("5380-WJKOV","Male","0","No","No","34","Yes","Yes","Fiber optic","No","Yes","Yes","No","Yes","Yes","Month-to-month","Yes","Electronic check","106.35","3549.25","Yes"); +INSERT INTO churn.train VALUES("8168-UQWWF","Female","0","No","No","11","Yes","Yes","Fiber optic","No","No","Yes","No","Yes","Yes","Month-to-month","Yes","Bank transfer (automatic)","97.85","1105.4","Yes"); +INSERT INTO churn.train VALUES("8865-TNMNX","Male","0","Yes","Yes","10","Yes","No","DSL","No","Yes","No","No","No","No","One year","No","Mailed check","49.55","475.7","No"); +INSERT INTO churn.train VALUES("9489-DEDVP","Female","0","Yes","Yes","70","Yes","Yes","DSL","Yes","Yes","No","No","Yes","No","Two year","Yes","Credit card (automatic)","69.2","4872.35","No"); +INSERT INTO churn.train VALUES("9867-JCZSP","Female","0","Yes","Yes","17","Yes","No","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","One year","No","Mailed check","20.75","418.25","No"); +INSERT INTO churn.train VALUES("4671-VJLCL","Female","0","No","No","63","Yes","Yes","DSL","Yes","Yes","Yes","Yes","Yes","No","Two year","Yes","Credit card (automatic)","79.85","4861.45","No"); +INSERT INTO churn.train VALUES("4080-IIARD","Female","0","Yes","No","13","Yes","Yes","DSL","Yes","Yes","No","Yes","Yes","No","Month-to-month","Yes","Electronic check","76.2","981.45","No"); +INSERT INTO churn.train VALUES("3714-NTNFO","Female","0","No","No","49","Yes","Yes","Fiber optic","No","No","No","No","No","Yes","Month-to-month","Yes","Electronic check","84.5","3906.7","No"); +INSERT INTO churn.train VALUES("5948-UJZLF","Male","0","No","No","2","Yes","No","DSL","No","Yes","No","No","No","No","Month-to-month","No","Mailed check","49.25","97","No"); +INSERT INTO churn.train VALUES("7760-OYPDY","Female","0","No","No","2","Yes","No","Fiber optic","No","No","No","No","Yes","No","Month-to-month","Yes","Electronic check","80.65","144.15","Yes"); +INSERT INTO churn.train VALUES("7639-LIAYI","Male","0","No","No","52","Yes","Yes","DSL","Yes","No","No","Yes","Yes","Yes","Two year","Yes","Credit card (automatic)","79.75","4217.8","No"); +INSERT INTO churn.train VALUES("2954-PIBKO","Female","0","Yes","Yes","69","Yes","Yes","DSL","Yes","No","Yes","Yes","No","No","Two year","Yes","Credit card (automatic)","64.15","4254.1","No"); +INSERT INTO churn.train VALUES("8012-SOUDQ","Female","1","No","No","43","Yes","Yes","Fiber optic","No","Yes","No","No","Yes","No","Month-to-month","Yes","Electronic check","90.25","3838.75","No"); +INSERT INTO churn.train VALUES("9420-LOJKX","Female","0","No","No","15","Yes","No","Fiber optic","Yes","Yes","No","No","Yes","Yes","Month-to-month","Yes","Credit card (automatic)","99.1","1426.4","Yes"); +INSERT INTO churn.train VALUES("6575-SUVOI","Female","1","Yes","No","25","Yes","Yes","DSL","Yes","No","No","Yes","Yes","No","Month-to-month","Yes","Credit card (automatic)","69.5","1752.65","No"); +INSERT INTO churn.train VALUES("7495-OOKFY","Female","1","Yes","No","8","Yes","Yes","Fiber optic","No","Yes","No","No","No","No","Month-to-month","Yes","Credit card (automatic)","80.65","633.3","Yes"); +INSERT INTO churn.train VALUES("4667-QONEA","Female","1","Yes","Yes","60","Yes","No","DSL","Yes","Yes","Yes","Yes","No","Yes","One year","Yes","Credit card (automatic)","74.85","4456.35","No"); +INSERT INTO churn.train VALUES("1658-BYGOY","Male","1","No","No","18","Yes","Yes","Fiber optic","No","No","No","No","Yes","Yes","Month-to-month","Yes","Electronic check","95.45","1752.55","Yes"); +INSERT INTO churn.train VALUES("8769-KKTPH","Female","0","Yes","Yes","63","Yes","Yes","Fiber optic","Yes","No","No","No","Yes","Yes","One year","Yes","Credit card (automatic)","99.65","6311.2","No"); +INSERT INTO churn.train VALUES("5067-XJQFU","Male","1","Yes","Yes","66","Yes","Yes","Fiber optic","No","Yes","Yes","Yes","Yes","Yes","One year","Yes","Electronic check","108.45","7076.35","No"); +INSERT INTO churn.train VALUES("3957-SQXML","Female","0","Yes","Yes","34","Yes","Yes","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","Two year","No","Credit card (automatic)","24.95","894.3","No"); +INSERT INTO churn.train VALUES("5954-BDFSG","Female","0","No","No","72","Yes","Yes","Fiber optic","No","No","Yes","Yes","Yes","Yes","Two year","Yes","Credit card (automatic)","107.5","7853.7","No"); +INSERT INTO churn.train VALUES("0434-CSFON","Female","0","Yes","No","47","Yes","Yes","Fiber optic","No","No","Yes","No","Yes","Yes","Month-to-month","Yes","Electronic check","100.5","4707.1","No"); +INSERT INTO churn.train VALUES("1215-FIGMP","Male","0","No","No","60","Yes","Yes","Fiber optic","No","Yes","No","No","Yes","No","Month-to-month","Yes","Bank transfer (automatic)","89.9","5450.7","No"); +INSERT INTO churn.train VALUES("0526-SXDJP","Male","0","Yes","No","72","No","No phone service","DSL","Yes","Yes","Yes","No","No","No","Two year","No","Bank transfer (automatic)","42.1","2962","No"); +INSERT INTO churn.train VALUES("0557-ASKVU","Female","0","Yes","Yes","18","Yes","No","DSL","No","No","Yes","Yes","No","No","One year","Yes","Credit card (automatic)","54.4","957.1","No"); +INSERT INTO churn.train VALUES("5698-BQJOH","Female","0","No","No","9","Yes","Yes","Fiber optic","No","No","No","No","Yes","Yes","Month-to-month","No","Electronic check","94.4","857.25","Yes"); +INSERT INTO churn.train VALUES("5122-CYFXA","Female","0","No","No","3","Yes","No","DSL","No","Yes","No","Yes","Yes","Yes","Month-to-month","Yes","Electronic check","75.3","244.1","No"); +INSERT INTO churn.train VALUES("8627-ZYGSZ","Male","0","Yes","No","47","Yes","Yes","Fiber optic","No","Yes","No","No","No","No","One year","Yes","Electronic check","78.9","3650.35","No"); +INSERT INTO churn.train VALUES("3410-YOQBQ","Female","0","No","No","31","Yes","No","DSL","No","Yes","Yes","Yes","Yes","Yes","Two year","No","Mailed check","79.2","2497.2","No"); +INSERT INTO churn.train VALUES("3170-NMYVV","Female","0","Yes","Yes","50","Yes","No","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","Two year","No","Bank transfer (automatic)","20.15","930.9","No"); +INSERT INTO churn.train VALUES("7410-OIEDU","Male","0","No","No","10","Yes","No","Fiber optic","Yes","No","Yes","No","No","No","Month-to-month","Yes","Mailed check","79.85","887.35","No"); +INSERT INTO churn.train VALUES("2273-QCKXA","Male","0","No","No","1","Yes","No","DSL","No","No","No","Yes","No","No","Month-to-month","No","Mailed check","49.05","49.05","No"); +INSERT INTO churn.train VALUES("0731-EBJQB","Female","0","Yes","Yes","52","Yes","No","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","One year","Yes","Electronic check","20.4","1090.65","No"); +INSERT INTO churn.train VALUES("1891-QRQSA","Male","1","Yes","Yes","64","Yes","Yes","Fiber optic","Yes","No","Yes","Yes","Yes","Yes","Two year","Yes","Bank transfer (automatic)","111.6","7099","No"); +INSERT INTO churn.train VALUES("8028-PNXHQ","Male","0","Yes","Yes","62","Yes","Yes","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","Two year","Yes","Bank transfer (automatic)","24.25","1424.6","No"); +INSERT INTO churn.train VALUES("5630-AHZIL","Female","0","No","Yes","3","Yes","No","DSL","Yes","No","No","Yes","No","Yes","Month-to-month","Yes","Bank transfer (automatic)","64.5","177.4","No"); +INSERT INTO churn.train VALUES("2673-CXQEU","Female","1","No","No","56","Yes","Yes","Fiber optic","Yes","Yes","Yes","No","Yes","Yes","One year","No","Electronic check","110.5","6139.5","No"); +INSERT INTO churn.train VALUES("6416-JNVRK","Female","0","No","No","46","Yes","No","DSL","No","No","No","No","No","Yes","One year","No","Credit card (automatic)","55.65","2688.85","No"); +INSERT INTO churn.train VALUES("5590-ZSKRV","Female","0","Yes","Yes","8","Yes","No","DSL","Yes","Yes","No","No","No","No","Month-to-month","No","Mailed check","54.65","482.25","No"); +INSERT INTO churn.train VALUES("0191-ZHSKZ","Male","1","No","No","30","Yes","No","DSL","Yes","Yes","No","No","Yes","Yes","Month-to-month","Yes","Electronic check","74.75","2111.3","No"); +INSERT INTO churn.train VALUES("3887-PBQAO","Female","0","Yes","Yes","45","Yes","Yes","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","One year","Yes","Credit card (automatic)","25.9","1216.6","No"); +INSERT INTO churn.train VALUES("5919-TMRGD","Female","0","No","Yes","1","Yes","No","Fiber optic","No","No","No","No","Yes","No","Month-to-month","Yes","Electronic check","79.35","79.35","Yes"); +INSERT INTO churn.train VALUES("8108-UXRQN","Female","0","Yes","Yes","11","No","No phone service","DSL","Yes","No","No","No","Yes","Yes","Month-to-month","No","Electronic check","50.55","565.35","No"); + +DROP TABLE IF EXISTS churn.test; +CREATE TABLE churn.test ( + customerID VARCHAR(255), + gender VARCHAR(255), + SeniorCitizen TINYINT, + Partner VARCHAR(255), + Dependents VARCHAR(255), + tenure INT, + PhoneService VARCHAR(255), + MultipleLines VARCHAR(255), + InternetService VARCHAR(255), + OnlineSecurity VARCHAR(255), + OnlineBackup VARCHAR(255), + DeviceProtection VARCHAR(255), + TechSupport VARCHAR(255), + StreamingTV VARCHAR(255), + StreamingMovies VARCHAR(255), + Contract VARCHAR(255), + PaperlessBilling VARCHAR(255), + PaymentMethod VARCHAR(255), + MonthlyCharges FLOAT, + TotalCharges FLOAT, + Churn VARCHAR(255), + PRIMARY KEY ( customerID ) +); +INSERT INTO churn.test VALUES("9191-MYQKX","Female","0","Yes","No","7","Yes","No","Fiber optic","No","No","Yes","No","No","No","Month-to-month","Yes","Bank transfer (automatic)","75.15","496.9","Yes"); +INSERT INTO churn.test VALUES("9919-YLNNG","Female","0","No","No","42","Yes","No","Fiber optic","No","Yes","Yes","Yes","Yes","Yes","Month-to-month","Yes","Bank transfer (automatic)","103.8","4327.5","No"); +INSERT INTO churn.test VALUES("0318-ZOPWS","Female","0","Yes","No","49","Yes","No","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","Two year","Yes","Bank transfer (automatic)","20.15","973.35","No"); +INSERT INTO churn.test VALUES("4445-ZJNMU","Male","0","No","No","9","Yes","Yes","Fiber optic","No","Yes","No","No","Yes","Yes","Month-to-month","Yes","Credit card (automatic)","99.3","918.75","No"); +INSERT INTO churn.test VALUES("4808-YNLEU","Female","0","Yes","No","35","Yes","No","DSL","Yes","No","No","No","Yes","No","One year","Yes","Bank transfer (automatic)","62.15","2215.45","No"); +INSERT INTO churn.test VALUES("1862-QRWPE","Female","0","Yes","Yes","48","Yes","No","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","Two year","No","Bank transfer (automatic)","20.65","1057","No"); +INSERT INTO churn.test VALUES("2796-NNUFI","Female","0","Yes","Yes","46","Yes","No","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","Two year","Yes","Mailed check","19.95","927.1","No"); +INSERT INTO churn.test VALUES("3016-KSVCP","Male","0","Yes","No","29","No","No phone service","DSL","No","No","No","No","Yes","No","Month-to-month","No","Mailed check","33.75","1009.25","No"); +INSERT INTO churn.test VALUES("4767-HZZHQ","Male","0","Yes","Yes","30","Yes","No","Fiber optic","No","Yes","Yes","No","No","No","Month-to-month","No","Bank transfer (automatic)","82.05","2570.2","No"); +INSERT INTO churn.test VALUES("2424-WVHPL","Male","1","No","No","1","Yes","No","Fiber optic","No","No","No","Yes","No","No","Month-to-month","No","Electronic check","74.7","74.7","No"); diff --git a/sqlflow/example/datasets/popularize_iris.sql b/sqlflow/example/datasets/popularize_iris.sql new file mode 100644 index 0000000..cc2b797 --- /dev/null +++ b/sqlflow/example/datasets/popularize_iris.sql @@ -0,0 +1,136 @@ +CREATE DATABASE IF NOT EXISTS iris; +DROP TABLE IF EXISTS iris.train; +CREATE TABLE iris.train ( + sepal_length float, + sepal_width float, + petal_length float, + petal_width float, + class int); +INSERT INTO iris.train VALUES(6.4,2.8,5.6,2.2,2); +INSERT INTO iris.train VALUES(5.0,2.3,3.3,1.0,1); +INSERT INTO iris.train VALUES(4.9,2.5,4.5,1.7,2); +INSERT INTO iris.train VALUES(4.9,3.1,1.5,0.1,0); +INSERT INTO iris.train VALUES(5.7,3.8,1.7,0.3,0); +INSERT INTO iris.train VALUES(4.4,3.2,1.3,0.2,0); +INSERT INTO iris.train VALUES(5.4,3.4,1.5,0.4,0); +INSERT INTO iris.train VALUES(6.9,3.1,5.1,2.3,2); +INSERT INTO iris.train VALUES(6.7,3.1,4.4,1.4,1); +INSERT INTO iris.train VALUES(5.1,3.7,1.5,0.4,0); +INSERT INTO iris.train VALUES(5.2,2.7,3.9,1.4,1); +INSERT INTO iris.train VALUES(6.9,3.1,4.9,1.5,1); +INSERT INTO iris.train VALUES(5.8,4.0,1.2,0.2,0); +INSERT INTO iris.train VALUES(5.4,3.9,1.7,0.4,0); +INSERT INTO iris.train VALUES(7.7,3.8,6.7,2.2,2); +INSERT INTO iris.train VALUES(6.3,3.3,4.7,1.6,1); +INSERT INTO iris.train VALUES(6.8,3.2,5.9,2.3,2); +INSERT INTO iris.train VALUES(7.6,3.0,6.6,2.1,2); +INSERT INTO iris.train VALUES(6.4,3.2,5.3,2.3,2); +INSERT INTO iris.train VALUES(5.7,4.4,1.5,0.4,0); +INSERT INTO iris.train VALUES(6.7,3.3,5.7,2.1,2); +INSERT INTO iris.train VALUES(6.4,2.8,5.6,2.1,2); +INSERT INTO iris.train VALUES(5.4,3.9,1.3,0.4,0); +INSERT INTO iris.train VALUES(6.1,2.6,5.6,1.4,2); +INSERT INTO iris.train VALUES(7.2,3.0,5.8,1.6,2); +INSERT INTO iris.train VALUES(5.2,3.5,1.5,0.2,0); +INSERT INTO iris.train VALUES(5.8,2.6,4.0,1.2,1); +INSERT INTO iris.train VALUES(5.9,3.0,5.1,1.8,2); +INSERT INTO iris.train VALUES(5.4,3.0,4.5,1.5,1); +INSERT INTO iris.train VALUES(6.7,3.0,5.0,1.7,1); +INSERT INTO iris.train VALUES(6.3,2.3,4.4,1.3,1); +INSERT INTO iris.train VALUES(5.1,2.5,3.0,1.1,1); +INSERT INTO iris.train VALUES(6.4,3.2,4.5,1.5,1); +INSERT INTO iris.train VALUES(6.8,3.0,5.5,2.1,2); +INSERT INTO iris.train VALUES(6.2,2.8,4.8,1.8,2); +INSERT INTO iris.train VALUES(6.9,3.2,5.7,2.3,2); +INSERT INTO iris.train VALUES(6.5,3.2,5.1,2.0,2); +INSERT INTO iris.train VALUES(5.8,2.8,5.1,2.4,2); +INSERT INTO iris.train VALUES(5.1,3.8,1.5,0.3,0); +INSERT INTO iris.train VALUES(4.8,3.0,1.4,0.3,0); +INSERT INTO iris.train VALUES(7.9,3.8,6.4,2.0,2); +INSERT INTO iris.train VALUES(5.8,2.7,5.1,1.9,2); +INSERT INTO iris.train VALUES(6.7,3.0,5.2,2.3,2); +INSERT INTO iris.train VALUES(5.1,3.8,1.9,0.4,0); +INSERT INTO iris.train VALUES(4.7,3.2,1.6,0.2,0); +INSERT INTO iris.train VALUES(6.0,2.2,5.0,1.5,2); +INSERT INTO iris.train VALUES(4.8,3.4,1.6,0.2,0); +INSERT INTO iris.train VALUES(7.7,2.6,6.9,2.3,2); +INSERT INTO iris.train VALUES(4.6,3.6,1.0,0.2,0); +INSERT INTO iris.train VALUES(7.2,3.2,6.0,1.8,2); +INSERT INTO iris.train VALUES(5.0,3.3,1.4,0.2,0); +INSERT INTO iris.train VALUES(6.6,3.0,4.4,1.4,1); +INSERT INTO iris.train VALUES(6.1,2.8,4.0,1.3,1); +INSERT INTO iris.train VALUES(5.0,3.2,1.2,0.2,0); +INSERT INTO iris.train VALUES(7.0,3.2,4.7,1.4,1); +INSERT INTO iris.train VALUES(6.0,3.0,4.8,1.8,2); +INSERT INTO iris.train VALUES(7.4,2.8,6.1,1.9,2); +INSERT INTO iris.train VALUES(5.8,2.7,5.1,1.9,2); +INSERT INTO iris.train VALUES(6.2,3.4,5.4,2.3,2); +INSERT INTO iris.train VALUES(5.0,2.0,3.5,1.0,1); +INSERT INTO iris.train VALUES(5.6,2.5,3.9,1.1,1); +INSERT INTO iris.train VALUES(6.7,3.1,5.6,2.4,2); +INSERT INTO iris.train VALUES(6.3,2.5,5.0,1.9,2); +INSERT INTO iris.train VALUES(6.4,3.1,5.5,1.8,2); +INSERT INTO iris.train VALUES(6.2,2.2,4.5,1.5,1); +INSERT INTO iris.train VALUES(7.3,2.9,6.3,1.8,2); +INSERT INTO iris.train VALUES(4.4,3.0,1.3,0.2,0); +INSERT INTO iris.train VALUES(7.2,3.6,6.1,2.5,2); +INSERT INTO iris.train VALUES(6.5,3.0,5.5,1.8,2); +INSERT INTO iris.train VALUES(5.0,3.4,1.5,0.2,0); +INSERT INTO iris.train VALUES(4.7,3.2,1.3,0.2,0); +INSERT INTO iris.train VALUES(6.6,2.9,4.6,1.3,1); +INSERT INTO iris.train VALUES(5.5,3.5,1.3,0.2,0); +INSERT INTO iris.train VALUES(7.7,3.0,6.1,2.3,2); +INSERT INTO iris.train VALUES(6.1,3.0,4.9,1.8,2); +INSERT INTO iris.train VALUES(4.9,3.1,1.5,0.1,0); +INSERT INTO iris.train VALUES(5.5,2.4,3.8,1.1,1); +INSERT INTO iris.train VALUES(5.7,2.9,4.2,1.3,1); +INSERT INTO iris.train VALUES(6.0,2.9,4.5,1.5,1); +INSERT INTO iris.train VALUES(6.4,2.7,5.3,1.9,2); +INSERT INTO iris.train VALUES(5.4,3.7,1.5,0.2,0); +INSERT INTO iris.train VALUES(6.1,2.9,4.7,1.4,1); +INSERT INTO iris.train VALUES(6.5,2.8,4.6,1.5,1); +INSERT INTO iris.train VALUES(5.6,2.7,4.2,1.3,1); +INSERT INTO iris.train VALUES(6.3,3.4,5.6,2.4,2); +INSERT INTO iris.train VALUES(4.9,3.1,1.5,0.1,0); +INSERT INTO iris.train VALUES(6.8,2.8,4.8,1.4,1); +INSERT INTO iris.train VALUES(5.7,2.8,4.5,1.3,1); +INSERT INTO iris.train VALUES(6.0,2.7,5.1,1.6,1); +INSERT INTO iris.train VALUES(5.0,3.5,1.3,0.3,0); +INSERT INTO iris.train VALUES(6.5,3.0,5.2,2.0,2); +INSERT INTO iris.train VALUES(6.1,2.8,4.7,1.2,1); +INSERT INTO iris.train VALUES(5.1,3.5,1.4,0.3,0); +INSERT INTO iris.train VALUES(4.6,3.1,1.5,0.2,0); +INSERT INTO iris.train VALUES(6.5,3.0,5.8,2.2,2); +INSERT INTO iris.train VALUES(4.6,3.4,1.4,0.3,0); +INSERT INTO iris.train VALUES(4.6,3.2,1.4,0.2,0); +INSERT INTO iris.train VALUES(7.7,2.8,6.7,2.0,2); +INSERT INTO iris.train VALUES(5.9,3.2,4.8,1.8,1); +INSERT INTO iris.train VALUES(5.1,3.8,1.6,0.2,0); +INSERT INTO iris.train VALUES(4.9,3.0,1.4,0.2,0); +INSERT INTO iris.train VALUES(4.9,2.4,3.3,1.0,1); +INSERT INTO iris.train VALUES(4.5,2.3,1.3,0.3,0); +INSERT INTO iris.train VALUES(5.8,2.7,4.1,1.0,1); +INSERT INTO iris.train VALUES(5.0,3.4,1.6,0.4,0); +INSERT INTO iris.train VALUES(5.2,3.4,1.4,0.2,0); +INSERT INTO iris.train VALUES(5.3,3.7,1.5,0.2,0); +INSERT INTO iris.train VALUES(5.0,3.6,1.4,0.2,0); +INSERT INTO iris.train VALUES(5.6,2.9,3.6,1.3,1); +INSERT INTO iris.train VALUES(4.8,3.1,1.6,0.2,0); + +DROP TABLE IF EXISTS iris.test; +CREATE TABLE iris.test ( + sepal_length float, + sepal_width float, + petal_length float, + petal_width float, + class int); +INSERT INTO iris.test VALUES(6.3,2.7,4.9,1.8,2); +INSERT INTO iris.test VALUES(5.7,2.8,4.1,1.3,1); +INSERT INTO iris.test VALUES(5.0,3.0,1.6,0.2,0); +INSERT INTO iris.test VALUES(6.3,3.3,6.0,2.5,2); +INSERT INTO iris.test VALUES(5.0,3.5,1.6,0.6,0); +INSERT INTO iris.test VALUES(5.5,2.6,4.4,1.2,1); +INSERT INTO iris.test VALUES(5.7,3.0,4.2,1.2,1); +INSERT INTO iris.test VALUES(4.4,2.9,1.4,0.2,0); +INSERT INTO iris.test VALUES(4.8,3.0,1.4,0.1,0); +INSERT INTO iris.test VALUES(5.5,2.4,3.7,1.0,1); diff --git a/sqlflow/example/fraud_detection/README.md b/sqlflow/example/fraud_detection/README.md new file mode 100644 index 0000000..8885fc9 --- /dev/null +++ b/sqlflow/example/fraud_detection/README.md @@ -0,0 +1,101 @@ +# Credit Card Fraud Detection + +Data collected from [here](https://www.kaggle.com/mlg-ulb/creditcardfraud) + +To run: `python premade_estimator.py`, which gives +``` +Adding numeric_columns: Time +Adding numeric_columns: V1 +Adding numeric_columns: V2 +Adding numeric_columns: V3 +Adding numeric_columns: V4 +Adding numeric_columns: V5 +Adding numeric_columns: V6 +Adding numeric_columns: V7 +Adding numeric_columns: V8 +Adding numeric_columns: V9 +Adding numeric_columns: V10 +Adding numeric_columns: V11 +Adding numeric_columns: V12 +Adding numeric_columns: V13 +Adding numeric_columns: V14 +Adding numeric_columns: V15 +Adding numeric_columns: V16 +Adding numeric_columns: V17 +Adding numeric_columns: V18 +Adding numeric_columns: V19 +Adding numeric_columns: V20 +Adding numeric_columns: V21 +Adding numeric_columns: V22 +Adding numeric_columns: V23 +Adding numeric_columns: V24 +Adding numeric_columns: V25 +Adding numeric_columns: V26 +Adding numeric_columns: V27 +Adding numeric_columns: V28 +Adding numeric_columns: Amount +INFO:tensorflow:Using default config. +WARNING:tensorflow:Using temporary folder as model directory: /tmp/tmpzj6rzutf +INFO:tensorflow:Using config: {'_num_worker_replicas': 1, '_global_id_in_cluster': 0, '_save_checkpoints_steps': None, '_tf_random_seed': None, '_evaluation_master': '', '_protocol': None, '_session_config': allow_soft_placement: true +graph_options { + rewrite_options { + meta_optimizer_iterations: ONE + } +} +, '_model_dir': '/tmp/tmpzj6rzutf', '_device_fn': None, '_cluster_spec': , '_save_summary_steps': 100, '_eval_distribute': None, '_keep_checkpoint_max': 5, '_master': '', '_train_distribute': None, '_log_step_count_steps': 100, '_num_ps_replicas': 0, '_experimental_distribute': None, '_save_checkpoints_secs': 600, '_service': None, '_task_type': 'worker', '_task_id': 0, '_keep_checkpoint_every_n_hours': 10000, '_is_chief': True} +INFO:tensorflow:Calling model_fn. +INFO:tensorflow:Done calling model_fn. +INFO:tensorflow:Create CheckpointSaverHook. +INFO:tensorflow:Graph was finalized. +INFO:tensorflow:Running local_init_op. +INFO:tensorflow:Done running local_init_op. +INFO:tensorflow:Saving checkpoints for 0 into /tmp/tmpzj6rzutf/model.ckpt. +INFO:tensorflow:loss = 892.3532, step = 1 +INFO:tensorflow:global_step/sec: 295.764 +INFO:tensorflow:loss = 0.00025717804, step = 101 (0.338 sec) +INFO:tensorflow:global_step/sec: 446.454 +INFO:tensorflow:loss = 0.00024299514, step = 201 (0.224 sec) +INFO:tensorflow:global_step/sec: 444.479 +INFO:tensorflow:loss = 0.00020988214, step = 301 (0.225 sec) +INFO:tensorflow:global_step/sec: 421.979 +INFO:tensorflow:loss = 0.00023327809, step = 401 (0.237 sec) +INFO:tensorflow:global_step/sec: 446.172 +INFO:tensorflow:loss = 0.00020405005, step = 501 (0.224 sec) +INFO:tensorflow:global_step/sec: 431.785 +INFO:tensorflow:loss = 0.00017921966, step = 601 (0.232 sec) +INFO:tensorflow:global_step/sec: 442.912 +INFO:tensorflow:loss = 0.00018488085, step = 701 (0.226 sec) +INFO:tensorflow:global_step/sec: 444.628 +INFO:tensorflow:loss = 0.00017795907, step = 801 (0.225 sec) +INFO:tensorflow:global_step/sec: 428.385 +INFO:tensorflow:loss = 0.00016896412, step = 901 (0.233 sec) +INFO:tensorflow:Saving checkpoints for 1000 into /tmp/tmpzj6rzutf/model.ckpt. +INFO:tensorflow:Loss for final step: 0.00016340206. +INFO:tensorflow:Calling model_fn. +WARNING:tensorflow:Trapezoidal rule is known to produce incorrect PR-AUCs; please switch to "careful_interpolation" instead. +WARNING:tensorflow:Trapezoidal rule is known to produce incorrect PR-AUCs; please switch to "careful_interpolation" instead. +INFO:tensorflow:Done calling model_fn. +INFO:tensorflow:Starting evaluation at 2018-10-16-01:02:49 +INFO:tensorflow:Graph was finalized. +INFO:tensorflow:Restoring parameters from /tmp/tmpzj6rzutf/model.ckpt-1000 +INFO:tensorflow:Running local_init_op. +INFO:tensorflow:Done running local_init_op. +INFO:tensorflow:Finished evaluation at 2018-10-16-01:02:49 +INFO:tensorflow:Saving dict for global step 1000: accuracy = 1.0, accuracy_baseline = 1.0, auc = 0.9999999, auc_precision_recall = 0.0, average_loss = 0.0001784083, global_step = 1000, label/mean = 0.0, loss = 0.0016056747, precision = 0.0, prediction/mean = 0.00017826515, recall = 0.0 +INFO:tensorflow:Saving 'checkpoint_path' summary for global step 1000: /tmp/tmpzj6rzutf/model.ckpt-1000 + +Test set accuracy: 1.00000 + +INFO:tensorflow:Calling model_fn. +INFO:tensorflow:Done calling model_fn. +INFO:tensorflow:Graph was finalized. +INFO:tensorflow:Restoring parameters from /tmp/tmpzj6rzutf/model.ckpt-1000 +INFO:tensorflow:Running local_init_op. +INFO:tensorflow:Done running local_init_op. + +Prediction is "Not Fraud" (100.0%), expected "Not Fraud" + +Prediction is "Not Fraud" (100.0%), expected "Not Fraud" + +Prediction is "Not Fraud" (100.0%), expected "Not Fraud" +``` diff --git a/sqlflow/example/fraud_detection/creditcard.csv.test b/sqlflow/example/fraud_detection/creditcard.csv.test new file mode 100644 index 0000000..134140e --- /dev/null +++ b/sqlflow/example/fraud_detection/creditcard.csv.test @@ -0,0 +1,10 @@ +"Time","V1","V2","V3","V4","V5","V6","V7","V8","V9","V10","V11","V12","V13","V14","V15","V16","V17","V18","V19","V20","V21","V22","V23","V24","V25","V26","V27","V28","Amount","Class" +0,1.19185711131486,0.26615071205963,0.16648011335321,0.448154078460911,0.0600176492822243,-0.0823608088155687,-0.0788029833323113,0.0851016549148104,-0.255425128109186,-0.166974414004614,1.61272666105479,1.06523531137287,0.48909501589608,-0.143772296441519,0.635558093258208,0.463917041022171,-0.114804663102346,-0.183361270123994,-0.145783041325259,-0.0690831352230203,-0.225775248033138,-0.638671952771851,0.101288021253234,-0.339846475529127,0.167170404418143,0.125894532368176,-0.00898309914322813,0.0147241691924927,2.69,"0" +1,-0.966271711572087,-0.185226008082898,1.79299333957872,-0.863291275036453,-0.0103088796030823,1.24720316752486,0.23760893977178,0.377435874652262,-1.38702406270197,-0.0549519224713749,-0.226487263835401,0.178228225877303,0.507756869957169,-0.28792374549456,-0.631418117709045,-1.0596472454325,-0.684092786345479,1.96577500349538,-1.2326219700892,-0.208037781160366,-0.108300452035545,0.00527359678253453,-0.190320518742841,-1.17557533186321,0.647376034602038,-0.221928844458407,0.0627228487293033,0.0614576285006353,123.5,"0" +7,-0.644269442348146,1.41796354547385,1.0743803763556,-0.492199018495015,0.948934094764157,0.428118462833089,1.12063135838353,-3.80786423873589,0.615374730667027,1.24937617815176,-0.619467796121913,0.291474353088705,1.75796421396042,-1.32386521970526,0.686132504394383,-0.0761269994382006,-1.2221273453247,-0.358221569869078,0.324504731321494,-0.156741852488285,1.94346533978412,-1.01545470979971,0.057503529867291,-0.649709005559993,-0.415266566234811,-0.0516342969262494,-1.20692108094258,-1.08533918832377,40.8,"0" +14,-5.40125766315825,-5.45014783420644,1.18630463143652,1.73623880012095,3.04910587764025,-1.76340557365201,-1.55973769907953,0.160841747266769,1.23308974041888,0.345172827050629,0.917229867699146,0.970116716069048,-0.266567764915222,-0.479129929276704,-0.526608502569153,0.47200411177674,-0.725480944982201,0.075081351540202,-0.406866573198217,-2.19684802485647,-0.503600328973703,0.984459785590244,2.45858857639219,0.0421188969891572,-0.481630823956716,-0.621272013713977,0.392053289557744,0.949594245504846,46.8,"0" +18,1.16661638244228,0.502120087854101,-0.0673003143663533,2.26156923949128,0.428804194630708,0.0894735167274599,0.241146579907281,0.138081705243702,-0.989162395259052,0.922174967079328,0.744785789437272,-0.531377250104474,-2.1053464503177,1.12687010488156,0.00307532251490036,0.424424506356606,-0.454475291663708,-0.0988706272108058,-0.816597306905994,-0.307168508741084,0.0187018719918659,-0.0619722672279421,-0.103854921794533,-0.370415176650599,0.603200338674873,0.108555872801682,-0.0405207055560295,-0.0114178152991164,2.28,"0" +23,1.17328461017079,0.353497876646456,0.28390506526532,1.1335633178795,-0.172577181842493,-0.916053706932115,0.36902484539083,-0.327260242196767,-0.24665102789456,-0.0461393022689538,-0.143418527309124,0.979350375717585,1.49228543545219,0.101417525935881,0.76147754528118,-0.014584082470356,-0.511640116553888,-0.32505635460906,-0.390933797720026,0.0278779087328671,0.0670033044874043,0.227811927638182,-0.150487224821759,0.43504510338369,0.724824579577241,-0.337082059875827,0.0163683792669046,0.0300411908366385,41.88,"0" +23,-0.414288810090829,0.905437322625407,1.72745294417921,1.47347126657189,0.00744274117322988,-0.200330677416199,0.740228319420026,-0.029247400012072,-0.593392019124765,-0.346188231273199,-0.0121421884961019,0.786796316060126,0.635953883297271,-0.0863244719500975,0.0768036871984215,-1.4059193336148,0.775591738395162,-0.94288892749919,0.543969461627011,0.0973075910928853,0.0772374339988386,0.457330598656119,-0.038499724582026,0.642521902774992,-0.183891335259639,-0.277464019495649,0.182687486490454,0.152664644968975,33,"0" +26,-0.53538776309446,0.865267807575272,1.35107628772237,0.14757547449791,0.433680212077009,0.0869829381161816,0.693039311115721,0.179742260742556,-0.285641861230583,-0.482474471178567,0.871799580440785,0.853447434991487,-0.571821891415037,0.102252103048142,-1.51999120265821,-0.285912497081499,-0.309633386824141,-0.403901992939884,-0.823742994489464,-0.283263782156003,0.049525686687599,0.206536543024476,-0.18710806982075,0.000753014257497798,0.0981166059546266,-0.553470968822753,-0.0783055021726955,0.0254273775211197,1.77,"0" +29,0.996369531566045,-0.12258878723806,0.546819472613086,0.706579541087689,0.134559557409452,1.15699511200606,-0.294561331432968,0.407429099419853,0.337862663634561,-0.408150492552647,0.682646123949549,1.25371968484033,0.391917044082342,-0.0375108583595551,0.795301519477008,-1.6682105913656,1.32299488209244,-2.83003522867318,-1.46343168833485,-0.203461209817417,-0.0763018365480559,0.108866913420892,0.162230940339532,-0.575624047818567,0.109794914589208,0.373813367278979,0.0505521972760614,0.00510555920011407,20.53,"0" diff --git a/sqlflow/example/fraud_detection/creditcard.csv.train b/sqlflow/example/fraud_detection/creditcard.csv.train new file mode 100644 index 0000000..b205dec --- /dev/null +++ b/sqlflow/example/fraud_detection/creditcard.csv.train @@ -0,0 +1,10 @@ +"Time","V1","V2","V3","V4","V5","V6","V7","V8","V9","V10","V11","V12","V13","V14","V15","V16","V17","V18","V19","V20","V21","V22","V23","V24","V25","V26","V27","V28","Amount","Class" +0,-1.3598071336738,-0.0727811733098497,2.53634673796914,1.37815522427443,-0.338320769942518,0.462387777762292,0.239598554061257,0.0986979012610507,0.363786969611213,0.0907941719789316,-0.551599533260813,-0.617800855762348,-0.991389847235408,-0.311169353699879,1.46817697209427,-0.470400525259478,0.207971241929242,0.0257905801985591,0.403992960255733,0.251412098239705,-0.018306777944153,0.277837575558899,-0.110473910188767,0.0669280749146731,0.128539358273528,-0.189114843888824,0.133558376740387,-0.0210530534538215,149.62,"0" +1,-1.35835406159823,-1.34016307473609,1.77320934263119,0.379779593034328,-0.503198133318193,1.80049938079263,0.791460956450422,0.247675786588991,-1.51465432260583,0.207642865216696,0.624501459424895,0.066083685268831,0.717292731410831,-0.165945922763554,2.34586494901581,-2.89008319444231,1.10996937869599,-0.121359313195888,-2.26185709530414,0.524979725224404,0.247998153469754,0.771679401917229,0.909412262347719,-0.689280956490685,-0.327641833735251,-0.139096571514147,-0.0553527940384261,-0.0597518405929204,378.66,"0" +2,-1.15823309349523,0.877736754848451,1.548717846511,0.403033933955121,-0.407193377311653,0.0959214624684256,0.592940745385545,-0.270532677192282,0.817739308235294,0.753074431976354,-0.822842877946363,0.53819555014995,1.3458515932154,-1.11966983471731,0.175121130008994,-0.451449182813529,-0.237033239362776,-0.0381947870352842,0.803486924960175,0.408542360392758,-0.00943069713232919,0.79827849458971,-0.137458079619063,0.141266983824769,-0.206009587619756,0.502292224181569,0.219422229513348,0.215153147499206,69.99,"0" +2,-0.425965884412454,0.960523044882985,1.14110934232219,-0.168252079760302,0.42098688077219,-0.0297275516639742,0.476200948720027,0.260314333074874,-0.56867137571251,-0.371407196834471,1.34126198001957,0.359893837038039,-0.358090652573631,-0.137133700217612,0.517616806555742,0.401725895589603,-0.0581328233640131,0.0686531494425432,-0.0331937877876282,0.0849676720682049,-0.208253514656728,-0.559824796253248,-0.0263976679795373,-0.371426583174346,-0.232793816737034,0.105914779097957,0.253844224739337,0.0810802569229443,3.67,"0" +4,1.22965763450793,0.141003507049326,0.0453707735899449,1.20261273673594,0.191880988597645,0.272708122899098,-0.00515900288250983,0.0812129398830894,0.464959994783886,-0.0992543211289237,-1.41690724314928,-0.153825826253651,-0.75106271556262,0.16737196252175,0.0501435942254188,-0.443586797916727,0.00282051247234708,-0.61198733994012,-0.0455750446637976,-0.21963255278686,-0.167716265815783,-0.270709726172363,-0.154103786809305,-0.780055415004671,0.75013693580659,-0.257236845917139,0.0345074297438413,0.00516776890624916,4.99,"0" +7,-0.89428608220282,0.286157196276544,-0.113192212729871,-0.271526130088604,2.6695986595986,3.72181806112751,0.370145127676916,0.851084443200905,-0.392047586798604,-0.410430432848439,-0.705116586646536,-0.110452261733098,-0.286253632470583,0.0743553603016731,-0.328783050303565,-0.210077268148783,-0.499767968800267,0.118764861004217,0.57032816746536,0.0527356691149697,-0.0734251001059225,-0.268091632235551,-0.204232669947878,1.0115918018785,0.373204680146282,-0.384157307702294,0.0117473564581996,0.14240432992147,93.2,"0" +9,-0.33826175242575,1.11959337641566,1.04436655157316,-0.222187276738296,0.49936080649727,-0.24676110061991,0.651583206489972,0.0695385865186387,-0.736727316364109,-0.366845639206541,1.01761446783262,0.836389570307029,1.00684351373408,-0.443522816876142,0.150219101422635,0.739452777052119,-0.540979921943059,0.47667726004282,0.451772964394125,0.203711454727929,-0.246913936910008,-0.633752642406113,-0.12079408408185,-0.385049925313426,-0.0697330460416923,0.0941988339514961,0.246219304619926,0.0830756493473326,3.68,"0" +10,1.44904378114715,-1.17633882535966,0.913859832832795,-1.37566665499943,-1.97138316545323,-0.62915213889734,-1.4232356010359,0.0484558879088564,-1.72040839292037,1.62665905834133,1.1996439495421,-0.671439778462005,-0.513947152539479,-0.0950450453999549,0.230930409124119,0.0319674667862076,0.253414715863197,0.854343814324194,-0.221365413645481,-0.387226474431156,-0.00930189652490052,0.313894410791098,0.0277401580170247,0.500512287104917,0.25136735874921,-0.129477953726618,0.0428498709381461,0.0162532619375515,7.8,"0" +10,0.38497821518095,0.616109459176472,-0.874299702595052,-0.0940186259679115,2.92458437838817,3.31702716826156,0.470454671805879,0.53824722837695,-0.558894612428441,0.30975539423728,-0.259115563735702,-0.326143233995877,-0.0900467227020648,0.362832368569793,0.928903660629178,-0.129486811402759,-0.809978925963589,0.359985390219981,0.70766382644648,0.12599157561542,0.049923685888971,0.238421512225103,0.00912986861262866,0.996710209581086,-0.767314827174801,-0.492208295340017,0.042472441919027,-0.0543373883732122,9.99,"0" diff --git a/sqlflow/example/fraud_detection/creditcard_data.py b/sqlflow/example/fraud_detection/creditcard_data.py new file mode 100644 index 0000000..439fea1 --- /dev/null +++ b/sqlflow/example/fraud_detection/creditcard_data.py @@ -0,0 +1,59 @@ +# Reference: https://github.com/tensorflow/models/blob/master/samples/core/get_started/iris_data.py +# which released under the Apache License 2.0 +import pandas as pd +import tensorflow as tf + +TRAIN_DATA_FILE = './creditcard.csv.train' +TEST_DATA_FILE = './creditcard.csv.test' + +CSV_COLUMN_NAMES = ["Time"] + ["V"+str(i) for i in range(1,29)] + ["Amount", "Class"] + + +def load_data(y_name='Class'): + """Returns the iris dataset as (train_x, train_y), (test_x, test_y).""" + import os + import errno + for filename in (TRAIN_DATA_FILE, TEST_DATA_FILE): + if not os.path.exists(filename): + raise FileNotFoundError(errno.ENOENT, os.strerror(errno.ENOENT), filename) + + train = pd.read_csv(TRAIN_DATA_FILE, names=CSV_COLUMN_NAMES, header=0) + train_x, train_y = train, train.pop(y_name) + + test = pd.read_csv(TEST_DATA_FILE, names=CSV_COLUMN_NAMES, header=0) + test_x, test_y = test, test.pop(y_name) + + return (train_x, train_y), (test_x, test_y) + + +def train_input_fn(features, labels, batch_size): + """An input function for training""" + + # Convert the inputs to a Dataset. + dataset = tf.data.Dataset.from_tensor_slices((dict(features), labels)) + + # Shuffle, repeat, and batch the examples. + dataset = dataset.shuffle(1000).repeat().batch(batch_size) + + # Return the dataset. + return dataset + + +def eval_input_fn(features, labels, batch_size): + """An input function for evaluation or prediction""" + features=dict(features) + if labels is None: + # No labels, use only features. + inputs = features + else: + inputs = (features, labels) + + # Convert the inputs to a Dataset. + dataset = tf.data.Dataset.from_tensor_slices(inputs) + + # Batch the examples + assert batch_size is not None, "batch_size must not be None" + dataset = dataset.batch(batch_size) + + # Return the dataset. + return dataset diff --git a/sqlflow/example/fraud_detection/premade_estimator.py b/sqlflow/example/fraud_detection/premade_estimator.py new file mode 100644 index 0000000..1bdcb5a --- /dev/null +++ b/sqlflow/example/fraud_detection/premade_estimator.py @@ -0,0 +1,115 @@ +# Copyright 2016 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +"""An Example of a DNNClassifier for the Iris dataset.""" +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import argparse +import tensorflow as tf +# import tensorflow.contrib.eager as tfe +# tf.enable_eager_execution() + +import creditcard_data + + +parser = argparse.ArgumentParser() +parser.add_argument('--batch_size', default=100, type=int, help='batch size') +parser.add_argument('--train_steps', default=1000, type=int, + help='number of training steps') + +def main(argv): + args = parser.parse_args(argv[1:]) + + # Fetch the data + (train_x, train_y), (test_x, test_y) = creditcard_data.load_data() + + # Feature columns describe how to use the input. + my_feature_columns = [] + for key in train_x.keys(): + print("Adding numeric_columns: {}".format(key)) + my_feature_columns.append(tf.feature_column.numeric_column(key=key)) + + # Build 2 hidden layer DNN with 100, 100 units respectively. + classifier = tf.estimator.DNNClassifier( + feature_columns=my_feature_columns, + # Two hidden layers of 10 nodes each. + hidden_units=[100, 100], + # The model must choose between 3 classes. + n_classes=2) + + # Train the Model. + classifier.train( + input_fn=lambda:creditcard_data.train_input_fn(train_x, train_y, + args.batch_size), + steps=args.train_steps) + + # Evaluate the model. + eval_result = classifier.evaluate( + input_fn=lambda:creditcard_data.eval_input_fn(test_x, test_y, + args.batch_size)) + + print('\nTest set accuracy: {accuracy:0.5f}\n'.format(**eval_result)) + + # # Generate predictions from the model + expected = [0, 0, 0] + predict_x = { + 'Time': [80.0, 80.0, 81.0], + 'V1': [-0.655264276212716, -3.00773905109524, 1.11104797028704], + 'V2': [0.40989854283685295, 0.929514494266081, 0.215351844382971], + 'V3': [1.28915643808596, -0.42372085549803296, 0.374072563622289], + 'V4': [-0.32504325923176003, -1.3929516201707801, 1.1500006534851999], + 'V5': [0.545669467098008, 0.0822467824081475, -0.329767568535767], + 'V6': [-0.349810664410475, -0.515781701182148, -0.898687821706977], + 'V7': [0.648240481592264, 0.104713792402636, 0.30990664398233], + 'V8': [0.0360628816033704, 0.8253105658956079, -0.25527729116611103], + 'V9': [0.0787014163023305, 0.38545311035542207, -0.110655138734984], + 'V10': [-0.829167267726513, 0.0640002245721828, -0.0575788321387636], + 'V11': [-1.78165873189534, -1.29310671179606, -0.21796149670612197], + 'V12': [-0.412806782520344, 0.0621999429338998, 0.600591209357443], + 'V13': [-0.6262202895658521, 0.27110701307160395, 0.686379564975428], + 'V14': [-0.113361054850318, 0.328058800506629, 0.262161286299107], + 'V15': [-0.714754632820629, 0.7240976288015201, 0.9226769790285508], + 'V16': [0.120701513866132, 0.6383894438325529, 0.135235271944387], + 'V17': [-0.47428167431379603, -0.43008221783924894, -0.490174913186185], + 'V18': [-0.240594669556031, -0.7235902240912719, -0.392752898979319], + 'V19': [0.0408234763497192, -0.562830692998269, -0.315464926788021], + 'V20': [-0.0347759899640066, -0.15171422957684697, 0.031671787802255], + 'V21': [-0.155727058998072, -0.40924608012503005, -0.132555303999709], + 'V22': [-0.47739031832143397, -0.588831627173022, -0.493866878067896], + 'V23': [-0.126524992708454, 0.0680450889350803, -0.0107821837368997], + 'V24': [-0.443627776856693, -0.890869739213917, 0.380388564169084], + 'V25': [-0.012118142098906, -0.047641092795032, 0.45621418060635105], + 'V26': [0.143172915176852, 0.717616984694838, -0.572346756048474], + 'V27': [0.049782585303356, -0.0540936594394297, 0.0109814630026346], + 'V28': [0.118280286710879, -0.42425971753603997, 0.039295013148069], + 'Amount': [32.51, 39.95, 68.74] + } + + predictions = classifier.predict( + input_fn=lambda:creditcard_data.eval_input_fn(features=predict_x, + labels=None, + batch_size=args.batch_size)) + + template = ('\nPrediction is "{}" ({:.1f}%), expected "{}"') + for pred_dict, expec in zip(predictions, expected): + class_id = pred_dict['class_ids'][0] + probability = pred_dict['probabilities'][class_id] + print(template.format(["Not Fraud", "Fraud"][class_id], + 100 * probability, ["Not Fraud", "Fraud"][expec])) + + +if __name__ == '__main__': + tf.logging.set_verbosity(tf.logging.INFO) + tf.app.run(main) diff --git a/sqlflow/example/jupyter/example.ipynb b/sqlflow/example/jupyter/example.ipynb new file mode 100644 index 0000000..7dac56f --- /dev/null +++ b/sqlflow/example/jupyter/example.ipynb @@ -0,0 +1,106 @@ +{ + "cells": [ + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "%reload_ext sqlflow.magic" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "%%sqlflow\n", + "show databases;" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "%%sqlflow\n", + "select *\n", + "from iris.train\n", + "limit 1" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "%%sqlflow\n", + "SELECT *\n", + "FROM iris.train\n", + "TRAIN DNNClassifier\n", + "WITH\n", + " n_classes = 3,\n", + " hidden_units = [10, 20]\n", + "COLUMN sepal_length, sepal_width, petal_length, petal_width\n", + "LABEL class\n", + "INTO sqlflow_models.my_dnn_model;" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "%%sqlflow\n", + "SELECT *\n", + "FROM iris.test\n", + "predict iris.predict.class\n", + "USING sqlflow_models.my_dnn_model;" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "%%sqlflow\n", + "SELECT *\n", + "FROM iris.predict\n", + "LIMIT 5;" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [] + } + ], + "metadata": { + "kernelspec": { + "display_name": "Python 3", + "language": "python", + "name": "python3" + }, + "language_info": { + "codemirror_mode": { + "name": "ipython", + "version": 3 + }, + "file_extension": ".py", + "mimetype": "text/x-python", + "name": "python", + "nbconvert_exporter": "python", + "pygments_lexer": "ipython3", + "version": "3.7.2" + } + }, + "nbformat": 4, + "nbformat_minor": 2 +} diff --git a/sqlflow/index.html b/sqlflow/index.html deleted file mode 100644 index 7237249..0000000 --- a/sqlflow/index.html +++ /dev/null @@ -1,13 +0,0 @@ - - - - - - - - -

    Redirecting to https://github.com/sql-machine-learning/sqlflow

    - - diff --git a/sqlflow/scripts/deploy.sh b/sqlflow/scripts/deploy.sh new file mode 100644 index 0000000..37f2d40 --- /dev/null +++ b/sqlflow/scripts/deploy.sh @@ -0,0 +1,13 @@ +#!/bin/bash +set -e + +echo "$DOCKER_PASSWORD" | docker login --username "$DOCKER_USERNAME" --password-stdin + +if [[ $TRAVIS_EVENT_TYPE == "cron" ]]; then + DOCKER_TAG="nightly" +else + DOCKER_TAG="latest" +fi + +docker build -t sqlflow/sqlflow:$DOCKER_TAG -f ./Dockerfile $GOPATH/bin +docker push sqlflow/sqlflow:$DOCKER_TAG diff --git a/sqlflow/scripts/test.sh b/sqlflow/scripts/test.sh new file mode 100644 index 0000000..f64c6d8 --- /dev/null +++ b/sqlflow/scripts/test.sh @@ -0,0 +1,9 @@ +#!/bin/bash +set -e + +service mysql start + +go generate ./... +go get -v -t ./... +go install ./... +go test -v ./... diff --git a/sqlflow/server/.gitignore b/sqlflow/server/.gitignore new file mode 100644 index 0000000..e6435d9 --- /dev/null +++ b/sqlflow/server/.gitignore @@ -0,0 +1,5 @@ +*~ +.idea/ +sqlflow.pb.go +*.vim +*.test diff --git a/sqlflow/server/.travis.yml b/sqlflow/server/.travis.yml new file mode 100644 index 0000000..2264573 --- /dev/null +++ b/sqlflow/server/.travis.yml @@ -0,0 +1,36 @@ +language: go + +# Force-enable Go modules. This will be unnecessary when Go 1.12 lands. +env: +- GO111MODULE=on + +# You don't need to test on very old version of the Go compiler. It's the user's +# responsibility to keep their compilers up to date. +go: +- 1.11.x + +# Only clone the most recent commit. +git: + depth: 1 + +services: +- docker + +# Skip the install step. Don't `go get` dependencies. Only build with the code +# in vendor/ +install: true + +# Don't email me the results of the test runs. +notifications: + email: false + +# Anything in before_script that returns a nonzero exit code will flunk the +# build and immediately stop. It's sorta like having set -e enabled in bash. +before_script: +- curl -sfL https://install.goreleaser.com/github.com/golangci/golangci-lint.sh | sh -s -- -b $(go env GOPATH)/bin + +# script always runs to completion (set +e). +script: +- go generate # generate proto files +- go test -v -race ./... # Run all the tests with the race detector enabled +- golangci-lint run # run a bunch of code checkers/linters in parallel diff --git a/sqlflow/server/README.md b/sqlflow/server/README.md new file mode 100644 index 0000000..ce2dfc0 --- /dev/null +++ b/sqlflow/server/README.md @@ -0,0 +1,14 @@ +# sqlflowserver +The gRPC proxy server of SQL engines + +## Prerequisite + +- Golang +- Docker +- GRPC + +## Running + +```bash +go generate && cd server && go run main.go +``` diff --git a/sqlflow/server/index.html b/sqlflow/server/index.html deleted file mode 100644 index 7237249..0000000 --- a/sqlflow/server/index.html +++ /dev/null @@ -1,13 +0,0 @@ - - - - - - - - -

    Redirecting to https://github.com/sql-machine-learning/sqlflow

    - - diff --git a/sqlflow/server/proto/sqlflow.proto b/sqlflow/server/proto/sqlflow.proto new file mode 100644 index 0000000..4e36995 --- /dev/null +++ b/sqlflow/server/proto/sqlflow.proto @@ -0,0 +1,60 @@ +syntax = "proto3"; + +import "google/protobuf/any.proto"; + +package proto; + +service SQLFlow { + // Run executes a sql statement + // + // SQL statements like `SELECT ...`, `DESCRIBE ...` returns a rowset. + // The rowset might be big. In such cases, Query returns a stream + // of RunResponse + // + // SQL statements like `USE database`, `DELETE` returns only a success + // message. + // + // SQL statement like `SELECT ... TRAIN/PREDICT ...` returns a stream of + // messages which indicates the training/predicting progress + rpc Run (Request) returns (stream Response); +} + +// SQL statements to run +// e.g. +// 1. `SELECT ...` +// 2. `USE ...`, `DELETE ...` +// 3. `SELECT ... TRAIN/PREDICT ...` +message Request { + string sql = 1; // The SQL statement to be executed. +} + +message Response { + oneof response { + Head head = 1; + Row row = 2; + Message message = 3; + } +} + +// SQL statements like `SELECT ...`, `DESCRIBE ...` returns a Head +// and a sequence of Rows +message Head { + repeated string column_names = 1; +} +message Row { + // Null is a special marker used in Structured Query Language to indicate + // that a data value does not exist in the database. + // We encoded this marker as message Null, and it is one possible type of + // google.protobuf.Any in the field data + message Null {} + repeated google.protobuf.Any data = 1; +} + +// SQL statements like `USE database`, `DELETE` returns only a success +// message. +// +// SQL statement like `SELECT ... TRAIN/PREDICT ...` returns a stream of +// messages which indicates the training/predicting progress +message Message { + string message = 1; +} diff --git a/sqlflow/server/sqlflowserver.go b/sqlflow/server/sqlflowserver.go new file mode 100644 index 0000000..8bf1d57 --- /dev/null +++ b/sqlflow/server/sqlflowserver.go @@ -0,0 +1,130 @@ +//go:generate protoc -I proto proto/sqlflow.proto --go_out=plugins=grpc:proto +package server + +import ( + "fmt" + "time" + + "github.com/golang/protobuf/proto" + "github.com/golang/protobuf/ptypes" + pyts "github.com/golang/protobuf/ptypes/timestamp" + "github.com/golang/protobuf/ptypes/wrappers" + + pb "github.com/sql-machine-learning/sqlflow/server/proto" + sf "github.com/sql-machine-learning/sqlflow/sql" +) + +// NewServer returns a server instance +func NewServer(run func(string, *sf.DB) *sf.PipeReader, db *sf.DB) *server { + return &server{run: run, db: db} +} + +type server struct { + run func(sql string, db *sf.DB) *sf.PipeReader + db *sf.DB +} + +// Run implements `rpc Run (Request) returns (stream Response)` +func (s *server) Run(req *pb.Request, stream pb.SQLFlow_RunServer) error { + pr := s.run(req.Sql, s.db) + defer pr.Close() + + for r := range pr.ReadAll() { + var res *pb.Response + var err error + switch s := r.(type) { + case error: + return s + case map[string]interface{}: + res, err = encodeHead(s) + case []interface{}: + res, err = encodeRow(s) + case string: + res, err = encodeMessage(s) + default: + return fmt.Errorf("unrecognize run channel return type %#v", s) + } + if err != nil { + return err + } + if err := stream.Send(res); err != nil { + return err + } + } + return nil +} + +func encodeHead(head map[string]interface{}) (*pb.Response, error) { + cn, ok := head["columnNames"] + if !ok { + return nil, fmt.Errorf("can't find field columnNames in head") + } + columnNames, ok := cn.([]string) + if !ok { + return nil, fmt.Errorf("head[\"columnNames\"] is of type %T, expected []string", cn) + } + return &pb.Response{Response: &pb.Response_Head{Head: &pb.Head{ColumnNames: columnNames}}}, nil +} + +func encodeRow(row []interface{}) (*pb.Response, error) { + encodedRow := &pb.Row{} + for _, element := range row { + pm, err := parse2Any(element) + if err != nil { + return nil, err + } + any, err := ptypes.MarshalAny(pm) + if err != nil { + return nil, err + } + encodedRow.Data = append(encodedRow.Data, any) + } + return &pb.Response{Response: &pb.Response_Row{Row: encodedRow}}, nil +} + +func encodeMessage(s string) (*pb.Response, error) { + return &pb.Response{Response: &pb.Response_Message{Message: &pb.Message{Message: s}}}, nil +} + +func parse2Any(val interface{}) (proto.Message, error) { + switch v := val.(type) { + case nil: + return &pb.Row_Null{}, nil + case bool: + return &wrappers.BoolValue{Value: v}, nil + case int8: + return &wrappers.Int32Value{Value: int32(v)}, nil + case int16: + return &wrappers.Int32Value{Value: int32(v)}, nil + case int32: + return &wrappers.Int32Value{Value: v}, nil + case int: + return &wrappers.Int64Value{Value: int64(v)}, nil + case int64: + return &wrappers.Int64Value{Value: v}, nil + case uint8: + return &wrappers.UInt32Value{Value: uint32(v)}, nil + case uint16: + return &wrappers.UInt32Value{Value: uint32(v)}, nil + case uint32: + return &wrappers.UInt32Value{Value: v}, nil + case uint: + return &wrappers.UInt64Value{Value: uint64(v)}, nil + case uint64: + return &wrappers.UInt64Value{Value: v}, nil + case float32: + return &wrappers.FloatValue{Value: v}, nil + case float64: + return &wrappers.DoubleValue{Value: v}, nil + case string: + return &wrappers.StringValue{Value: v}, nil + case []byte: + return &wrappers.BytesValue{Value: v}, nil + case time.Time: + return &pyts.Timestamp{ + Seconds: int64(v.Second()), + Nanos: int32(v.Nanosecond())}, nil + default: + return nil, fmt.Errorf("can't convert %#v to protobuf.Any", val) + } +} diff --git a/sqlflow/server/sqlflowserver_test.go b/sqlflow/server/sqlflowserver_test.go new file mode 100644 index 0000000..029c19b --- /dev/null +++ b/sqlflow/server/sqlflowserver_test.go @@ -0,0 +1,139 @@ +package server + +import ( + "context" + "fmt" + "io" + "log" + "net" + "os" + "testing" + "time" + + "github.com/fortytw2/leaktest" + "github.com/stretchr/testify/assert" + "google.golang.org/grpc" + "google.golang.org/grpc/codes" + "google.golang.org/grpc/reflection" + "google.golang.org/grpc/status" + + pb "github.com/sql-machine-learning/sqlflow/server/proto" + sf "github.com/sql-machine-learning/sqlflow/sql" +) + +const ( + testErrorSQL = "ERROR ..." + testQuerySQL = "SELECT ..." + testExecuteSQL = "INSERT ..." + testExtendedSQL = "SELECT ... TRAIN ..." +) + +var testServerAddress string + +func mockRun(sql string, db *sf.DB) *sf.PipeReader { + rd, wr := sf.Pipe() + go func() { + defer wr.Close() + + switch sql { + case testErrorSQL: + wr.Write(fmt.Errorf("run error: %v", testErrorSQL)) + case testQuerySQL: + m := make(map[string]interface{}) + m["columnNames"] = []string{"X", "Y"} + wr.Write(m) + wr.Write([]interface{}{true, false, "hello", []byte("world")}) + wr.Write([]interface{}{int8(1), int16(1), int32(1), int(1), int64(1)}) + wr.Write([]interface{}{uint8(1), uint16(1), uint32(1), uint(1), uint64(1)}) + wr.Write([]interface{}{float32(1), float64(1)}) + wr.Write([]interface{}{time.Now(), nil}) + case testExecuteSQL: + wr.Write("success; 0 rows affected") + case testExtendedSQL: + wr.Write("log 0") + wr.Write("log 1") + } + }() + return rd +} + +func startServer(done chan bool) { + listener, err := net.Listen("tcp", ":0") + if err != nil { + panic(err) + } + testServerAddress = fmt.Sprintf("localhost:%v", listener.Addr().(*net.TCPAddr).Port) + done <- true + + s := grpc.NewServer() + s.GetServiceInfo() + pb.RegisterSQLFlowServer(s, &server{run: mockRun, db: nil}) + reflection.Register(s) + if err := s.Serve(listener); err != nil { + log.Fatalf("failed to serve: %v", err) + } +} + +func createRudeClient() { + conn, _ := grpc.Dial(testServerAddress, grpc.WithInsecure()) + c := pb.NewSQLFlowClient(conn) + time.Sleep(time.Second) + + ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second) + defer cancel() + _, err := c.Run(ctx, &pb.Request{Sql: testQuerySQL}) + if err != nil { + log.Fatalf("Run encounts err:%v", err) + } + + // conn closed without *any* stream.Recv(), act as rude client + conn.Close() +} + +func TestSQL(t *testing.T) { + a := assert.New(t) + // Set up a connection to the server. + conn, err := grpc.Dial(testServerAddress, grpc.WithInsecure()) + if err != nil { + log.Fatalf("did not connect: %v", err) + } + defer conn.Close() + c := pb.NewSQLFlowClient(conn) + + ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second) + defer cancel() + + stream, err := c.Run(ctx, &pb.Request{Sql: testErrorSQL}) + a.NoError(err) + _, err = stream.Recv() + a.Equal(status.Error(codes.Unknown, fmt.Sprintf("run error: %v", testErrorSQL)), err) + + for _, s := range []string{testQuerySQL, testExecuteSQL, testExtendedSQL} { + stream, err := c.Run(ctx, &pb.Request{Sql: s}) + a.NoError(err) + for { + _, err := stream.Recv() + if err == io.EOF { + break + } + a.NoError(err) + } + } +} + +func TestGoroutineLeaky(t *testing.T) { + defer leaktest.CheckTimeout(t, 10*time.Second)() + for i := 0; i < 50; i++ { + go func() { + createRudeClient() + }() + } +} + +func TestMain(m *testing.M) { + done := make(chan bool) + go startServer(done) + <-done + + os.Exit(m.Run()) +} diff --git a/sqlflow/sql/.gitignore b/sqlflow/sql/.gitignore new file mode 100644 index 0000000..721ba44 --- /dev/null +++ b/sqlflow/sql/.gitignore @@ -0,0 +1,2 @@ +y.* +*.output diff --git a/sqlflow/sql/README.md b/sqlflow/sql/README.md new file mode 100644 index 0000000..9771697 --- /dev/null +++ b/sqlflow/sql/README.md @@ -0,0 +1,13 @@ +# SQLFlow + +## Build, Run, Test + +Before running the unit tests, we need to build and run a Docker container that +hosts a MySQL database following [this guide](../example/datasets/README.md). + +To build the parser using `goyacc` and run all unit tests, use the following +command: + +```bash +go get -d ./... && goyacc -p sql -o parser.go sql.y && go test -v +``` diff --git a/sqlflow/sql/cmd.go b/sqlflow/sql/cmd.go new file mode 100644 index 0000000..59df333 --- /dev/null +++ b/sqlflow/sql/cmd.go @@ -0,0 +1,54 @@ +package sql + +import ( + "fmt" + "os/exec" +) + +func tryRun(cmd string, args ...string) bool { + return exec.Command(cmd, args...).Run() == nil +} + +func hasPython() bool { + return tryRun("python", "-V") +} + +func hasTensorFlow() bool { + return tryRun("python", "-c", "import tensorflow") +} + +func hasMySQLConnector() bool { + return tryRun("python", "-c", "import mysql.connector") +} + +func hasDocker() bool { + return tryRun("docker", "version") +} + +func hasDockerImage(image string) bool { + b, e := exec.Command("docker", "images", "-q", image).Output() + if e != nil || len(b) == 0 { + return false + } + return true +} + +func tensorflowCmd(cwd string) (cmd *exec.Cmd) { + if hasPython() && hasTensorFlow() && hasMySQLConnector() { + log.Printf("tensorflowCmd: run locally") + cmd = exec.Command("python") + cmd.Dir = cwd + } else if hasDocker() { + log.Printf("tensorflowCmd: run in Docker container") + const tfImg = "sqlflow/sqlflow" + if !hasDockerImage(tfImg) { + log.Printf("No local Docker image %s. It will take a long time to pull.", tfImg) + } + cmd = exec.Command("docker", "run", "--rm", + fmt.Sprintf("-v%s:/work", cwd), + "-w/work", "--network=host", "-i", tfImg, "python") + } else { + log.Fatalf("No local TensorFlow or Docker. No way to run TensorFlow programs") + } + return cmd +} diff --git a/sqlflow/sql/codegen.go b/sqlflow/sql/codegen.go new file mode 100644 index 0000000..bc5dfb9 --- /dev/null +++ b/sqlflow/sql/codegen.go @@ -0,0 +1,213 @@ +package sql + +import ( + "fmt" + "io" + "strings" + "text/template" + + "github.com/go-sql-driver/mysql" +) + +// TODO(tonyyang): This is currently a quick hack to map from SQL +// field types to feature types. We will enhance it to support more +// complex cases like cross features. +var fieldTypeFeatureType = map[string]string{"FLOAT": "numeric_column"} + +type columnType struct { + Name string + Type string +} + +type connectionConfig struct { + User string + Password string + Host string + Port string + Database string +} + +type modelConfig struct { + Estimator string + Attrs map[string]string + Save string +} + +type filler struct { + Train bool + Driver string + StandardSelect string + modelConfig + X []columnType + Y columnType + TableName string + connectionConfig +} + +func newFiller(pr *extendedSelect, fts fieldTypes, db *DB) (*filler, error) { + r := &filler{ + Train: pr.train, + StandardSelect: pr.standardSelect.String(), + modelConfig: modelConfig{ + Estimator: pr.estimator, + Attrs: make(map[string]string), + Save: pr.save}} + for k, v := range pr.attrs { + r.Attrs[k] = v.String() + } + + for _, c := range pr.columns { + typ, ok := fts.get(c.val) + if !ok { + return nil, fmt.Errorf("genTF: Cannot find type of field %s", c.val) + } + ct := columnType{Name: c.val, Type: fieldTypeFeatureType[typ]} + r.X = append(r.X, ct) + } + typ, ok := fts.get(pr.label) + if !ok { + return nil, fmt.Errorf("genTF: Cannot find type of label %s", pr.label) + } + r.Y = columnType{Name: pr.label, Type: fieldTypeFeatureType[typ]} + + if !pr.train { + r.TableName = strings.Join(strings.Split(pr.into, ".")[:2], ".") + } + + switch db.driverName { + case "mysql": + cfg, err := mysql.ParseDSN(db.dataSourceName) + if err != nil { + return nil, err + } + r.Driver = "mysql" + r.User = cfg.User + r.Password = cfg.Passwd + r.Host = strings.Split(cfg.Addr, ":")[0] + r.Port = strings.Split(cfg.Addr, ":")[1] + case "sqlite3": + r.Driver = "sqlite3" + r.Database = db.dataSourceName + default: + return nil, fmt.Errorf("sqlfow currently doesn't support DB %v", db.driverName) + } + + return r, nil +} + +func genTF(w io.Writer, pr *extendedSelect, fts fieldTypes, db *DB) error { + r, e := newFiller(pr, fts, db) + if e != nil { + return e + } + if e = codegenTemplate.Execute(w, r); e != nil { + return fmt.Errorf("genTF: failed executing template: %v", e) + } + return nil +} + +const codegenTemplateText = ` +import os +# Disable Tensorflow INFO and WARNING logs +os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3' + +import sys, json +import tensorflow as tf +import mysql.connector + +# Disable Tensorflow INFO and WARNING logs +tf.logging.set_verbosity(tf.logging.ERROR) +` + + // TODO(tonyyang-svail): remove hard coded BATCHSIZE, STEP + ` +BATCHSIZE = 1 +STEP = 1000 + +{{if eq .Driver "mysql"}} +db = mysql.connector.connect(user="{{.User}}", + passwd="{{.Password}}", + host="{{.Host}}", + port={{.Port}}{{if eq .Database ""}}{{- else}}, database="{{.DATABASE}}"{{end}}) +{{else}} +{{if eq .Driver "sqlite3"}} +db = sqlite3.connect({{.Database}}) +{{else}} +raise ValueError("unrecognized database driver: {{.Driver}}") +{{end}} +{{end}} + +cursor = db.cursor() +cursor.execute("""{{.StandardSelect}}""") +field_names = [i[0] for i in cursor.description] +columns = list(map(list, zip(*cursor.fetchall()))) + +feature_columns = [{{range .X}}tf.feature_column.{{.Type}}(key="{{.Name}}"), + {{end}}] +feature_column_names = [{{range .X}}"{{.Name}}", + {{end}}] + +X = {name: columns[field_names.index(name)] for name in feature_column_names} +{{if .Train}} +Y = columns[field_names.index("{{.Y.Name}}")] +{{- end}} + +classifier = tf.estimator.{{.Estimator}}( + feature_columns=feature_columns,{{range $key, $value := .Attrs}} + {{$key}} = {{$value}},{{end}} + model_dir= "{{.Save}}") + +{{if .Train}} +def train_input_fn(features, labels, batch_size): + dataset = tf.data.Dataset.from_tensor_slices((dict(features), labels)) + dataset = dataset.shuffle(1000).repeat().batch(batch_size) + return dataset + +classifier.train( + input_fn=lambda:train_input_fn(X, Y, BATCHSIZE), + steps=STEP) + +def eval_input_fn(features, labels, batch_size): + dataset = tf.data.Dataset.from_tensor_slices((dict(features), labels)) + dataset = dataset.batch(batch_size) + return dataset + +eval_result = classifier.evaluate( + input_fn=lambda:eval_input_fn(X, Y, BATCHSIZE), + steps=STEP) +print("\nTraining set accuracy: {accuracy:0.5f}\n".format(**eval_result)) + +print("Done training") +{{- else}} +def eval_input_fn(features, batch_size): + dataset = tf.data.Dataset.from_tensor_slices(dict(features)) + dataset = dataset.batch(batch_size) + return dataset + +predictions = classifier.predict( + input_fn=lambda:eval_input_fn(X, BATCHSIZE)) + +X["{{.Y.Name}}"] = [p['class_ids'][0] for p in predictions] + +def insert(table_name, X, db): + length = [len(X[key]) for key in X] + assert len(set(length)) == 1, "All the fields should have the same length" + + field_names = [key for key in X] + # FIXME(tony): HIVE and ODPS use INSERT INTO TABLE ... + sql = "INSERT INTO {} ({}) VALUES ({})".format( + table_name, ",".join(field_names), ",".join(["%s" for _ in field_names])) + val = [] + for i in range(length[0]): + val.append(tuple([str(X[f][i]) for f in field_names])) + + cursor = db.cursor() + cursor.executemany(sql, val) + db.commit() + +insert("{{.TableName}}", X, db) + +print("Done predicting. Predict Table : {{.TableName}}") +{{- end}} +` + +var codegenTemplate = template.Must(template.New("codegen").Parse(codegenTemplateText)) diff --git a/sqlflow/sql/codegen_test.go b/sqlflow/sql/codegen_test.go new file mode 100644 index 0000000..d5ee4ee --- /dev/null +++ b/sqlflow/sql/codegen_test.go @@ -0,0 +1,59 @@ +package sql + +import ( + "io/ioutil" + "testing" + + "github.com/stretchr/testify/assert" +) + +const ( + testStandardExecutiveSQLStatement = `DELETE FROM iris.train WHERE class = 4;` + testSelectIris = ` +SELECT * +FROM iris.train +` + testTrainSelectIris = testSelectIris + ` +TRAIN DNNClassifier +WITH + n_classes = 3, + hidden_units = [10, 20] +COLUMN sepal_length, sepal_width, petal_length, petal_width +LABEL class +INTO sqlflow_models.my_dnn_model +; +` + testPredictSelectIris = ` +SELECT * +FROM iris.test +predict iris.predict.class +USING sqlflow_models.my_dnn_model; +` +) + +func TestCodeGenTrain(t *testing.T) { + a := assert.New(t) + r, e := newParser().Parse(testTrainSelectIris) + a.NoError(e) + + fts, e := verify(r, testDB) + a.NoError(e) + + a.NoError(genTF(ioutil.Discard, r, fts, testDB)) +} + +func TestCodeGenPredict(t *testing.T) { + a := assert.New(t) + r, e := newParser().Parse(testTrainSelectIris) + a.NoError(e) + tc := r.trainClause + + r, e = newParser().Parse(testPredictSelectIris) + a.NoError(e) + r.trainClause = tc + + fts, e := verify(r, testDB) + a.NoError(e) + + a.NoError(genTF(ioutil.Discard, r, fts, testDB)) +} diff --git a/sqlflow/sql/column_type.go b/sqlflow/sql/column_type.go new file mode 100644 index 0000000..0c42ce9 --- /dev/null +++ b/sqlflow/sql/column_type.go @@ -0,0 +1,157 @@ +package sql + +import ( + "database/sql" + "fmt" + "reflect" + "time" + + "github.com/go-sql-driver/mysql" +) + +var ( + // Column types: https://golang.org/pkg/database/sql/#Rows.Scan + sqlNullBool = reflect.TypeOf(sql.NullBool{}) + sqlNullInt64 = reflect.TypeOf(sql.NullInt64{}) + sqlNullFloat64 = reflect.TypeOf(sql.NullFloat64{}) + sqlRawBytes = reflect.TypeOf(sql.RawBytes{}) + sqlNullString = reflect.TypeOf(sql.NullString{}) + mysqlNullTime = reflect.TypeOf(mysql.NullTime{}) + // builtin type supports sql like `select 1;` or `select count(*) from ...` + builtIntBytes = reflect.TypeOf([]byte("")) + builtinString = reflect.TypeOf(string("")) + builtinInt = reflect.TypeOf(int(0)) + builtinInt8 = reflect.TypeOf(int8(0)) + builtinInt16 = reflect.TypeOf(int16(0)) + builtinInt32 = reflect.TypeOf(int32(0)) + builtinInt64 = reflect.TypeOf(int64(0)) + builtinUint = reflect.TypeOf(uint(0)) + builtinUint8 = reflect.TypeOf(uint8(0)) + builtinUint16 = reflect.TypeOf(uint16(0)) + builtinUint32 = reflect.TypeOf(uint32(0)) + builtinUint64 = reflect.TypeOf(uint64(0)) + builtinFloat32 = reflect.TypeOf(float32(0)) + builtinFloat64 = reflect.TypeOf(float64(0)) + builtinTime = reflect.TypeOf(time.Time{}) +) + +func createByType(rt reflect.Type) (interface{}, error) { + switch rt { + case sqlNullBool: + return new(sql.NullBool), nil + case sqlNullInt64: + return new(sql.NullInt64), nil + case sqlNullFloat64: + return new(sql.NullFloat64), nil + case sqlRawBytes: + return new(sql.RawBytes), nil + case sqlNullString: + return new(sql.NullString), nil + case mysqlNullTime: + return new(mysql.NullTime), nil + case builtinTime: + return new(time.Time), nil + case builtIntBytes: + return new([]byte), nil + case builtinString: + return new(string), nil + case builtinInt: + return new(int), nil + case builtinInt8: + return new(int8), nil + case builtinInt16: + return new(int16), nil + case builtinInt32: + return new(int32), nil + case builtinInt64: + return new(int64), nil + case builtinUint: + return new(uint), nil + case builtinUint8: + return new(uint8), nil + case builtinUint16: + return new(uint16), nil + case builtinUint32: + return new(uint32), nil + case builtinUint64: + return new(uint64), nil + case builtinFloat32: + return new(float32), nil + case builtinFloat64: + return new(float64), nil + default: + return nil, fmt.Errorf("unrecognized column scan type %v", rt) + } +} + +func parseVal(val interface{}) (interface{}, error) { + switch v := val.(type) { + case *sql.NullBool: + if (*v).Valid { + return (*v).Bool, nil + } + return nil, nil + case *sql.NullInt64: + if (*v).Valid { + return (*v).Int64, nil + } + return nil, nil + case *sql.NullFloat64: + if (*v).Valid { + return (*v).Float64, nil + } + return nil, nil + case *sql.RawBytes: + if *v == nil { + return nil, nil + } + return string(*v), nil + case *sql.NullString: + if (*v).Valid { + return (*v).String, nil + } + return nil, nil + case *mysql.NullTime: + if (*v).Valid { + return (*v).Time, nil + } + return nil, nil + case *(time.Time): + return *v, nil + case *([]byte): + if *v == nil { + return nil, nil + } + return *v, nil + case *bool: + return *v, nil + case *string: + return *v, nil + case *int: + return *v, nil + case *int8: + return *v, nil + case *int16: + return *v, nil + case *int32: + return *v, nil + case *int64: + return *v, nil + case *uint: + return *v, nil + case *uint8: + return *v, nil + case *uint16: + return *v, nil + case *uint32: + return *v, nil + case *uint64: + return *v, nil + case *float32: + return *v, nil + case *float64: + return *v, nil + default: + return nil, fmt.Errorf("unrecogized type %v", v) + } +} diff --git a/sqlflow/sql/database.go b/sqlflow/sql/database.go new file mode 100644 index 0000000..923c8d3 --- /dev/null +++ b/sqlflow/sql/database.go @@ -0,0 +1,36 @@ +package sql + +import ( + "database/sql" + "fmt" + + _ "github.com/go-sql-driver/mysql" + _ "github.com/mattn/go-sqlite3" +) + +// DB extends sql.DB +type DB struct { + driverName string + dataSourceName string + *sql.DB +} + +// Open opens a database specified by its database driver name and a +// driver-specific data source name, usually consisting of at least a +// database name and connection information. +// +// In addition to sql.Open, it also does the book keeping on driverName and +// dataSourceName +func Open(driverName, dataSourceName string) (*DB, error) { + db := &DB{driverName: driverName, dataSourceName: dataSourceName} + + var err error + switch driverName { + case "sqlite3", "mysql": + db.DB, err = sql.Open(driverName, dataSourceName) + default: + db.DB, err = nil, fmt.Errorf("sqlfow currently doesn't support DB %v", driverName) + } + + return db, err +} diff --git a/sqlflow/sql/database_test.go b/sqlflow/sql/database_test.go new file mode 100644 index 0000000..70d7bf9 --- /dev/null +++ b/sqlflow/sql/database_test.go @@ -0,0 +1,33 @@ +package sql + +import ( + "testing" + + "github.com/go-sql-driver/mysql" + "github.com/stretchr/testify/assert" +) + +func TestDatabaseOpenMysql(t *testing.T) { + a := assert.New(t) + cfg := &mysql.Config{ + User: "root", + Passwd: "root", + Net: "tcp", + Addr: "localhost:3306", + AllowNativePasswords: true, + } + db, e := Open("mysql", cfg.FormatDSN()) + a.NoError(e) + defer db.Close() + + _, e = db.Exec("show databases") + a.NoError(e) +} + +func TestDatabaseOpenSQLite3(t *testing.T) { + a := assert.New(t) + db, e := Open("sqlite3", "test") + a.NoError(e) + defer db.Close() + // TODO: need more tests +} diff --git a/sqlflow/sql/db_test.go b/sqlflow/sql/db_test.go new file mode 100644 index 0000000..8a88375 --- /dev/null +++ b/sqlflow/sql/db_test.go @@ -0,0 +1,96 @@ +package sql + +import ( + "bufio" + "fmt" + "os" + "testing" + + "github.com/go-sql-driver/mysql" + _ "github.com/mattn/go-sqlite3" +) + +var ( + testDB *DB +) + +func TestMain(m *testing.M) { + dbms := getEnv("SQLFLOW_TEST_DB", "mysql") + + var e error + switch dbms { + case "sqlite3": + testDB, e = Open("sqlite3", ":memory:") + assertNoErr(e) + // attach an In-Memory Database in SQLite + for _, name := range []string{"iris", "churn"} { + _, e = testDB.Exec(fmt.Sprintf("ATTACH DATABASE ':memory:' AS %s;", name)) + assertNoErr(e) + } + defer testDB.Close() + case "mysql": + cfg := &mysql.Config{ + User: getEnv("SQLFLOW_TEST_DB_MYSQL_USER", "root"), + Passwd: getEnv("SQLFLOW_TEST_DB_MYSQL_PASSWD", "root"), + Net: getEnv("SQLFLOW_TEST_DB_MYSQL_NET", "tcp"), + Addr: getEnv("SQLFLOW_TEST_DB_MYSQL_ADDR", "127.0.0.1:3306"), + AllowNativePasswords: true, + } + testDB, e = Open("mysql", cfg.FormatDSN()) + assertNoErr(e) + _, e = testDB.Exec("CREATE DATABASE IF NOT EXISTS iris;") + assertNoErr(e) + _, e = testDB.Exec("CREATE DATABASE IF NOT EXISTS churn;") + assertNoErr(e) + _, e = testDB.Exec("CREATE DATABASE IF NOT EXISTS sqlflow_models;") + assertNoErr(e) + defer testDB.Close() + default: + e = fmt.Errorf("unrecognized environment variable SQLFLOW_TEST_DB %s", dbms) + } + assertNoErr(e) + + assertNoErr(popularize(testDB, "testdata/iris.sql")) + assertNoErr(popularize(testDB, "testdata/churn.sql")) + + os.Exit(m.Run()) +} + +// assertNoError prints the error if there is any in TestMain, which +// log doesn't work. +func assertNoErr(e error) { + if e != nil { + fmt.Println(e) + os.Exit(-1) + } +} + +// popularize reads SQL statements from the file named sqlfile in the +// ./testdata directory, and runs each SQL statement with db. +func popularize(db *DB, sqlfile string) error { + f, e := os.Open(sqlfile) + if e != nil { + return e + } + defer f.Close() + + onSemicolon := func(data []byte, atEOF bool) (advance int, token []byte, err error) { + for i := 0; i < len(data); i++ { + if data[i] == ';' { + return i + 1, data[:i], nil + } + } + return 0, nil, nil + } + + scanner := bufio.NewScanner(f) + scanner.Split(onSemicolon) + + for scanner.Scan() { + _, e := db.Exec(scanner.Text()) + if e != nil { + return e + } + } + return scanner.Err() +} diff --git a/sqlflow/sql/executor.go b/sqlflow/sql/executor.go new file mode 100644 index 0000000..6f841b0 --- /dev/null +++ b/sqlflow/sql/executor.go @@ -0,0 +1,346 @@ +package sql + +import ( + "bytes" + "fmt" + "io/ioutil" + "os" + "strings" + "sync" + "time" +) + +// Run executes a SQL query and returns a stream of row or message +func Run(slct string, db *DB) *PipeReader { + slctUpper := strings.ToUpper(slct) + if strings.Contains(slctUpper, "TRAIN") || strings.Contains(slctUpper, "PREDICT") { + pr, e := newParser().Parse(slct) + if e == nil && pr.extended { + return runExtendedSQL(slct, db, pr) + } + } + return runStandardSQL(slct, db) +} + +// TODO(weiguo): isQuery is a hacky way to decide which API to call: +// https://golang.org/pkg/database/sql/#DB.Exec . +// We will need to extend our parser to be a full SQL parser in the future. +func isQuery(slct string) bool { + s := strings.ToUpper(strings.TrimSpace(slct)) + has := strings.Contains + if strings.HasPrefix(s, "SELECT") && !has(s, "INTO") { + return true + } + if strings.HasPrefix(s, "SHOW") && (has(s, "CREATE") || has(s, "DATABASES") || has(s, "TABLES")) { + return true + } + if strings.HasPrefix(s, "DESCRIBE") { + return true + } + return false +} + +func runStandardSQL(slct string, db *DB) *PipeReader { + if isQuery(slct) { + return runQuery(slct, db) + } + return runExec(slct, db) +} + +// FIXME(tony): how to deal with large tables? +// TODO(tony): test on null table elements +func runQuery(slct string, db *DB) *PipeReader { + rd, wr := Pipe() + go func() { + defer wr.Close() + + err := func() error { + startAt := time.Now() + log.Infof("Starting runStanrardSQL:%s", slct) + + rows, err := db.Query(slct) + if err != nil { + return fmt.Errorf("runQuery failed: %v", err) + } + defer rows.Close() + + cols, err := rows.Columns() + if err != nil { + return fmt.Errorf("failed to get columns: %v", err) + } + + header := make(map[string]interface{}) + header["columnNames"] = cols + if e := wr.Write(header); e != nil { + return e + } + + for rows.Next() { + // Since we don't know the table schema in advance, need to + // create an slice of empty interface and adds column type + // at runtime. Some databases support dynamic types between + // rows, such as sqlite's affinity. So we move columnTypes inside + // the row.Next() loop. + count := len(cols) + values := make([]interface{}, count) + columnTypes, err := rows.ColumnTypes() + if err != nil { + return fmt.Errorf("failed to get columnTypes: %v", err) + } + for i, ct := range columnTypes { + v, e := createByType(ct.ScanType()) + if e != nil { + return e + } + values[i] = v + } + + if err := rows.Scan(values...); err != nil { + return err + } + + row := make([]interface{}, count) + for i, val := range values { + v, e := parseVal(val) + if e != nil { + return e + } + row[i] = v + } + if e := wr.Write(row); e != nil { + return e + } + } + log.Infof("runQuery finished, elapsed: %v", time.Since(startAt)) + return nil + }() + + if err != nil { + log.Errorf("runQuery error:%v", err) + if err != ErrClosedPipe { + if err := wr.Write(err); err != nil { + log.Errorf("runQuery error(piping):%v", err) + } + } + } + }() + return rd +} + +func runExec(slct string, db *DB) *PipeReader { + rd, wr := Pipe() + go func() { + defer wr.Close() + + err := func() error { + startAt := time.Now() + log.Infof("Starting runStanrardSQL1:%s", slct) + + res, e := db.Exec(slct) + if e != nil { + return fmt.Errorf("runExec failed: %v", e) + } + affected, e := res.RowsAffected() + if e != nil { + return fmt.Errorf("failed to get affected row number: %v", e) + } + var msg string + if affected > 1 { + msg = fmt.Sprintf("%d rows affected", affected) + } else { + msg = fmt.Sprintf("%d row affected", affected) + } + if e := wr.Write(msg); e != nil { + return e + } + log.Infof("runExec finished, elapsed: %v", time.Since(startAt)) + return nil + }() + if err != nil { + log.Errorf("runExec error:%v", err) + if err != ErrClosedPipe { + if err := wr.Write(err); err != nil { + log.Errorf("runExec error(piping):%v", err) + } + } + } + }() + return rd +} + +func runExtendedSQL(slct string, db *DB, pr *extendedSelect) *PipeReader { + rd, wr := Pipe() + go func() { + defer wr.Close() + + err := func() error { + startAt := time.Now() + log.Infof("Starting runExtendedSQL:%s", slct) + + // NOTE: the temporary directory must be in a host directory + // which can be mounted to Docker containers. If I don't + // specify the "/tmp" prefix, ioutil.TempDir would by default + // generate a directory in /private/tmp for macOS, which + // cannot be mounted by Docker into the container. For more + // detailed, please refer to + // https://docs.docker.com/docker-for-mac/osxfs/#namespaces. + cwd, e := ioutil.TempDir("/tmp", "sqlflow") + if e != nil { + return e + } + defer os.RemoveAll(cwd) + + if pr.train { + e = train(pr, slct, db, cwd, wr) + } else { + e = pred(pr, db, cwd, wr) + } + log.Infof("runExtendedSQL finished, elapsed:%v", time.Since(startAt)) + return e + }() + + if err != nil { + log.Errorf("runExtendedSQL error:%v", err) + if err != ErrClosedPipe { + if err := wr.Write(err); err != nil { + log.Errorf("runExtendedSQL error(piping):%v", err) + } + } + } + }() + return rd +} + +type logChanWriter struct { + wr *PipeWriter + + m sync.Mutex + buf bytes.Buffer + prev string +} + +func (cw *logChanWriter) Write(p []byte) (n int, err error) { + // Both cmd.Stdout and cmd.Stderr are writing to cw + cw.m.Lock() + defer cw.m.Unlock() + + n, err = cw.buf.Write(p) + if err != nil { + return n, err + } + + for { + line, err := cw.buf.ReadString('\n') + cw.prev = cw.prev + line + // ReadString returns err != nil if and only if the returned Data + // does not end in delim. + if err != nil { + break + } + + if err := cw.wr.Write(cw.prev); err != nil { + return len(cw.prev), err + } + cw.prev = "" + } + return n, nil +} + +func train(tr *extendedSelect, slct string, db *DB, cwd string, wr *PipeWriter) error { + fts, e := verify(tr, db) + if e != nil { + return e + } + + var program bytes.Buffer + if e := genTF(&program, tr, fts, db); e != nil { + return e + } + + cw := &logChanWriter{wr: wr} + cmd := tensorflowCmd(cwd) + cmd.Stdin = &program + cmd.Stdout = cw + cmd.Stderr = cw + if e := cmd.Run(); e != nil { + return fmt.Errorf("training failed %v", e) + } + + m := model{workDir: cwd, TrainSelect: slct} + return m.save(db, tr.save) +} + +func pred(pr *extendedSelect, db *DB, cwd string, wr *PipeWriter) error { + m, e := load(db, pr.model, cwd) + if e != nil { + return e + } + + // Parse the training SELECT statement used to train + // the model for the prediction. + tr, e := newParser().Parse(m.TrainSelect) + if e != nil { + return e + } + + if e := verifyColumnNameAndType(tr, pr, db); e != nil { + return e + } + + if e := createPredictionTable(tr, pr, db); e != nil { + return e + } + + pr.trainClause = tr.trainClause + fts, e := verify(pr, db) + + var buf bytes.Buffer + if e := genTF(&buf, pr, fts, db); e != nil { + return e + } + + cw := &logChanWriter{wr: wr} + cmd := tensorflowCmd(cwd) + cmd.Stdin = &buf + cmd.Stdout = cw + cmd.Stderr = cw + return cmd.Run() +} + +// Create prediction table with appropriate column type. +// If prediction table already exists, it will be overwritten. +func createPredictionTable(trainParsed, predParsed *extendedSelect, db *DB) error { + if len(strings.Split(predParsed.into, ".")) != 3 { + return fmt.Errorf("invalid predParsed.into %s. should be DBName.TableName.ColumnName", predParsed.into) + } + tableName := strings.Join(strings.Split(predParsed.into, ".")[:2], ".") + columnName := strings.Split(predParsed.into, ".")[2] + + dropStmt := fmt.Sprintf("drop table if exists %s;", tableName) + if _, e := db.Query(dropStmt); e != nil { + return fmt.Errorf("failed executing %s: %q", dropStmt, e) + } + + fts, e := verify(trainParsed, db) + if e != nil { + return e + } + + var b bytes.Buffer + fmt.Fprintf(&b, "create table %s (", tableName) + for _, c := range trainParsed.columns { + typ, ok := fts.get(c.val) + if !ok { + return fmt.Errorf("createPredictionTable: Cannot find type of field %s", c.val) + } + fmt.Fprintf(&b, "%s %s, ", c.val, typ) + } + tpy, _ := fts.get(trainParsed.label) + fmt.Fprintf(&b, "%s %s);", columnName, tpy) + + createStmt := b.String() + if _, e := db.Query(createStmt); e != nil { + return fmt.Errorf("failed executing %s: %q", createStmt, e) + } + return nil +} diff --git a/sqlflow/sql/executor_test.go b/sqlflow/sql/executor_test.go new file mode 100644 index 0000000..e183068 --- /dev/null +++ b/sqlflow/sql/executor_test.go @@ -0,0 +1,125 @@ +package sql + +import ( + "testing" + + "github.com/stretchr/testify/assert" +) + +const ( + testTrainBoostedTreeOnIris = ` +SELECT * +FROM iris.train +WHERE class = 0 OR class = 1 +TRAIN BoostedTreesClassifier +WITH + n_batches_per_layer = 20 +COLUMN sepal_length, sepal_width, petal_length, petal_width +LABEL class +INTO sqlflow_models.my_boosted_tree_model; +` + testPredBoostedTreeOnIris = ` +SELECT * +FROM iris.test +WHERE class = 0 OR class = 1 +predict iris.predict.class +USING sqlflow_models.my_boosted_tree_model; +` +) + +func goodStream(stream chan interface{}) bool { + for rsp := range stream { + switch rsp.(type) { + case error: + return false + } + } + return true +} + +func TestExecutorTrainAndPredictDNN(t *testing.T) { + a := assert.New(t) + a.NotPanics(func() { + pr, e := newParser().Parse(testTrainSelectIris) + a.NoError(e) + stream := runExtendedSQL(testTrainSelectIris, testDB, pr) + a.True(goodStream(stream.ReadAll())) + + pr, e = newParser().Parse(testPredictSelectIris) + a.NoError(e) + stream = runExtendedSQL(testPredictSelectIris, testDB, pr) + a.True(goodStream(stream.ReadAll())) + }) +} + +func TestExecutorTrainAndPredictBoostedTree(t *testing.T) { + a := assert.New(t) + a.NotPanics(func() { + a.True(goodStream(Run(testTrainBoostedTreeOnIris, testDB).ReadAll())) + a.True(goodStream(Run(testPredBoostedTreeOnIris, testDB).ReadAll())) + }) +} + +func TestStandardSQL(t *testing.T) { + a := assert.New(t) + a.NotPanics(func() { + stream := runStandardSQL(testSelectIris, testDB) + a.True(goodStream(stream.ReadAll())) + }) + a.NotPanics(func() { + stream := runStandardSQL(testStandardExecutiveSQLStatement, testDB) + a.True(goodStream(stream.ReadAll())) + }) + a.NotPanics(func() { + stream := runStandardSQL("SELECT * FROM iris.iris_empty LIMIT 10;", testDB) + a.True(goodStream(stream.ReadAll())) + }) +} + +func TestCreatePredictionTable(t *testing.T) { + a := assert.New(t) + trainParsed, e := newParser().Parse(testTrainSelectIris) + a.NoError(e) + predParsed, e := newParser().Parse(testPredictSelectIris) + a.NoError(e) + a.NoError(createPredictionTable(trainParsed, predParsed, testDB)) +} + +func TestIsQuery(t *testing.T) { + a := assert.New(t) + a.True(isQuery("select * from iris.iris")) + a.True(isQuery("show create table iris.iris")) + a.True(isQuery("show databases")) + a.True(isQuery("show tables")) + a.True(isQuery("describe iris.iris")) + + a.False(isQuery("select * from iris.iris limit 10 into iris.tmp")) + a.False(isQuery("insert into iris.iris values ...")) + a.False(isQuery("delete from iris.iris where ...")) + a.False(isQuery("update iris.iris where ...")) + a.False(isQuery("drop table")) +} + +func TestLogChanWriter_Write(t *testing.T) { + a := assert.New(t) + rd, wr := Pipe() + go func() { + defer wr.Close() + cw := &logChanWriter{wr: wr} + cw.Write([]byte("hello\n世界")) + cw.Write([]byte("hello\n世界")) + cw.Write([]byte("\n")) + cw.Write([]byte("世界\n世界\n世界\n")) + }() + + c := rd.ReadAll() + + a.Equal("hello\n", <-c) + a.Equal("世界hello\n", <-c) + a.Equal("世界\n", <-c) + a.Equal("世界\n", <-c) + a.Equal("世界\n", <-c) + a.Equal("世界\n", <-c) + _, more := <-c + a.False(more) +} diff --git a/sqlflow/sql/index.html b/sqlflow/sql/index.html deleted file mode 100644 index 7237249..0000000 --- a/sqlflow/sql/index.html +++ /dev/null @@ -1,13 +0,0 @@ - - - - - - - - -

    Redirecting to https://github.com/sql-machine-learning/sqlflow

    - - diff --git a/sqlflow/sql/lexer.go b/sqlflow/sql/lexer.go new file mode 100644 index 0000000..58cbe83 --- /dev/null +++ b/sqlflow/sql/lexer.go @@ -0,0 +1,163 @@ +package sql + +import ( + "regexp" + "strings" + "unicode" + "unicode/utf8" +) + +const ( + eof rune = 0 // a special rune +) + +// The lexer in this package is inspired by Rob Pike's 2011 talk on +// writing lexers manually https://talks.golang.org/2011/lex.slide#1. +// It makes a significant simplification of the idea and doesn't use +// goroutines and channels. +type lexer struct { + input string // the string being scanned + start int // start position of this item + pos int // current position in the input + width int // width of last rune read from input +} + +func newLexer(input string) *lexer { + return &lexer{input: input} +} + +func (l *lexer) Error(e string) { + log.Panicf("start=%d, pos=%d : %s near %.10q\n", l.start, l.pos, e, l.input[l.start:]) +} + +func (l *lexer) emit(lval *sqlSymType, typ int) int { + lval.val = l.input[l.start:l.pos] + l.start = l.pos + return typ +} + +func (l *lexer) next() (r rune) { + if l.pos >= len(l.input) { + l.width = 0 + return eof + } + r, l.width = utf8.DecodeRuneInString(l.input[l.pos:]) + l.pos += l.width + return r +} + +func (l *lexer) backup() { + l.pos -= l.width // when next() return eof, l.width==0. +} + +func (l *lexer) peek() rune { + r := l.next() + l.backup() + return r +} + +func (l *lexer) skipSpaces() { + for r := l.next(); unicode.IsSpace(r); r = l.next() { + } + l.backup() + l.start = l.pos +} + +func (l *lexer) Lex(lval *sqlSymType) int { + l.skipSpaces() + r := l.peek() + switch { + case unicode.IsLetter(r): + return l.lexIdentOrKeyword(lval) + case unicode.IsDigit(r): + return l.lexNumber(lval) + case r == '"': + return l.lexString(lval) + case strings.IndexRune("+-*/%<>=()[]{},;", r) >= 0: + return l.lexOperator(lval) + case r == eof: + return 0 // indicate the end of lexing. + } + log.Panicf("Lex: Unknown problem %s", l.input[l.start:]) + return -1 // indicate an error +} + +func (l *lexer) lexIdentOrKeyword(lval *sqlSymType) int { + // lexToken ensures that the first rune is a letter. + r := l.next() + for { + for unicode.IsLetter(r) || unicode.IsNumber(r) || r == '_' { + r = l.next() + } + if r != '.' { // The dot cannot be the last rune. + break + } else { + r = l.next() + } + } + l.backup() + + return l.emitIdentOrKeyword(lval) +} + +func (l *lexer) emitIdentOrKeyword(lval *sqlSymType) int { + keywds := map[string]int{ + "SELECT": SELECT, + "FROM": FROM, + "WHERE": WHERE, + "LIMIT": LIMIT, + "TRAIN": TRAIN, + "PREDICT": PREDICT, + "USING": USING, + "WITH": WITH, + "COLUMN": COLUMN, + "LABEL": LABEL, + "INTO": INTO, + "AND": AND, + "OR": OR, + "NOT": NOT, + } + if typ, ok := keywds[strings.ToUpper(l.input[l.start:l.pos])]; ok { + return l.emit(lval, typ) + } + return l.emit(lval, IDENT) +} + +var ( + // Stolen https://www.regular-expressions.info/floatingpoint.html. + reNumber = regexp.MustCompile("[-+]?[0-9]*.?[0-9]+([eE][-+]?[0-9]+)?") +) + +func (l *lexer) lexNumber(lval *sqlSymType) int { + m := reNumber.FindStringIndex(l.input[l.pos:]) + if m == nil || m[0] != 0 { + log.Panicf("Expecting a number, but see %.10q", l.input[l.pos:]) + } + l.pos += m[1] + return l.emit(lval, NUMBER) +} + +func (l *lexer) lexOperator(lval *sqlSymType) int { + r := l.next() + if r == '*' && l.peek() == '*' { + l.next() + return l.emit(lval, POWER) + } else if r == '<' && l.peek() == '=' { + l.next() + return l.emit(lval, LE) + } else if r == '>' && l.peek() == '=' { + l.next() + return l.emit(lval, GE) + } + return l.emit(lval, int(r)) +} + +func (l *lexer) lexString(lval *sqlSymType) int { + l.next() // the left quote + for r := l.next(); r != '"'; r = l.next() { + if r == '\\' { + l.next() + } + } + return l.emit(lval, STRING) +} diff --git a/sqlflow/sql/lexer_test.go b/sqlflow/sql/lexer_test.go new file mode 100644 index 0000000..35fff77 --- /dev/null +++ b/sqlflow/sql/lexer_test.go @@ -0,0 +1,102 @@ +package sql + +import ( + "testing" + + "github.com/stretchr/testify/assert" +) + +func TestNewLexer(t *testing.T) { + a := assert.New(t) + l := newLexer("") + var n sqlSymType + a.Equal(0, l.Lex(&n)) +} + +func TestNextAndBackup(t *testing.T) { + a := assert.New(t) + l := newLexer("ab") + a.Equal('a', l.next()) + l.backup() + a.Equal('a', l.next()) + a.Equal('b', l.next()) + a.Equal(eof, l.next()) + a.Equal(eof, l.next()) + l.backup() + a.Equal(eof, l.next()) +} + +func TestSkipSpaces(t *testing.T) { + a := assert.New(t) + l := newLexer("ab") + l.skipSpaces() + a.Equal('a', rune(l.input[l.start])) + a.Equal('a', l.next()) + l.skipSpaces() + a.Equal('b', rune(l.input[l.start])) + a.Equal('b', l.next()) +} + +func TestLexNumber(t *testing.T) { + a := assert.New(t) + l := newLexer("123.4") + var n sqlSymType + a.Equal(NUMBER, l.Lex(&n)) + a.Equal("123.4", n.val) +} + +func TestLexString(t *testing.T) { + a := assert.New(t) + l := newLexer(` "\"" `) + var n sqlSymType + a.Equal(STRING, l.Lex(&n)) + a.Equal(`"\""`, n.val) +} + +func TestLexOperator(t *testing.T) { + a := assert.New(t) + l := newLexer(`+-***/%()[]{}<<==,;`) + + typs := []int{ + '+', '-', POWER, '*', '/', '%', '(', ')', '[', ']', '{', '}', + '<', LE, '=', ',', ';'} + vals := []string{ + "+", "-", "**", "*", "/", "%", "(", ")", "[", "]", + "{", "}", "<", "<=", "=", ",", ";"} + i := 0 + var n sqlSymType + for typ := l.Lex(&n); typ != 0; typ = l.Lex(&n) { + a.Equal(typs[i], typ) + a.Equal(vals[i], n.val) + i++ + } +} + +func TestLexIdentOrKeyword(t *testing.T) { + a := assert.New(t) + vals := []string{"a1_2b", "x.y", "x.y.z", "Select", "froM", "where", "tRain", "colUmn", + "and", "or", "not"} + typs := []int{IDENT, IDENT, IDENT, SELECT, FROM, WHERE, TRAIN, COLUMN, + AND, OR, NOT} + var n sqlSymType + for i, it := range vals { + l := newLexer(it) + a.Equal(typs[i], l.Lex(&n)) + a.Equal(vals[i], n.val) + } +} + +func TestLexSQL(t *testing.T) { + a := assert.New(t) + l := newLexer(" Select * from a_table where a_table.col_1 > 100;") + typs := []int{ + SELECT, '*', FROM, IDENT, WHERE, IDENT, '>', NUMBER, ';'} + vals := []string{ + "Select", "*", "from", "a_table", "where", + "a_table.col_1", ">", "100", ";"} + var n sqlSymType + for i := range typs { + a.Equal(typs[i], l.Lex(&n)) + a.Equal(vals[i], n.val) + } +} diff --git a/sqlflow/sql/log.go b/sqlflow/sql/log.go new file mode 100644 index 0000000..422b3b0 --- /dev/null +++ b/sqlflow/sql/log.go @@ -0,0 +1,44 @@ +package sql + +import ( + "fmt" + "os" + "path" + + "github.com/sirupsen/logrus" +) + +var log *logrus.Entry + +func getEnv(key, fallback string) string { + value := os.Getenv(key) + if len(value) == 0 { + return fallback + } + return value +} + +func init() { + logDir := getEnv("SQLFLOW_log_dir", "logs") + logLevel := getEnv("SQLFLOW_log_level", "info") + + ll, e := logrus.ParseLevel(logLevel) + if e != nil { + ll = logrus.InfoLevel + } + + e = os.MkdirAll(logDir, 0744) + if e != nil { + log.Panicf("create log directory failed: %v", e) + } + + f, e := os.Create(path.Join(logDir, fmt.Sprintf("sqlflow-%d.log", os.Getpid()))) + if e != nil { + log.Panicf("open log file failed: %v", e) + } + + lg := logrus.New() + lg.SetOutput(f) + lg.SetLevel(ll) + log = lg.WithFields(logrus.Fields{"package": "sql"}) +} diff --git a/sqlflow/sql/model.go b/sqlflow/sql/model.go new file mode 100644 index 0000000..d490b81 --- /dev/null +++ b/sqlflow/sql/model.go @@ -0,0 +1,65 @@ +package sql + +import ( + "bytes" + "encoding/gob" + "fmt" + "os/exec" + + "github.com/sql-machine-learning/sqlflow/sqlfs" +) + +type model struct { + workDir string // We don't expose and gob workDir; instead we tar it. + TrainSelect string +} + +// save creates a sqlfs table if it doesn't yet exist, and writes the +// train select statement into the table, followed by the tar-gzipped +// SQLFlow working directory, which contains the TensorFlow working +// directory and the trained TenosrFlow model. +func (m *model) save(db *DB, table string) (e error) { + sqlf, e := sqlfs.Create(db.DB, table) + if e != nil { + return fmt.Errorf("Cannot create sqlfs file %s: %v", table, e) + } + defer sqlf.Close() + + // Use a bytes.Buffer as the gob message container to separate + // the message from the following tarball. + var buf bytes.Buffer + if e := gob.NewEncoder(&buf).Encode(m); e != nil { + return fmt.Errorf("model.save: gob-encoding model failed: %v", e) + } + if _, e := buf.WriteTo(sqlf); e != nil { + return fmt.Errorf("model.save: write the buffer failed: %v", e) + } + + cmd := exec.Command("tar", "czf", "-", "-C", m.workDir, ".") + cmd.Stdout = sqlf + return cmd.Run() +} + +// load reads from the given sqlfs table for the train select +// statement, and untar the SQLFlow working directory, which contains +// the TenosrFlow model, into directory cwd. +func load(db *DB, table, cwd string) (m *model, e error) { + sqlf, e := sqlfs.Open(db.DB, table) + if e != nil { + return nil, fmt.Errorf("Cannot open sqlfs file %s: %v", table, e) + } + defer sqlf.Close() + + var buf bytes.Buffer + if _, e := buf.ReadFrom(sqlf); e != nil { + return nil, e + } + m = &model{} + if e := gob.NewDecoder(&buf).Decode(m); e != nil { + return nil, fmt.Errorf("gob-decoding train select failed: %v", e) + } + + cmd := exec.Command("tar", "xzf", "-", "-C", cwd) + cmd.Stdin = &buf + return m, cmd.Run() +} diff --git a/sqlflow/sql/parser.go b/sqlflow/sql/parser.go new file mode 100644 index 0000000..669977d --- /dev/null +++ b/sqlflow/sql/parser.go @@ -0,0 +1,1114 @@ +// Code generated by goyacc -p sql -o parser.go sql.y. DO NOT EDIT. + +//line sql.y:2 +package sql + +import __yyfmt__ "fmt" + +//line sql.y:2 + +import ( + "fmt" + "sort" + "strings" + "sync" +) + +/* expr defines an expression as a Lisp list. If len(val)>0, + it is an atomic expression, in particular, NUMBER, IDENT, + or STRING, defined by typ and val; otherwise, it is a + Lisp S-expression. */ +type expr struct { + typ int + val string + sexp exprlist +} + +type exprlist []*expr + +/* construct an atomic expr */ +func atomic(typ int, val string) *expr { + return &expr{ + typ: typ, + val: val, + } +} + +/* construct a funcall expr */ +func funcall(name string, oprd exprlist) *expr { + return &expr{ + sexp: append(exprlist{atomic(IDENT, name)}, oprd...), + } +} + +/* construct a unary expr */ +func unary(typ int, op string, od1 *expr) *expr { + return &expr{ + sexp: append(exprlist{atomic(typ, op)}, od1), + } +} + +/* construct a binary expr */ +func binary(typ int, od1 *expr, op string, od2 *expr) *expr { + return &expr{ + sexp: append(exprlist{atomic(typ, op)}, od1, od2), + } +} + +/* construct a variadic expr */ +func variadic(typ int, op string, ods exprlist) *expr { + return &expr{ + sexp: append(exprlist{atomic(typ, op)}, ods...), + } +} + +type extendedSelect struct { + extended bool + train bool + standardSelect + trainClause + predictClause +} + +type standardSelect struct { + fields []string + tables []string + where *expr + limit string +} + +type trainClause struct { + estimator string + attrs attrs + columns exprlist + label string + save string +} + +type attrs map[string]*expr + +type predictClause struct { + model string + into string +} + +var parseResult *extendedSelect + +func attrsUnion(as1, as2 attrs) attrs { + for k, v := range as2 { + if _, ok := as1[k]; ok { + log.Panicf("attr %q already specified", as2) + } + as1[k] = v + } + return as1 +} + +//line sql.y:102 +type sqlSymType struct { + yys int + val string /* NUMBER, IDENT, STRING, and keywords */ + flds []string + tbls []string + expr *expr + expl exprlist + atrs attrs + eslt extendedSelect + slct standardSelect + tran trainClause + infr predictClause +} + +const SELECT = 57346 +const FROM = 57347 +const WHERE = 57348 +const LIMIT = 57349 +const TRAIN = 57350 +const PREDICT = 57351 +const WITH = 57352 +const COLUMN = 57353 +const LABEL = 57354 +const USING = 57355 +const INTO = 57356 +const IDENT = 57357 +const NUMBER = 57358 +const STRING = 57359 +const AND = 57360 +const OR = 57361 +const GE = 57362 +const LE = 57363 +const NOT = 57364 +const POWER = 57365 +const UMINUS = 57366 + +var sqlToknames = [...]string{ + "$end", + "error", + "$unk", + "SELECT", + "FROM", + "WHERE", + "LIMIT", + "TRAIN", + "PREDICT", + "WITH", + "COLUMN", + "LABEL", + "USING", + "INTO", + "IDENT", + "NUMBER", + "STRING", + "AND", + "OR", + "'>'", + "'<'", + "'='", + "GE", + "LE", + "'+'", + "'-'", + "'*'", + "'/'", + "'%'", + "NOT", + "POWER", + "UMINUS", + "';'", + "','", + "'('", + "')'", + "'['", + "']'", +} +var sqlStatenames = [...]string{} + +const sqlEofCode = 1 +const sqlErrCode = 2 +const sqlInitialStackSize = 16 + +//line sql.y:254 + +/* Like Lisp's builtin function cdr. */ +func (e *expr) cdr() (r []string) { + for i := 1; i < len(e.sexp); i++ { + r = append(r, e.sexp[i].String()) + } + return r +} + +func (e *expr) String() string { + if e.typ == 0 { /* a compound expression */ + switch e.sexp[0].typ { + case '+', '*', '/', '%', '=', '<', '>', LE, GE, AND, OR: + if len(e.sexp) != 3 { + log.Panicf("Expecting binary expression, got %.10q", e.sexp) + } + return fmt.Sprintf("%s %s %s", e.sexp[1], e.sexp[0].val, e.sexp[2]) + case '-': + switch len(e.sexp) { + case 2: + return fmt.Sprintf(" -%s", e.sexp[1]) + case 3: + return fmt.Sprintf("%s - %s", e.sexp[1], e.sexp[2]) + default: + log.Panicf("Expecting either unary or binary -, got %.10q", e.sexp) + } + case '(': + if len(e.sexp) != 2 { + log.Panicf("Expecting ( ) as unary operator, got %.10q", e.sexp) + } + return fmt.Sprintf("(%s)", e.sexp[1]) + case '[': + return "[" + strings.Join(e.cdr(), ", ") + "]" + case NOT: + return fmt.Sprintf("NOT %s", e.sexp[1]) + case IDENT: /* function call */ + return e.sexp[0].val + "(" + strings.Join(e.cdr(), ", ") + ")" + } + } else { + return fmt.Sprintf("%s", e.val) + } + + log.Panicf("Cannot print an unknown expression") + return "" +} + +func (s standardSelect) String() string { + r := "SELECT " + if len(s.fields) == 0 { + r += "*" + } else { + r += strings.Join(s.fields, ", ") + } + r += "\nFROM " + strings.Join(s.tables, ", ") + if s.where != nil { + r += fmt.Sprintf("\nWHERE %s", s.where) + } + if len(s.limit) > 0 { + r += fmt.Sprintf("\nLIMIT %s", s.limit) + } + return r + ";" +} + +func jsonString(s string) string { + return strings.Replace( + strings.Replace( + strings.Replace(s, "\n", "\\n", -1), + "\r", "\\r", -1), + "\"", "\\\"", -1) +} + +func (ats attrs) JSON() string { + ks := []string{} + for k := range ats { + ks = append(ks, k) + } + sort.Strings(ks) /* Remove the randomness of map traversal. */ + + for i, k := range ks { + ks[i] = fmt.Sprintf(`"%s": "%s"`, k, jsonString(ats[k].String())) + } + return "{\n" + strings.Join(ks, ",\n") + "\n}" +} + +func (el exprlist) JSON() string { + ks := []string{} + for _, e := range el { + ks = append(ks, "\""+jsonString(e.String())+"\"") + } + return "[\n" + strings.Join(ks, ",\n") + "\n]" +} + +func (s trainClause) JSON() string { + fmter := `{ +"estimator": "%s", +"attrs": %s, +"columns": %s, +"save": "%s" +}` + return fmt.Sprintf(fmter, s.estimator, s.attrs.JSON(), s.columns.JSON(), s.save) +} + +func (s predictClause) JSON() string { + fmter := `{ +"model":"%s" +}` + return fmt.Sprintf(fmter, "\""+s.model+"\"") +} + +func (s extendedSelect) JSON() string { + bf := `{ +"extended": %t, +"train": %t, +"standardSelect": "%s" +}` + tf := `{ +"extended": %t, +"train": %t, +"standardSelect": "%s", +"trainClause": %s +}` + nf := `{ +"extended": %t, +"train": %t, +"standardSelect": "%s", +"predictClause": %s +}` + if s.extended { + if s.train { + return fmt.Sprintf(tf, s.extended, s.train, + jsonString(s.standardSelect.String()), s.trainClause.JSON()) + } else { + return fmt.Sprintf(nf, s.extended, s.train, + jsonString(s.standardSelect.String()), s.predictClause.JSON()) + } + } + return fmt.Sprintf(bf, s.extended, s.train, jsonString(s.standardSelect.String())) +} + +// sqlReentrantParser makes sqlParser, generated by goyacc and using a +// global variable parseResult to return the result, reentrant. +type sqlSyncParser struct { + pr sqlParser +} + +func newParser() *sqlSyncParser { + return &sqlSyncParser{sqlNewParser()} +} + +var mu sync.Mutex + +func (p *sqlSyncParser) Parse(s string) (r *extendedSelect, e error) { + defer func() { + if r := recover(); r != nil { + var ok bool + e, ok = r.(error) + if !ok { + e = fmt.Errorf("%v", r) + } + } + }() + + mu.Lock() + defer mu.Unlock() + + p.pr.Parse(newLexer(s)) + return parseResult, nil +} + +//line yacctab:1 +var sqlExca = [...]int{ + -1, 1, + 1, -1, + -2, 0, +} + +const sqlPrivate = 57344 + +const sqlLast = 141 + +var sqlAct = [...]int{ + + 52, 26, 84, 75, 51, 72, 72, 33, 78, 73, + 20, 46, 44, 45, 41, 40, 39, 43, 42, 34, + 35, 36, 37, 38, 32, 16, 47, 82, 48, 49, + 71, 15, 19, 95, 80, 57, 58, 59, 60, 61, + 62, 63, 64, 65, 66, 67, 68, 22, 21, 23, + 90, 70, 7, 9, 8, 10, 11, 81, 28, 86, + 92, 76, 27, 36, 37, 38, 14, 25, 77, 29, + 50, 85, 91, 79, 56, 22, 21, 23, 13, 55, + 4, 31, 87, 89, 30, 88, 28, 22, 21, 23, + 27, 18, 94, 87, 93, 25, 69, 29, 28, 54, + 53, 3, 27, 74, 83, 24, 17, 25, 12, 29, + 44, 45, 41, 40, 39, 43, 42, 34, 35, 36, + 37, 38, 41, 40, 39, 43, 42, 34, 35, 36, + 37, 38, 34, 35, 36, 37, 38, 6, 5, 2, + 1, +} +var sqlPact = [...]int{ + + 97, -1000, 47, 51, -1000, -2, -8, 76, 16, 72, + 69, 66, -10, -1000, -1000, -1000, -1000, -27, -1000, -1000, + 92, -1000, -24, -1000, -1000, 72, -1000, 72, 72, 32, + 90, 86, 64, 59, 72, 72, 72, 72, 72, 72, + 72, 72, 72, 72, 72, 72, 60, -6, -1000, -1000, + -1000, -29, 92, 46, 53, -1000, -1000, 36, 36, -1000, + -1000, -1000, 107, 107, 107, 107, 107, 102, 102, -1000, + -28, -1000, 72, -1000, 23, -1000, 5, -1000, -1000, 92, + 44, 46, 72, 38, -1000, -1000, -24, -1000, -1000, 92, + 45, 44, 78, -1000, 18, -1000, +} +var sqlPgo = [...]int{ + + 0, 140, 139, 138, 137, 108, 106, 0, 1, 2, + 4, 105, 104, 3, 103, +} +var sqlR1 = [...]int{ + + 0, 1, 1, 1, 2, 2, 2, 2, 3, 4, + 5, 5, 5, 9, 9, 9, 12, 12, 6, 6, + 13, 14, 14, 8, 8, 10, 10, 11, 11, 7, + 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, + 7, 7, 7, 7, 7, 7, 7, 7, 7, +} +var sqlR2 = [...]int{ + + 0, 2, 3, 3, 2, 3, 3, 3, 10, 4, + 1, 1, 3, 1, 1, 1, 1, 3, 1, 3, + 3, 1, 3, 3, 4, 1, 3, 2, 3, 1, + 1, 1, 1, 3, 1, 3, 3, 3, 3, 3, + 3, 3, 3, 3, 3, 3, 3, 2, 2, +} +var sqlChk = [...]int{ + + -1000, -1, -2, 4, 33, -3, -4, 5, 7, 6, + 8, 9, -5, 27, 15, 33, 33, -6, 15, 16, + -7, 16, 15, 17, -11, 35, -8, 30, 26, 37, + 15, 15, 34, 34, 25, 26, 27, 28, 29, 22, + 21, 20, 24, 23, 18, 19, 35, -7, -7, -7, + 38, -10, -7, 10, 13, 15, 15, -7, -7, -7, + -7, -7, -7, -7, -7, -7, -7, -7, -7, 36, + -10, 36, 34, 38, -14, -13, 15, 15, 36, -7, + 11, 34, 22, -12, -9, 27, 15, -8, -13, -7, + 12, 34, 15, -9, 14, 15, +} +var sqlDef = [...]int{ + + 0, -2, 0, 0, 1, 0, 0, 0, 0, 0, + 0, 0, 4, 10, 11, 2, 3, 5, 18, 6, + 7, 29, 30, 31, 32, 0, 34, 0, 0, 0, + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, + 0, 0, 0, 0, 0, 0, 0, 0, 47, 48, + 27, 0, 25, 0, 0, 12, 19, 35, 36, 37, + 38, 39, 40, 41, 42, 43, 44, 45, 46, 23, + 0, 33, 0, 28, 0, 21, 0, 9, 24, 26, + 0, 0, 0, 0, 16, 13, 14, 15, 22, 20, + 0, 0, 0, 17, 0, 8, +} +var sqlTok1 = [...]int{ + + 1, 3, 3, 3, 3, 3, 3, 3, 3, 3, + 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, + 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, + 3, 3, 3, 3, 3, 3, 3, 29, 3, 3, + 35, 36, 27, 25, 34, 26, 3, 28, 3, 3, + 3, 3, 3, 3, 3, 3, 3, 3, 3, 33, + 21, 22, 20, 3, 3, 3, 3, 3, 3, 3, + 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, + 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, + 3, 37, 3, 38, +} +var sqlTok2 = [...]int{ + + 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, + 12, 13, 14, 15, 16, 17, 18, 19, 23, 24, + 30, 31, 32, +} +var sqlTok3 = [...]int{ + 0, +} + +var sqlErrorMessages = [...]struct { + state int + token int + msg string +}{} + +//line yaccpar:1 + +/* parser for yacc output */ + +var ( + sqlDebug = 0 + sqlErrorVerbose = false +) + +type sqlLexer interface { + Lex(lval *sqlSymType) int + Error(s string) +} + +type sqlParser interface { + Parse(sqlLexer) int + Lookahead() int +} + +type sqlParserImpl struct { + lval sqlSymType + stack [sqlInitialStackSize]sqlSymType + char int +} + +func (p *sqlParserImpl) Lookahead() int { + return p.char +} + +func sqlNewParser() sqlParser { + return &sqlParserImpl{} +} + +const sqlFlag = -1000 + +func sqlTokname(c int) string { + if c >= 1 && c-1 < len(sqlToknames) { + if sqlToknames[c-1] != "" { + return sqlToknames[c-1] + } + } + return __yyfmt__.Sprintf("tok-%v", c) +} + +func sqlStatname(s int) string { + if s >= 0 && s < len(sqlStatenames) { + if sqlStatenames[s] != "" { + return sqlStatenames[s] + } + } + return __yyfmt__.Sprintf("state-%v", s) +} + +func sqlErrorMessage(state, lookAhead int) string { + const TOKSTART = 4 + + if !sqlErrorVerbose { + return "syntax error" + } + + for _, e := range sqlErrorMessages { + if e.state == state && e.token == lookAhead { + return "syntax error: " + e.msg + } + } + + res := "syntax error: unexpected " + sqlTokname(lookAhead) + + // To match Bison, suggest at most four expected tokens. + expected := make([]int, 0, 4) + + // Look for shiftable tokens. + base := sqlPact[state] + for tok := TOKSTART; tok-1 < len(sqlToknames); tok++ { + if n := base + tok; n >= 0 && n < sqlLast && sqlChk[sqlAct[n]] == tok { + if len(expected) == cap(expected) { + return res + } + expected = append(expected, tok) + } + } + + if sqlDef[state] == -2 { + i := 0 + for sqlExca[i] != -1 || sqlExca[i+1] != state { + i += 2 + } + + // Look for tokens that we accept or reduce. + for i += 2; sqlExca[i] >= 0; i += 2 { + tok := sqlExca[i] + if tok < TOKSTART || sqlExca[i+1] == 0 { + continue + } + if len(expected) == cap(expected) { + return res + } + expected = append(expected, tok) + } + + // If the default action is to accept or reduce, give up. + if sqlExca[i+1] != 0 { + return res + } + } + + for i, tok := range expected { + if i == 0 { + res += ", expecting " + } else { + res += " or " + } + res += sqlTokname(tok) + } + return res +} + +func sqllex1(lex sqlLexer, lval *sqlSymType) (char, token int) { + token = 0 + char = lex.Lex(lval) + if char <= 0 { + token = sqlTok1[0] + goto out + } + if char < len(sqlTok1) { + token = sqlTok1[char] + goto out + } + if char >= sqlPrivate { + if char < sqlPrivate+len(sqlTok2) { + token = sqlTok2[char-sqlPrivate] + goto out + } + } + for i := 0; i < len(sqlTok3); i += 2 { + token = sqlTok3[i+0] + if token == char { + token = sqlTok3[i+1] + goto out + } + } + +out: + if token == 0 { + token = sqlTok2[1] /* unknown char */ + } + if sqlDebug >= 3 { + __yyfmt__.Printf("lex %s(%d)\n", sqlTokname(token), uint(char)) + } + return char, token +} + +func sqlParse(sqllex sqlLexer) int { + return sqlNewParser().Parse(sqllex) +} + +func (sqlrcvr *sqlParserImpl) Parse(sqllex sqlLexer) int { + var sqln int + var sqlVAL sqlSymType + var sqlDollar []sqlSymType + _ = sqlDollar // silence set and not used + sqlS := sqlrcvr.stack[:] + + Nerrs := 0 /* number of errors */ + Errflag := 0 /* error recovery flag */ + sqlstate := 0 + sqlrcvr.char = -1 + sqltoken := -1 // sqlrcvr.char translated into internal numbering + defer func() { + // Make sure we report no lookahead when not parsing. + sqlstate = -1 + sqlrcvr.char = -1 + sqltoken = -1 + }() + sqlp := -1 + goto sqlstack + +ret0: + return 0 + +ret1: + return 1 + +sqlstack: + /* put a state and value onto the stack */ + if sqlDebug >= 4 { + __yyfmt__.Printf("char %v in %v\n", sqlTokname(sqltoken), sqlStatname(sqlstate)) + } + + sqlp++ + if sqlp >= len(sqlS) { + nyys := make([]sqlSymType, len(sqlS)*2) + copy(nyys, sqlS) + sqlS = nyys + } + sqlS[sqlp] = sqlVAL + sqlS[sqlp].yys = sqlstate + +sqlnewstate: + sqln = sqlPact[sqlstate] + if sqln <= sqlFlag { + goto sqldefault /* simple state */ + } + if sqlrcvr.char < 0 { + sqlrcvr.char, sqltoken = sqllex1(sqllex, &sqlrcvr.lval) + } + sqln += sqltoken + if sqln < 0 || sqln >= sqlLast { + goto sqldefault + } + sqln = sqlAct[sqln] + if sqlChk[sqln] == sqltoken { /* valid shift */ + sqlrcvr.char = -1 + sqltoken = -1 + sqlVAL = sqlrcvr.lval + sqlstate = sqln + if Errflag > 0 { + Errflag-- + } + goto sqlstack + } + +sqldefault: + /* default state action */ + sqln = sqlDef[sqlstate] + if sqln == -2 { + if sqlrcvr.char < 0 { + sqlrcvr.char, sqltoken = sqllex1(sqllex, &sqlrcvr.lval) + } + + /* look through exception table */ + xi := 0 + for { + if sqlExca[xi+0] == -1 && sqlExca[xi+1] == sqlstate { + break + } + xi += 2 + } + for xi += 2; ; xi += 2 { + sqln = sqlExca[xi+0] + if sqln < 0 || sqln == sqltoken { + break + } + } + sqln = sqlExca[xi+1] + if sqln < 0 { + goto ret0 + } + } + if sqln == 0 { + /* error ... attempt to resume parsing */ + switch Errflag { + case 0: /* brand new error */ + sqllex.Error(sqlErrorMessage(sqlstate, sqltoken)) + Nerrs++ + if sqlDebug >= 1 { + __yyfmt__.Printf("%s", sqlStatname(sqlstate)) + __yyfmt__.Printf(" saw %s\n", sqlTokname(sqltoken)) + } + fallthrough + + case 1, 2: /* incompletely recovered error ... try again */ + Errflag = 3 + + /* find a state where "error" is a legal shift action */ + for sqlp >= 0 { + sqln = sqlPact[sqlS[sqlp].yys] + sqlErrCode + if sqln >= 0 && sqln < sqlLast { + sqlstate = sqlAct[sqln] /* simulate a shift of "error" */ + if sqlChk[sqlstate] == sqlErrCode { + goto sqlstack + } + } + + /* the current p has no shift on "error", pop stack */ + if sqlDebug >= 2 { + __yyfmt__.Printf("error recovery pops state %d\n", sqlS[sqlp].yys) + } + sqlp-- + } + /* there is no state on the stack with an error shift ... abort */ + goto ret1 + + case 3: /* no shift yet; clobber input char */ + if sqlDebug >= 2 { + __yyfmt__.Printf("error recovery discards %s\n", sqlTokname(sqltoken)) + } + if sqltoken == sqlEofCode { + goto ret1 + } + sqlrcvr.char = -1 + sqltoken = -1 + goto sqlnewstate /* try again in the same state */ + } + } + + /* reduction by production sqln */ + if sqlDebug >= 2 { + __yyfmt__.Printf("reduce %v in:\n\t%v\n", sqln, sqlStatname(sqlstate)) + } + + sqlnt := sqln + sqlpt := sqlp + _ = sqlpt // guard against "declared and not used" + + sqlp -= sqlR2[sqln] + // sqlp is now the index of $0. Perform the default action. Iff the + // reduced production is ε, $1 is possibly out of range. + if sqlp+1 >= len(sqlS) { + nyys := make([]sqlSymType, len(sqlS)*2) + copy(nyys, sqlS) + sqlS = nyys + } + sqlVAL = sqlS[sqlp+1] + + /* consult goto table to find next state */ + sqln = sqlR1[sqln] + sqlg := sqlPgo[sqln] + sqlj := sqlg + sqlS[sqlp].yys + 1 + + if sqlj >= sqlLast { + sqlstate = sqlAct[sqlg] + } else { + sqlstate = sqlAct[sqlj] + if sqlChk[sqlstate] != -sqln { + sqlstate = sqlAct[sqlg] + } + } + // dummy call; replaced with literal code + switch sqlnt { + + case 1: + sqlDollar = sqlS[sqlpt-2 : sqlpt+1] +//line sql.y:140 + { + parseResult = &extendedSelect{ + extended: false, + standardSelect: sqlDollar[1].slct} + } + case 2: + sqlDollar = sqlS[sqlpt-3 : sqlpt+1] +//line sql.y:145 + { + parseResult = &extendedSelect{ + extended: true, + train: true, + standardSelect: sqlDollar[1].slct, + trainClause: sqlDollar[2].tran} + } + case 3: + sqlDollar = sqlS[sqlpt-3 : sqlpt+1] +//line sql.y:152 + { + parseResult = &extendedSelect{ + extended: true, + train: false, + standardSelect: sqlDollar[1].slct, + predictClause: sqlDollar[2].infr} + } + case 4: + sqlDollar = sqlS[sqlpt-2 : sqlpt+1] +//line sql.y:162 + { + sqlVAL.slct.fields = sqlDollar[2].flds + } + case 5: + sqlDollar = sqlS[sqlpt-3 : sqlpt+1] +//line sql.y:163 + { + sqlVAL.slct.tables = sqlDollar[3].tbls + } + case 6: + sqlDollar = sqlS[sqlpt-3 : sqlpt+1] +//line sql.y:164 + { + sqlVAL.slct.limit = sqlDollar[3].val + } + case 7: + sqlDollar = sqlS[sqlpt-3 : sqlpt+1] +//line sql.y:165 + { + sqlVAL.slct.where = sqlDollar[3].expr + } + case 8: + sqlDollar = sqlS[sqlpt-10 : sqlpt+1] +//line sql.y:169 + { + sqlVAL.tran.estimator = sqlDollar[2].val + sqlVAL.tran.attrs = sqlDollar[4].atrs + sqlVAL.tran.columns = sqlDollar[6].expl + sqlVAL.tran.label = sqlDollar[8].val + sqlVAL.tran.save = sqlDollar[10].val + } + case 9: + sqlDollar = sqlS[sqlpt-4 : sqlpt+1] +//line sql.y:179 + { + sqlVAL.infr.into = sqlDollar[2].val + sqlVAL.infr.model = sqlDollar[4].val + } + case 10: + sqlDollar = sqlS[sqlpt-1 : sqlpt+1] +//line sql.y:186 + { + sqlVAL.flds = append(sqlVAL.flds, sqlDollar[1].val) + } + case 11: + sqlDollar = sqlS[sqlpt-1 : sqlpt+1] +//line sql.y:187 + { + sqlVAL.flds = append(sqlVAL.flds, sqlDollar[1].val) + } + case 12: + sqlDollar = sqlS[sqlpt-3 : sqlpt+1] +//line sql.y:188 + { + sqlVAL.flds = append(sqlVAL.flds, sqlDollar[3].val) + } + case 13: + sqlDollar = sqlS[sqlpt-1 : sqlpt+1] +//line sql.y:192 + { + sqlVAL.expr = atomic(IDENT, "*") + } + case 14: + sqlDollar = sqlS[sqlpt-1 : sqlpt+1] +//line sql.y:193 + { + sqlVAL.expr = atomic(IDENT, sqlDollar[1].val) + } + case 15: + sqlDollar = sqlS[sqlpt-1 : sqlpt+1] +//line sql.y:194 + { + sqlVAL.expr = sqlDollar[1].expr + } + case 16: + sqlDollar = sqlS[sqlpt-1 : sqlpt+1] +//line sql.y:198 + { + sqlVAL.expl = exprlist{sqlDollar[1].expr} + } + case 17: + sqlDollar = sqlS[sqlpt-3 : sqlpt+1] +//line sql.y:199 + { + sqlVAL.expl = append(sqlDollar[1].expl, sqlDollar[3].expr) + } + case 18: + sqlDollar = sqlS[sqlpt-1 : sqlpt+1] +//line sql.y:203 + { + sqlVAL.tbls = []string{sqlDollar[1].val} + } + case 19: + sqlDollar = sqlS[sqlpt-3 : sqlpt+1] +//line sql.y:204 + { + sqlVAL.tbls = append(sqlDollar[1].tbls, sqlDollar[3].val) + } + case 20: + sqlDollar = sqlS[sqlpt-3 : sqlpt+1] +//line sql.y:208 + { + sqlVAL.atrs = attrs{sqlDollar[1].val: sqlDollar[3].expr} + } + case 21: + sqlDollar = sqlS[sqlpt-1 : sqlpt+1] +//line sql.y:212 + { + sqlVAL.atrs = sqlDollar[1].atrs + } + case 22: + sqlDollar = sqlS[sqlpt-3 : sqlpt+1] +//line sql.y:213 + { + sqlVAL.atrs = attrsUnion(sqlDollar[1].atrs, sqlDollar[3].atrs) + } + case 23: + sqlDollar = sqlS[sqlpt-3 : sqlpt+1] +//line sql.y:217 + { + sqlVAL.expr = funcall(sqlDollar[1].val, nil) + } + case 24: + sqlDollar = sqlS[sqlpt-4 : sqlpt+1] +//line sql.y:218 + { + sqlVAL.expr = funcall(sqlDollar[1].val, sqlDollar[3].expl) + } + case 25: + sqlDollar = sqlS[sqlpt-1 : sqlpt+1] +//line sql.y:222 + { + sqlVAL.expl = exprlist{sqlDollar[1].expr} + } + case 26: + sqlDollar = sqlS[sqlpt-3 : sqlpt+1] +//line sql.y:223 + { + sqlVAL.expl = append(sqlDollar[1].expl, sqlDollar[3].expr) + } + case 27: + sqlDollar = sqlS[sqlpt-2 : sqlpt+1] +//line sql.y:227 + { + sqlVAL.expl = nil + } + case 28: + sqlDollar = sqlS[sqlpt-3 : sqlpt+1] +//line sql.y:228 + { + sqlVAL.expl = sqlDollar[2].expl + } + case 29: + sqlDollar = sqlS[sqlpt-1 : sqlpt+1] +//line sql.y:232 + { + sqlVAL.expr = atomic(NUMBER, sqlDollar[1].val) + } + case 30: + sqlDollar = sqlS[sqlpt-1 : sqlpt+1] +//line sql.y:233 + { + sqlVAL.expr = atomic(IDENT, sqlDollar[1].val) + } + case 31: + sqlDollar = sqlS[sqlpt-1 : sqlpt+1] +//line sql.y:234 + { + sqlVAL.expr = atomic(STRING, sqlDollar[1].val) + } + case 32: + sqlDollar = sqlS[sqlpt-1 : sqlpt+1] +//line sql.y:235 + { + sqlVAL.expr = variadic('[', "square", sqlDollar[1].expl) + } + case 33: + sqlDollar = sqlS[sqlpt-3 : sqlpt+1] +//line sql.y:236 + { + sqlVAL.expr = unary('(', "paren", sqlDollar[2].expr) + } + case 34: + sqlDollar = sqlS[sqlpt-1 : sqlpt+1] +//line sql.y:237 + { + sqlVAL.expr = sqlDollar[1].expr + } + case 35: + sqlDollar = sqlS[sqlpt-3 : sqlpt+1] +//line sql.y:238 + { + sqlVAL.expr = binary('+', sqlDollar[1].expr, sqlDollar[2].val, sqlDollar[3].expr) + } + case 36: + sqlDollar = sqlS[sqlpt-3 : sqlpt+1] +//line sql.y:239 + { + sqlVAL.expr = binary('-', sqlDollar[1].expr, sqlDollar[2].val, sqlDollar[3].expr) + } + case 37: + sqlDollar = sqlS[sqlpt-3 : sqlpt+1] +//line sql.y:240 + { + sqlVAL.expr = binary('*', sqlDollar[1].expr, sqlDollar[2].val, sqlDollar[3].expr) + } + case 38: + sqlDollar = sqlS[sqlpt-3 : sqlpt+1] +//line sql.y:241 + { + sqlVAL.expr = binary('/', sqlDollar[1].expr, sqlDollar[2].val, sqlDollar[3].expr) + } + case 39: + sqlDollar = sqlS[sqlpt-3 : sqlpt+1] +//line sql.y:242 + { + sqlVAL.expr = binary('%', sqlDollar[1].expr, sqlDollar[2].val, sqlDollar[3].expr) + } + case 40: + sqlDollar = sqlS[sqlpt-3 : sqlpt+1] +//line sql.y:243 + { + sqlVAL.expr = binary('=', sqlDollar[1].expr, sqlDollar[2].val, sqlDollar[3].expr) + } + case 41: + sqlDollar = sqlS[sqlpt-3 : sqlpt+1] +//line sql.y:244 + { + sqlVAL.expr = binary('<', sqlDollar[1].expr, sqlDollar[2].val, sqlDollar[3].expr) + } + case 42: + sqlDollar = sqlS[sqlpt-3 : sqlpt+1] +//line sql.y:245 + { + sqlVAL.expr = binary('>', sqlDollar[1].expr, sqlDollar[2].val, sqlDollar[3].expr) + } + case 43: + sqlDollar = sqlS[sqlpt-3 : sqlpt+1] +//line sql.y:246 + { + sqlVAL.expr = binary(LE, sqlDollar[1].expr, sqlDollar[2].val, sqlDollar[3].expr) + } + case 44: + sqlDollar = sqlS[sqlpt-3 : sqlpt+1] +//line sql.y:247 + { + sqlVAL.expr = binary(GE, sqlDollar[1].expr, sqlDollar[2].val, sqlDollar[3].expr) + } + case 45: + sqlDollar = sqlS[sqlpt-3 : sqlpt+1] +//line sql.y:248 + { + sqlVAL.expr = binary(AND, sqlDollar[1].expr, sqlDollar[2].val, sqlDollar[3].expr) + } + case 46: + sqlDollar = sqlS[sqlpt-3 : sqlpt+1] +//line sql.y:249 + { + sqlVAL.expr = binary(OR, sqlDollar[1].expr, sqlDollar[2].val, sqlDollar[3].expr) + } + case 47: + sqlDollar = sqlS[sqlpt-2 : sqlpt+1] +//line sql.y:250 + { + sqlVAL.expr = unary(NOT, sqlDollar[1].val, sqlDollar[2].expr) + } + case 48: + sqlDollar = sqlS[sqlpt-2 : sqlpt+1] +//line sql.y:251 + { + sqlVAL.expr = unary('-', sqlDollar[1].val, sqlDollar[2].expr) + } + } + goto sqlstack /* stack new state and value */ +} diff --git a/sqlflow/sql/parser_test.go b/sqlflow/sql/parser_test.go new file mode 100644 index 0000000..7c5e315 --- /dev/null +++ b/sqlflow/sql/parser_test.go @@ -0,0 +1,96 @@ +package sql + +import ( + "testing" + + "github.com/stretchr/testify/assert" +) + +const ( + testStandardSelectStmt = ` +SELECT employee.age, last_name, salary +FROM employee +LIMIT 100 +WHERE + employee.age % 10 < (salary / 10000) + AND + strings.Upper(last_name) = "WANG" +` + testTrainSelect = testStandardSelectStmt + `TRAIN DNNClassifier +WITH + n_classes = 3, + hidden_units = [10, 20] +COLUMN + employee.name, + bucketize(last_name, 1000), + cross(embedding(emplyoee.name), bucketize(last_name, 1000)) +LABEL employee.salary +INTO sqlflow_models.my_dnn_model; +` + testPredictSelect = testStandardSelectStmt + `PREDICT db.table.field +USING sqlflow_models.my_dnn_model;` +) + +func TestStandardSelect(t *testing.T) { + a := assert.New(t) + r, e := newParser().Parse(testStandardSelectStmt + ";") + a.NoError(e) + a.False(r.extended) + a.Equal([]string{"employee.age", "last_name", "salary"}, + r.fields) + a.Equal([]string{"employee"}, r.tables) + a.Equal("100", r.limit) + a.Equal(AND, r.where.sexp[0].typ) + a.Equal('<', rune(r.where.sexp[1].sexp[0].typ)) + a.Equal('=', rune(r.where.sexp[2].sexp[0].typ)) + a.Equal(`employee.age % 10 < (salary / 10000) AND `+ + `strings.Upper(last_name) = "WANG"`, + r.where.String()) +} + +func TestTrainParser(t *testing.T) { + a := assert.New(t) + r, e := newParser().Parse(testTrainSelect) + a.NoError(e) + a.True(r.extended) + a.True(r.train) + a.Equal("DNNClassifier", r.estimator) + a.Equal("[10, 20]", r.attrs["hidden_units"].String()) + a.Equal("3", r.attrs["n_classes"].String()) + a.Equal(`employee.name`, + r.columns[0].String()) + a.Equal(`bucketize(last_name, 1000)`, + r.columns[1].String()) + a.Equal( + `cross(embedding(emplyoee.name), bucketize(last_name, 1000))`, + r.columns[2].String()) + a.Equal("employee.salary", r.label) + a.Equal("sqlflow_models.my_dnn_model", r.save) +} + +func TestPredictParser(t *testing.T) { + a := assert.New(t) + r, e := newParser().Parse(testPredictSelect) + a.NoError(e) + a.True(r.extended) + a.False(r.train) + a.Equal("sqlflow_models.my_dnn_model", r.model) + a.Equal("db.table.field", r.into) +} + +func TestSelectStarAndPrint(t *testing.T) { + a := assert.New(t) + r, e := newParser().Parse(`SELECT *, b FROM a LIMIT 10;`) + a.NoError(e) + a.Equal(2, len(r.fields)) + a.Equal("*", r.fields[0]) + a.False(r.extended) + a.False(r.train) + a.Equal("SELECT *, b\nFROM a\nLIMIT 10;", r.standardSelect.String()) +} + +func TestStandardDropTable(t *testing.T) { + if _, e := newParser().Parse(`DROP TABLE PREDICT`); e != nil { + t.Skipf("[FIXME]`drop table` expected no error, but got:%v", e) + } +} diff --git a/sqlflow/sql/pipe.go b/sqlflow/sql/pipe.go new file mode 100644 index 0000000..87a180e --- /dev/null +++ b/sqlflow/sql/pipe.go @@ -0,0 +1,65 @@ +package sql + +import ( + "errors" +) + +var ErrClosedPipe = errors.New("pipe: write on closed pipe") + +// pipe follows the design at https://blog.golang.org/pipelines +// - wrCh: chan for piping data +// - done: chan for signaling Close from Reader to Writer +type pipe struct { + wrCh chan interface{} + done chan struct{} +} + +// PipeReader reads real data +type PipeReader struct { + p *pipe +} + +// PipeWriter writes real data +type PipeWriter struct { + p *pipe +} + +// Pipe creates a synchronous in-memory pipe. +// +// It is safe to call Read and Write in parallel with each other or with Close. +// Parallel calls to Read and parallel calls to Write are also safe: +// the individual calls will be gated sequentially. +func Pipe() (*PipeReader, *PipeWriter) { + p := &pipe{ + wrCh: make(chan interface{}), + done: make(chan struct{})} + return &PipeReader{p}, &PipeWriter{p} +} + +// Close closes the reader; subsequent writes to the +func (r *PipeReader) Close() { + close(r.p.done) +} + +// ReadAll returns the data chan. The caller should +// use it as `for r := range pr.ReadAll()` +func (r *PipeReader) ReadAll() chan interface{} { + return r.p.wrCh +} + +// Close closes the writer; subsequent ReadAll from the +// read half of the pipe will return a closed channel. +func (w *PipeWriter) Close() { + close(w.p.wrCh) +} + +// Write writes the item to the underlying data stream. +// It returns ErrClosedPipe when the data stream is closed. +func (w *PipeWriter) Write(item interface{}) error { + select { + case w.p.wrCh <- item: + return nil + case <-w.p.done: + return ErrClosedPipe + } +} diff --git a/sqlflow/sql/pipe_test.go b/sqlflow/sql/pipe_test.go new file mode 100644 index 0000000..ab6db98 --- /dev/null +++ b/sqlflow/sql/pipe_test.go @@ -0,0 +1,59 @@ +package sql + +import ( + "testing" + "time" + + "github.com/stretchr/testify/assert" +) + +func TestPipeWriteAndRead(t *testing.T) { + a := assert.New(t) + rd, wr := Pipe() + loop := 10 + + // writer + go func(n int) { + defer wr.Close() + for i := 0; i < n; i++ { + e := wr.Write(i) + a.NoError(e) + time.Sleep(5 * time.Millisecond) + } + }(loop) + + // reader + i := 0 + for data := range rd.ReadAll() { + a.Equal(data, i) + i++ + } + a.Equal(i, loop) + rd.Close() +} + +func TestPipeReaderClose(t *testing.T) { + a := assert.New(t) + rd, wr := Pipe() + + writeReturns := make(chan bool) + // writer + go func() { + defer wr.Close() + defer func() { + writeReturns <- true + }() + for { + if e := wr.Write(1); e != nil { + return + } + } + }() + + rd.Close() + select { + case <-writeReturns: + case <-time.After(time.Second): + a.True(false, "time out on writer return") + } +} diff --git a/sqlflow/sql/python/.gitignore b/sqlflow/sql/python/.gitignore new file mode 100644 index 0000000..0d20b64 --- /dev/null +++ b/sqlflow/sql/python/.gitignore @@ -0,0 +1 @@ +*.pyc diff --git a/sqlflow/sql/python/README.md b/sqlflow/sql/python/README.md new file mode 100644 index 0000000..a5e59e6 --- /dev/null +++ b/sqlflow/sql/python/README.md @@ -0,0 +1,13 @@ +# Python Code Template + +To run a test, say, `test_sql_data.py`, we need to + +1. Start a container that runs MySQL server with populated data following [this guide](https://github.com/sql-machine-learning/sqlflow/blob/develop/example/datasets/README.md). + +1. Run the tests in a SQLFlow container that has TensorFlow and `mysql_connector` installed: + + ```bash + docker run --rm -it --network="host" -v $PWD:/work -w /work sqlflow/sqlflow python test_sql_data.py + ``` + + where `--network="host"` allows processes running in the container to access the host's network, where the MySQL server container exposes its port. diff --git a/sqlflow/sql/python/sql_data.py b/sqlflow/sql/python/sql_data.py new file mode 100644 index 0000000..49040ed --- /dev/null +++ b/sqlflow/sql/python/sql_data.py @@ -0,0 +1,57 @@ +import tensorflow as tf +import mysql.connector + + +def connect(user, passwd, host, port): + """connect is a convenient shortcut to mysql.connector.connect. Also, + it makes it reaonable to import mysql.connector in this file, so + to make it self-complete as a template. + + """ + return mysql.connector.connect(user=user, + passwd=passwd, + host=host, + port=port) + + +def load(db, slct, label, features): + """load returns the training features and the labels. The features is +a dict from field names to data columns, and the label is a dict from +the label field name to the label data column. + + Args: + db: returned from mysql.connector.connect() + slct (str): SQL SELECT statement + label (str): the label field name as a string. + features (list of str or None): feature field names. + + Returns: + (features, label): maps from feature/label field names to columns. + + """ + cursor = db.cursor() + cursor.execute(slct) + f = [i[0] for i in cursor.description] # get field names. + c = list(zip(*cursor.fetchall())) # transpose rows into columns. + d = dict(zip(f, c)) # dict from field names to columns. + l = d.pop(label) + if features != None: + d = dict((k, d[k]) for k in features) + return d, l + + +def feature_columns(features): + """feature_columns returns a list of tf.feature_column. + + Note: currently, we are using a quick-and-hacky implementation + that assumes all fields are numeric_columns, and not supporting + feature derivations like tf.feature_column.cross_column. + + Args: + features: returned from load + + Returns: + list of tf.feature_column + + """ + return [tf.feature_column.numeric_column(key=k) for k in features.keys()] diff --git a/sqlflow/sql/python/test_sql_data.py b/sqlflow/sql/python/test_sql_data.py new file mode 100644 index 0000000..00ac8a1 --- /dev/null +++ b/sqlflow/sql/python/test_sql_data.py @@ -0,0 +1,32 @@ +import sql_data +import unittest + + +class TestSQLData(unittest.TestCase): + def __init__(self, *args, **kwargs): + super(TestSQLData, self).__init__(*args, **kwargs) + self.db = sql_data.connect('root', 'root', 'localhost', 3306) + self.assertIsNotNone(self.db) + + def test_load(self): + f, l = sql_data.load( + self.db, 'SELECT * FROM iris.train LIMIT 3', 'class', None) + self.assertEqual(4, len(f.keys())) # 4 features + self.assertEqual(3, len(l)) # label column length + + def test_load_with_filter(self): + fs = ['sepal_length', 'petal_width'] + f, l = sql_data.load( + self.db, 'SELECT * FROM iris.train LIMIT 3', 'class', fs) + self.assertEqual(len(fs), len(f)) + self.assertEqual(3, len(l)) # label column length + + def test_feature_columns(self): + f, l = sql_data.load( + self.db, 'SELECT * FROM iris.train LIMIT 3', 'class', None) + c = sql_data.feature_columns(f) + self.assertEqual(4, len(c)) # 4 features + + +if __name__ == '__main__': + unittest.main() diff --git a/sqlflow/sql/sql.y b/sqlflow/sql/sql.y new file mode 100644 index 0000000..cbd8202 --- /dev/null +++ b/sqlflow/sql/sql.y @@ -0,0 +1,422 @@ +%{ + package sql + + import ( + "fmt" + "sort" + "strings" + "sync" + ) + + /* expr defines an expression as a Lisp list. If len(val)>0, + it is an atomic expression, in particular, NUMBER, IDENT, + or STRING, defined by typ and val; otherwise, it is a + Lisp S-expression. */ + type expr struct { + typ int + val string + sexp exprlist + } + + type exprlist []*expr + + /* construct an atomic expr */ + func atomic(typ int, val string) *expr { + return &expr{ + typ : typ, + val : val, + } + } + + /* construct a funcall expr */ + func funcall(name string, oprd exprlist) *expr { + return &expr{ + sexp : append(exprlist{atomic(IDENT, name)}, oprd...), + } + } + + /* construct a unary expr */ + func unary(typ int, op string, od1 *expr) *expr { + return &expr{ + sexp : append(exprlist{atomic(typ, op)}, od1), + } + } + + /* construct a binary expr */ + func binary(typ int, od1 *expr, op string, od2 *expr) *expr { + return &expr{ + sexp : append(exprlist{atomic(typ, op)}, od1, od2), + } + } + + /* construct a variadic expr */ + func variadic(typ int, op string, ods exprlist) *expr { + return &expr{ + sexp : append(exprlist{atomic(typ, op)}, ods...), + } + } + + type extendedSelect struct { + extended bool + train bool + standardSelect + trainClause + predictClause + } + + type standardSelect struct { + fields []string + tables []string + where *expr + limit string + } + + type trainClause struct { + estimator string + attrs attrs + columns exprlist + label string + save string + } + + type attrs map[string]*expr + + type predictClause struct { + model string + into string + } + + var parseResult *extendedSelect + + func attrsUnion(as1, as2 attrs) attrs { + for k, v := range as2 { + if _, ok := as1[k]; ok { + log.Panicf("attr %q already specified", as2) + } + as1[k] = v + } + return as1 + } +%} + +%union { + val string /* NUMBER, IDENT, STRING, and keywords */ + flds []string + tbls []string + expr *expr + expl exprlist + atrs attrs + eslt extendedSelect + slct standardSelect + tran trainClause + infr predictClause +} + +%type select_stmt +%type select +%type train_clause +%type predict_clause +%type fields +%type tables +%type expr funcall column +%type exprlist pythonlist columns +%type attr +%type attrs + +%token SELECT FROM WHERE LIMIT TRAIN PREDICT WITH COLUMN LABEL USING INTO +%token IDENT NUMBER STRING + +%left AND OR +%left '>' '<' '=' GE LE +%left '+' '-' +%left '*' '/' '%' +%left NOT +%left POWER /* think about the example "NOT base ** -3" */ +%left UMINUS + +%% + +select_stmt +: select ';' { + parseResult = &extendedSelect{ + extended: false, + standardSelect: $1} + } +| select train_clause ';' { + parseResult = &extendedSelect{ + extended: true, + train: true, + standardSelect: $1, + trainClause: $2} + } +| select predict_clause ';' { + parseResult = &extendedSelect{ + extended: true, + train: false, + standardSelect: $1, + predictClause: $2} + } +; + +select +: SELECT fields { $$.fields = $2 } +| select FROM tables { $$.tables = $3 } +| select LIMIT NUMBER { $$.limit = $3 } +| select WHERE expr { $$.where = $3 } +; + +train_clause +: TRAIN IDENT WITH attrs COLUMN columns LABEL IDENT INTO IDENT { + $$.estimator = $2 + $$.attrs = $4 + $$.columns = $6 + $$.label = $8 + $$.save = $10 + } +; + +predict_clause +: PREDICT IDENT USING IDENT { + $$.into = $2 + $$.model = $4 +} +; + +fields +: '*' { $$ = append($$, $1) } +| IDENT { $$ = append($$, $1) } +| fields ',' IDENT { $$ = append($$, $3) } +; + +column +: '*' { $$ = atomic(IDENT, "*") } +| IDENT { $$ = atomic(IDENT, $1) } +| funcall { $$ = $1 } +; + +columns +: column { $$ = exprlist{$1} } +| columns ',' column { $$ = append($1, $3) } +; + +tables +: IDENT { $$ = []string{$1} } +| tables ',' IDENT { $$ = append($1, $3) } +; + +attr +: IDENT '=' expr { $$ = attrs{$1 : $3} } +; + +attrs +: attr { $$ = $1 } +| attrs ',' attr { $$ = attrsUnion($1, $3) } +; + +funcall +: IDENT '(' ')' { $$ = funcall($1, nil) } +| IDENT '(' exprlist ')' { $$ = funcall($1, $3) } +; + +exprlist +: expr { $$ = exprlist{$1} } +| exprlist ',' expr { $$ = append($1, $3) } +; + +pythonlist +: '[' ']' { $$ = nil } +| '[' exprlist ']' { $$ = $2 } +; + +expr +: NUMBER { $$ = atomic(NUMBER, $1) } +| IDENT { $$ = atomic(IDENT, $1) } +| STRING { $$ = atomic(STRING, $1) } +| pythonlist { $$ = variadic('[', "square", $1) } +| '(' expr ')' { $$ = unary('(', "paren", $2) } /* take '(' as the operator */ +| funcall { $$ = $1 } +| expr '+' expr { $$ = binary('+', $1, $2, $3) } +| expr '-' expr { $$ = binary('-', $1, $2, $3) } +| expr '*' expr { $$ = binary('*', $1, $2, $3) } +| expr '/' expr { $$ = binary('/', $1, $2, $3) } +| expr '%' expr { $$ = binary('%', $1, $2, $3) } +| expr '=' expr { $$ = binary('=', $1, $2, $3) } +| expr '<' expr { $$ = binary('<', $1, $2, $3) } +| expr '>' expr { $$ = binary('>', $1, $2, $3) } +| expr LE expr { $$ = binary(LE, $1, $2, $3) } +| expr GE expr { $$ = binary(GE, $1, $2, $3) } +| expr AND expr { $$ = binary(AND, $1, $2, $3) } +| expr OR expr { $$ = binary(OR, $1, $2, $3) } +| NOT expr %prec NOT { $$ = unary(NOT, $1, $2) } +| '-' expr %prec UMINUS { $$ = unary('-', $1, $2) } +; + +%% + +/* Like Lisp's builtin function cdr. */ +func (e *expr) cdr() (r []string) { + for i := 1; i < len(e.sexp); i++ { + r = append(r, e.sexp[i].String()) + } + return r +} + +func (e *expr) String() string { + if e.typ == 0 { /* a compound expression */ + switch e.sexp[0].typ { + case '+', '*', '/', '%', '=', '<', '>', LE, GE, AND, OR: + if len(e.sexp) != 3 { + log.Panicf("Expecting binary expression, got %.10q", e.sexp) + } + return fmt.Sprintf("%s %s %s", e.sexp[1], e.sexp[0].val, e.sexp[2]) + case '-': + switch len(e.sexp) { + case 2: + return fmt.Sprintf(" -%s", e.sexp[1]) + case 3: + return fmt.Sprintf("%s - %s", e.sexp[1], e.sexp[2]) + default: + log.Panicf("Expecting either unary or binary -, got %.10q", e.sexp) + } + case '(': + if len(e.sexp) != 2 { + log.Panicf("Expecting ( ) as unary operator, got %.10q", e.sexp) + } + return fmt.Sprintf("(%s)", e.sexp[1]) + case '[': + return "[" + strings.Join(e.cdr(), ", ") + "]" + case NOT: + return fmt.Sprintf("NOT %s", e.sexp[1]) + case IDENT: /* function call */ + return e.sexp[0].val + "(" + strings.Join(e.cdr(), ", ") + ")" + } + } else { + return fmt.Sprintf("%s", e.val) + } + + log.Panicf("Cannot print an unknown expression") + return "" +} + +func (s standardSelect) String() string { + r := "SELECT " + if len(s.fields) == 0 { + r += "*" + } else { + r += strings.Join(s.fields, ", ") + } + r += "\nFROM " + strings.Join(s.tables, ", ") + if s.where != nil { + r += fmt.Sprintf("\nWHERE %s", s.where) + } + if len(s.limit) > 0 { + r += fmt.Sprintf("\nLIMIT %s", s.limit) + } + return r + ";" +} + +func jsonString(s string) string { + return strings.Replace( + strings.Replace( + strings.Replace(s, "\n", "\\n", -1), + "\r", "\\r", -1), + "\"", "\\\"", -1) +} + +func (ats attrs) JSON() string { + ks := []string{} + for k := range ats { + ks = append(ks, k) + } + sort.Strings(ks) /* Remove the randomness of map traversal. */ + + for i, k := range ks { + ks[i] = fmt.Sprintf(`"%s": "%s"`, k, jsonString(ats[k].String())) + } + return "{\n" + strings.Join(ks, ",\n") + "\n}" +} + +func (el exprlist) JSON() string { + ks := []string{} + for _, e := range el { + ks = append(ks, "\"" + jsonString(e.String()) + "\"") + } + return "[\n" + strings.Join(ks, ",\n") + "\n]" +} + +func (s trainClause) JSON() string { + fmter := `{ +"estimator": "%s", +"attrs": %s, +"columns": %s, +"save": "%s" +}` + return fmt.Sprintf(fmter, s.estimator, s.attrs.JSON(), s.columns.JSON(), s.save) +} + +func (s predictClause) JSON() string { + fmter := `{ +"model":"%s" +}` + return fmt.Sprintf(fmter, "\"" + s.model + "\"") +} + +func (s extendedSelect) JSON() string { + bf := `{ +"extended": %t, +"train": %t, +"standardSelect": "%s" +}` + tf := `{ +"extended": %t, +"train": %t, +"standardSelect": "%s", +"trainClause": %s +}` + nf := `{ +"extended": %t, +"train": %t, +"standardSelect": "%s", +"predictClause": %s +}` + if s.extended { + if s.train { + return fmt.Sprintf(tf, s.extended, s.train, + jsonString(s.standardSelect.String()), s.trainClause.JSON()) + } else { + return fmt.Sprintf(nf, s.extended, s.train, + jsonString(s.standardSelect.String()), s.predictClause.JSON()) + } + } + return fmt.Sprintf(bf, s.extended, s.train, jsonString(s.standardSelect.String())) +} + +// sqlReentrantParser makes sqlParser, generated by goyacc and using a +// global variable parseResult to return the result, reentrant. +type sqlSyncParser struct { + pr sqlParser +} + +func newParser() *sqlSyncParser { + return &sqlSyncParser{sqlNewParser()} +} + +var mu sync.Mutex + +func (p *sqlSyncParser) Parse(s string) (r *extendedSelect, e error) { + defer func() { + if r := recover(); r != nil { + var ok bool + e, ok = r.(error) + if !ok { + e = fmt.Errorf("%v", r) + } + } + }() + + mu.Lock() + defer mu.Unlock() + + p.pr.Parse(newLexer(s)) + return parseResult, nil +} diff --git a/sqlflow/sql/test b/sqlflow/sql/test new file mode 100644 index 0000000..e69de29 diff --git a/sqlflow/sql/testdata/churn.sql b/sqlflow/sql/testdata/churn.sql new file mode 100644 index 0000000..4d25135 --- /dev/null +++ b/sqlflow/sql/testdata/churn.sql @@ -0,0 +1,143 @@ +DROP TABLE IF EXISTS churn.train; +CREATE TABLE churn.train ( + customerID VARCHAR(255), + gender VARCHAR(255), + SeniorCitizen TINYINT, + Partner VARCHAR(255), + Dependents VARCHAR(255), + tenure INT, + PhoneService VARCHAR(255), + MultipleLines VARCHAR(255), + InternetService VARCHAR(255), + OnlineSecurity VARCHAR(255), + OnlineBackup VARCHAR(255), + DeviceProtection VARCHAR(255), + TechSupport VARCHAR(255), + StreamingTV VARCHAR(255), + StreamingMovies VARCHAR(255), + Contract VARCHAR(255), + PaperlessBilling VARCHAR(255), + PaymentMethod VARCHAR(255), + MonthlyCharges FLOAT, + TotalCharges FLOAT, + Churn VARCHAR(255), + PRIMARY KEY ( customerID ) +); +INSERT INTO churn.train VALUES("7590-VHVEG","Female","0","Yes","No","1","No","No phone service","DSL","No","Yes","No","No","No","No","Month-to-month","Yes","Electronic check","29.85","29.85","No"); +INSERT INTO churn.train VALUES("5575-GNVDE","Male","0","No","No","34","Yes","No","DSL","Yes","No","Yes","No","No","No","One year","No","Mailed check","56.95","1889.5","No"); +INSERT INTO churn.train VALUES("3668-QPYBK","Male","0","No","No","2","Yes","No","DSL","Yes","Yes","No","No","No","No","Month-to-month","Yes","Mailed check","53.85","108.15","Yes"); +INSERT INTO churn.train VALUES("7795-CFOCW","Male","0","No","No","45","No","No phone service","DSL","Yes","No","Yes","Yes","No","No","One year","No","Bank transfer (automatic)","42.3","1840.75","No"); +INSERT INTO churn.train VALUES("9237-HQITU","Female","0","No","No","2","Yes","No","Fiber optic","No","No","No","No","No","No","Month-to-month","Yes","Electronic check","70.7","151.65","Yes"); +INSERT INTO churn.train VALUES("9305-CDSKC","Female","0","No","No","8","Yes","Yes","Fiber optic","No","No","Yes","No","Yes","Yes","Month-to-month","Yes","Electronic check","99.65","820.5","Yes"); +INSERT INTO churn.train VALUES("1452-KIOVK","Male","0","No","Yes","22","Yes","Yes","Fiber optic","No","Yes","No","No","Yes","No","Month-to-month","Yes","Credit card (automatic)","89.1","1949.4","No"); +INSERT INTO churn.train VALUES("6713-OKOMC","Female","0","No","No","10","No","No phone service","DSL","Yes","No","No","No","No","No","Month-to-month","No","Mailed check","29.75","301.9","No"); +INSERT INTO churn.train VALUES("7892-POOKP","Female","0","Yes","No","28","Yes","Yes","Fiber optic","No","No","Yes","Yes","Yes","Yes","Month-to-month","Yes","Electronic check","104.8","3046.05","Yes"); +INSERT INTO churn.train VALUES("6388-TABGU","Male","0","No","Yes","62","Yes","No","DSL","Yes","Yes","No","No","No","No","One year","No","Bank transfer (automatic)","56.15","3487.95","No"); +INSERT INTO churn.train VALUES("9763-GRSKD","Male","0","Yes","Yes","13","Yes","No","DSL","Yes","No","No","No","No","No","Month-to-month","Yes","Mailed check","49.95","587.45","No"); +INSERT INTO churn.train VALUES("7469-LKBCI","Male","0","No","No","16","Yes","No","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","Two year","No","Credit card (automatic)","18.95","326.8","No"); +INSERT INTO churn.train VALUES("8091-TTVAX","Male","0","Yes","No","58","Yes","Yes","Fiber optic","No","No","Yes","No","Yes","Yes","One year","No","Credit card (automatic)","100.35","5681.1","No"); +INSERT INTO churn.train VALUES("0280-XJGEX","Male","0","No","No","49","Yes","Yes","Fiber optic","No","Yes","Yes","No","Yes","Yes","Month-to-month","Yes","Bank transfer (automatic)","103.7","5036.3","Yes"); +INSERT INTO churn.train VALUES("5129-JLPIS","Male","0","No","No","25","Yes","No","Fiber optic","Yes","No","Yes","Yes","Yes","Yes","Month-to-month","Yes","Electronic check","105.5","2686.05","No"); +INSERT INTO churn.train VALUES("3655-SNQYZ","Female","0","Yes","Yes","69","Yes","Yes","Fiber optic","Yes","Yes","Yes","Yes","Yes","Yes","Two year","No","Credit card (automatic)","113.25","7895.15","No"); +INSERT INTO churn.train VALUES("8191-XWSZG","Female","0","No","No","52","Yes","No","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","One year","No","Mailed check","20.65","1022.95","No"); +INSERT INTO churn.train VALUES("9959-WOFKT","Male","0","No","Yes","71","Yes","Yes","Fiber optic","Yes","No","Yes","No","Yes","Yes","Two year","No","Bank transfer (automatic)","106.7","7382.25","No"); +INSERT INTO churn.train VALUES("4190-MFLUW","Female","0","Yes","Yes","10","Yes","No","DSL","No","No","Yes","Yes","No","No","Month-to-month","No","Credit card (automatic)","55.2","528.35","Yes"); +INSERT INTO churn.train VALUES("4183-MYFRB","Female","0","No","No","21","Yes","No","Fiber optic","No","Yes","Yes","No","No","Yes","Month-to-month","Yes","Electronic check","90.05","1862.9","No"); +INSERT INTO churn.train VALUES("8779-QRDMV","Male","1","No","No","1","No","No phone service","DSL","No","No","Yes","No","No","Yes","Month-to-month","Yes","Electronic check","39.65","39.65","Yes"); +INSERT INTO churn.train VALUES("1680-VDCWW","Male","0","Yes","No","12","Yes","No","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","One year","No","Bank transfer (automatic)","19.8","202.25","No"); +INSERT INTO churn.train VALUES("1066-JKSGK","Male","0","No","No","1","Yes","No","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","Month-to-month","No","Mailed check","20.15","20.15","Yes"); +INSERT INTO churn.train VALUES("3638-WEABW","Female","0","Yes","No","58","Yes","Yes","DSL","No","Yes","No","Yes","No","No","Two year","Yes","Credit card (automatic)","59.9","3505.1","No"); +INSERT INTO churn.train VALUES("6322-HRPFA","Male","0","Yes","Yes","49","Yes","No","DSL","Yes","Yes","No","Yes","No","No","Month-to-month","No","Credit card (automatic)","59.6","2970.3","No"); +INSERT INTO churn.train VALUES("6865-JZNKO","Female","0","No","No","30","Yes","No","DSL","Yes","Yes","No","No","No","No","Month-to-month","Yes","Bank transfer (automatic)","55.3","1530.6","No"); +INSERT INTO churn.train VALUES("6467-CHFZW","Male","0","Yes","Yes","47","Yes","Yes","Fiber optic","No","Yes","No","No","Yes","Yes","Month-to-month","Yes","Electronic check","99.35","4749.15","Yes"); +INSERT INTO churn.train VALUES("8665-UTDHZ","Male","0","Yes","Yes","1","No","No phone service","DSL","No","Yes","No","No","No","No","Month-to-month","No","Electronic check","30.2","30.2","Yes"); +INSERT INTO churn.train VALUES("5248-YGIJN","Male","0","Yes","No","72","Yes","Yes","DSL","Yes","Yes","Yes","Yes","Yes","Yes","Two year","Yes","Credit card (automatic)","90.25","6369.45","No"); +INSERT INTO churn.train VALUES("8773-HHUOZ","Female","0","No","Yes","17","Yes","No","DSL","No","No","No","No","Yes","Yes","Month-to-month","Yes","Mailed check","64.7","1093.1","Yes"); +INSERT INTO churn.train VALUES("3841-NFECX","Female","1","Yes","No","71","Yes","Yes","Fiber optic","Yes","Yes","Yes","Yes","No","No","Two year","Yes","Credit card (automatic)","96.35","6766.95","No"); +INSERT INTO churn.train VALUES("4929-XIHVW","Male","1","Yes","No","2","Yes","No","Fiber optic","No","No","Yes","No","Yes","Yes","Month-to-month","Yes","Credit card (automatic)","95.5","181.65","No"); +INSERT INTO churn.train VALUES("6827-IEAUQ","Female","0","Yes","Yes","27","Yes","No","DSL","Yes","Yes","Yes","Yes","No","No","One year","No","Mailed check","66.15","1874.45","No"); +INSERT INTO churn.train VALUES("7310-EGVHZ","Male","0","No","No","1","Yes","No","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","Month-to-month","No","Bank transfer (automatic)","20.2","20.2","No"); +INSERT INTO churn.train VALUES("3413-BMNZE","Male","1","No","No","1","Yes","No","DSL","No","No","No","No","No","No","Month-to-month","No","Bank transfer (automatic)","45.25","45.25","No"); +INSERT INTO churn.train VALUES("6234-RAAPL","Female","0","Yes","Yes","72","Yes","Yes","Fiber optic","Yes","Yes","No","Yes","Yes","No","Two year","No","Bank transfer (automatic)","99.9","7251.7","No"); +INSERT INTO churn.train VALUES("6047-YHPVI","Male","0","No","No","5","Yes","No","Fiber optic","No","No","No","No","No","No","Month-to-month","Yes","Electronic check","69.7","316.9","Yes"); +INSERT INTO churn.train VALUES("6572-ADKRS","Female","0","No","No","46","Yes","No","Fiber optic","No","No","Yes","No","No","No","Month-to-month","Yes","Credit card (automatic)","74.8","3548.3","No"); +INSERT INTO churn.train VALUES("5380-WJKOV","Male","0","No","No","34","Yes","Yes","Fiber optic","No","Yes","Yes","No","Yes","Yes","Month-to-month","Yes","Electronic check","106.35","3549.25","Yes"); +INSERT INTO churn.train VALUES("8168-UQWWF","Female","0","No","No","11","Yes","Yes","Fiber optic","No","No","Yes","No","Yes","Yes","Month-to-month","Yes","Bank transfer (automatic)","97.85","1105.4","Yes"); +INSERT INTO churn.train VALUES("8865-TNMNX","Male","0","Yes","Yes","10","Yes","No","DSL","No","Yes","No","No","No","No","One year","No","Mailed check","49.55","475.7","No"); +INSERT INTO churn.train VALUES("9489-DEDVP","Female","0","Yes","Yes","70","Yes","Yes","DSL","Yes","Yes","No","No","Yes","No","Two year","Yes","Credit card (automatic)","69.2","4872.35","No"); +INSERT INTO churn.train VALUES("9867-JCZSP","Female","0","Yes","Yes","17","Yes","No","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","One year","No","Mailed check","20.75","418.25","No"); +INSERT INTO churn.train VALUES("4671-VJLCL","Female","0","No","No","63","Yes","Yes","DSL","Yes","Yes","Yes","Yes","Yes","No","Two year","Yes","Credit card (automatic)","79.85","4861.45","No"); +INSERT INTO churn.train VALUES("4080-IIARD","Female","0","Yes","No","13","Yes","Yes","DSL","Yes","Yes","No","Yes","Yes","No","Month-to-month","Yes","Electronic check","76.2","981.45","No"); +INSERT INTO churn.train VALUES("3714-NTNFO","Female","0","No","No","49","Yes","Yes","Fiber optic","No","No","No","No","No","Yes","Month-to-month","Yes","Electronic check","84.5","3906.7","No"); +INSERT INTO churn.train VALUES("5948-UJZLF","Male","0","No","No","2","Yes","No","DSL","No","Yes","No","No","No","No","Month-to-month","No","Mailed check","49.25","97","No"); +INSERT INTO churn.train VALUES("7760-OYPDY","Female","0","No","No","2","Yes","No","Fiber optic","No","No","No","No","Yes","No","Month-to-month","Yes","Electronic check","80.65","144.15","Yes"); +INSERT INTO churn.train VALUES("7639-LIAYI","Male","0","No","No","52","Yes","Yes","DSL","Yes","No","No","Yes","Yes","Yes","Two year","Yes","Credit card (automatic)","79.75","4217.8","No"); +INSERT INTO churn.train VALUES("2954-PIBKO","Female","0","Yes","Yes","69","Yes","Yes","DSL","Yes","No","Yes","Yes","No","No","Two year","Yes","Credit card (automatic)","64.15","4254.1","No"); +INSERT INTO churn.train VALUES("8012-SOUDQ","Female","1","No","No","43","Yes","Yes","Fiber optic","No","Yes","No","No","Yes","No","Month-to-month","Yes","Electronic check","90.25","3838.75","No"); +INSERT INTO churn.train VALUES("9420-LOJKX","Female","0","No","No","15","Yes","No","Fiber optic","Yes","Yes","No","No","Yes","Yes","Month-to-month","Yes","Credit card (automatic)","99.1","1426.4","Yes"); +INSERT INTO churn.train VALUES("6575-SUVOI","Female","1","Yes","No","25","Yes","Yes","DSL","Yes","No","No","Yes","Yes","No","Month-to-month","Yes","Credit card (automatic)","69.5","1752.65","No"); +INSERT INTO churn.train VALUES("7495-OOKFY","Female","1","Yes","No","8","Yes","Yes","Fiber optic","No","Yes","No","No","No","No","Month-to-month","Yes","Credit card (automatic)","80.65","633.3","Yes"); +INSERT INTO churn.train VALUES("4667-QONEA","Female","1","Yes","Yes","60","Yes","No","DSL","Yes","Yes","Yes","Yes","No","Yes","One year","Yes","Credit card (automatic)","74.85","4456.35","No"); +INSERT INTO churn.train VALUES("1658-BYGOY","Male","1","No","No","18","Yes","Yes","Fiber optic","No","No","No","No","Yes","Yes","Month-to-month","Yes","Electronic check","95.45","1752.55","Yes"); +INSERT INTO churn.train VALUES("8769-KKTPH","Female","0","Yes","Yes","63","Yes","Yes","Fiber optic","Yes","No","No","No","Yes","Yes","One year","Yes","Credit card (automatic)","99.65","6311.2","No"); +INSERT INTO churn.train VALUES("5067-XJQFU","Male","1","Yes","Yes","66","Yes","Yes","Fiber optic","No","Yes","Yes","Yes","Yes","Yes","One year","Yes","Electronic check","108.45","7076.35","No"); +INSERT INTO churn.train VALUES("3957-SQXML","Female","0","Yes","Yes","34","Yes","Yes","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","Two year","No","Credit card (automatic)","24.95","894.3","No"); +INSERT INTO churn.train VALUES("5954-BDFSG","Female","0","No","No","72","Yes","Yes","Fiber optic","No","No","Yes","Yes","Yes","Yes","Two year","Yes","Credit card (automatic)","107.5","7853.7","No"); +INSERT INTO churn.train VALUES("0434-CSFON","Female","0","Yes","No","47","Yes","Yes","Fiber optic","No","No","Yes","No","Yes","Yes","Month-to-month","Yes","Electronic check","100.5","4707.1","No"); +INSERT INTO churn.train VALUES("1215-FIGMP","Male","0","No","No","60","Yes","Yes","Fiber optic","No","Yes","No","No","Yes","No","Month-to-month","Yes","Bank transfer (automatic)","89.9","5450.7","No"); +INSERT INTO churn.train VALUES("0526-SXDJP","Male","0","Yes","No","72","No","No phone service","DSL","Yes","Yes","Yes","No","No","No","Two year","No","Bank transfer (automatic)","42.1","2962","No"); +INSERT INTO churn.train VALUES("0557-ASKVU","Female","0","Yes","Yes","18","Yes","No","DSL","No","No","Yes","Yes","No","No","One year","Yes","Credit card (automatic)","54.4","957.1","No"); +INSERT INTO churn.train VALUES("5698-BQJOH","Female","0","No","No","9","Yes","Yes","Fiber optic","No","No","No","No","Yes","Yes","Month-to-month","No","Electronic check","94.4","857.25","Yes"); +INSERT INTO churn.train VALUES("5122-CYFXA","Female","0","No","No","3","Yes","No","DSL","No","Yes","No","Yes","Yes","Yes","Month-to-month","Yes","Electronic check","75.3","244.1","No"); +INSERT INTO churn.train VALUES("8627-ZYGSZ","Male","0","Yes","No","47","Yes","Yes","Fiber optic","No","Yes","No","No","No","No","One year","Yes","Electronic check","78.9","3650.35","No"); +INSERT INTO churn.train VALUES("3410-YOQBQ","Female","0","No","No","31","Yes","No","DSL","No","Yes","Yes","Yes","Yes","Yes","Two year","No","Mailed check","79.2","2497.2","No"); +INSERT INTO churn.train VALUES("3170-NMYVV","Female","0","Yes","Yes","50","Yes","No","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","Two year","No","Bank transfer (automatic)","20.15","930.9","No"); +INSERT INTO churn.train VALUES("7410-OIEDU","Male","0","No","No","10","Yes","No","Fiber optic","Yes","No","Yes","No","No","No","Month-to-month","Yes","Mailed check","79.85","887.35","No"); +INSERT INTO churn.train VALUES("2273-QCKXA","Male","0","No","No","1","Yes","No","DSL","No","No","No","Yes","No","No","Month-to-month","No","Mailed check","49.05","49.05","No"); +INSERT INTO churn.train VALUES("0731-EBJQB","Female","0","Yes","Yes","52","Yes","No","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","One year","Yes","Electronic check","20.4","1090.65","No"); +INSERT INTO churn.train VALUES("1891-QRQSA","Male","1","Yes","Yes","64","Yes","Yes","Fiber optic","Yes","No","Yes","Yes","Yes","Yes","Two year","Yes","Bank transfer (automatic)","111.6","7099","No"); +INSERT INTO churn.train VALUES("8028-PNXHQ","Male","0","Yes","Yes","62","Yes","Yes","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","Two year","Yes","Bank transfer (automatic)","24.25","1424.6","No"); +INSERT INTO churn.train VALUES("5630-AHZIL","Female","0","No","Yes","3","Yes","No","DSL","Yes","No","No","Yes","No","Yes","Month-to-month","Yes","Bank transfer (automatic)","64.5","177.4","No"); +INSERT INTO churn.train VALUES("2673-CXQEU","Female","1","No","No","56","Yes","Yes","Fiber optic","Yes","Yes","Yes","No","Yes","Yes","One year","No","Electronic check","110.5","6139.5","No"); +INSERT INTO churn.train VALUES("6416-JNVRK","Female","0","No","No","46","Yes","No","DSL","No","No","No","No","No","Yes","One year","No","Credit card (automatic)","55.65","2688.85","No"); +INSERT INTO churn.train VALUES("5590-ZSKRV","Female","0","Yes","Yes","8","Yes","No","DSL","Yes","Yes","No","No","No","No","Month-to-month","No","Mailed check","54.65","482.25","No"); +INSERT INTO churn.train VALUES("0191-ZHSKZ","Male","1","No","No","30","Yes","No","DSL","Yes","Yes","No","No","Yes","Yes","Month-to-month","Yes","Electronic check","74.75","2111.3","No"); +INSERT INTO churn.train VALUES("3887-PBQAO","Female","0","Yes","Yes","45","Yes","Yes","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","One year","Yes","Credit card (automatic)","25.9","1216.6","No"); +INSERT INTO churn.train VALUES("5919-TMRGD","Female","0","No","Yes","1","Yes","No","Fiber optic","No","No","No","No","Yes","No","Month-to-month","Yes","Electronic check","79.35","79.35","Yes"); +INSERT INTO churn.train VALUES("8108-UXRQN","Female","0","Yes","Yes","11","No","No phone service","DSL","Yes","No","No","No","Yes","Yes","Month-to-month","No","Electronic check","50.55","565.35","No"); + +DROP TABLE IF EXISTS churn.test; +CREATE TABLE churn.test ( + customerID VARCHAR(255), + gender VARCHAR(255), + SeniorCitizen TINYINT, + Partner VARCHAR(255), + Dependents VARCHAR(255), + tenure INT, + PhoneService VARCHAR(255), + MultipleLines VARCHAR(255), + InternetService VARCHAR(255), + OnlineSecurity VARCHAR(255), + OnlineBackup VARCHAR(255), + DeviceProtection VARCHAR(255), + TechSupport VARCHAR(255), + StreamingTV VARCHAR(255), + StreamingMovies VARCHAR(255), + Contract VARCHAR(255), + PaperlessBilling VARCHAR(255), + PaymentMethod VARCHAR(255), + MonthlyCharges FLOAT, + TotalCharges FLOAT, + Churn VARCHAR(255), + PRIMARY KEY ( customerID ) +); +INSERT INTO churn.test VALUES("9191-MYQKX","Female","0","Yes","No","7","Yes","No","Fiber optic","No","No","Yes","No","No","No","Month-to-month","Yes","Bank transfer (automatic)","75.15","496.9","Yes"); +INSERT INTO churn.test VALUES("9919-YLNNG","Female","0","No","No","42","Yes","No","Fiber optic","No","Yes","Yes","Yes","Yes","Yes","Month-to-month","Yes","Bank transfer (automatic)","103.8","4327.5","No"); +INSERT INTO churn.test VALUES("0318-ZOPWS","Female","0","Yes","No","49","Yes","No","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","Two year","Yes","Bank transfer (automatic)","20.15","973.35","No"); +INSERT INTO churn.test VALUES("4445-ZJNMU","Male","0","No","No","9","Yes","Yes","Fiber optic","No","Yes","No","No","Yes","Yes","Month-to-month","Yes","Credit card (automatic)","99.3","918.75","No"); +INSERT INTO churn.test VALUES("4808-YNLEU","Female","0","Yes","No","35","Yes","No","DSL","Yes","No","No","No","Yes","No","One year","Yes","Bank transfer (automatic)","62.15","2215.45","No"); +INSERT INTO churn.test VALUES("1862-QRWPE","Female","0","Yes","Yes","48","Yes","No","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","Two year","No","Bank transfer (automatic)","20.65","1057","No"); +INSERT INTO churn.test VALUES("2796-NNUFI","Female","0","Yes","Yes","46","Yes","No","No","No internet service","No internet service","No internet service","No internet service","No internet service","No internet service","Two year","Yes","Mailed check","19.95","927.1","No"); +INSERT INTO churn.test VALUES("3016-KSVCP","Male","0","Yes","No","29","No","No phone service","DSL","No","No","No","No","Yes","No","Month-to-month","No","Mailed check","33.75","1009.25","No"); +INSERT INTO churn.test VALUES("4767-HZZHQ","Male","0","Yes","Yes","30","Yes","No","Fiber optic","No","Yes","Yes","No","No","No","Month-to-month","No","Bank transfer (automatic)","82.05","2570.2","No"); +INSERT INTO churn.test VALUES("2424-WVHPL","Male","1","No","No","1","Yes","No","Fiber optic","No","No","No","Yes","No","No","Month-to-month","No","Electronic check","74.7","74.7","No"); diff --git a/sqlflow/sql/testdata/iris.sql b/sqlflow/sql/testdata/iris.sql new file mode 100644 index 0000000..fdef1ba --- /dev/null +++ b/sqlflow/sql/testdata/iris.sql @@ -0,0 +1,143 @@ +DROP TABLE IF EXISTS iris.train; +CREATE TABLE iris.train ( + sepal_length float, + sepal_width float, + petal_length float, + petal_width float, + class int); +INSERT INTO iris.train VALUES(6.4,2.8,5.6,2.2,2); +INSERT INTO iris.train VALUES(5.0,2.3,3.3,1.0,1); +INSERT INTO iris.train VALUES(4.9,2.5,4.5,1.7,2); +INSERT INTO iris.train VALUES(4.9,3.1,1.5,0.1,0); +INSERT INTO iris.train VALUES(5.7,3.8,1.7,0.3,0); +INSERT INTO iris.train VALUES(4.4,3.2,1.3,0.2,0); +INSERT INTO iris.train VALUES(5.4,3.4,1.5,0.4,0); +INSERT INTO iris.train VALUES(6.9,3.1,5.1,2.3,2); +INSERT INTO iris.train VALUES(6.7,3.1,4.4,1.4,1); +INSERT INTO iris.train VALUES(5.1,3.7,1.5,0.4,0); +INSERT INTO iris.train VALUES(5.2,2.7,3.9,1.4,1); +INSERT INTO iris.train VALUES(6.9,3.1,4.9,1.5,1); +INSERT INTO iris.train VALUES(5.8,4.0,1.2,0.2,0); +INSERT INTO iris.train VALUES(5.4,3.9,1.7,0.4,0); +INSERT INTO iris.train VALUES(7.7,3.8,6.7,2.2,2); +INSERT INTO iris.train VALUES(6.3,3.3,4.7,1.6,1); +INSERT INTO iris.train VALUES(6.8,3.2,5.9,2.3,2); +INSERT INTO iris.train VALUES(7.6,3.0,6.6,2.1,2); +INSERT INTO iris.train VALUES(6.4,3.2,5.3,2.3,2); +INSERT INTO iris.train VALUES(5.7,4.4,1.5,0.4,0); +INSERT INTO iris.train VALUES(6.7,3.3,5.7,2.1,2); +INSERT INTO iris.train VALUES(6.4,2.8,5.6,2.1,2); +INSERT INTO iris.train VALUES(5.4,3.9,1.3,0.4,0); +INSERT INTO iris.train VALUES(6.1,2.6,5.6,1.4,2); +INSERT INTO iris.train VALUES(7.2,3.0,5.8,1.6,2); +INSERT INTO iris.train VALUES(5.2,3.5,1.5,0.2,0); +INSERT INTO iris.train VALUES(5.8,2.6,4.0,1.2,1); +INSERT INTO iris.train VALUES(5.9,3.0,5.1,1.8,2); +INSERT INTO iris.train VALUES(5.4,3.0,4.5,1.5,1); +INSERT INTO iris.train VALUES(6.7,3.0,5.0,1.7,1); +INSERT INTO iris.train VALUES(6.3,2.3,4.4,1.3,1); +INSERT INTO iris.train VALUES(5.1,2.5,3.0,1.1,1); +INSERT INTO iris.train VALUES(6.4,3.2,4.5,1.5,1); +INSERT INTO iris.train VALUES(6.8,3.0,5.5,2.1,2); +INSERT INTO iris.train VALUES(6.2,2.8,4.8,1.8,2); +INSERT INTO iris.train VALUES(6.9,3.2,5.7,2.3,2); +INSERT INTO iris.train VALUES(6.5,3.2,5.1,2.0,2); +INSERT INTO iris.train VALUES(5.8,2.8,5.1,2.4,2); +INSERT INTO iris.train VALUES(5.1,3.8,1.5,0.3,0); +INSERT INTO iris.train VALUES(4.8,3.0,1.4,0.3,0); +INSERT INTO iris.train VALUES(7.9,3.8,6.4,2.0,2); +INSERT INTO iris.train VALUES(5.8,2.7,5.1,1.9,2); +INSERT INTO iris.train VALUES(6.7,3.0,5.2,2.3,2); +INSERT INTO iris.train VALUES(5.1,3.8,1.9,0.4,0); +INSERT INTO iris.train VALUES(4.7,3.2,1.6,0.2,0); +INSERT INTO iris.train VALUES(6.0,2.2,5.0,1.5,2); +INSERT INTO iris.train VALUES(4.8,3.4,1.6,0.2,0); +INSERT INTO iris.train VALUES(7.7,2.6,6.9,2.3,2); +INSERT INTO iris.train VALUES(4.6,3.6,1.0,0.2,0); +INSERT INTO iris.train VALUES(7.2,3.2,6.0,1.8,2); +INSERT INTO iris.train VALUES(5.0,3.3,1.4,0.2,0); +INSERT INTO iris.train VALUES(6.6,3.0,4.4,1.4,1); +INSERT INTO iris.train VALUES(6.1,2.8,4.0,1.3,1); +INSERT INTO iris.train VALUES(5.0,3.2,1.2,0.2,0); +INSERT INTO iris.train VALUES(7.0,3.2,4.7,1.4,1); +INSERT INTO iris.train VALUES(6.0,3.0,4.8,1.8,2); +INSERT INTO iris.train VALUES(7.4,2.8,6.1,1.9,2); +INSERT INTO iris.train VALUES(5.8,2.7,5.1,1.9,2); +INSERT INTO iris.train VALUES(6.2,3.4,5.4,2.3,2); +INSERT INTO iris.train VALUES(5.0,2.0,3.5,1.0,1); +INSERT INTO iris.train VALUES(5.6,2.5,3.9,1.1,1); +INSERT INTO iris.train VALUES(6.7,3.1,5.6,2.4,2); +INSERT INTO iris.train VALUES(6.3,2.5,5.0,1.9,2); +INSERT INTO iris.train VALUES(6.4,3.1,5.5,1.8,2); +INSERT INTO iris.train VALUES(6.2,2.2,4.5,1.5,1); +INSERT INTO iris.train VALUES(7.3,2.9,6.3,1.8,2); +INSERT INTO iris.train VALUES(4.4,3.0,1.3,0.2,0); +INSERT INTO iris.train VALUES(7.2,3.6,6.1,2.5,2); +INSERT INTO iris.train VALUES(6.5,3.0,5.5,1.8,2); +INSERT INTO iris.train VALUES(5.0,3.4,1.5,0.2,0); +INSERT INTO iris.train VALUES(4.7,3.2,1.3,0.2,0); +INSERT INTO iris.train VALUES(6.6,2.9,4.6,1.3,1); +INSERT INTO iris.train VALUES(5.5,3.5,1.3,0.2,0); +INSERT INTO iris.train VALUES(7.7,3.0,6.1,2.3,2); +INSERT INTO iris.train VALUES(6.1,3.0,4.9,1.8,2); +INSERT INTO iris.train VALUES(4.9,3.1,1.5,0.1,0); +INSERT INTO iris.train VALUES(5.5,2.4,3.8,1.1,1); +INSERT INTO iris.train VALUES(5.7,2.9,4.2,1.3,1); +INSERT INTO iris.train VALUES(6.0,2.9,4.5,1.5,1); +INSERT INTO iris.train VALUES(6.4,2.7,5.3,1.9,2); +INSERT INTO iris.train VALUES(5.4,3.7,1.5,0.2,0); +INSERT INTO iris.train VALUES(6.1,2.9,4.7,1.4,1); +INSERT INTO iris.train VALUES(6.5,2.8,4.6,1.5,1); +INSERT INTO iris.train VALUES(5.6,2.7,4.2,1.3,1); +INSERT INTO iris.train VALUES(6.3,3.4,5.6,2.4,2); +INSERT INTO iris.train VALUES(4.9,3.1,1.5,0.1,0); +INSERT INTO iris.train VALUES(6.8,2.8,4.8,1.4,1); +INSERT INTO iris.train VALUES(5.7,2.8,4.5,1.3,1); +INSERT INTO iris.train VALUES(6.0,2.7,5.1,1.6,1); +INSERT INTO iris.train VALUES(5.0,3.5,1.3,0.3,0); +INSERT INTO iris.train VALUES(6.5,3.0,5.2,2.0,2); +INSERT INTO iris.train VALUES(6.1,2.8,4.7,1.2,1); +INSERT INTO iris.train VALUES(5.1,3.5,1.4,0.3,0); +INSERT INTO iris.train VALUES(4.6,3.1,1.5,0.2,0); +INSERT INTO iris.train VALUES(6.5,3.0,5.8,2.2,2); +INSERT INTO iris.train VALUES(4.6,3.4,1.4,0.3,0); +INSERT INTO iris.train VALUES(4.6,3.2,1.4,0.2,0); +INSERT INTO iris.train VALUES(7.7,2.8,6.7,2.0,2); +INSERT INTO iris.train VALUES(5.9,3.2,4.8,1.8,1); +INSERT INTO iris.train VALUES(5.1,3.8,1.6,0.2,0); +INSERT INTO iris.train VALUES(4.9,3.0,1.4,0.2,0); +INSERT INTO iris.train VALUES(4.9,2.4,3.3,1.0,1); +INSERT INTO iris.train VALUES(4.5,2.3,1.3,0.3,0); +INSERT INTO iris.train VALUES(5.8,2.7,4.1,1.0,1); +INSERT INTO iris.train VALUES(5.0,3.4,1.6,0.4,0); +INSERT INTO iris.train VALUES(5.2,3.4,1.4,0.2,0); +INSERT INTO iris.train VALUES(5.3,3.7,1.5,0.2,0); +INSERT INTO iris.train VALUES(5.0,3.6,1.4,0.2,0); +INSERT INTO iris.train VALUES(5.6,2.9,3.6,1.3,1); +INSERT INTO iris.train VALUES(4.8,3.1,1.6,0.2,0); + +DROP TABLE IF EXISTS iris.test; +CREATE TABLE iris.test ( + sepal_length float, + sepal_width float, + petal_length float, + petal_width float, + class int); +INSERT INTO iris.test VALUES(6.3,2.7,4.9,1.8,2); +INSERT INTO iris.test VALUES(5.7,2.8,4.1,1.3,1); +INSERT INTO iris.test VALUES(5.0,3.0,1.6,0.2,0); +INSERT INTO iris.test VALUES(6.3,3.3,6.0,2.5,2); +INSERT INTO iris.test VALUES(5.0,3.5,1.6,0.6,0); +INSERT INTO iris.test VALUES(5.5,2.6,4.4,1.2,1); +INSERT INTO iris.test VALUES(5.7,3.0,4.2,1.2,1); +INSERT INTO iris.test VALUES(4.4,2.9,1.4,0.2,0); +INSERT INTO iris.test VALUES(4.8,3.0,1.4,0.1,0); +INSERT INTO iris.test VALUES(5.5,2.4,3.7,1.0,1); + +DROP TABLE IF EXISTS iris.iris_empty; +CREATE TABLE iris.iris_empty ( + sepal_length float, + sepal_width float, + petal_length float, + petal_width float, + class int); diff --git a/sqlflow/sql/verifier.go b/sqlflow/sql/verifier.go new file mode 100644 index 0000000..5dc93ed --- /dev/null +++ b/sqlflow/sql/verifier.go @@ -0,0 +1,164 @@ +package sql + +import ( + "fmt" + "strings" +) + +// fieldTypes[field][table]type. For more information, please check +// verifier_test.go. +type fieldTypes map[string]map[string]string + +// verify checks the following: +// +// 1. The standard SELECT part is syntactically and logically legal. +// +// 2. TODO(yi): The COLUMN clause refers to only fields in the SELECT +// clause. Please be aware that both SELECT and COLUMN might have +// star '*'. +// +// It returns a fieldTypes describing types of fields in SELECT. +func verify(slct *extendedSelect, db *DB) (ft fieldTypes, e error) { + if e := dryRunSelect(slct, db); e != nil { + return nil, e + } + return describeTables(slct, db) +} + +func dryRunSelect(slct *extendedSelect, db *DB) error { + oldLimit := slct.standardSelect.limit + defer func() { + slct.standardSelect.limit = oldLimit + }() + + slct.standardSelect.limit = "1" + stmt := slct.standardSelect.String() + rows, e := db.Query(stmt) + if e != nil { + return fmt.Errorf("dryRunSelect failed executing %s: %q", stmt, e) + } + defer rows.Close() + + return rows.Err() +} + +func (ft fieldTypes) get(ident string) (string, bool) { + tbl, fld := decomp(ident) + tbls, ok := ft[fld] + if !ok { + return "", false + } + if len(tbl) == 0 && len(tbls) == 1 { + for _, typ := range tbls { + return typ, true + } + } + typ, ok := tbls[tbl] + return typ, ok +} + +// decomp returns the table name and field name in the given +// identifier: t.f=>(t,f), db.t.f=>(db.t,f), f=>("",f). +func decomp(ident string) (tbl string, fld string) { + s := strings.Split(ident, ".") + return strings.Join(s[:len(s)-1], "."), s[len(s)-1] +} + +// Retrieve the type of fields mentioned in SELECT. +func describeTables(slct *extendedSelect, db *DB) (ft fieldTypes, e error) { + ft = indexSelectFields(slct) + hasStar := len(ft) == 0 + for _, tn := range slct.tables { + slct := "SELECT * from " + tn + " limit 1" + rows, err := db.Query(slct) + if err != nil { + return nil, err + } + defer rows.Close() + + cols, err := rows.Columns() + if err != nil { + return nil, err + } + + if !rows.Next() { + return nil, fmt.Errorf("table is Empty. table name: %s", tn) + } + + if rows.Err() != nil { + return nil, e + } + + columnTypes, err := rows.ColumnTypes() + if err != nil { + return nil, err + } + for i, ct := range columnTypes { + fld := cols[i] + typeName := ct.DatabaseTypeName() + + if hasStar { + if _, ok := ft[fld]; !ok { + ft[fld] = make(map[string]string) + } + ft[fld][tn] = typeName + } else { + if tbls, ok := ft[fld]; ok { + if len(tbls) == 0 { + tbls[tn] = typeName + } else if _, ok := tbls[tn]; ok { + tbls[tn] = typeName + } + } + } + } + + } + return ft, nil +} + +// Index fields in the SELECT clause. For `SELECT f`, returns {f:{}}. +// For `SELECT t.f`, returns {f:{t:1}}. For `SELECT t1.f, t2.f`, +// returns {f:{t1:1,t2:1}}. For `SELECT ... * ...`, returns {}. +func indexSelectFields(slct *extendedSelect) (ft fieldTypes) { + ft = make(fieldTypes) + for _, f := range slct.fields { + if f == "*" { + return fieldTypes{} + } + tbl, fld := decomp(f) + if _, ok := ft[fld]; !ok { + ft[fld] = make(map[string]string) + } + if len(tbl) > 0 { + ft[fld][tbl] = "" + } + } + return ft +} + +// Check train and pred clause uses has the same feature columns +// 1. every column field in the training clause is selected in the pred clause, and they are of the same type +func verifyColumnNameAndType(trainParsed, predParsed *extendedSelect, db *DB) error { + trainFields, e := verify(trainParsed, db) + if e != nil { + return e + } + + predFields, e := verify(predParsed, db) + if e != nil { + return e + } + + for _, c := range trainParsed.columns { + it, ok := predFields.get(c.val) + if !ok { + return fmt.Errorf("predFields doesn't contain column %s", c.val) + } + tt, _ := trainFields.get(c.val) + if it != tt { + return fmt.Errorf("field %s type dismatch %s(pred) vs %s(train)", c.val, it, tt) + } + } + return nil +} diff --git a/sqlflow/sql/verifier_test.go b/sqlflow/sql/verifier_test.go new file mode 100644 index 0000000..19b0d71 --- /dev/null +++ b/sqlflow/sql/verifier_test.go @@ -0,0 +1,113 @@ +package sql + +import ( + "testing" + + "github.com/stretchr/testify/assert" +) + +func TestDryRunSelect(t *testing.T) { + a := assert.New(t) + r, e := newParser().Parse(`SELECT * FROM churn.train LIMIT 10;`) + a.NoError(e) + a.Nil(dryRunSelect(r, testDB)) +} + +func TestDescribeTables(t *testing.T) { + a := assert.New(t) + r, e := newParser().Parse(`SELECT * FROM churn.train LIMIT 10;`) + a.NoError(e) + fts, e := describeTables(r, testDB) + a.NoError(e) + a.Equal(21, len(fts)) + + r, e = newParser().Parse(`SELECT Churn, churn.train.Partner,TotalCharges FROM churn.train LIMIT 10;`) + a.NoError(e) + fts, e = describeTables(r, testDB) + a.NoError(e) + a.Equal(3, len(fts)) + a.Contains([]string{"VARCHAR(255)", "VARCHAR"}, fts["Churn"]["churn.train"]) + a.Contains([]string{"VARCHAR(255)", "VARCHAR"}, fts["Partner"]["churn.train"]) + a.Equal("FLOAT", fts["TotalCharges"]["churn.train"]) +} + +func TestIndexSelectFields(t *testing.T) { + a := assert.New(t) + r, e := newParser().Parse(`SELECT * FROM churn.train LIMIT 10;`) + a.NoError(e) + f := indexSelectFields(r) + a.Equal(0, len(f)) + + r, e = newParser().Parse(`SELECT f FROM churn.train LIMIT 10;`) + a.NoError(e) + f = indexSelectFields(r) + a.Equal(1, len(f)) + a.Equal(map[string]string{}, f["f"]) + + r, e = newParser().Parse(`SELECT t1.f, t2.f, g FROM churn.train LIMIT 10;`) + a.NoError(e) + f = indexSelectFields(r) + a.Equal(2, len(f)) + a.Equal(map[string]string{}, f["g"]) + a.Equal("", f["f"]["t1"]) + a.Equal("", f["f"]["t2"]) +} + +func TestVerify(t *testing.T) { + a := assert.New(t) + r, e := newParser().Parse(`SELECT Churn, churn.train.Partner FROM churn.train LIMIT 10;`) + a.NoError(e) + fts, e := verify(r, testDB) + a.NoError(e) + a.Equal(2, len(fts)) + typ, ok := fts.get("Churn") + a.Equal(true, ok) + a.Contains([]string{"VARCHAR(255)", "VARCHAR"}, typ) + + typ, ok = fts.get("churn.train.Partner") + a.Equal(true, ok) + a.Contains([]string{"VARCHAR(255)", "VARCHAR"}, typ) + + _, ok = fts.get("churn.train.gender") + a.Equal(false, ok) + + _, ok = fts.get("gender") + a.Equal(false, ok) +} + +func TestVerifyColumnNameAndType(t *testing.T) { + a := assert.New(t) + trainParse, e := newParser().Parse(`SELECT gender, tenure, TotalCharges +FROM churn.train LIMIT 10 +TRAIN DNNClassifier +WITH + n_classes = 3, + hidden_units = [10, 20] +COLUMN gender, tenure, TotalCharges +LABEL class +INTO sqlflow_models.my_dnn_model;`) + a.NoError(e) + + predParse, e := newParser().Parse(`SELECT gender, tenure, TotalCharges +FROM churn.train LIMIT 10 +PREDICT iris.predict.class +USING sqlflow_models.my_dnn_model;`) + a.NoError(e) + a.NoError(verifyColumnNameAndType(trainParse, predParse, testDB)) + + predParse, e = newParser().Parse(`SELECT gender, tenure +FROM churn.train LIMIT 10 +PREDICT iris.predict.class +USING sqlflow_models.my_dnn_model;`) + a.NoError(e) + a.EqualError(verifyColumnNameAndType(trainParse, predParse, testDB), + "predFields doesn't contain column TotalCharges") +} + +func TestDescribeEmptyTables(t *testing.T) { + a := assert.New(t) + r, e := newParser().Parse(`SELECT * FROM iris.iris_empty LIMIT 10;`) + a.NoError(e) + _, e = describeTables(r, testDB) + a.EqualError(e, "table is Empty. table name: iris.iris_empty") +} diff --git a/sqlflow/sqlfs/.gitignore b/sqlflow/sqlfs/.gitignore new file mode 100644 index 0000000..e69de29 diff --git a/sqlflow/sqlfs/README.md b/sqlflow/sqlfs/README.md new file mode 100644 index 0000000..01b7e3a --- /dev/null +++ b/sqlflow/sqlfs/README.md @@ -0,0 +1,56 @@ +# `sqlfs` + +The package `sqlfs` provides an `io.ReadCloser` and an `io.WriteCloser` that treats a SQL database a filesystem, where each table in the database is like a file. The schema of the table is very simple -- it has only one column of BLOB type. All the rows consist the storage. + +`sqlfs` provides the following features. + +## Create a file + +To create a table named "hello" in a database "mydb" for writing, we can call `Create`. + +```go +f, e := sqlfs.Create(db, "mydb.hello") +f.Write([]byte("hello world!\n")) +f.Close() +``` + +where `db` comes from a call to `database/sql.Open`. + +## Append to a file + +```go +f, e := sqlfs.Append(db, "mydb.hello") +f.Write([]byte("hello world!\n") +f.Close() +``` + +## Read from a file + +```go +f, e := sqlfs.Open(db, "mydb.hello") +buf := make([]byte, 1024) +f.Read(buf) +f.Close() +``` + +## Remove a file + +```go +DropTable(db, "mydb.hello") +``` + +## Check if a file exists + +```go +HasTable(db, "mydb.hello") +``` + +## Other I/O operations + +Feel free to use standard packages `io`, `ioutil`, etc with `sqlfs`. For example, we can call `io.Copy` to copy everything from the standard input to a table. + +```go +f, e := sqlfs.Create(db, "mydb.hello") +io.Copy(f, os.Stdin) +f.Close() +``` diff --git a/sqlflow/sqlfs/index.html b/sqlflow/sqlfs/index.html deleted file mode 100644 index 7237249..0000000 --- a/sqlflow/sqlfs/index.html +++ /dev/null @@ -1,13 +0,0 @@ - - - - - - - - -

    Redirecting to https://github.com/sql-machine-learning/sqlflow

    - - diff --git a/sqlflow/sqlfs/reader.go b/sqlflow/sqlfs/reader.go new file mode 100644 index 0000000..bc4b594 --- /dev/null +++ b/sqlflow/sqlfs/reader.go @@ -0,0 +1,73 @@ +package sqlfs + +import ( + "database/sql" + "fmt" + "io" +) + +// Reader implements io.ReadCloser +type Reader struct { + db *sql.DB + table string + buf []byte + rows *sql.Rows +} + +// Open returns a reader to read from the given table in db. +func Open(db *sql.DB, table string) (*Reader, error) { + has, e := hasTable(db, table) + if !has { + return nil, fmt.Errorf("open: table %s doesn't exist", table) + } + if e != nil { + return nil, fmt.Errorf("open: hasTable failed with %v", e) + } + + r := &Reader{ + db: db, + table: table, + buf: nil, + rows: nil} + + r.rows, e = r.db.Query(fmt.Sprintf("SELECT block FROM %s ORDER BY id", table)) + if e != nil { + return nil, fmt.Errorf("open: failed to query: %v", e) + } + return r, nil +} + +func (r *Reader) Read(p []byte) (n int, e error) { + if r.db == nil { + return 0, fmt.Errorf("read from a closed reader") + } + n = 0 + for n < len(p) { + m := copy(p[n:], r.buf) + n += m + r.buf = r.buf[m:] + if len(r.buf) <= 0 { + if r.rows.Next() { + e = r.rows.Scan(&r.buf) + if e != nil { + break + } + } else { + e = io.EOF + break + } + } + } + return n, e +} + +func (r *Reader) Close() error { + if r.rows != nil { + if e := r.rows.Close(); e != nil { + return e + } + r.rows = nil + } + r.db = nil // Mark closed. + return nil +} diff --git a/sqlflow/sqlfs/sqlfile_test.go b/sqlflow/sqlfs/sqlfile_test.go new file mode 100644 index 0000000..7fbe8c1 --- /dev/null +++ b/sqlflow/sqlfs/sqlfile_test.go @@ -0,0 +1,151 @@ +package sqlfs + +import ( + "database/sql" + "fmt" + "io" + "math/rand" + "os" + "strings" + "testing" + + "github.com/go-sql-driver/mysql" + _ "github.com/mattn/go-sqlite3" + + "github.com/stretchr/testify/assert" +) + +var ( + testCfg *mysql.Config + testDB *sql.DB +) + +const testDatabaseName = `sqlfs_test` + +func TestCreateHasDropTable(t *testing.T) { + a := assert.New(t) + + fn := fmt.Sprintf("%s.unitest%d", testDatabaseName, rand.Int()) + a.NoError(createTable(testDB, fn)) + has, e := hasTable(testDB, fn) + a.NoError(e) + a.True(has) + a.NoError(dropTable(testDB, fn)) +} + +func TestWriterCreate(t *testing.T) { + a := assert.New(t) + + fn := fmt.Sprintf("%s.unitest%d", testDatabaseName, rand.Int()) + w, e := Create(testDB, fn) + a.NoError(e) + a.NotNil(w) + defer w.Close() + + has, e1 := hasTable(testDB, fn) + a.NoError(e1) + a.True(has) + + a.NoError(dropTable(testDB, fn)) +} + +func TestWriteAndRead(t *testing.T) { + a := assert.New(t) + + fn := fmt.Sprintf("%s.unitest%d", testDatabaseName, rand.Int()) + + w, e := Create(testDB, fn) + a.NoError(e) + a.NotNil(w) + + // A small output. + buf := []byte("\n\n\n") + n, e := w.Write(buf) + a.NoError(e) + a.Equal(len(buf), n) + + // A big output. + buf = make([]byte, kBufSize+1) + for i := range buf { + buf[i] = 'x' + } + n, e = w.Write(buf) + a.NoError(e) + a.Equal(len(buf), n) + + a.NoError(w.Close()) + + r, e := Open(testDB, fn) + a.NoError(e) + a.NotNil(r) + + // A small read + buf = make([]byte, 2) + n, e = r.Read(buf) + a.NoError(e) + a.Equal(2, n) + a.Equal(2, strings.Count(string(buf), "\n")) + + // A big read of rest + buf = make([]byte, kBufSize*2) + n, e = r.Read(buf) + a.Equal(io.EOF, e) + a.Equal(kBufSize+2, n) + a.Equal(1, strings.Count(string(buf), "\n")) + a.Equal(kBufSize+1, strings.Count(string(buf), "x")) + + // Another big read + n, e = r.Read(buf) + a.Equal(io.EOF, e) + a.Equal(0, n) + a.NoError(r.Close()) + + a.NoError(dropTable(testDB, fn)) +} + +// assertNoError prints the error if there is any in TestMain, which +// log doesn't work. +func assertNoErr(e error) { + if e != nil { + fmt.Println(e) + os.Exit(-1) + } +} + +func getEnv(key, fallback string) string { + value := os.Getenv(key) + if len(value) == 0 { + return fallback + } + return value +} + +func TestMain(m *testing.M) { + dbms := getEnv("SQLFLOW_TEST_DB", "mysql") + + var e error + switch dbms { + case "sqlite3": + testDB, e = sql.Open("sqlite3", ":memory:") + assertNoErr(e) + _, e = testDB.Exec(fmt.Sprintf("ATTACH DATABASE ':memory:' AS %s;", testDatabaseName)) + assertNoErr(e) + defer testDB.Close() + case "mysql": + cfg := &mysql.Config{ + User: getEnv("SQLFLOW_TEST_DB_MYSQL_USER", "root"), + Passwd: getEnv("SQLFLOW_TEST_DB_MYSQL_PASSWD", "root"), + Net: getEnv("SQLFLOW_TEST_DB_MYSQL_NET", "tcp"), + Addr: getEnv("SQLFLOW_TEST_DB_MYSQL_ADDR", "127.0.0.1:3306"), + AllowNativePasswords: true, + } + testDB, e = sql.Open("mysql", cfg.FormatDSN()) + assertNoErr(e) + _, e = testDB.Exec(fmt.Sprintf("CREATE DATABASE IF NOT EXISTS %s;", testDatabaseName)) + assertNoErr(e) + defer testDB.Close() + default: + assertNoErr(fmt.Errorf("unrecognized environment variable SQLFLOW_TEST_DB %s", dbms)) + } + os.Exit(m.Run()) +} diff --git a/sqlflow/sqlfs/table.go b/sqlflow/sqlfs/table.go new file mode 100644 index 0000000..01f3b97 --- /dev/null +++ b/sqlflow/sqlfs/table.go @@ -0,0 +1,49 @@ +package sqlfs + +import ( + "database/sql" + "fmt" +) + +// createTable creates a table, if it doesn't exist. If the table +// name includes the database name, e.g., "db.tbl", it creates the +// database if necessary. +func createTable(db *sql.DB, table string) error { + // FIXME(tony): HIVE and ODPS don't support AUTO_INCREMENT + // FIXME(tony): Hive and ODPS don't support BLOB, use BINARY instead + stmt := fmt.Sprintf("CREATE TABLE IF NOT EXISTS %s (id INT AUTO_INCREMENT, block BLOB, PRIMARY KEY (id))", table) + if _, e := db.Exec(stmt); e != nil { + return fmt.Errorf("createTable cannot create table %s: %v", table, e) + } + + // NOTE: a double-check of hasTable is necessary. For example, + // MySQL doesn't allow '-' in table names; however, if there + // is, the db.Exec wouldn't return any error. + has, e1 := hasTable(db, table) + if e1 != nil { + return fmt.Errorf("createTable cannot verify the creation: %v", e1) + } + if !has { + return fmt.Errorf("createTable verified table not created") + } + return nil +} + +// dropTable removes a table if it exists. If the table name includes +// the database name, e.g., "db.tbl", it doesn't try to remove the +// database. +func dropTable(db *sql.DB, table string) error { + stmt := fmt.Sprintf("DROP TABLE IF EXISTS %s", table) + if _, e := db.Exec(stmt); e != nil { + return fmt.Errorf("dropTable %s: %v", table, e) + } + return nil +} + +// hasTable checks if a table exists. +func hasTable(db *sql.DB, table string) (bool, error) { + if _, e := db.Exec(fmt.Sprintf("SELECT 1 FROM %s LIMIT 1;", table)); e != nil { + return false, fmt.Errorf("hasTable %s failed: %v", table, e) + } + return true, nil +} diff --git a/sqlflow/sqlfs/writer.go b/sqlflow/sqlfs/writer.go new file mode 100644 index 0000000..8c84e5e --- /dev/null +++ b/sqlflow/sqlfs/writer.go @@ -0,0 +1,84 @@ +package sqlfs + +import ( + "database/sql" + "fmt" +) + +var kBufSize = 4 * 1024 + +// Writer implements io.WriteCloser. +type Writer struct { + db *sql.DB + table string + buf []byte + insert *sql.Stmt +} + +// Create creates a new table or truncates an existing table and +// returns a writer. +func Create(db *sql.DB, table string) (*Writer, error) { + if e := dropTable(db, table); e != nil { + return nil, fmt.Errorf("create: %v", e) + } + return Append(db, table) +} + +// Append returns a writer to append to an existing table. It creates +// the table if it doesn't exist. +func Append(db *sql.DB, table string) (*Writer, error) { + if e := createTable(db, table); e != nil { + return nil, fmt.Errorf("create: %v", e) + } + return &Writer{db, table, make([]byte, 0, kBufSize), nil}, nil +} + +func (w *Writer) Write(p []byte) (n int, e error) { + n = 0 + for len(p) > 0 { + fill := kBufSize - len(w.buf) + if fill > len(p) { + fill = len(p) + } + w.buf = append(w.buf, p[:fill]...) + p = p[fill:] + n += fill + if len(w.buf) >= kBufSize { + if e = w.flush(); e != nil { + return n, fmt.Errorf("writer flush failed: %v", e) + } + } + } + return n, nil +} + +func (w *Writer) Close() error { + if e := w.flush(); e != nil { + return fmt.Errorf("writer flush failed: %v", e) + } + if w.insert != nil { + if e := w.insert.Close(); e != nil { + return e + } + w.insert = nil // Mark closed. + } + w.db = nil // Mark closed. + return nil +} + +func (w *Writer) flush() error { + var e error + if w.insert == nil { + w.insert, e = w.db.Prepare( + fmt.Sprintf("INSERT INTO %s (block) VALUES(?)", w.table)) + if e != nil { + return fmt.Errorf("flush failed to prepare insert %s: %v", w.table, e) + } + } + _, e = w.insert.Exec(w.buf) + if e != nil { + return fmt.Errorf("flush failed to execute insert: %v", e) + } + w.buf = w.buf[:0] + return nil +} From 2e42834796e7ff366cde02239f3b1f4f525b538b Mon Sep 17 00:00:00 2001 From: mxz96102 Date: Mon, 15 Apr 2019 16:24:02 +0800 Subject: [PATCH 2/9] update to using git submodule to maintain docs --- docs/assets/js/search-data.json | 165 ++------------------------------ docs/assets/select.gif | Bin 0 -> 71982 bytes docs/assets/sqlflow-logo.svg | 44 +++++++++ docs/assets/train.gif | Bin 0 -> 70492 bytes 4 files changed, 54 insertions(+), 155 deletions(-) create mode 100644 docs/assets/select.gif create mode 100644 docs/assets/sqlflow-logo.svg create mode 100644 docs/assets/train.gif diff --git a/docs/assets/js/search-data.json b/docs/assets/js/search-data.json index 5fa146f..a01adeb 100644 --- a/docs/assets/js/search-data.json +++ b/docs/assets/js/search-data.json @@ -1,157 +1,12 @@ +--- +--- { - "0": { - "id": "0", - "title": "About Us", - "content": "About Add info about the product. Add info about the team. Add info about the contact.", - "url": "/pages/about.html", - "relUrl": "/pages/about.html" - }, - "1": { - "id": "1", - "title": "SQLFlow: Bridging Data and AI", - "content": "", - "url": "/", - "relUrl": "/" - }, - "2": { - "id": "2", - "title": "Build TensorFlow from Source Code using Docker", - "content": "Build TensorFlow from Source Code using Docker To contribute to TensorFlow, we need to build TensorFlow from source code. The official guide is great. However, it interleaves the native building process and that using Docker and makes it confusing because packages needed by the former are not by the latter. Also, we found some useful tricks to start Docker containers in practices that are not in the official guide. Hence this document. Build the Pip Package in TensorFlow Development Container On either Mac or Linux, or any other OS, we don’t have to install and configure the building tools; instead, we can use a Docker image where all tools have been installed and properly configured. Get the Docker image containing all the building tools: docker pull tensorflow/tensorflow:latest-devel Then, let’s get the source code. On any OS, please install git using the native package manager. For example, on Ubuntu, please sudo apt-get install git or, on Mac, brew install git Then, use the git just installed, let’s clone tensorflow source code: git clone --recursive https://github.com/tensorflow/tensorflow cd tensorflow By default, we will be on the master branch. Feel free to do you change in your feature branches, or switch to a release branch, for example: git checkout v1.11.0 git checkout -b v1.11.0 Then, let us start a Docker container running the tensorflow/tensorflow:latest-devel image: docker run --rm -it -w /tensorflow -v $PWD:/tensorflow -v $HOME/.cache:/root/.cache -e "HOST_PERMS=$(id -u):$(id -g)" tensorflow/tensorflow:latest-devel /bin/bash -w /tensorflow brings us to the /tensorflow directory in the container once after we start it. -v $PWD:/tensorflow maps the current directory, which is the just cloned TensorFlow source directory on the host, to /tensorflow in the container. -v $HOME/.cache:/root/.cache maps the Bazel temporary directory on the host into the container, so the intermediate files generated by Bazel running in the container are actually saved on the host. This allows us to interrupt the container during the build and restart it later to resume the building. e "HOST_PERMS=$(id -u):$(id -g)" passes the user identification on the host into the container as an environment variable. We can reset the mode of files generated in the container to this user identity. From now on, we will be working in the container. Let us first configure the building: ./configure Usually, I would simply choose all the default options by hitting enter all the way down. Build the pip package builder: bazel build --config=opt //tensorflow/tools/pip_package:build_pip_package Build the pip package and save it into /tensorflow: ./bazel-bin/tensorflow/tools/pip_package/build_pip_package /tensorflow Change the generated wheel file into the usual mode as on the host: chown $HOST_PERMS /mnt/tensorflow-*.whl Install the Pip Pacakge in TensorFlow Development Container Let’s now try the new pip package. First, we need to uninstall the current tensorflow pip package, then we install the newly built one: pip uninstall tensorflow pip install /tensorflow/tensorflow-*.whl Now, we can verify if the new package works. First, we need to switch to a directory out from /tensorflow, so we don’t import from the source directory: cd /tmp # other directories also work then, we can try import the newly installed tensorflow package and verify it: python >>> import tensorflow >>> print(tensorflow.__version__) Now, let us quit from Python and from the Docker container. We should see the tensorflow-*.whl file on the host in the current directory. Install the Pip Package in a Clean Python Package After we quit from the development container, we should see the wheel file in the TensorFlow source directory on the host. Now, we can start a Python container and install the Pip package in it. Start the Python container docker run --rm -it -v $PWD:/tensorflow python:2.7 bash Install the pip pacakge in the container pip install /tensorflow/tensorflow*.whl Try TensorFlow by starting Python in the container python >>> import tensorflow as tf >>> print(tf.__version__)", - "url": "/sqlflow/doc/build-tensorflow.html", - "relUrl": "/sqlflow/doc/build-tensorflow.html" - }, - "3": { - "id": "3", - "title": "Canonical Development Environment", - "content": "Canonical Development Environment Referring to this example, we create a canonical development environment for Go and Python programmers using Docker. Editing on Host When we use this Docker image for daily development work, the source code relies on the host computer instead of the container. The source code includes this repo and all its dependencies, for example, the Go package google.golang.org/grpc. Code-on-the-host allows us to run our favorite editors (Emacs, VIM, Eclipse, and more) on the host. Please free to rely on editors add-ons to analyze the source code for auto-completion. Building in Container We build a Docker image that contains development tools: The Python interpreter The Go compiler The protobuf compiler The protobuf to Go compiler extension The protobuf to Python compiler extension Because this repo contains Go code, please make sure that you have the directory structure required by Go. On my laptop computer, I have export GOPATH=$HOME/go You could have your $GOPATH pointing to any directory you like. Given $GOPATH$ set, we could git clone the source code of our project by running: go get github.com/sql-machine-learning/sqlflow Change the directory to our project root, and we can use go get to retrieve and update Go dependencies. cd $GOPATH/src/github.com/sql-machine-learning/sqlflow go get -u -t ./... Note -t instructs get to also download the packages required to build the tests for the specified packages. As all Git users would do, we run git pull from time to time to sync up with others’ work. If somebody added new dependencies, we might need to run go -u ./... after git pull to update dependencies. To build this project, we need the protobuf compiler, Go compiler, Python interpreter, gRPC extension to the protobuf compiler. To ease the installation and configuration of these tools, we provided a Dockerfile to install them into a Docker image. To build the Docker image: docker build -t sqlflow:dev -f Dockerfile.dev . Development Build and Test We build and test the project inside the docker container. To run the container, we need to map the $GOPATH directory on the host into the /go directory in the container, because the Dockerfile configures /go as the $GOPATH in the container: docker run --rm -it -v $GOPATH:/go -w /go/src/github.com/sql-machine-learning/sqlflow sqlflow:dev bash Inside the Docker container, start a MySQL server in the background service mysql start& run all the tests as go generate ./... go install ./... go test -v ./... where go generate invokes the protoc command to translate server/sqlflow.proto into server/sqlflow.pb.go and go test -v builds and run unit tests. Release The above build process currently generates two binary files in $GOPATH/bin on the host. To package them into a Docker image, please run docker build -t sqlflow -f ./Dockerfile $GOPATH/bin To publish the released Docker image to our official DockerHub docker tag sqlflow sqlflow/sqlflow:latest docker push sqlflow/sqlflow:latest Demo: Command line Prompt The demo requires a MySQL server instance with populated data. If we don’t, we could follow example/datasets/README.md to start one on the host. After setting up MySQL, run the following inside the Docker container go run cmd/demo/demo.go --db_user root --db_password root --db_address host.docker.internal:3306 You should be able to see the following prompt sqlflow>", - "url": "/sqlflow/doc/build.html", - "relUrl": "/sqlflow/doc/build.html" - }, - "4": { - "id": "4", - "title": "Closing the producer goroutine from the consumer", - "content": "Closing the producer goroutine from the consumer The producer-and-consumer pattern is well used in Go concurrent programming. When the consumer stops, we want to gracefully stop the producer as well. Problem When a gRPC server receives a streaming request, it usually calls a function that returns a channel, reads the result from that channel and send the result to the client one by one. Take the following code for instance: upon receiving a request, the main goroutine Service calls launchJob. launchJob starts a separate goroutine as an anonymous function call and returns a channel. In the anonymous function, items will be sent to channel. And Service on the otherside of the channel will reads from it. func Service(req *Request, stream *StreamResponse) error { result := launchJob(req.Content) for r := range result { if e := stream.Send(result); e != nil { // should we signal the running goroutine so it will stop sending? return e } } } func launchJob(content string) chan Item { c := make(chan Item) go func() { defer close(c) acquireScarceResources() defer releaseScarceResources() ... // if stream.Send(result) returns an error and the Service returns, this will be blocked c <- Item{} ... }() return c } There is a major problem in this implementation. As pointed out by the comment, if the Send in Service returns an error, the Service function will return, leaving the anonymous function being blocked on c <- Item{} forever. This problem is important because the leaking goroutine usually owns scarce system resources such as network connection and memory. Solution: pipeline explicit cancellation Inspired by this blog post section Explicit cancellation, we can signal the cancellation via closing on a separate channel. And we can follow the terminology as io.Pipe. package sql import ( "errors" ) var ErrClosedPipe = errors.New("pipe: write on closed pipe") // pipe follows the design at https://blog.golang.org/pipelines // - wrCh: chan for piping data // - done: chan for signaling Close from Reader to Writer type pipe struct { wrCh chan interface{} done chan struct{} } // PipeReader reads real data type PipeReader struct { p *pipe } // PipeWriter writes real data type PipeWriter struct { p *pipe } // Pipe creates a synchronous in-memory pipe. // // It is safe to call Read and Write in parallel with each other or with Close. // Parallel calls to Read and parallel calls to Write are also safe: // the individual calls will be gated sequentially. func Pipe() (*PipeReader, *PipeWriter) { p := &pipe{ wrCh: make(chan interface{}), done: make(chan struct{})} return &PipeReader{p}, &PipeWriter{p} } // Close closes the reader; subsequent writes to the func (r *PipeReader) Close() { close(r.p.done) } // ReadAll returns the data chan. The caller should // use it as `for r := range pr.ReadAll()` func (r *PipeReader) ReadAll() chan interface{} { return r.p.wrCh } // Close closes the writer; subsequent ReadAll from the // read half of the pipe will return a closed channel. func (w *PipeWriter) Close() { close(w.p.wrCh) } // Write writes the item to the underlying data stream. // It returns ErrClosedPipe when the data stream is closed. func (w *PipeWriter) Write(item interface{}) error { select { case w.p.wrCh <- item: return nil case <-w.p.done: return ErrClosedPipe } } And the consumer and producer be can implemented as func Service(req *Request, stream *StreamResponse) error { pr := launchJob(req.Content) defer pr.Close() for r := range pr.ReadAll() { if e := stream.Send(r); e != nil { return e } } } func launchJob(content string) PipeReader { pr, pw := Pipe() go func() { defer pw.Close() if err := pw.Write(Item{}); err != nil { return } } return pr } Further Reading Google Form: Channel send timeout Go by Example: Timeouts Google I/O 2013 - Advanced Go Concurrency Patterns Go Concurrency Patterns Talk Go Concurrency Patterns: Pipelines and cancellation", - "url": "/sqlflow/doc/close_producer_from_consumer.html", - "relUrl": "/sqlflow/doc/close_producer_from_consumer.html" - }, - "5": { - "id": "5", - "title": "Compatibility with Various SQL Engines", - "content": "Compatibility with Various SQL Engines SQLFlow interacts with SQL engines like MySQL and Hive, while different SQL engines use variants of SQL syntax, it is important for SQLFlow to have an abstraction layer that hides such differences. SQLFlow calls Go’s standard database API. The submitter programs generated by SQLFlow call Python’s database API. Both APIs abstract the interface to various SQL engines; however, they are insufficient for SQLFlow to work. In this document, we examine all interactions between SQLFlow and the SQL engine so to identify what SQLFlow authors have to abstract in addition to calling Go’s and Python’s database APIs. Data Operations in Go Data Retrieval The basic idea of SQLFlow is to extend the SELECT statement of SQL to have the TRAIN and PREDICT clauses. For more discussion, please refer to the syntax design. SQLFlow translates such “extended SQL statements” into submitter programs, which forward the part from SELECT to TRAIN or PREDICT, which we call the “standard part”, to the SQL engine. SQLFlow also accepts the SELECT statement without TRAIN or PREDICt clauses and would forward such “standard statements” to the engine. It is noticeable that the “standard part” or “standard statements” are not standardized. For example, various engines use different syntax for joining. MySQL: SELECT pet.name, comment FROM pet, event WHERE pet.name =event.name; with keyword WHERE . Hive: SELECT pet.name, comment FROM pet JOIN event ON (pet.name =event.name) with keyword JOIN and ON. ODPS and SQLite use either INNER JOIN or OUTER JOIN. Fortunately, as SQLFlow forwards the above parts to the engine, it doesn’t have to care much about the differences above. Metadata Retrieval To verify the semantics of users’ inputs, SQLFlow needs to retrieve the schema of tables. For example, the input might be SELECT name, age, income FROM employee TRAIN DNNRegressor WITH hidden_layers=[10,50,10] COLUMN name, agee LABEL income; In the above example, the user misspelled the field name age in the COLUMN clause as “agee”. SQLFlow must be able to find that out. To do that, SQLFlow needs to query the field names from the SQL engine. However, different engines use various syntax. For example: MySQL: DESCRIBE/DESC employee; Hive: DESCRIBE FORMATTED employee; ODPS: DESC employee; SQLite: PRAGMA table_info([employee]); The returned data format varies too. Our solution to avoid such differences is not-to-use-them; instead, SQLFlow retrieves the table schema by running a query like SELECT * FROM employee LIMIT 1; and inferring field types using the mechanism called DatabaseTypeName provided by SQL engines drivers beneath the Go’s standard database API. Prepare Prediction Table A SQLFlow prediction job writes its prediction results into a table. It prepares the prediction table by Dropping previous prediction table DROP TABLE IF EXISTS my_table; Creating table with schema CREATE TABLE my_table (name1, type1, name2 type2); Most SQL engines, including MySQL, Hive, ODPS, SQLite, support both statements. Translate Database Column Type to TensorFlow Feature Column Type After retrieving database column type name through DatabaseTypeName, we can derive TensorFlow’s feature column type via a mapping such as {"FLOAT", "DOUBLE"} -> tf.numeric_column. Save Model SQLFlow saves trained ML model by dumping the serialized the model directory into a table. It first creates a table by CREATE TABLE IF NOT EXISTS %s (id INT AUTO_INCREMENT, block BLOB, PRIMARY KEY (id)) and insert blobs by INSERT INTO %s (block) VALUES(?). Note that Hive and ODPS doesn’t have BLOB type, we need to use BINARY (docs at ODPS, Hive) instead. Also, note that Hive and ODPS doesn’t support AUTO_INCREMENT, we need to implemented auto increment logic in sqlfs. Load Model SQLFlow loads trained ML model by reading rows in a table and deserializing the blob to a model directory. It reads rows by running SELECT block FROM %s ORDER BY id, which is supported by most databases. Data Operations in Python Connect to SQL Engines Thanks to the Python database API, connecting to different databases follows a similar API. conn = mysql.connector.connect(user='scott', password='password', host='127.0.0.1', database='employees') conn = sqlite3.connect('path/to/your/sqlite/file') conn = pyhive.connect('localhost') cursor = conn.cursor() cursor.execute('select * from my_table;') Insert Prediction Result into Prediction Table Python database API provides execute_many(sql, value) to insert multiple values at once. So one can prepare the following insertion statement. Please be aware that MySQL and SQLite use INSERT INTO to insert rows while Hive and ODPS use INSERT INTO TABLE. -- MySQL, SQLite INSERT INTO table_name VALUES (value1, value2, value3, ...); -- Hive, ODPS INSERT INTO TABLE table_name VALUES (value1, value2, value3, ...);", - "url": "/sqlflow/doc/database_abstraction_layer.html", - "relUrl": "/sqlflow/doc/database_abstraction_layer.html" - }, - "6": { - "id": "6", - "title": "Run MySQL Server and Client in Docker Containers", - "content": "Run MySQL Server and Client in Docker Containers The document explains how to setup MySQL in our development environment. Run MySQL Server in a Docker Container docker run --rm -v /tmp/test1:/var/lib/mysql --name mysql01 -e MYSQL_ROOT_PASSWORD=root -e MYSQL_ROOT_HOST='%' -p 3306:3306 -d mysql/mysql-server:8.0 the -v option ensures that the database is saved on the host. The default directory where MySQL saves the database is /var/lib/mysql. This directory can be configured in /etc/mysql/my.cnf, as explained in this post. By overlaying the directory /tmp/test1 on the host to /var/lib/mysql, we “cheat” MySQL to save databases on the host. So, we can kill the container and restart it, and the database is still there. Please be aware that the directory on the host must be empty the first time we run the above command; otherwise, MySQL would fail to initialize. I figured out this problem after several failures using docker logs. the -e option sets the root password of MySQL to “root”. Feel free to set it to any password you like. the second -e options sets MYSQL_ROOT_HOST to a wildcard so to allow clients connecting to the server via TCP/IP as the user “root”. This trick works with MySQL 5.7 and 8.0, but not the most recent under-development version. the --name option names the container to mysql01, which can be used to refer to this container. the -p option maps the port 3306, on which the MySQL server listens, to the same port on the host, so that clients could connect to the server via TCP/IP. Run MySQL Client in the Server Container docker exec -it mysql01 mysql -uroot -p This command executes the command mysql, which is the command line tool of MySQL, in the container named mysql01. The command line flags of mysql include -u, which specifies the username of MySQL, and -p, which makes MySQL prompts for the password. For this example, we should type the password “root”, which was set in the previous command. Please wait for a few seconds after the starting of the MySQL server container before we execute the client; otherwise, the startup of the client might fail. Once we get into the MySQL client, we can type SQL commands, e.g., show databases; create database yi; Run Client in a Different Container on the Same Host docker run --rm -it -v /tmp/test1:/var/lib/mysql mysql/mysql-server:8.0 mysql -uroot -p The -v option maps the database directory on the host to the client container. This mapping is necessary because, by default, the client talks to the server via Unix socket /var/lib/mysql/mysql.sock, which is /tmp/test1/mysql.sock on the host. Run Client in a Container on a Remote Host docker run --rm -it mysql/mysql-server:8.0 mysql -h 192.168.1.3 -P 3306 -uroot -p the -h option tells the client where the server is running on. In this example, the given IP is the one of the host where I ran the MySQL server container. Please be aware that the above command works only if the server allows remote connections. Run Python Client in a Container To write a Python client, we need to install the Python package mysql-connector-python. FROM python:2.7 RUN pip install mysql-connector-python Please be aware that some documents says that we need to install mysql-connector. I tried; but the mysql.connector.connect call failed with the error mysql.connector.errors.NotSupportedError: Authentication plugin 'caching_sha2_password' is not supported. Build the Docker image: docker build -t sqlflow . Run the image: docker run --rm -it sqlflow bash and we can start Python and run the following Python code snippet >>> import mysql.connector >>> db = mysql.connector.connect(user='root', passwd='root', host='192.168.1.3') >>> print(db) <mysql.connector.connection_cext.CMySQLConnection object at 0x7fbab9f3fed0> Run a Go Client In order to connect to a database, you need to import the database’s driver first. export GOPATH=$HOME/go go get -u github.com/go-sql-driver/mysql go run the following file package main import ( "database/sql" "github.com/go-sql-driver/mysql" "log" ) func main() { testConfig := &mysql.Config{ User: "root", Passwd: "root", Net: "tcp", Addr: "localhost:3306", } db, e := sql.Open("mysql", testConfig.FormatDSN()) if e != nil { log.Fatal(e) } defer db.Close() }", - "url": "/sqlflow/doc/mysql-setup.html", - "relUrl": "/sqlflow/doc/mysql-setup.html" - }, - "7": { - "id": "7", - "title": "Piping Responses", - "content": "Piping Responses Streaming Responses As described in the overall design, a SQLFlow job could be a standard or an extended SQL statemnt, where an extended SQL statement will be translated into a Python program. Therefore, each job might generate up to the following data streams: standard output, where each element is a line of text, standard error, where each element is a line of text, data rows, where the first element consists of fields name/types, and each of the rest is a row of data, status, where the element could be pending, failed, and succeeded. To create good user experience, we need to pipe these responses from SQLFlow jobs to Jupyter Notebook in real-time. Stages in the Pipe The pipe that streams outputs from SQLFlow jobs to the Jupyter Notebook consists of the following stages: Web browser ↑ | HTTP ↓ Jupyter Notebook server ↑ | ZeroMQ streams: Shell, IOPub, stdin, Controls, Heartbeat ↓ iPython kernel ↑ | IPython magic command framework ↓ SQLFlow magic command for Jupyter ↑ | gRPC ↓ SQLFlow server ↑ | Go channels ↓ SQLFlow job manager (Go functions) ↑ | IPC with Go's standard library ↓ SQLFlow jobs In the above figure, from the SQLFlow magic command to the bottom layer are our work. Streaming We have two alternative ideas: multiple streams and a multiplexing stream. We decided to use a multiplexing stream because we had a unsuccessful trial with the multiple streams idea: we make the job writes to various Go channels and forward each Go channel to a streaming gRPC call, as the following: Multiple streams The above figure shows that there are multiple streams between the Jupyter Notebook server and Jupyter kernels. According to the document, there are five: Shell, IOPub, stdin, Control, and Heartbeat. These streams are ZeroMQ streams. We don’t use ZeroMQ, but we can take the idea of having multiple parallel streams in the pipe. service SQLFlow { rpc File(string sql) returns (int id) {} rpc ReadStdout(int id) returns (stream string) {} rpc ReadStderr(int id) returns (stream string) {} rpc ReadData(int id) returns (stream Row) {} rpc ReadStatus(int id) returns (stream int) {} } However, we realized that if the user doesn’t call any one of the SQLFlow.Read... call, there would be no forwarding from the Go channel to Jupyter, thus the job would block forever at writing. A Multiplexing Stream Another idea is multiplexing all streams into one. For example, we can have only one ZeroMQ stream, where each element is a polymorphic type – could be a text string or a data row. service SQLFlow { rpc Run(string sql) returns (stream Response) {} } // Only one of the following fields should be set. message Response { oneof record { repeated string head = 1; // Column names. repeated google.protobuf.Any row = 2; // Cells in a row. string log = 3; // A line from stderr or stdout. } }", - "url": "/sqlflow/doc/pipe.html", - "relUrl": "/sqlflow/doc/pipe.html" - }, - "8": { - "id": "8", - "title": "Quick start", - "content": "Quick start SQLFlow is currently under active development. For those who are interested in trying it out, we have provided several demos. Play around with it. Any bug report and issue are welcomed. :) Setup Install Docker. Set up a MySQL server following example/datasets/README.md. Pull the latest SQLFlow Docker image: docker pull sqlflow/sqlflow:latest. Demo 1: Jupyter Notebook Start a Docker container that runs sqlflowserver and Jupyter Notebook. If you are using Docker for Linux, please change host.docker.internal:3306 to localhost:3306. docker run --rm -it -p 8888:8888 sqlflow/sqlflow:latest bash -c "sqlflowserver --db_user root --db_password root --db_address host.docker.internal:3306 & SQLFLOW_SERVER=localhost:50051 jupyter notebook --ip=0.0.0.0 --port=8888 --allow-root" If you are using Docker for Mac, please be aware the option --db_address host.docker.internal:3306 where host.docker.internal translates to the host ip address as recommended here. If you are running MySQL on remote, please be aware that MySQL only allows connections from localhost by default. Fix can be found here. Open a Web browser and direct to localhost:8888 and input the token. Then you can create notebooks. In a cell, you should be able to type %%sqlflow select 1 Explore more examples at example.ipynb Demo 2: Command Line Prompt Start a Docker container that runs SQLFlow command line prompt. If you are using Docker for Linux, please change host.docker.internal:3306 to localhost:3306. docker run -it --rm --net=host sqlflow/sqlflow:latest demo --db_user root --db_password root --db_address host.docker.internal:3306 You should be able to see the following prompt. sqlflow> Example Select data sqlflow> select * from iris.train limit 2; -- +--+-+--+-+-+ | SEPAL LENGTH | SEPAL WIDTH | PETAL LENGTH | PETAL WIDTH | CLASS | +--+-+--+-+-+ | 6.4 | 2.8 | 5.6 | 2.2 | 2 | | 5 | 2.3 | 3.3 | 1 | 1 | +--+-+--+-+-+ Train a Tensorflow DNNClassifier sqlflow> SELECT * FROM iris.train TRAIN DNNClassifier WITH n_classes = 3, hidden_units = [10, 20] COLUMN sepal_length, sepal_width, petal_length, petal_width LABEL class INTO sqlflow_models.my_dnn_model; -- ... Training set accuracy: 0.96721 Done training Prediction using a trained model sqlflow> SELECT * FROM iris.test predict iris.predict.class USING sqlflow_models.my_dnn_model; Checkout prediction result sqlflow> select * from iris.predict limit 10;", - "url": "/sqlflow/doc/quickstart.html", - "relUrl": "/sqlflow/doc/quickstart.html" - }, - "9": { - "id": "9", - "title": "Extended SQL Parser Design", - "content": "Extended SQL Parser Design This documentation explains the technical decision made in building a SQL parser in Go. It is used to parsed the extended SELECT syntax of SQL that integrates TensorFlow Estimators. Related Work Lexer and Parser Generator In 2001, when I was in graduate school, I defined an extended SQL syntax for querying distributed relational databases, as part of the course project of Distributed RDBMS by Prof. Li-zhu Zhou. I wrote the parser using bison (a modern version of yacc) and flex (a modern version of lex). yacc and lex generate C code; bison and flex generate C++ code. However, this time, I’d use Go. I surveyed goyacc, a standard Go tool. The usage is very similar to that of yacc and bison. However, the Go toolchain doesn’t provide a tool like lex/flex. Google revealed golex, which is out of maintenance. The Medium post recommends Ragel, which is a C++ program and could generate Go lexer; however, it lacks documents. Handwritten Lexer and Parser Some documents, including this one recommends handwriting lexers. However, it doesn’t explain how to write the parser. GoAcademy always provides high-quality tech blog posts. This one is from the author of InfluxDB. However, I stopped at where it explains wrapping a SQL statement as a string by an io.Reader, because it is obvious that we should keep the string as a string so that that token strings could refer to the same memory storage of the SQL statement. Following a link in the above GoAcademy post, I found Rob Pike’s excellent talk on how to write a lexer in Go in 2011. Many works after that change Rob’s implementation somehow but always lead to longer and less comprehensible codebases. The Choice Therefore, I wrote the lexer and parser both following Rob Pike’s idea. After few days work, I realized that: I should borrow the idea from Rob to represent SQL statements as strings, but not io.Reader as other work do, but no need to use channels and goroutines at all, and it is technically intractable to write a SQL lexer/parser manually. So, I switched to write a lexer manually, and to generate the parser using goyacc. During my work, I referred to this example and the official yacc manual for details about operator association and precedence.", - "url": "/sqlflow/doc/sql_parser.html", - "relUrl": "/sqlflow/doc/sql_parser.html" - }, - "10": { - "id": "10", - "title": "Submitter", - "content": "Submitter A submitter is a pluggable module in SQLFlow that is used to submit an ML job to a third party computation service. Workflow When a user types in an extended SQL statement, SQLFlow first parses and semantically verifies the statement. Then SQLFlow either runs the ML job locally or submits the ML job to a third party computation service. In the latter case, SQLFlow produces a job description (TrainDescription or PredictDescription) and hands it over to the submitter. For a training SQL, SQLFlow produces TrainDescription; for prediction SQL, SQLFlow produces PredDescription. The concrete definition of the description looks like the following {% raw %} type ColumnType struct { Name string // e.g. sepal_length DatabaseTypeName string // e.g. FLOAT } // SELECT * // FROM iris.train // TRAIN DNNClassifier // WITH // n_classes = 3, // hidden_units = [10, 20] // COLUMN sepal_length, sepal_width, petal_length, petal_width // LABEL class // INTO sqlflow_models.my_dnn_model; type TrainDescription struct { StandardSelect string // e.g. SELECT * FROM iris.train Estimator string // e.g. DNNClassifier Attrs map[string]string // e.g. {{"n_classes", "3"}, {"hidden_units", "[10, 20]"}} X []ColumnType // e.g. {{"sepal_length", "FLOAT"}, ...} Y ColumnType // e.g. {"class", "INT"} ModelName string // e.g. my_dnn_model } // SELECT * // FROM iris.test // PREDICT iris.predict.class // USING sqlflow_models.my_dnn_model; type PredDescription struct { StandardSelect string // e.g. SELECT * FROM iris.test TableName string // e.g. iris.predict ModelName string // e.g. my_dnn_model } {% endraw %} Submitter Interface The submitter interface should provide two functions Train and Predict. The detailed definition can be the following type Submitter interface { // Train executes a ML training job and streams job's response through writer. // A typical Train function should include // - Loading the training data // - Initializing the model // - model.train // - Saving the trained model to a persistent storage Train(desc TrainDescription, writer PipeWriter) error // Predict executes a ML predicting job and streams job's response through writer // A typical Predict function should include // - Loading the model from a persistent storage // - Loading the prediction data // - model.predict // - Writing the prediction result to a table Predict(desc PredictDescription, writer PipeWriter) error } Register a submitter A new submitter can be added as import ( ".../my_submitter" ".../sqlflow/sql" ) func main() { // ... sql.Register(my_submitter.NewSubmitter()) // ... for { sql := recv() sql.Run(sql) } } where sql.Register will put my_submitter instance to package level registry. During sql.Run, it will check whether there is a submitter registered. If there is, sql.Run will run either submitter.Train or submitter.Predict.", - "url": "/sqlflow/doc/submitter.html", - "relUrl": "/sqlflow/doc/submitter.html" - }, - "11": { - "id": "11", - "title": "SQLFlow: Design Doc", - "content": "SQLFlow: Design Doc What is SQLFlow SQLFlow is a bridge that connects a SQL engine, for example, MySQL, Hive, SparkSQL, Oracle, or SQL Server, and TensorFlow and other machine learning toolkits. SQLFlow extends the SQL syntax to enable model training and inference. Related Work We could write simple machine learning prediction (or scoring) algorithms in SQL using operators like DOT_PRODUCT. However, this requires copy-n-pasting model parameters from the training program into SQL statements. Some proprietary SQL engines provide extensions to support machine learning. Microsoft SQL Server Microsoft SQL Server has the machine learning service that runs machine learning programs in R or Python as an external script: CREATE PROCEDURE generate_linear_model AS BEGIN EXEC sp_execute_external_script @language = N'R' , @script = N'lrmodel <- rxLinMod(formula = distance ~ speed, data = CarsData); trained_model <- data.frame(payload = as.raw(serialize(lrmodel, connection=NULL)));' , @input_data_1 = N'SELECT [speed], [distance] FROM CarSpeed' , @input_data_1_name = N'CarsData' , @output_data_1_name = N'trained_model' WITH RESULT SETS ((model varbinary(max))); END; A challenge to the users is that they need to know not only SQL but also R or Python, and they must be capable of writing machine learning programs in R or Python. Teradata SQL for DL Teradata also provides a RESTful service, which is callable from the extended SQL SELECT syntax. SELECT * FROM deep_learning_scorer( ON (SELECT * FROM cc_data LIMIT 100) URL('http://localhost:8000/api/v1/request') ModelName('cc') ModelVersion('1') RequestType('predict') columns('v1', 'v2', ..., 'amount') ) The above syntax couples the deployment of the service (the URL in the above SQL statement) with the algorithm. Google BigQuery Google BigQuery enables machine learning in SQL by introducing the CREATE MODEL statement. CREATE MODEL dataset.model_name OPTIONS(model_type='linear_reg', input_label_cols=['input_label']) AS SELECT * FROM input_table; Currently, BigQuery only supports two simple models: linear regression and logistic regression. Design Goal None of the above meets our requirement. First of all, we want to build an open source software. Also, we want it to be extensible: We want it extensible to many SQL engines, instead of targeting any one of them. Therefore, we don’t want to build our syntax extension on top of user-defined functions (UDFs); otherwise, we’d have to implement them for each SQL engine. We want the system extensible to support sophisticated machine learning models and toolkits, including TensorFlow for deep learning and xgboost for trees. Another challenge is that we want SQLFlow to be flexible enough to configure and run cutting-edge algorithms, including specifying feature crosses. At the same time, we want SQLFlow easy to learn – at least, no Python or R code embedded in the SQL statements, and integrate hyperparameter estimation. We understand that a key to address the above challenges is the syntax of the SQL extension. To craft a highly-effective and easy-to-learn syntax, we need user feedback and fast iteration. Therefore, we’d start from a prototype that supports only MySQL and TensorFlow. We plan to support more SQL engines and machine learning toolkits later. Design Decisions As the beginning of the iteration, we propose an extension to the SQL SELECT statement. We are not going a new statement way like that BigQuery provides CREATE MODEL, because we want to maintain a loose couple between our system and the underlying SQL engine, and we cannot create the new data type for the SQL engine, like CREATE MODEL requires. We highly appreciate the work of TensorFlow Estimator, a high-level API for deep learning. The basic idea behind Estimator is to implement each deep learning model, and related training/testing/evaluating algorithms as a Python class derived from tf.estimator.Estimator. As we want to keep our SQL syntax simple, we would make the system extensible by calling estimators contributed by machine learning experts and written in Python. The SQL syntax must allow users to set Estimator attributes (parameters of the Python class’ constructor, and those of train, evaluate, or predict). Users can choose to use default values. We have a plan to integrate our hyperparameter estimation research into the system to optimize the default values. Though estimators derived from tf.estimator.Estimator run algorithms as TensorFlow graphs; SQLFlow doesn’t restrict that the underlying machine learning toolkit has to be TensorFlow. Indeed, as long as an estimator provides methods of train, evaluate, and predict, SQLFlow doesn’t care if it calls TensorFlow or xgboost. Precisely, what SQLFlow expect is an interface like the following: class AnEstimatorClass: __init__(self, **kwargs) train(self, **kwargs) evaluate(self, **kwargs) predict(self, **kwargs) We also want to reuse the feature columns API from Estimator, which allows users to columns of tables in a SQL engine to features to the model. Extended SQL Syntax Again, just as the beginning of the iteration, we propose the syntax for training as SELECT * FROM kaggle_credit_fraud_training_data LIMIT 1000 TRAIN DNNClassifier /* a pre-defined TensorFlow estimator, tf.estimator.DNNClassifier */ WITH layers=[100, 200], /* a parameter of the Estimator class constructor */ train.batch_size = 8 /* a parameter of the Estimator.train method */ COLUMN *, /* all columns as raw features */ cross(v1, v9, v28) /* plus a derived (crossed) column */ LABEL class INTO sqlflow_models.my_model_table; /* saves trained model parameters and features into a table */ We see the redundancy of * in two clauses: SELECT and COLUMN. The following alternative can avoid the redundancy, but cannot specify the label. SELECT * /* raw features or the label? */ corss(v1, v9, v28) /* derived featuers */ FROM kaggle_credit_fraud_training_data Please be aware that we save the trained models into tables, instead of a variable maintained by the underlying SQL engine. To invent a new variable type to hold trained models, we’d make our system tightly integrated with the SQL engine, and harms the extensibility to other engines. The result table should include the following information: The estimator name, e.g., DNNClassifier in this case. Estimator attributes, e.g., layer and train.batch_size. The feature mapping, e.g., * and cross(v1, v9, v28). Similarly, to infer the class (fraud or regular), we could SELECT * FROM kaggle_credit_fraud_development_data PREDICT kaggle_credit_fraud_development_data.class USING sqlflow_models.my_model_table; System Architecture A Conceptual Overview In the prototype, we use the following architecture: SQL statement -> our SQL parser --standard SQL-> MySQL -extended SQL-> code generator -> execution engine In the prototype, the code generator generates a Python program that trains or predicts. In either case, it retrieves the data from MySQL via MySQL Connector Python API, optionally, retrieves the model from MySQL, trains the model or predicts using the trained model by calling the user specified TensorFlow estimator, and writes the trained model or prediction results into a table. Working with Jupyter Notebook and Kubernetes The following figures shows the system components and their runtime environment. The left part shows how to run the system on a PC/laptop, the right part shows how to run it on a Kubernetes cluster.", - "url": "/sqlflow/doc/syntax.html", - "relUrl": "/sqlflow/doc/syntax.html" - }, - "12": { - "id": "12", - "title": "SQLFlow: Code Walkthrough", - "content": "SQLFlow: Code Walkthrough User Experience SQLFlow allows users to write SQL programs with extended syntax in Jupyter Notebook or a command-line tool. The following SQL statements train a TensorFlow model named DNNClassifier, which is a Python class derived from tf.estimator.Estimator: SELECT * FROM a_table TRAIN DNNClassifier WITH learning_rate=0.01 INTO sqlflow_models.my_model; And the following statement uses the trained model for prediction. SELECT * FROM b_table PREDICT b_table.predicted_label USING sqlflow_models.my_model; Please be aware that the part in the above statements before the extended keyword TRAIN and PREDICT is a standard SQL statement. This feature simplifies the implementation of the SQLFlow system. System Implementation If a SQL statement is of the standard syntax, SQLFlow throws it to the SQL engine and redirects the output to the user; otherwise, SQLFlow translates the statement of extended syntax into a Python program. Currently, it generates a program that throws the standard-syntax part of SELECT to MySQL, reads the results in the train-loop of a TensorFlow program. We will talk about how to extend SQLFlow to connect more SQL engines like Oracle, Hive, and SparkSQL, and generates more types of machine learning programs that calls distributed TensorFlow, PyTorch, and xgboost later. Before that, let us explain the system components. SQLFlow as a gRPC Server SQLFlow is a gRPC server, which can connect with multiple clients. A typical client is pysqlflow, the SQLFlow plugin for Jupyter Notebook server. Another once is a text-based client /cmd/sqlflowserver/main.go. Jupyter Notebook (SQL statements)--> SQLFlow gRPC server (SQLFlow magic command) <--(a stream of messages)-- The protobuf definition of the gRPC service is at /server/proto/sqlflow.proto. The return of the method SQLFlow.Run is a stream of Reponses, where each represents either a table header, a row, or a log message. The header and rows are usually from a standard SQL statement, for example, SELECT or DESCRIBE, and the log messages are usually from the running of a generated Python program. SQLFlow in the gRPC Server Once the SQLFlow server receives a batch of SQL statements via a gRPC call, it runs the following steps for each statement: the parser to generate parsing result, the verifier to verify the semantics given the parsing result, the code generator to generate a Python program, or the submitter, from the parsing result, the executor that runs the submitter locally. Step 3. and 4. are only for a SQL statement of extended syntax; otherwise, SQLFlow server proxies the standard-syntax statement to the SQL engine. The executor calls Go’s standard package that captures the stdout and stderr from the submitter process and passing the result back to the gRPC client. Therefore, it is the responsibility of the submitter to print log messages to its stderr and stdout. Minimal Viable Product In the minimal viable product (MVP) of SQLFlow, the code generator generates a Python program consists of two parts: throw the standard SELECT part in the extended-syntax statement to MySQL via ODBC, and a loop that reads outputs from the run of the SELECT statement and trains the model (or, using a trained model to predict). The training part calls TensorFlow to update the parameters of the model specified in the TRAIN clause. Extensibility By writing more code generators, we could extend SQLFlow to support more SQL engines, e.g., Hive and Oracle, and use machine learning toolkits, e.g., PyTorch and xgboost, in addition to TensorFlow, on various computing platforms. You are welcome to add more code generators such as codegen_distributed_tf.go to generate a submitter program similar to the MVP but runs a distributed TensorFlow training job. codegen_kubernetes_tf.go to launch a distributed TensorFlow job on a Kubernetes cluster, other than running locally, in the same container as where SQLFlow gRPC server resides. codegen_gcloud_pytorch.go to launch a submitter that calls PyTorch instead of TensorFlow for training on the Google Cloud. Job Management The current design of the gRPC interface assumes that the connection between the client, e.g., the Jupyter Notebook, and the SQLFlow server keeps alive during the running of the training program. This assumption is reasonable because even if the user closes her/his Web browser and disconnect to the Jupyter Notebook server, the connection between Jupyter to SQLFlow server might keep alive. However, this might not be robust enough if the Jupyter Notebook server runs on a user’s laptop and gets killed. In such a case, the gRPC server cannot stream the messages back to the client and would cause the failure of the submitter. A solution is to change the gRPC interface of SQLFlow server to have a method that files a job and returns immediately, and another method to get a batch of recent messages given a job ID. We will make a design for that soon.", - "url": "/sqlflow/doc/walkthrough.html", - "relUrl": "/sqlflow/doc/walkthrough.html" - }, - "13": { - "id": "13", - "title": "SQLFlow", - "content": "SQLFlow SQLFlow enriches SQL systems, e.g., MySQL, Hive, SparkSQL, with the capability of machine learning, using TensorFlow. We slightly extend the syntax of the SELECT statement of SQL to support model training and prediction. How to Use Quick start Extended SQL syntax How to Contribute Build from source code. The walkthrough of the source code The choice of parser generator Questions and Feedback Your feedback is our motivation to move on. Please let us know your questions, concerns, and issues by filing Github Issues. License Apache License 2.0", - "url": "/sqlflow/", - "relUrl": "/sqlflow/" - }, - "14": { - "id": "14", - "title": "SQLFlow Demo", - "content": "SQLFlow Demo Before you start, you need to set up test MySQL server as described here Start SQLFlow as go get -d ./... && go run main.go -logdir="/path/to/logs/" -loglevel="info" Regular SQL Statements sqlflow> SHOW DATABASES; +--+ | Database | +--+ | churn | | information_schema | | iris | | mysql | | performance_schema | | sqlflow_models | | sqlfs | | sys | +--+ sqlflow> SELECT * FROM iris.train LIMIT 1; +--+-+--+-+-+ | sepal_length | sepal_width | petal_length | petal_width | class | +--+-+--+-+-+ | 6.4 | 2.8 | 5.6 | 2.2 | 2 | +--+-+--+-+-+ Train sqlflow> SELECT * FROM iris.train TRAIN DNNClassifier WITH n_classes = 3, hidden_units = [10, 20] COLUMN sepal_length, sepal_width, petal_length, petal_width LABEL class INTO sqlflow_models.my_dnn_model; -- 2018/12/16 15:03:54 tensorflowCmd: run in Docker container Job success Predict sqlflow> SELECT * FROM iris.test PREDICT iris.predict.class USING sqlflow_models.my_dnn_model; -- 2018/12/16 15:05:58 tensorflowCmd: run in Docker container Job success The prediction is stored at iris.predict, you can look them up at sqlflow> SELECT * FROM iris.predict LIMIT 10; -- +--+-+--+-+-+ | sepal_length | sepal_width | petal_length | petal_width | class | +--+-+--+-+-+ | 6.4 | 2.8 | 5.6 | 2.2 | 2 | | 5 | 2.3 | 3.3 | 1 | 1 | | 4.9 | 2.5 | 4.5 | 1.7 | 2 | | 4.9 | 3.1 | 1.5 | 0.1 | 0 | | 5.7 | 3.8 | 1.7 | 0.3 | 0 | | 4.4 | 3.2 | 1.3 | 0.2 | 0 | | 5.4 | 3.4 | 1.5 | 0.4 | 0 | | 6.9 | 3.1 | 5.1 | 2.3 | 2 | | 6.7 | 3.1 | 4.4 | 1.4 | 1 | | 5.1 | 3.7 | 1.5 | 0.4 | 0 | +--+-+--+-+-+", - "url": "/sqlflow/cmd/demo/", - "relUrl": "/sqlflow/cmd/demo/" - }, - "15": { - "id": "15", - "title": "The Credit Card Fraud Detection Example", - "content": "The Credit Card Fraud Detection Example The downsample/creditcard.csv file in this directory comes from Kaggle. Down-sample To make it an example that can be checked into a git repo, I randomly picked up 2000 out from the 284,808 lines from the original dataset using the following command: gshuf -n 5000 /tmp/creditcard.csv > ./downsample/creditcard.csv where /tmp/creditcard.csv is the one downloaded from Kaggle. If you want to down-sample by yourself on macOS, you might need to install gshuf. brew install coreutils Import into MySQL To develop/test/debug the TensorFlow program automatically generated by SQLFlow, which is supposed to be able to read data from MySQL, we’d import the downsample/creditcard.csv into MySQL. Run MySQL Server For more about running MySQL in Docker containers, please refer to Run MySQL Server and Client in Docker Containers. The following command starts the MySQL server: docker run --rm -v $HOME/work/creditcarddb:/var/lib/mysql --name mysql01 -e MYSQL_ROOT_PASSWORD=root -e MYSQL_ROOT_HOST='%' -p 3306:3306 -d mysql/mysql-server:8.0 mysqld --secure-file-priv="" Please be aware that the above command doesn’t use the default entrypoint of the Docker image; instead, it explicitly start mysqld in the container so that it could provide the command line option --secure-file-priv="", which is required by mysqlimport. Run the following command to start a MySQL client: docker run --rm -it -v $HOME/work/creditcarddb:/var/lib/mysql mysql/mysql-server:8.0 mysql -uroot -p Remember to type the password root when prompted. Then input the following SQL program in the client to create the table creditcard.creditcard: CREATE DATABASE creditcard; USE creditcard; CREATE TABLE IF NOT EXISTS creditcard ( time INT, v1 float, v2 float, v3 float, v4 float, v5 float, v6 float, v7 float, v8 float, v9 float, v10 float, v11 float, v12 float, v13 float, v14 float, v15 float, v16 float, v17 float, v18 float, v19 float, v20 float, v21 float, v22 float, v23 float, v24 float, v25 float, v26 float, v27 float, v28 float, amount float, class varchar(255)); Before we could import the CSV file, we must follow the security policy of MySQL, which is wired and hard to understand, and copy the CSV file into the database directory: cp downsample/creditcard.csv ~/work/creditcarddb/creditcard/ The following command runs mysqlimport in the mysql/mysql Docker image to import the data: docker run --rm -it -v $HOME/work/creditcarddb:/var/lib/mysql -v $HOME/work/creditcard:/var/lib/mysql-files mysql:8.0 mysqlimport --ignore-lines=1 --fields-terminated-by=, --socket /var/lib/mysql/mysql.sock -u root -p creditcard creditcard.csv Please be aware that mysqlimport doesn’t exist in the mysql/mysql-server image. A message likes the following indicates the success of importing: creditcard.creditcard: Records: 1999 Deleted: 0 Skipped: 0 Warnings: 0 We can verify the imported result by running a SELECT statement like the following docker run --rm -it -v $HOME/work/creditcarddb:/var/lib/mysql mysql/mysql-server:8.0 mysql -uroot -p -e 'use creditcard; select time, v1, v28, amount, class from creditcard limit 10;' which should print something like the following +--+--++--+-+ | time | v1 | v28 | amount | class | +--+--++--+-+ | 123590 | -0.341154 | 0.130741 | 69 | "0" | | 135797 | -0.118853 | 0.11599 | 11.79 | "0" | | 161249 | -1.04129 | -0.362025 | 7.38 | "0" | | 72471 | 0.603151 | 0.0551408 | 352.04 | "0" | | 120182 | 2.10678 | -0.0566284 | 0.89 | "0" | | 160476 | 2.37223 | -0.0724277 | 6 | "0" | | 135023 | 1.87987 | 0.0214597 | 20.05 | "0" | | 117329 | -0.375213 | -0.130152 | 159 | "0" | | 148637 | 1.83178 | -0.0378951 | 38.37 | "0" | | 48293 | 1.26094 | 0.0275234 | 12 | "0" | +--+--++--+-+", - "url": "/sqlflow/example/creditcard/", - "relUrl": "/sqlflow/example/creditcard/" - }, - "16": { - "id": "16", - "title": "The MySQL Server Container for Testing", - "content": "The MySQL Server Container for Testing This directory contains a Dockerfile that builds a Docker image derived the MySQL Server 8.0 image, and includes SQL programs that popularize the following datasets: Churn from Kaggle Irises classfication from TensorFlow We can run a Docker container of it for unit testing. Build cd example/datasets docker build -t sqlflow:data . Run docker run --rm -d --name sqlflowdata -p 3306:3306 -e MYSQL_ROOT_PASSWORD=root -e MYSQL_ROOT_HOST=% sqlflow:data Popularize Datasets We need to manually popularize the databases and tables: docker exec -it sqlflowdata bash To popularize the Churn dataset into churn: cat /popularize_churn.sql | mysql -uroot -proot To popularize the Irises dataset into iris: cat /popularize_iris.sql | mysql -uroot -proot To prepare database for storing machine learning models: echo "CREATE DATABASE IF NOT EXISTS sqlflow_models;" | mysql -uroot -proot Query In the container, run echo "select count(*) from churn.test;" | mysql -uroot -proot should print the number of rows as the following count(*) 10 Trouble shooting: It usually takes about 15 seconds to bring up the MySQL Server. If you try to connect it before that, you may see the following error ERROR 1045 (28000): Access denied for user 'root'@'localhost' (using password: YES)", - "url": "/sqlflow/example/datasets/", - "relUrl": "/sqlflow/example/datasets/" - }, - "17": { - "id": "17", - "title": "Credit Card Fraud Detection", - "content": "Credit Card Fraud Detection Data collected from here To run: python premade_estimator.py, which gives Adding numeric_columns: Time Adding numeric_columns: V1 Adding numeric_columns: V2 Adding numeric_columns: V3 Adding numeric_columns: V4 Adding numeric_columns: V5 Adding numeric_columns: V6 Adding numeric_columns: V7 Adding numeric_columns: V8 Adding numeric_columns: V9 Adding numeric_columns: V10 Adding numeric_columns: V11 Adding numeric_columns: V12 Adding numeric_columns: V13 Adding numeric_columns: V14 Adding numeric_columns: V15 Adding numeric_columns: V16 Adding numeric_columns: V17 Adding numeric_columns: V18 Adding numeric_columns: V19 Adding numeric_columns: V20 Adding numeric_columns: V21 Adding numeric_columns: V22 Adding numeric_columns: V23 Adding numeric_columns: V24 Adding numeric_columns: V25 Adding numeric_columns: V26 Adding numeric_columns: V27 Adding numeric_columns: V28 Adding numeric_columns: Amount INFO:tensorflow:Using default config. WARNING:tensorflow:Using temporary folder as model directory: /tmp/tmpzj6rzutf INFO:tensorflow:Using config: {'_num_worker_replicas': 1, '_global_id_in_cluster': 0, '_save_checkpoints_steps': None, '_tf_random_seed': None, '_evaluation_master': '', '_protocol': None, '_session_config': allow_soft_placement: true graph_options { rewrite_options { meta_optimizer_iterations: ONE } } , '_model_dir': '/tmp/tmpzj6rzutf', '_device_fn': None, '_cluster_spec': <tensorflow.python.training.server_lib.ClusterSpec object at 0x7f40f9b40438>, '_save_summary_steps': 100, '_eval_distribute': None, '_keep_checkpoint_max': 5, '_master': '', '_train_distribute': None, '_log_step_count_steps': 100, '_num_ps_replicas': 0, '_experimental_distribute': None, '_save_checkpoints_secs': 600, '_service': None, '_task_type': 'worker', '_task_id': 0, '_keep_checkpoint_every_n_hours': 10000, '_is_chief': True} INFO:tensorflow:Calling model_fn. INFO:tensorflow:Done calling model_fn. INFO:tensorflow:Create CheckpointSaverHook. INFO:tensorflow:Graph was finalized. INFO:tensorflow:Running local_init_op. INFO:tensorflow:Done running local_init_op. INFO:tensorflow:Saving checkpoints for 0 into /tmp/tmpzj6rzutf/model.ckpt. INFO:tensorflow:loss = 892.3532, step = 1 INFO:tensorflow:global_step/sec: 295.764 INFO:tensorflow:loss = 0.00025717804, step = 101 (0.338 sec) INFO:tensorflow:global_step/sec: 446.454 INFO:tensorflow:loss = 0.00024299514, step = 201 (0.224 sec) INFO:tensorflow:global_step/sec: 444.479 INFO:tensorflow:loss = 0.00020988214, step = 301 (0.225 sec) INFO:tensorflow:global_step/sec: 421.979 INFO:tensorflow:loss = 0.00023327809, step = 401 (0.237 sec) INFO:tensorflow:global_step/sec: 446.172 INFO:tensorflow:loss = 0.00020405005, step = 501 (0.224 sec) INFO:tensorflow:global_step/sec: 431.785 INFO:tensorflow:loss = 0.00017921966, step = 601 (0.232 sec) INFO:tensorflow:global_step/sec: 442.912 INFO:tensorflow:loss = 0.00018488085, step = 701 (0.226 sec) INFO:tensorflow:global_step/sec: 444.628 INFO:tensorflow:loss = 0.00017795907, step = 801 (0.225 sec) INFO:tensorflow:global_step/sec: 428.385 INFO:tensorflow:loss = 0.00016896412, step = 901 (0.233 sec) INFO:tensorflow:Saving checkpoints for 1000 into /tmp/tmpzj6rzutf/model.ckpt. INFO:tensorflow:Loss for final step: 0.00016340206. INFO:tensorflow:Calling model_fn. WARNING:tensorflow:Trapezoidal rule is known to produce incorrect PR-AUCs; please switch to "careful_interpolation" instead. WARNING:tensorflow:Trapezoidal rule is known to produce incorrect PR-AUCs; please switch to "careful_interpolation" instead. INFO:tensorflow:Done calling model_fn. INFO:tensorflow:Starting evaluation at 2018-10-16-01:02:49 INFO:tensorflow:Graph was finalized. INFO:tensorflow:Restoring parameters from /tmp/tmpzj6rzutf/model.ckpt-1000 INFO:tensorflow:Running local_init_op. INFO:tensorflow:Done running local_init_op. INFO:tensorflow:Finished evaluation at 2018-10-16-01:02:49 INFO:tensorflow:Saving dict for global step 1000: accuracy = 1.0, accuracy_baseline = 1.0, auc = 0.9999999, auc_precision_recall = 0.0, average_loss = 0.0001784083, global_step = 1000, label/mean = 0.0, loss = 0.0016056747, precision = 0.0, prediction/mean = 0.00017826515, recall = 0.0 INFO:tensorflow:Saving 'checkpoint_path' summary for global step 1000: /tmp/tmpzj6rzutf/model.ckpt-1000 Test set accuracy: 1.00000 INFO:tensorflow:Calling model_fn. INFO:tensorflow:Done calling model_fn. INFO:tensorflow:Graph was finalized. INFO:tensorflow:Restoring parameters from /tmp/tmpzj6rzutf/model.ckpt-1000 INFO:tensorflow:Running local_init_op. INFO:tensorflow:Done running local_init_op. Prediction is "Not Fraud" (100.0%), expected "Not Fraud" Prediction is "Not Fraud" (100.0%), expected "Not Fraud" Prediction is "Not Fraud" (100.0%), expected "Not Fraud"", - "url": "/sqlflow/example/fraud_detection/", - "relUrl": "/sqlflow/example/fraud_detection/" - }, - "18": { - "id": "18", - "title": "sqlflowserver", - "content": "sqlflowserver The gRPC proxy server of SQL engines Prerequisite Golang Docker GRPC Running go generate && cd server && go run main.go", - "url": "/sqlflow/server/", - "relUrl": "/sqlflow/server/" - }, - "19": { - "id": "19", - "title": "SQLFlow", - "content": "SQLFlow Build, Run, Test Before running the unit tests, we need to build and run a Docker container that hosts a MySQL database following this guide. To build the parser using goyacc and run all unit tests, use the following command: go get -d ./... && goyacc -p sql -o parser.go sql.y && go test -v", - "url": "/sqlflow/sql/", - "relUrl": "/sqlflow/sql/" - }, - "20": { - "id": "20", - "title": "Python Code Template", - "content": "Python Code Template To run a test, say, test_sql_data.py, we need to Start a container that runs MySQL server with populated data following this guide. Run the tests in a SQLFlow container that has TensorFlow and mysql_connector installed: docker run --rm -it --network="host" -v $PWD:/work -w /work sqlflow/sqlflow python test_sql_data.py where --network="host" allows processes running in the container to access the host’s network, where the MySQL server container exposes its port.", - "url": "/sqlflow/sql/python/", - "relUrl": "/sqlflow/sql/python/" - }, - "21": { - "id": "21", - "title": "sqlfs", - "content": "sqlfs The package sqlfs provides an io.ReadCloser and an io.WriteCloser that treats a SQL database a filesystem, where each table in the database is like a file. The schema of the table is very simple – it has only one column of BLOB type. All the rows consist the storage. sqlfs provides the following features. Create a file To create a table named “hello” in a database “mydb” for writing, we can call Create. f, e := sqlfs.Create(db, "mydb.hello") f.Write([]byte("hello world! n")) f.Close() where db comes from a call to database/sql.Open. Append to a file f, e := sqlfs.Append(db, "mydb.hello") f.Write([]byte("hello world! n") f.Close() Read from a file f, e := sqlfs.Open(db, "mydb.hello") buf := make([]byte, 1024) f.Read(buf) f.Close() Remove a file DropTable(db, "mydb.hello") Check if a file exists HasTable(db, "mydb.hello") Other I/O operations Feel free to use standard packages io, ioutil, etc with sqlfs. For example, we can call io.Copy to copy everything from the standard input to a table. f, e := sqlfs.Create(db, "mydb.hello") io.Copy(f, os.Stdin) f.Close()", - "url": "/sqlflow/sqlfs/", - "relUrl": "/sqlflow/sqlfs/" - } - + {% for page in site.html_pages %}{% if page.search_exclude != true %}"{{ forloop.index0 }}": { + "id": "{{ forloop.index0 }}", + "title": "{{ page.title | replace: '&', '&' }}", + "content": "{{ page.content | markdownify | strip_html | escape_once | remove: 'Table of contents' | remove: '```' | remove: '---' | replace: '\', ' ' | replace: '"', '\"' | normalize_whitespace }}", + "url": "{{ page.url | absolute_url }}", + "relUrl": "{{ page.url }}" + }{% unless forloop.last %},{% endunless %} + {% endif %}{% endfor %} } diff --git a/docs/assets/select.gif b/docs/assets/select.gif new file mode 100644 index 0000000000000000000000000000000000000000..037549c7244f2b0e5509335bca12e89f71e71219 GIT binary patch literal 71982 zcmce-Wq2FSmbNQ1b8N>DbIi=lF+0YXnVDl|W{jDcnVFfHnVBt9tdnJ z*U=w!m)2TU)q1Y#lKOtcCB-bx-+^&2i`bM?}#@v67|7vC+A^MLHCrdsOp}#&L`m0Y7TL)tz zHhOkCLnc->A`UKkCUzE9E-qRkW=1AX21ZT>CKfs-c5YTiZf0hpf4oTk^yXk>!mTJO z_K&{)obi#EIXT&JGcdTgxX`<>(AzqgGB9y*aWOD5GcYsL{fVG+bhB~Nccrs&B>lUB zsIjA=gSnlPxvdS+UlsKYY=1iOk^C9ye@?;L?w@6C9O?gKI`oFN)(o!tb_`7PjDJn( z??74E|2fp!`k&E`PKw6=HQxWv!j8&rcE${f#*Vf>9Sn{Cgp>Z&l^wT;gR#Dot%I_y zt<~RGQQpkf$=1=#){aO-Dw3@ON#Q5{Ha54 zZf?ZQBFfDAP57G_8wUp$6Vo>?u5awZoUF`jj7-8Dj6$Mff9HzY8ve94wsHD9*XWk`fuy~BiHEPj^*38f1AtjXD|$Z zeZBwkb^qP;M?-%d{!_evPW~x>W1BykclaZ2Fwn>E_qW%VXW-N0!~Nau&Gpsg#rfIk z$??(Q!T#Rv&i2;k#`@ao%JS0U!u;Iq%=Fac#Q50g$nenMK!0CvPj^>mM|)dqOLJ3W zLw#LsO?6deMR{3iNpVqOL4ICtPIgviMtWLmN^(--uY~xx*qG?3$cXT;(2(GuzyNPL2-tcD6RwR+bj#W~L^_MurCZdb&E=TACW_YN{$fl$8_}5+|ujjp~&{%JED0?wuM}8T$0; zlLq%moA&LS7SHe1>(|d6-qY7_AYVWKfWV;OkkGL3h{&kunAo`ZgkOnC$tkI6=^2?> z**Up+`2~eV#U-U>l>R}+dI2^`v-?d$0w&}=NFe(*EhF!_YaRxz~`6OxA)&4AOHkBzC=w< zZvZR?wbn>YZhr_WsZ@qUZQfwSXBLZ%k=p#>7$T7;YQegK(F95*P?^?fUEz2#qv2SF zWPQ3iXVxsWR>HrponZJIlLQ_1vn>Hva3g&GF{y?QUNP1Ol0sn%(|T3>uw@mfHQ{ zSW@Y%3y$N8j9)C4TNADI$J03?5d^Yp_miru(&ajnZH?#4HHPEPNRF5jt1XV_Ta)e0 z*W2N`No`0SEw}q42{gJkEye-hd0Ncb-#gnL&sT)1-+$c+0+4VaLkZ=&I$rKi=F4@b zyE@;1H~Zt+a@}3O-=6Q!x2L4}3{lf0qQL^pav~ASsLDCK z&&~nC@zfRI0_ZtswVzPoNaMC=Pj@PK*wu;}An5N#EG zth$W)_%m}H<~EKHcs#+3K79}FXTF^Q0$_&W3_MQ;*&Heh!egRLvE<#7Ymtn} zj9gvCa^ZE}LUIS!l+1lJ2Li%l7f<=?J~#?3Dqx=RXwn=2!-&rs$w1!&=wk$*GT##` zN62anqp)L+K@O&EUEmT%_1!`O&m!_ma9jt1ntVNg&k#jQS2->^Pm^{56_&m(XB_Sy zE<5AgRxPLSwo}}#JucD@{=PgmJbjqSSviZ01XV=+Gu_V*zJ5M1u3J03p&pGWYCeD` zPK)dv$kpkq!mVHyGtN^Ovn@u*O&@Ldrs#xrE{F?}ox%Xu{Sgqv!o2P!RxEda>$DWu zP5MgJPrzExxQeHcCb8lkcGK>Nbr!DLMqndp+ajGhpGME|XTgS8eQEnT77a*9wc1$q zi$I(R-|fC*-ql8AJ&^xy4E$0f@uZy@_QQU{>QBQk|KVjKl_l4deqNwnv_%Tr)p*$OSkXk zY4BjpM3^L2j`(jIU@6ezaoWgEh`3o;erg3*y;JVrf8|B%zzAZae|H|wPs6k<7M9_^ zi$3GoPgW`EBB`T~qzTZ2+=POKe;NeGy1;~n$>~LkE(;)R(2NNl?$O~#)=XDM{lJ2i zxcOBRvU$yUY|CdqfW{|7we{tq5eGeL-vO3R=t!|6`co|W3fpYwHd$Z z(w0nWYQQY6y}v)rDmZZ+Sa1L2xkMNTOW$c66XmRh+VlJ@9a0(X%JsUkJi#>x6=&^! zb8yf|WS74iM;A68QOuq_Dr=z#T1a1dD}jbg$xry<;0MmP?ed;FW6~Py{$Q0kzd==C zvGysg(Xn){FLfYzbqrk4I&~a8m%SUnmQ?4YNyow2tW{MeF4B%Z%dWy@6kh9>VWkJL z2vg_ORTX%8aq5o+@MFA-HN0`mh8=|`03=;!a|jpRb*L^~58U&>=?OwSau4QPTjmt7 z>_4&I%&eU@N^2+V`fq5+^*M~i2#lwYXaaq4Z5)ch*v#}!&HJZo%k*4506ZSc{FO~q z)swF)yx1OevuP@k_I!i^=n)eQiVe2LlWPwp+)Pr6w$X@om%C2q0+~8ghI~l4vVNZc zuq#a!Gm8c+T!jr;^c)hPM))%4X?A6SPm3irzfu+bFzGarX~A1B)Olgfe(2FxnIoO4 zFjfEh{(8ikYS@V;`D0m}wDm{M^P@?eDExOFw^}cf_miO&kMDMumDRlNr|#-_^QXHE z9tQVQF02&1(Z&}>u;cmChATQhhZz0v>n-cwS9#)IBL^|kZ#)eV^zo{mdOxV`9j8ZW zlEf0bdJaBUpP3G3*UqQFw_PQCSJ|`AVlswTgFA9su;Ya8ABCm5bJ6V8~m8nl}w1}rO@-SAq$tIm%6x{Qq_It@uC1zk= zb|swQLgbC#z2IBhxotC4{%cc1>Y?TK2du>CYunT)(3_=Sf=A82Yr)>Ivc(ZdJ~7e8 z1Qw8(n~V&*ZN{vnA560iApy@Bk6{;&IFsF`FwAJJf#!}x=2FdO)w!tMd}xzpmUla( z7`d(59Wr^w3__wWOnOW8+o{jx23_P|IQXtF4>h^eR$x~UHk$63pt|`jAk2bs*b=8vAOUrM{&=UleQ3$x*&bm3#xtTOQ#P_X!xp(nEn#KGeW-DLOzc!zSw$T?WWewem+Pg zKHAIP8O?qG2|vsduq`*^UAs;j|AEFfxh8FJ>Q|v`(Kj;=e3|DBfUVpk0Fb>!N zbYcLUuRm*x)K>+V3mD~v5HPnyfB+}NqyZR34n6h^V6qO17d7xvI8f|!fMknbKZV!E zpsxd}SGtS8B4^MKN(d-BFkIiDuX=QDY~YS+A_K0 zEJLH;Vf-#&9ePd#$aciOtbXqeB^?bVg<*_bWl0)vrJJ4w*Lx0EK1Ee=OVm~mh~i8% z*^xpojplm})_{s6i;Z|PgYfyB?5UojaO(e)Dgu)W3^B}Ioyz|>C1AtZ-yaz4{u2;b z3htu}MUz4r@sa}a4Mbv1MRJSKfP+C!NlkxAM0!Z2ij6}eOT#itL-0$>D@+Tg^o(pt zjtNcGcJs%CbEt>&xh+nIBnD`FiwJe~1uh3x;lz1|ByfdgjE-h(P(gz02`Z%o?x*-3 zs-+nUX85iK443}MTFtzAf!Kb@=t@oOQ_uW5`*Wo=6>~Lp{Uz%o72=jE8)V!P$SRqc z?3TT(o-vfFdLfw(ddZCvQgz^lEx-B|nCGX2?UJPzQoJM{*jxuLjF;r{;Ol%eP zWzo`QnUG#dP>1h5uEjZ^zROe0JxYbdb$uZ}JaOKq-8ueRaC0ra$$?35s zJUK2A%@^TN24nesQjHQkZ4e&ij2^WV?KKh_fVIXFY`ZESjm8@3DhA6+Xb=BiA?_H^HBGn!6D?7^L&s@#@zsz%22gxk@@{xWoEi+h%Ee7O6WVT>)l1nWY|3 zIAhI?uN_~`+IR8V>&DyhM>Q>g79J&;mMxM}UcIlK zzt6j!;d;Qdx>3}!f+DivCrX;vdXOhj_oefWqtFb0Hq;TAVb^|Mu=aqxCPSknWUtq5? zqJHyilGW^S^XN$|&bdCT&BGfcU>YonNK|s~;FIY~DJxAnA0z}D!lD^!EN_;1Z|!&+ zu;>_~mKh|{$~v`#+8*rGizvg<8rJn3*7vNHtQd&rX{dkgH}mZC$8QPT82TVO_?05|g?J81 zSB&_CjVhgw)bsi@R`e5xc9Oh9nIMjL(hPLJ_snb7x<$4zL}IhI_c(P-^lVJHT8)^8 zj)zr@Cr^wtosBPf2KH7=uH#Qu;|JkejP3KHWxY>)+8Be)$f?jA+uN85Qk$~DAG_-q z8PS@2#z)(i7NkL(X0dGFdY{ys$O7We;8#r3dCu5n%*Y~EBVEqyxDLm^7t(Kx)oV0# zZVZ0@J+q=Uf+jnvgEy?BUanO!LAsfNs$D=i+06{eABCW~y&u8us^W4gMT` z*=@5_cB*A=0WP;em9vCaGhqCE9^(-jbeqs! z^NX7sH&L5+bX#aTC;ty*`4yrJsSSq zuR6PMxLZ{8`%kj_N7nmq+U1|C_E~gR;3v|C#&^rsn+h`ag|m=pZPw_c*JQtJK5QQB z+#IyiL3JV>$k-en`5q2#4wPB1|M`8W=e@~+yl??X~x=y<-)PtQT3iB)$8B zwD2KyILdcC(KED2lh8ZbuTZt!BUMDVdF)+|o^hzaxQ?N zxi#zfcJr)h>wM$pyqw^|p8sN?3h7YxgsO5blxE>a*Tsx{HKcae*Ygus?+dHUwYjR3 zg%5;#3)ZLKXT{^+-W7i$)^a(x6seH|$VbwiP3?jS0IhF1? zi@-yl{DU{r!)(-*)HOUvMqu?l>wcc#v5WEXOY}W6!-JXtqDJ`=&>F3>5c!X z!{*6n>t5gHN@(gy^!mx{W8Vnn(ZL7aY|_M>q&u&D+pUHNp4{OovrYmRic@%;=P z1Pm5H@Ml>6o_!MIgBUlx?&Ska)V+5C9Yxwcx^}|4MNsA3%*_ZWdgeCsPAG{~q0?%-jGqgx+le!=IGr!Zv zTl*P4viK|37(k#n0C*oW{cv2)XXo1FM~y^U?4O%EtzGw>Acz2NO>6Q1bq#7nYI~w6 z$Udq@i%x_3HM-{W-~Ro3Os!V@`SZN9CvFy18uTe~YV!G)P=-+Cx_K$Q-*z!^2}sDP z>6zKN`Guit8#})aG!hL?tbDULy(Ok9g+xYo&dMt)tEy{i>r#|WQZ5@IzP0ar^~4RT z18}FtN}OWh5t35SFtCiJ^=MB$E*?^^`wfx(X5$h z;$DGw;`&vhl`nqiM;0{Cus;<2^E(|H<4Hp=l%X@Wg;M2Zc? zcj6R~p7RG%L7b}xz6!xlGS-5eAnZ9Tb?(+hFe2N#$z(WF*Hr)9Fp8(a9goRDiX8oe zC|81TmN~Pe@SVZKJx^V0b`&a@rOySivG(E-Kk(16c9NBZ_atDk)OUMlpDm zkUH6dC4UN2RKY1qdDbdj3_zWui17lBvKOV>(&z6QjkBHTCrLA_tI7&A!4+rwv?-E% zq73C1xgq+&mE=W8{W!?D+AS$ya$Djc_Kq}2qayLIb2Z6zzE#gF%8z%4E_QATJ&w@j zax=>=?8BwOE3AY~jUp8F_is42Xaul61|kIbk$nIMNn`C;nWk>4&%;w;iU) zSF26Jb}^q%9?8S6_orSTbErbKrLCWWozuFiJPw95?&ZPg)Sd1VPM;*CUxsj;l{0i{ zpzo*Wv!0SSjLKN3k>9KwA~*fgWUbjfhn76x4p*>Zi7nnlqfK#ixo9 zeR&h8<5C~r>{a{Yq4&kwvFzQ^z+Yp*xw zy_-(Y=OvpEALV#h4#~%N^jJ8npA5#@=6Tv@A^lCdknp|h)talkiF>)HCjwxxdU122OdM7GawPDbUeB#m2J5L6FI&lnIjqf2@e z&rn?`7eL|(34^oSivn$GOR7Z-LDi58h79kxai;R9X4W4LiAs4L@vHuTLx8Wa(wPH4 z4@as6wgpNdR8-60b7TqZ2UT2{V#W{Tb^k&ubs0T-%dF?3lR-L8CN%j5&^GpXNjx=u zq^~+F5t4}65J%Q*P|Rx#WUE-O%3^hd-9|nNd9eu47xBlHzXIOk(6Gn{d=w+1Iv9(K zAiV)JVu z5ec!FmDsvO9o|CXp%QWJwD^SBwmk}pXAxc!)Rf8tV`9ZsP8n*J1go4e2wUQD^`#2r z`T!GFG%i^fc)O70fg)f-il~ZnLfE#aK82=|IM0Wj`R=$ex9#(oI1)>6Kt}G?f!MTr zon5A5MS?{S6RpB4Movh_F$7rYkl5WlcG7F1a0um$#gb$!bSEh+2uXdGS%)>_#c4(? z=0$@2_kBK2205L4PN0ky3;wTo3XUIb(ozLPP>GrfLdFs~5Y$0=>|cnbw{M1>b*c&k zl#Fi`xs^e+pR$n~BRL;xDjm@(?Bp`{lv~iE;+W@(N$QN{Gusv)zcuE{-WAAP!;1Rn z-dDhX5^_nY@h=ZesFcz^W8qN(2U8QQlCBWc_y#|cqLP^ZMcWq|oJfrXS`dQs2aQm` zw{ci1p*(?bUuXgzQ6$lvDw0W2BD(ZR7o2cz@=z;C-f`h6Xu$^V0f71DYCoIFmB}V2 zz!VQpZTVvwJPxXo@hm)Ua$=x??M+7<3Jov?Pt-P-N&mCl60~`4z|p?8N$=8rzIpA> z(Xsc1!EM}f>pqX8^Gut;WBq&!*ry7}iK;Zyex1AQ#Q2r-e2UdA>JI>L_JAYXLBd<@ zK#+3w!s{>w;9cy%h;a6yXE6rRSnVPha`xkHF^2G5?4ks44v-Kq`NLQl0C>h){HQK% z#(F!Zh%6u}*lYc=rT6gmIfuEom|{FG_6Z@lMudxQ9HH?2aIhttR5k|mw0M=|!^or% zQkbGFc@Jm|xyCiNn3LNt4jJTHM#RJ%kbkC$b&@X?P#{_08Oa<{2XIX}WU*wPTOISj zs8866*+a>W3fMbJp`?eb%z)V>whLOznx_ zcPNjrzAz-^SrsLB4m0dLR>yk;(3nMN?!<48i%B^8wHQaTXj=ea)`teLfTg3_SJ2(= zy{9T1^?(r-6noyO=63V+vLJNi_JwjNLaA z=2s3V%e;piAiF7?F2|nLQ%Ip?hh z!qwbMtzC~entZeR*dEK6$xW#=zvuG?)}z$}+; z559BXVfSqqA0P&(jo@nzCtY-xl0iJnWNsJafHNa`01M2Cm7C2Rb43)+ zgRN|V-{-cVw|xQ7>$U*s4dnw6u{|&e0>@V#>{Cv1Xv7pR9j!u&*2UC)9?cF5(t>=| zh0@r8GSGvz*nxJ~`7Nt=b-4BRn5KZU#CZD&oP1~cCNzOH8P8V>fJ+D%(|eyJLdyFb zznBAgl!LsLGjK_?_nQdyNiWZ9X9kKm5i@t>Yc4v$ZarHr2J-F}UG67-{ZiT8%OAOD zV){qgyE{K~vB8NDpH^vO^}t9*!Njr>etv^^hgPf5{cf!(pXmRFBn&IIgYCbIl}13| zppOMkKxjii1f56tluPiDO-!(NJDE*XML^0am$nY<8ZipB!Edg-i^4o=VGzk>WgbnF0f{2fN2kHVVjkTP37V9_-sRqag27jf zyb^`1IrBYhzqD386%%#&{n=dvY8|RQ0xB`mD_K&ukK`}m#MyBgZ2Y7g&iicqJM}qk z{)R%{hCFMeUxd@M$e51?q(cVHZ$QL9z@X6hm7|WN$ zEfCH-U?(_`swj_-rxeN!Yf1d3unJxQ?-kMxm+IA(=(t-p{XU2FB@L2WlpTYHft-Sh{Nb zM4C;7>fHINQ@g@_#{6x@Kb8swwv2hNzUm$s3q5_+E;>{PKhljh(mN{@!7coH_DQD| z5dVo3p1=^TMjwwH%&uc7uBTVbqCf#WAC^5$SjN80@*|>&B&a<#A2$z zvWeVkt-va;zrB;6s!kQ!yKpnUUR?Pn3~g%sk0U{Pza&7*G89OFYF&0>7U)s+2(d zF!MAp3l^pfQ7j43pbWL440Waq^DYSsrVNj!3{Rqr$SR2_qKvGfjC4NnZOZkjf(iDTz6bq_QhxIskS;OEUwF_KGI4YssF@hXXpVwl-NK9q^QDH_3KQT?5 zKH^``r#Gdgv#XTx7Dm}k+o}9!xnk)C%2WXaXKCbRy5GwDsj2gXB^M-16+f5eY*7~5 zoO+s27urzg8&G>$fX)h&P783U{n<uucg(%GIJ1$i*E>_X~op!9Qd`v^Q7T0Qmg!V0$rgx0CwU2fZx^gC*X6A{uRbdrK-v`dOwr8 z%XOW~^*p+bIqS_SYr7olwX@5uHS6tH>#eKHosY{M^r{_Pn;m${9desp@~d6?t9{%* ze|lH#v4ica&>yN)9S&U{Ib3bJ({G8{oP4L>i>E)(xjJR9IxV|8HmEw-Dqh-B^{!K+ zcp6xChjQcn+LTp!5$eK%C)M3%`GX0_~4@xzN&GPF|l)0$c$-B6SH+^p?yor>!w zD(g*#k0Vg^%T@L2tL+Q2-8DGl$6fUYmK``@4T#weZ21~O#0~(whQzCZ60ZRm-Yluy zKx5RvDAj-m+<>C(pk3_X)~jLjYal&pxIJq??2;MoX*_Q6Oj!eO?oe4pw(J$UupD-9 z^^EXuHK;r_&=|L@vqhjgP9XG!hU|%;7;dSPV>9x=i@ewOHd!pO)>>@;@luFmXmeQvJ|3@~Bt@KGGTzS)08s3VuNBX?vb@~Qh8;y}@TOKHoD zm;j|C zPDI}xHIJ*bDIRgEueAe0e4g&Lvzc^N8+DiO^+}&}fsY0hO*%a6hBVGbcI>v zqi9PMXLBICEeuenpWP0{#S*327Q)%#jm?3D!yfcxA<%3m$Kk}_@{`8J(6ITZCC58A z8(lFp9rm#GzOxM;(3uD5>q1Q@8YZE8nh1#f^ZGCe3k+}Mt^Q`|NQL7 z0psuTY#{d>X4~RZ{v6KZ8X@B9-~a4C?HVBN8gEKqR+ZIpAT@1lpV#%Ez!JUgT50{(lo+syCX6RlP z)0WZR7VG0)5yD;B{aUf!R(9+jZ~I#1;$EHXt_6ZCduyx0eydJ+t;J}sljg3!cdz4V zui$WRz~gD$Zp+&T_*stCX?m1vzBN|AHgmMM2=G)Sv^RvjwR*g@VgzeJEVuW2bd0}s z%zJdMzjf|=be+F--FtMuy>&x)_8`3XV0iZ8z4wxO_R+lev3U0Ly!VTE4oJTbD0vQQ zz7HCD4q3hrIeHFzyblL>jzqkVBzTUdzmKL*#pHMNstQ7uypJJxj)(jU=ZX*b4p0Nf z{fl$mp@{xBoNMd>4z5>JzQezAt`*Mz%DFzUyx^4lC+C{K^7J2^Yn;!&a;}U}QU6QM zHPyBG|4+^}9AEPP%DD!>fBEm6t9-uHe{rs3jphI3TvzMO=Ktng_s24&o2xe4UGM)# z&Q+$h?r=Qizj3Z@4X5)ZO8>>VX34fUU9SI=bL|O0ApHI}=Q`EVdVf5VFZ~bBwcc`j zs&adO;}GS3BVD3;{xaamf$@9O`%4+SXTcB zi5)Xi@INT@e53Fx16Vj zK#2mfm+{}MZ*77S=<>t7-Xfn<4D98<^<$viW`2GaO+&JyUpX+Ls%SG|E5 z`LrhmvSXzT%8N;qrpffU zWZarz5-hgwCrzaKQYu8^x(Ln^u7G7|N=qBJ1goYeXq&Xjoqri#!F~ZgI_zYF)`=mI z;fND-1KU41MvMifCf_wEkDLFk_WEc&n)*SwQf%e=%c#>I5}Pj&;vWj-jy_%bLt_7} zP!aQ<^yaJW7A8H3`SupM9cdj(@hQ)s+r7yIvy|_h+7IU>S->G1qgP+Bh1P5b4Y79H zTPErm%ADPHkMonbbb;?AJx-5ow5+o}{vcRg!sq5ba5OVJA!S_OB!7@Eo0;Gi{6#D9 zi|6fc_H8&QwH&Yj+M40O5OP$tzTn46C?S6s0H-LVDHzxu62cC)NFchR8nI317gPdU zIQix!7_x9*zZhW^mtH6+XjCvY5k(ih1YAzTY$$9QpkFkUg&;h%B(yw=2TXX5kRMb$ zDjB@Gm>seQH2I5A%siLUrjvf!jc=U4vD?nCAMQ$EnUV%vn*v%EsJ_XB@2f?*w%kpXD z+Xcb*oGkaNwTJRgz$ZT3eW9eHbM!NhqUA4OlE0^vdLqTP_CO(RxR-cYLQ&o11|!o{ zcr(n+e+SGp1yuNjU0J|~;-;Zq`zU0+_*GL|@%jfmNzo>8P>Mm`4Zqf^)uzCFa51}~ zX?fCjeTv^uPFhmvh@)Z-DYqD8+v>Q+hQI5@gMoC%sL9p&;9>TGVW796~ZD+8?ucqD--)Rytu0#YJpxgdUET_4n13vpxdHIx!>f z^iGdua23yvdqBZAn8hnT`yv083bTgCAtjjtx2IjfyXOyyPY`-41SXovSiq9+u_JU+ zeth~U78Vfp%1^?hta2qd_T0u&j+e!eOrbpL4%y#|?=4jW3+E36zxBTyGnVF9i3J1K@ zD~`-xlvYp^(!Ci~%9*Y=*IrU%=~GfU-6%D-9`lyysR!8H68bk^nz=*0pa*+h3(RG# z@}|JSpM}s?hFM?g6kkRJF=KiMc{4jk0P3-v(GI@+Ixrj!H#Nm2YrCYEH8ey2??`bD`2@g`VgA?K7+=H!A{faLVcfgeP?-V}?9mZ1?z_0y$O1A7$x$hdyZDre z0&pp1wTB|k$K{rc2VV8$5P8}YmZf~ljS5jU`PCaTJ*SbFnJ1K1|LGFj!^JidA54s72+s&*Z=)Od}>hCAZ3m0Kte ze5eC<-b7V<9!qb(t^vEAC#$`ICJE?~g5uyf5Q#YwUEz+1qYKL)CxD|AQqP$o&JWjKoeDpUZ9^0%^hC%L})ZLRi+ zErVZ%1D96=yx>TjarXB$I!P!Apn1VR4>fB+soJ}8eltf4Ax;h@a)+DK8GH>I(AaZ= z?;|a;5XAW@VVj~ZC(1sX#2%MPu{Zlke?#=3iR=^dfQGXz+&RB|(icjM^MFB?=IESF zUh?o*5w7wRj!*Au{Vmo`ns?FLQOS*Wq`SBR=3(26f`W3AJEz0wXvkqfUqHwZbj!-p zhkRgq^0TSUIw)#x#{urmNNCx}XdkKro>U`wB3?Y67=heC?ZQ6wTN%oJfA$ed5)(ep zhuxo?Ev698=wJ2l>=H-iRh96k1otDD1p)bR{Bx)VaW*fDq?ig0c_i3MIE^X6RQC2O z(Xza;a*(5DzL;qOY4Y*@gou(NN&ADV>H-y$ZVPa1vN-*A)QE@2iA9%G_$X|@F{X3G zJR8W+$m`+`P!SpiFx5q5jPvD#UIdGfe5dS)5c&AT{1qbeocmQ958yBnb9GXLg%Z}t zu!~4}aYUtK2;|Ep$=(eYAR0$jOT_8da^j)LVxRr$alT38e5#eAfd|zunT!Y2a?F%g zGOu{kT(umm?O3;-6qR4KT{NYmHD3+URC8goYKjza2l8BZy%+hDAil=S+jc`@BaSkG z_84n=*Ccvs7-#rtt09p+YYGR_1jYNhXZ9kP*uZ=uV~3gbZbSz5KB(`;c+4`u$1rMX z?oOMF6Bod#_7Y292x)FUkOSRi^= z7lBiu_r8hjdQyIO*(Ck-TXHD^B-mP}^AS5#@Sx`3r9K`<*Du_4_>^Q4ILSjA}Ri zHQNG3=i^3mmltq{rK0xK!Yan=`J`D|t);E_`eO|D`1i*9y*^zqK{zz6qy zo((W&XFEf>@x6QAh6NApfxeUWMvf$6ZzAu7e~R|Qn%Lo&DC$K{iU=S;%xw~+>_ayR z3?!Gy<(KCKhb_Aa8rRE(4TR~(9jbw5ZOBCqBJTh4T7$#)o{JiWHb9Kag#C>-4_!5G z;43%UCt1(DPaJ69@MUkq!%OmD1w#huhlC<{Z}UFGfenF}(rQtRdGm3PV~5z0!5Px>$6V%5J^Svs-d%foqrK64r;iAI@@fMJT#EyL9o`Z?zDgDtdiN0WUeTQ0auRb6x7x1Q~0-`KY;~Qv6a!5ovv?>^Fz{^!A6Z zoL2(+^X5Z8>y#p|JWa9psix8M>a)rg!1!)Ydp+)yqCVb>$$(Tdr05Mf zBD|EsJ|#0jS2elFjl`m)@cm(m%0XD2C2%Zd3TO_XhJ1o|N(nZ~xj2oQ0`f#6DW$WX z#JW#@^j@a`L+bgD*@mLGcB=39#q*g{)y4cj3FQOAyK>PnJ;nH_f$?PvXH1Qy(;n2y zsT7N))YWAwUkFs{*V-!`emd*?rateZUK);TtkC13(MT;?YP_qcw8_NL+8?WHo%>ao9qo0|Mnbddq|j7HI8Y(PW;rnK z@T8gb+p0;JM#DU=sewZDQshT@|9qWuKy_uVr6Ens>Ri&7rp`((hjLAao2KTqXF9uD z;*DL!{Fb3aXom!Zs6*ytoGGs=<~$m05*1(pXt|Q;XMK6Zr2{th*6%91J{leJBsO4>)s>m_niy+$8TFQaM~Sl)RvwwPDX!}+#M+F zyk2Ai$UxFaF?bi-KCC5(V3r9nWS+(TnRUj{ZK6H2=G*~t9f)ux8F4J##lay>rpSJx zeH@?MA@(cqXvYb0f`G;0%5|pLVxj|L@!S!yVA=TjHY0M5)JnNvhhN*=e~d%wRjNY{ zN#||G^e#Bny3md(KyDM}BBb`$pI)Ee7h#|#u6Q|@K;T9+ZmKlHXKKM<$R z2ihrzg2zl4d#x@+%qfqh-OPcwbv8Q?)X+jecdDd<+mKA4UnJgcZmIA(mtEyl!Xs

    Hb-6+V!2@W;(y3}G~^X7mQrG5XpO#Q%CQ!8|?kCMMKlmtPG znRH>+6Y1Ay(}bBc8GtY4f&5&K3)63GJ*t(Cu6r&v;%Op?wqoeGE!Ir3z;^))x#OV4 zI)t&eS!#P;YAU0~R#Wq2Fj*=`lIZjKPV z$;^V^GX&IM-;&rm03+N#**f(PPu(R8c+T;3y+|tP9H(5rBk~hEPR{#2)P;i6+XVR3Z8hYpPxr0;%T{{%rCE)xLx>^_hS z1iVf-wqKV``Il=@AkR#*AEj=((9vU?V&T@8ymT@3g70FWT72O4uK9sss34++?T>qo zzTZJLJ|D+~aKD?~oQWAhQi8&-D#X5~kcVOUNsS8yW1!#v!`@p)#hI?#+E7U0P`CvMZo%E%Lkg1MNeJ!`AV9Dn1r+WCcXtc! z?iPYWaCZnA!l`7f-rc=d_g-@L+TT87e1m@se$*K6TRitO=RL1khloZe&Wj?!*$&|v z_ls5t^y%}Dg$90z_tUHkOumAoUIixf1Sjh(kKIgS(0ShUkIj>j7VKp4=y?QMeFSbGYQv`p|1khwJsKOB66c-e zzJj8sR|H-Gatj!RFd;IHDhkOIHOtAgviyblSQKwglp+q`R{@O*HsIU+Zon7f`}+ev z3}x_piTA$}(Ef@t@Ft@nkctlffie(8O8hgF0hIJ33}s*oYWy3NK|RLkf1nJ^aONej zmzP#6YuEk|W#B?iczX|*Vuu1l8A!U|;?j#iHv!UySky|6$K{!_c(`m+Cwb$~)o^Xh*{yly#% z;FiCj434>@{QrV7FjTXkR8Cb9+*lldNxX#b-c^}OdWr`h;=QZh5lDIRKP>Uy|9>m- z+VwDPCI1FxaIv$M=7r3>oem+i*v^2`7j9?%f->02j?u8#$%!{9+{sOL+TF=Z_h;VC z&yKU$Ey&L;+$}8rE6QN6xO&QBucU6XaIduKVt22s4Vh)Xyo=CszoL)6XuooZZ*RY9 zOpfKCdQ!vkVC>_JNzp;gg45nX?Xv%m0YA?25H{fNshaH~x7##|H=)~NpN?1_HD7J+ zDL3C->>ahj0a=gR5Q(gg+l@|wk2~%!Rv{f2(nH5x_-j-ZXjso#0S21qxZa&@eCa2s z)Zuzk{q%WW_5;&w<=q1)o$fLKR-zTI7VXc84#QCLa_&)y!HCbJbT|jOqhd)>aKmY& zqVhe6ugK3Qn+-TBKd0(H=B>%OL*wm!W&7x2#%O-<&8%5ob4N3QLJ2zq(n4(FoFk>k zWtYtsOBK<>-`wH_;P~N00iwU&;y)LZTK4`8ujE@!&_At`b0d2F_wh=)DPXtwH=`4U zNYgV&pXV1|U@b1M=zLk-*i@g}+TByw**`iKIXpT4%5`>eeM5V7`~BA2`5%DpUqY0%URoF)$X7A)9bq>YuQ9xDgv*EFJ|= zTW42DrcM(~p@Q}fjgGaBj8A^9`ZQHb3Z0)_8t8bxSiAB?e9d)RcEfFd^N0@N_{;8D zUBAlVh0M+E)T!h#(hh&fGO@20&ILH%FQYYpfqaTe3p@k&goN zC|RkW-baY^SJK7Auo(`-tYU2r!B{*}R&CLFsuF~rrMrsHU@RUUD6{N9wK-}IIAyh+ z!>90>0yy~PL6v^HhOD-Sr|75W{+~fFdk?h&>5H{K8cobaDEpLP1#%t7jiLJu-1mHr6`WkZ6W+*Q!8G&m(;1@t;^>!%gea8 z|FlO*DeLw4o{W?gy6j16MlBD)zA=@AgEzyGghQ=+1Mg32ueIVyjELjy4X=y#;aqje3S~Rhq6rmhaDfkIq~I`kkM9YFk0dqoYL9x%7wC(rupG%|91ZO|+V>V6?P+F|GdzvVF#R~x>BB|9sl)k<>yy^BnDbeq?B>Y1{-6s`GxmE5;w%_ z;ev^!LJ`WjFNGqc6G_`v)5Bwp64X7l=@JY!ml{9l@7m{i}EHqXb+HyuTP!6llETyo4An+Tq|o& z6VLc*k!&kAPZbHOOHb=}kB#@DZC*v&l^n$+$IK|}lKHCD;-5HZeXy%hkdN$HUxhBq zEDca0HcC{*!&_<)p9e_K{Dyw)sWg6Ce&h(ro`Y4J9<)%tJH*}0G1qc2v}-2oe!8Ab-biLK zObaMXy-^nYsQ$(ACR9X-a5Sub4>zF@AD%6vK#L!erI-svmO{g&*3*-w#{P#Kpvoso z0>|UC!iL#r9~UNU!V@Joh{_mhL2;`eEn&jeDPs=*%$~HDjibFPe>rhxGKVkV zVb%M`O45qTih>Lgt?mhy#snzOi}7U$0Gpo>yF#`^vR*SgdctmpE+&a)vTcD;A$n-Z zQ#W$@)q~IO;Zl3_wz*k3#;lL`t-Q$7#qy1x*dygW!&f4;!X;IYWLwiz%fc0Ul3kEe z@=U7bGzidaPn~r)$j>vJ`#4xFC#QShdE$0dqycd}GYQkPHFwQIEkfVy0f&P6aztKC zR9^|?jUw7(X057|$;PiJ@6!I8nh|OPBT%o zbkXo%IY2;i-3wd+EM&1`GTh>2K90I}OBW8vC#n_Hw2TfyHkd*s;*aTY)j4&{0d0mU zR1PqOO0^j9?%-8jH?~mU~{`3)W@od?^W64x|H{0LV^GFdMXc_TsZ+G zaBxUPunnYjtyr_d+{N8LI^F;2$@TA&_8_2)zmoR6vVPg`5G{dSPjkp&r+ao&Ye#2% zdslB?SWo}ZFl2CKeBv$9r>W^lnuouB{vV`3c-Rjd62dakZ?4^cJGIRITc;KR*whmD zV`_;_6iWZIspXf=DmsG9j-KAOuD-#c=5E+TUN$=MD?v~d;mc2g;N}W!B8L$Kt9Oo0 z&xDE2FRz|oTz|V$xcv@5JBNo&YDxP(v!vLEG zxyYqrx+kn&Dy;Y|mvM0%yY6st0`u00y11`DvupLIiQI%t)e=Wu%k0uu|FO67SAyWb zXS3=bzjiZuj#7McCXtXzsyskdpW`F{TviRX81xpdg7idn{M@}O;^ zAhtQ#tS}pR^K5Mur5%E}mpZb!i9NG_a_#2oH7~Bl;<79eSO_Y@Y&cYsE^8TpvnuDn zP_{zrp>UI`9EfYPs-h7=KLAch=~_Rlnr2QqsQ#O4_fKDK@38T>hvlg0eCp@5y9nm1 z{r=jG!cctNfx*9j+=(O4`X&xt<(p(T*Ke-}f1q#u0|xfLqHp~jRHrp*=`W~GW%&9g zp2p@jUhH;Npkyl-sw#*VG}Oi2IrgzfcVglhhS6NijPigfk_dtkd=F^pGt)9h;{r_| z^DfsJPtV!h&{w{*?U`F9ELKMNV}xRCfT0Kt3PMMKH@aGRZ~#ER-J6eGOc*?xgz^@L zLSF==4?uOICBOGX#63lX_%NPxlAjD^1$r>cL#~`oeb8YmG3?Ac-Brggav*fkowo=IOGGd|@L_T&+Era! zQdU(``^$=WkCWFnDPe!-V6ocJCH7G+uTp#qox%q7sY){%(6+T3%7=v5xdhfWP=oWl zKH5^OnT1i#c@;jb3~6D0mXyxyVs7OkXkWN7bmVN&!N8mZ_xLSspi3+sxg%CW7ssL^ zq1X3g%SOthC!0(_M*9q|fHDWYY);&cwsRo?b%r;iub1#AHt#hapEz!j?51~WyJC;Y zrit7S)B4;SzTy@?c9kj5=N%k<2-?KCQ5XQ$R`&7;mJ<=$p^n}O<#2L(x#f##l5B-L zIY%BS2oVx4ag#aa$eC@Ji2HV49+@6(TE=lpaPVafOUG-)yok_se?gZs496!gl)i|D zy{pr|FEf3o8LpJlV3eZ$Ar?pNC6?eNuIQxDtQqRV*hC-C>J>pBZL5Rd)(j*Vg{14>>urFGq5pq|`E|*m>67DN%yKdt5RL^?IoiX~66v6u^eY78} zPWu^)9@-DE)~X;QgS}V*Lu}Jl4is={5@*9AmnyQsqQ!bXLt;eMJUGHxa6BL)9FOyd z%Hj`#$J8~gKVm744oOcwoh*cWg6!)kjA&9uy`AE>6YUt#AN9Jx3R1I{&ZXTSqU#!J z3I%l0ZI`Ib^F~-BAiFm{{)9C%yj(qPj4gFG;`)A^Po4t?X|(_%C`e5TJ{`V7wcqI-=>aS|4>41Ey2!6I@~fiIYCD0_3&E9KJ>x}OD4g4wU2 zLZz-13d>_tkdVR*Y~0{tNow|fMX;G+_ytuq+QoFgddEYrIL;jAH}C_`8=^2vk|m44 zcJSxTr*WjoQD5!tT0rOa4@ZeH%$lgnEMLpLZ0^e8GUb(h&}E-6Cey{8rTCZ?!!Qx0 zLxZ<=7TGE{C~0}$1TT?f#2(O*w5hTM8hb~hR(O{3N}mC39d*?8&?}X`El>24sf!-6 z{aQ9ilIz-B)dTlDU7I7H&?<1uKlkB#EWccm==3q@j|H(5BXWfVHG0+ibCmV{e1#(L zc;wX6jPlidMH-!n*ajYf8`lDrTfvcpx$02f<{t}UPuZ%ge=dl9IR2C|SDl}P3(q~X zl46SStYDU<{b|PTprL<%K1s2eLIv8B68q1&5lrleDp?vAhpu|2K9!l{}d2{Nm< zmb#J5n!&CgOG!fasJ4SN{Zc-HbPtM#ub5|icX@;Tj0LtxhfqEAIj%$2pxkKNhb-GpHaV$;o+1O@ITUftiOs#?FlHSpqW_yQkZ z-loo5PDg^Hg;WHt#dv#6$6EwJRwfPzI3D!tCZ5N9h1pJlY#Lz7v5J#))%CvI4j(RL zRU<#1;>7$PydF${?N(KN7?YLU&h%8!TchRlzOr~@=S}@!(Jj}H>%nTwb*O>Z*{6wO zldOvRVKF!Ez0;p8gWQ7w8tk_=nDlYeZ%32B(jSN&M1{ATz0PXn7*K=;v}=vc#4 z-N?sM=i$AVw>tQ>-#?aDknO)FyPfG{p&GYb7uY^g65sG`4y+2dAuT+WgkVFih z#H6`n)$yD&^c@_d9)0^FoPYgn7w$}qzkNigmBB>DWuqrT1eKLvUd&$bBrvbRPYwHg zqPLEUT95STDV)I6tyqjCam#o@IeNF{!W9on=IKX&p)p^w15uppo##}yku2D^qITcT zjUxz7!_kg4D8w zAoyD!L-%T-@$rF)F!y>U5awQ|69;9&+-t3%T$pTuR6G*|DZL6RCl0O@2+kk# zdI;nK$sl;gBgf=BKaLJwpYm#qM@;HMc%h&i=VDK22jw{i^fI8wK<&fL;Kvn^KNG(t zhzK1u1J2DLbk~JGw+mfnMEN2R_G&CN)HH0@1--W}Ono^-U<__!J4`tublWWCJU;y9 zO75U8{P9@$cPNZV;%4->#`BBc==T3MDUhg7OV485(FucvR2 zA7iK#o)><6^z)SQ$n=~HruzW=tmFXQngou(mh28S{Kf_+d>1t~=L!?)_$2iO(~Ze@ zx~l{n0o%|fDgY`UD$$B>2P>uklTMz`T!S+;N0crgV+Hhn@@hj`lGCpzghB0IBqY)v zSpt_BFVPIuS4Y{EIi`fR@=UrPB>r?pcZ<_3#*CsiWt#+%;Xw)q4{5Aq7332$y9a=z z7r$qc3V2D>a?QY6?FMZ7PJa~KRZ>y)eP`%GS=Z)%B9S8CuFfRK44-yCf* zP!w!D;F<|)ZJwAAkq}|Nz(Bd7033SP`=1bxy6sPaJU)h`YrbV)%=gLUnA%su0Ro!7 zgt;&jCWRosJPbUQpFpw`(%v!TIR3=`hEG(^6Q6;^_o)zM(LYjbCUiOe-iP6Q8UcZh zmtV?_TIJ&>!AG3nSn*O!T0h_1d|)DFZRj=-NS}L!=*K+gZ4f*KMKNS$AAx0yitY5_ zVGMl*rBE?sCR!RI3>ZT{7(ehAhJIAJA&jA4-xH(pn{3fFG(0iL!at;v6JWU{)h2g@GVa^1@fPaXH{mkocJ3T^W5aQI`h1A8(-u6 zY#%<0UyNtcoxIW>3-sd3pT_g9gk@<{X{BXZoBye0d6xy3RYjkez18oH=STP0uTg2R zIp;ki<|My#@BwP+j^7>pvRV4KH(=}Z{vXD(NU#Ik(557+82i$d68c36p)p73ST?|kszqpvU z>wR>&;5co4x#+T0a=GMwd2so~3&?i03?Z_)T7fc@Uaf}kA6~6R$g^Fq$80ez2e94TZ)bn7x!K9jE4|q*u06cjEAM5y-LIauxjm@cD!n~yx;(r+Y6G%= zJMJQ~{dUsF0NX4b;)iXPj>)s%olk21s@oskT`mW(f4^G)VEg@gJFo2f&3^6C_uJ!M z_WN(=)3*0_S3fsPFOOg~4jAchK!r8{>9H4zN;-h3uMJrU{MOEuikzBJk{?{|V^v3m z{24}^&&=?p;6XrJ#jnC8^sWf9JG(N=!!npkk(u@U>90CMClN$yZWO0ZHnek9X zW-G4fbQOQ>oXdW&fkDi)G#wgNzr13iI%4U1Cp?w})^ktxxUwN*Lal-}#49d%x8)6N z_iJ(eUEC;Qg-y=LH(a=S%aP-5o7}f!{l+yXAyY#p#LR7vUt9b~a^ow(a`H?yEQa^1 z&M5#V0M`uVFOWlcea6MwH*>1b!#8XjlyVX(Ngr2ofOb|1*aNlUa0vqJC*a zp^HgL&k+8N;QHf({8ez(2mMBH{q64w{^a-nH-Aq96ROn#oWCAtue|@L5{4xo`Y#IA z9&-LaDq$edpHW#W`JX`v^|+)|0Z@tPI=Yi%96b5zf{7slxeiL zKI()C)%~nJkwXK2QVEkyvAMGwy1nJ&`3c9~m!lXv(|Fsh4e@j%he$D z*V<+ZH9h$IY<}AB*WLu+RN&%Zhf>tPdMN$()pW^Y+K^WMsOb`{F|TGo`puE`BV*?X z(Es)5F#AU`cE3A2|Iv>r{9^|Kl=x@gv0TJoC<|5f?XWhnCQu>v^Fiw2 zUYHjM)EJ$pO?985ndd>ZoODB{TwCBFHD00XXZk|D%XQ2+zi4%uJL|nqzrlSex4*lA z!;%_-N5zY=$_*PAX^51p>N-=B|PqaplI zOn)v>;J=BPL~|n>alJy(7|pkyf#;2fIvB%qsOEbO(NQBoPK(>u&iNvqE55K(p6|0U zLHSDXy9-(;m0U4aExD|a|6alyyJ(1XuoQ)MG)}>SAxDe_ZWY5qi}vl{F3MYEBo^vq z*#|Fh!|N9aR`~SiSoU+V0#&-G5GjK|Ul#|6X~X(W7+bhcNPr?iUR6{Ock11+2~VJu z$)zp%))YS$qE{XpV0Vx%%vbb{1UK5U`e_O3u!il|JxS91axWNV;l*67C*_kRqOYfa zX%iC-M?YK}3|M~+YZD9EU4ylW5ng_W`$<^{%X|)_ENnyX>Cj-51q@&i93|+Uq$B-d z@@Kmx-|jl@*M6c3dA*{bFuDQ%zBdZN0nx?qnK&eg#|B88g|MDVIeoC2B(vCRGA{d4 zz$ydX1u^hRAx1qUN{(+CCTccQc169&un6L`HSJuGm4;F!Cf~)ULp>QxT!f(s+ga5jMvNFC zd}up43f|i0sGfz(jJfHhndWGATtvGCJvj1GdGhW-yG6(pR2#UZ9TqH%A5+_QGionf zl}jbBnXR&)!wK(K5EBOPR~8fN`c@6e%>%1Bz$^#%ADso5O*QLZ@0HNc0d5ZJ5+Fq# z*%(^NhYibc6$dDdl`QNuBuWgcO;?_{gjf^Jd*&^0=%;I#m&jJftam|f#~n0>5yw@F zT&ySLBsWQ2vCR6(K0SDby0+N6POO#0*wP}W{TDKN((R13LqJp3MplO!p)QfLVW`$V zXhh^aJa?Zcums6e=!o@vd|*?*N?d{eQTM12WdD59=~Nh{_t`w!+bO;2igRMh*NY#^ zb<+edX0_ZyFXmq7-`V0AREfTswVxLCpU<5waaQL&I;hq2xPRml?=3HSwelP<h-2UcmMTP;5^m!{R0p!9{=|3c&N~3VVUYXqpZAbq3QB7 zs;~QOy{dqWx<);>qv#{c&W*PDMb{Hw%J8P;NEpoNRExO`cr>=g;PJIb*oNss>ecAo zTjO44c$X9|lj76iKV~4HbRktUD*cNvex2fX0S4Z!;?!H^Tf)ku>g&!IA zN?=)bWb~D|W-Rx{3DM7NVrXyJDfM}fMTM-J+m8ADi35LnDzG2912j?wp-$VZ9e#FrEQ< zyh8VD^W`8d^5OBQZDeK*H7!zNFJhow_kF>e3e1<^sHuYJWyuAwk#m?O$)13{IKm=B zOeWjOB01#AvnvAWQQ8PDiRA^QavV=qy^+N-y(v@EBRI^mF%b*A$rQPwfIjO$W*oS+ z^z<-iRzv*SUVm}6$S}e&FM4G#{G;6TC}VRYzy%6H%A@7bx0CIU0-WKW)Az<<4A5aH zIm2Pvnd3U)XE7!@^{DeL!+{vraD(5;`d+Cc61N!$Rt9#X>r*Adqrf>%z3DcyDNgP; z%R#Q2jX*RzL(pzo&6ZnM#Z*y9`NETnU8Y5EW&lFE#@G^=d=-)ADI7eegc>4=cRZa}{0evrhEVl-hTI0g5QUSj(ry^bu2e0UYy*oLjegiwh+w=j zmm36q&S2YsOJ3-GDvvYlE{7{)NY^VwrXvK(wwEHt$(DExmQUe1N1(zqr})SZrIY(4 zKx76d-N`il;k#-13J<-sgB0rP!t_#t$MEusIJ4OLv0NapTp6fRN%rsaOrIhq3YQbv{aA#!inKu@)v?@Wmh$3>DqIbxwjPf3>hnPpA^ufT(=z*S z*p&ebhs89}9!$RR%0osoRWvKLnp@m>zUrDJ{70GC}@qnZlY(pD~?v+6j5 zpD8JJ2)Nb-trG;de|(5`S1#jL=osCcjm(htM(qQkett7~01`FFIC>wC+=>yqFF=IB zPz>(DUe_UNHTtNVd2L}yHR2PdJ^f4@qL;B zHdj}`No6(_%G=M zLYT^|UP`TX18$K#DB&0P$=Fzrkrt=M!^GvtbH??Rv^ZyL@QwAdW8>vm{9WH71P!7V)oa5k+VxC(S3(4GWd=B*Nb!S3Pq zObc~z7o)g$ahGB{y@+U9-xCxZ_>^@T`yp?GjS_pv@GvFe6S**7>$|u1r4HH0OHoI>hEHjy%}hY>Q*QZEs*>W#R>cj5NjnjH%Bp`OGSo1BIKTUk#eG5@0y9XP% z^(LjQ#iO+Bj>WF^_W70ARUXy%6Zwp9o8fvlv%cTIa%bEfY68BkjmqC-nSQ@u2Yf#h zhuocsqkO-7FOK7N4=|%gf;s@9UZ`~r!Z^n0E@W6 zD{m@dpDk)1S}h-X7aztx)}$z(;Bq)z1twNvUoHV(9xdOtV;??Zv@?7JmMg@(XkQJg z2clPqy@(J~2`{2}1zIYI+!c!OHoNo;BooI^T>(Y74?*F|tAGzs^vqAc4IeBZ@>0Ox z+1`)74PtbKN-gDQj)IDb!>xVg^@h=py3%j&lfO}$zl*T~Q#!;YAIan@Aj}J1X~xg7 z8^N0qsu}@x5|9tfhbkRIRjH_>Py*Yf5S`2dpR6HxDj+4G5Msx;T%84m(;>v>2R_k< z%HhxwjRgkfYl#)bBeFpGOJ@WK<4yd)9Ci~yWi!Dp41nzI;0_m-V*8*@MnqRenqDYU z-Xx+jUdW&VLgOZ?=vBxVaVSe0Vt+hfnlWT?8<7wvtPmQ?eH@yj5T>mk#+hy=4~9Ry z3Oj^`y+{u`jSt}B2>2QwZXkhrlpk)k9v)C03Ks|j2u3_LictTpQ9%kP0cQ^o`OSF# z`?6n=#Ph$D{f1Xj{wKs#@KgIAWxsmoQEo78x@ziUK{)p(*0S*Ji*G7f3_a#90=DnW zJ>*c`)x7=1Lc+p1f^lf0S%Ya}sS_fSogAalV^TzuvU2l8a&!yxOJHEpr{xuS#k@8A zA9$<4?;EP~-bWnWw z#C$wYF^wE~Is8?AF94gydlmG@Yf@ z#lg$qr5Fzcbiz}9*_C$a2E_||feORA&?Avlm21C2SVGSBZ4s>OmnKp?xQARvW@kJf zGgWms@yM9wh3=q9qifBY37tsb=-s$XmCz{Iy}hIBOrM$obs%BrFM zJz*2=b~<^ZDC0xsi^paeGz2g6UsI5TfmZynoET-?sp5zOAJ8V|Q)BxD&U6H@sLmLK z@T9I7gbFZ6{jTiS)yeQj*)J_W5$jI2k}x?;7JqgTZ}3+L8XB$f2U$ zW>h3s?9e~8?P2PSwZDOyz{-B>jp1D{0s(2B$U=a0&g?gN>FXd@>dY6+q@5Wd{#*Iz z{&IyI*>OFA=DG3gNScWA7x9axqxD|(^=tSVO@4O%&khzZYxOhl1b-+vr?5nSRhXH%qCyF?1Ksp|+; zh9%@qjcgifN(M`<4qFCwnocw2Y@0I|gO6I&PGJeTPXwS0$BqdSahEh+4S=i_qHevT8ekA```s^TT3lT>n#zshR7dZ0F?WBU1{pWgc) z3Aqo}7qEm}$;F)25t-B97!}wjPG3&GVO|OOVE zHl9;0;xe|QQS|zhrc)vF_qCIom%D+;RDpu(=^*4{3Ce|+Xpj3k2pf+5s?mqhiWpdk z;|%@Lx?aP-Rgj<%vgH-S(8sY>kZHBfg9@r(QS4IE)uWaOQ^#$ns6{`h)BVa z#&uhD!3%P36XZ3xblnp@iwsGDTf@(Wwsz}#hgvTH$PMFVcqphb99UCGO~>Gb?G;Gg z4JDHq+VwCXPS|$~(QaG_Aj>M*X#1Yipi|%|igtBHdg)Hn(Y^D@;)ntzt<&~16ArMs zRz_8z(9wT%Qu^dx`LM(>SVNT`2qx0`{qdsM5p8J-d11?IKj#Vn@BLi;!I zbJ!euXT^75gkZBa<-y8-^u@vhfr#GSrF^x=dsg+ojiVX-ze7dM3ncjMk~Amm{S<;H z`E|tE;?pqO3)*KHb+rECQhMZmbpO{8=kJ+U8d!73b?&NV^G8=rODeOs_kWbKkn;E~ zdaLXciTh{6f+ELHYV6jR2#3b;CxSei-Fzc;rvat3QMQ+Tnnsu})a$Rz8T`OuIFPJZl{*Z}Q0^}DSLKbSN&F$_b#5;d zrRbTZ{;-fV!Zhj2)&Ufj8Y@inM}3K+8BW%|A0Lf%#8!Omoc{>{7_2S(()t$&K>121 z^e+&AwcZFC7zE(o%XyW+`i2Sr>;EHm*z>{1?xA2M?+m@Jt23Tll%~C|74uNZv9K4pmtMJ=k{% zSZv*7|662;V*CgH86=C7T_-s;Dp#I2cs49ow;mXYKxxOM}-kYDoCCgfxi7pP;OugBQ zWVix`Hum!DJzDQ6jqVthWrk+ProB*3arFOg`UKZoGLVEB_yYBoxR)A(jg4V3i181( zz9b9Ya`XkT-5VQ-Q8ZYNX#=)<^Cw*22kP=`LC)Hj)L04Yne>VO5cDBY4k1Q^dB-r; z#MRs|L8pPqAjLIt1%SRemPU~-!la%bVaM>aQGUCH3tp=)bS#6A4}{k)h>5uMtv-#i(5{s6emt$)-ap z`DjCC4cV6xmMx7*==N=Xg$XKaJd0yVtLbQ-qb*9Km$Pe6w+Q+*eewspH^nin)VKIW z$ibpMl>h;1Sa$5#im>`jBKPQ+TLl z$U>_p!emM$$DiRDe4W&@x~rCi=Q_}P1H=a}%o4?WI3b#s4cE+>X)Zg>TYd`VTj-um z;a{>H9j;ySnD&65ke3J7d~q0jarq^vH|1(An1YQ~=%tA6b*f6L`&+EWUl z`w4H&CBbd0S1+&NQ$M6S?-8|qzub>0Q#APUcCU30>k6qb`VkMND8DBU zvOQ&qH}a71z5DqD_fhiMq~VuuXCT`0yR)S<567#|npC3aUo0hBF1FGbrA`l$-izKW zJhAn>9<2(0|9vs9RqX0?@YO{!x5UlK6LZl$iEpR;EpT(;-k@>{gl0`~gw3P(yNzkz z*|sIn`RqLJiqdaK47l&j%e}vyr30m1+VD)k5H`J6Zz_Rp_!Bs;gG?*mUJA71RYv#< zXX|5=Kszt*iYa;5W&k?~V%cUMk8DYd{VF!;s*%cM>jTTktA>etTS1f_P z51n!U0GLoBHbpDb!(e#;+;bXNVr0bNMWXnkv@)($#hCs{TfbZ(R~%)&af3e2V451& zhpNeJURiU1R;4~_D~1UjiFuVCArv_$djnK|?Tr{8g*YsdCxvnCgQQRfCl^goK-K57 zH%pxca)*~_jeN58s7)}PcpjL2(S(2>;|6qDLP`xNtLkb?o$h1+UG6lYoZm-hQZq6V z)z^+~Y3m0dLCwBPc&mLcaLm;7h1b~cZ$-nO6hzl&RJLBD=TS(C)Az$G!j>U;)QoJIYJk!DIFZr6&j<<`nqJfD1#NS#^AGa$W zmHB&xsrdX70)IkKU$D1a74tmYP-adA$DEzv>sO%aTEtpE#Wzh&A!gPV zmhiS*+Jn;%!gEi%qXdZp_1Zs=@I|2Eo$Cs+ELe@|KH1rc9o`+jPx#Qw>cjPZd5Bh_ zDs{g2nSgv~M__4HfZE{JV|iD^XtojG;bg$syK6H))$p^e5d-ihU7)|`8I8)Uf! zcItHR>$w1^I)X7oblq9=52~xYxdSIiyY&cbeV?%I6BLc&_qVhGaEL9Ix`px1_~ht| zAJ|!EVdUP0R!trbYk7VVLK;REU_@McDMTjHolsuwvHN0)CcJS9pTW8HG0xhcr~cB| z%j^nnfisZm(}()i+vxOYc%LrL(xU0Sk=CD0M@Z)kWj(%-y1x$wY;~1Zrk9p>%h#Ws zCh|opweC5xKXyFU(4iPD+AZi^{35Z}xXRyi#S>HN_~r?I?aTM|A{I%Gy0mV_s?KkX zo|Z1gCh6}@W9=kcUtBcPJYatP>U_4@v^s3C;fRRc+x)S+=DVuUZ4ujI{}M4-5~gW@9i6-nriDfN+bku8;IhJb$@J9s%;2}yfpx~! z$h<{7es|*K>NlxqU$25OZ;vd+Q$E}Bq~W{YA!dlXZL57j92$MauXT?xyTzVg{2}0Y zt*gzzHtqJB*1pBI6mU)X$YTqg`LSe>+sC=@R685*go#Y=jmFRj^;H^h5X^ne-*G5B z<4}cUbMeAG7Kl(Vk@q4Zen-CI<=Eyu(IFqP?A=ENg1-`@V)Pcl^D&_Ev0A5~oAD7n z@!_b$gNe?9CqC?1*xU#xg4;5}eXQaNNTM^?0>lu14v2~o#Lo*7&<4?2h7hjs3KF5H z(D1%8L{Zm5)pkMA%}06>5Bg;Ctk_qmLfXH@K`4_(B+=JD_mlrS8a`D#F>?WtYVLry z+x`~BBAwjU@eTr3@xsOe0j4g(RtUoW`2hip0%dZ1fzSYN1!y=5)EPzOeSE;%&VXh^ zf2e~0dxQXmr@qK1Tx{z$GLF*OeG<85I{5-YoY_J5fb}5k>Y$vNpd1BD&aR-U_~6RA zU;rezNkO9g%C><~qDCO3VkWpJUuq;SB=0JypHYVVJfvGdrj;=S4?kqsC1kEox=%}d zRzdvCFzC}vDA_aVCKRz5V(H6G5ICP(35t3|jFM6dK`~?d>H>NiFRP1>cD)U}6F|CG zKu1r?ZX81(f5kXTaHw>10%t^ zoaZ6)eSN`!qXy5xO?8dvepCvO>j0{{XS76LNg8=Z+B>>VHZ3|~UrRMl70a?M&fvG%} zG#}Zl-(FrDN&ou&z83JT(~IWV$<{koMqdnG%1D4W&I~6^FG1X~;H90uPL4smrywFU zAzD=(IGQD(r-*T8=3|RvL%b!TsSWhLh@$}}gj|2Xc%5K(jhLw&^}zeXL^+BygX3~*DI#QJdw;yJ{ZRkZ%Z*+FxqmN95{nc zjN~jC7iSH|(SRsVWkmZcV>ImM<2ygvJYp+bN^h~R^OjVqS48}~bOqNQu z4}lH}7E&~_5Hi0u^GmrW_T|wWIN2`?fStVY5GEu=*X%cnC`tv{SaZoT@ARI|W|>Dk z1^TfA-{wpT0-c`AXfY!O<-iNXXMsuKOav3p+HmHzGo6leUya9k=)X4HB@~p;q|rg~ zdYAXQJl?w?-A)NOLW1D8oA)FkPbxoOw18KXH~l3zap8070+W1<4!}<~&wWf(^#(u0 zH8lY2@=6D}MezgmR;rzZ5v|0B`@*1n_mcuO{oE9tSrMO#770D_ZHAyG!&6ZmIC4L+r)=~ZoR<2N718d5!6OKgt8{fP)myM=el z#ZTZeRvVBp2a+8kO2rjQHI_>cT_JpPkTd48*EnU$T4mP_k(91w^fzU2WaR)i#sh)O zlNm1{I)rttEbT+NC%9Z>yZkM4MVMkmqd`S2*zb)ND(N7v-V6^#BM(($3}mc=)(wfC zC2MUtFrg~YDPG{^rmB3j6!iRf`Do$6p27(J7C*1P>|HOl#H1AcoH8*AU7` ztku=D5Qk1~hg1@WnJ9;CLTjDS#JBRL_`7OyiE9z@Yu^QjneBzz4u;yf)!EP2%@fy| zv(y<3)VdDV#3|QV3)d4XW8x%1(A~I`$bcy7Hf27_{W;UBO$Yw#d>ht6vV4}cB%foTleb*%%Ooa-#izENIauYrUK@?gm%1IVBCO+( z%r~n#R826ylNlB|@yaDDl$)~iTeVIVbP1EJ2N# zqYQ!48q3+Sl04 zDnotALltj^1b*)hvcWk4Ui~7an5UX0{+sxjf8h(Lu*b$HB*H)*sein@=ZE-%pV`>t z32J5}wZ&{J0;<98?KJ(*TRDci-*ind4H`{{&wdsfUyvVVUZP&3+o0T8~9yqkkR0bXofP=nF?5iP%AI({I} z+yfLUE7U#%hQ)iUBzU#a0c>=mCR(+~hY^T$Z@a1mXmpfZv{ymoYG+?c<0LxnFeZ*9 zRD6ZqJC3z{n3XA&1+fy$+2uv^C4(m7Q8m60SI~`3^%V|%fhO}rSfIi2A?dNoe$e4< zpXz-<$xazGb49;Ug`$^)3K4$Z7!6`3vG$R`f|84adzPz%NR{CUmnz-f-BeaZKQ7!w!-_OcO= zhstuVMKJn8G|3$#UNHs<^dJg>X%Hix;atNbFka*%I*FuT`o7Vz-p*o1o!Rz_*7(-F zjzVIsEgw#)nh-26r=yn3$FtlAju(vyq$O1fLIoyRp=u?EZbcR_Hw8iRD>iioik~kR?v-V3&sgBL04(`QcpwHkYP$L^xmr)jGY@LI zMkZG*`gTJW)rP{)#p;8~&pOb-NW6dC6(gb2)r@Z%BF9CD?OoABl{8)v5a7>x+?T?7 zQPE$Rx!^0$(wifd9b3s-I>@)Rk23OjviNLNdJ_G7EdEsGeEj_tE6#*U(nB6RPX0%4 zK}t`oKaxIS2Yvml$8E&Z@FGA|c1p?0Ln=EaP}KG_b9~9ge9?ndx&@Nx6a`zePu2yC z5&OKgOAV)@zSTz4(2x5ItD)*BsqzCN7`sDqZZsM=kn4De^DPGi;;o5#!U zG-$OO%on?F_nEyG+g5^I_FHfF*YEaqT)KQ3RDN=a(W9qqdo=l~4=Bigi!8l651cZH zs^j!#&Zvz`Tw6AGM6W(C?IHSnKJ35kc6Bja=5}p;_R0zZj~fzG6lmt6O2-b z2Pz>EG{tExg6hCignIDtZjPvqZK)Nv49Nn2XFOU}n77C<&iQG1LMs?uh~8y>!;rT1 z!(A!{O!?*=i(EFO;~+nq0|QhVD2XTp47fXt>80!CJt3v{pfS?)W2bVDdulkRYpkUb^vGQdwkn_w`P+u5VF_sO#)4r zTAwaqkekCg$8z9o2jkw^Ng&mmVMw`P3>SdnFZ1Y*1UKd1LNoJM(V)Y6y#1K=B_nUg zEmpPk%ye3H*2Dt8?Q80ELi;eO7Z5|?I~rI`$p#`3sY@8j(dv!+Y) z@?#f<@<-Hik`~saJl5YOlt<)B#w@Bt(C0>i1`d9I2DgiO$Tas^2#Ak7?jgP%Ssj=*ui9Zle^{3qXi7 zg{VS?Lhy$rcTaEYhA5%Jdv6fqh44DGsoY>1)dVtD6T_;i!tk8YjE-`0FmkjKi-E}k zKXdW}J$yxxpbm`5i&4|&Q6x%FMtJ2_xQR9#E{cIZ4%Z~_m&JZH1{(NwcT1--eAf;2 z%(ijNJ<3wqMXRru(LoMSmfl-pkV!n~rbC{r6LY%r_*MJN6!~vPXt9tU48;)BE%ZEW zUQfHTF6m<*wWsZKNcxuZ$ou}lV&lFcWc?Hwq&rWuZlrS>)t1OEI;G1j)&MSHmy4uV zwTOX;V_XgKnhhm0V-QNt$m>FTtiY%CEE~EG?~p9mf(@t^@eA_XxrwnupVl+uzK!p= zo?r{7dbE%?)E_C%V22@rOvA=@c1w8%RgAqW19oyOhu1$bwxQ)?B|cd5&3kL1ZLB%M6{V@FzVljox4mQ%{sSig*Q*MXJajWEmm_RHRE@NQ zO@B$h*r%RX*Y`OHw;1I%;s=BUy|Z8%%+K*~(_&v6W}kQ&Q&SNjzH-=LUzYplUaI3%FG+$Hd@=$JH;QUA*jp<`FR=UU$r5n`_o7K9snFDvEWiuN*l4AvIlGlr)@M|mYq3+f$cZ!2_L?s2DVSdR0LzqMT{y8}OKVhS@k ztUG_NyaDf#+Y8|(ROGVZX@|?kJIsy6C(?j&tQEFJsg#rJszyS6PwH9w@%eIE8-r{u+>}-5AK&S#(lIK!A(27TRph;YU7;8 z#T7}i*w^~@V(E|Gqd02(d8V>jl?>&hgP}XZQiaI|k zdwc1DG;=wH>if)DBO-^M%d)QKQL0XN$%8))Nr6YJKAEGo zxjYFA-X}{y3*(``gmI99|LL@|ZrtG}gy7KIU^&FCrKz^&z{(7GXgr_|)K>}6 z5s|y*8@J}j34Ev@$fKg2;_uF>3Q~3iLAv{stYW2`K79=jjeQiD%^46;jq9laBW@fV zj7PP$5TuIdzSW9TWdnf)T|c0B9KdPFCz#hC#0-iqA+Cy?pAD+QS>p78vE0OB$k*g=cOaY-j zt8Rp}RulFB>%^lZK7!>q;;x;K7#-o7N;IW=v}}N!Qd_iQSCsOE``s=wC474o^O$wj zfEdlRX!N)k;s-G}YcV?b)DQ7p4Dn-QRAcdEVQ+Umto8`h$G0&xkCn@aZQzN6^NDkE z;&liJaCVAY8;(=Pj1v`)HyH>B14V;8Ak)d?lWb$8=He9$;?=|xVh|E4yb==5;wGT) zMam+Uzz3|=AY8C(-vLwOxl^R`B^Jt(#^plePsA4CBbEnvRV2U!b_Jm9CQ{6WR-Yy9 zmPIt4K^C_sJrL2ZMM!?dn%s>5vzQCjt_jt`m(n|t{5S@t-#kS%JVgjSWilXT7C)8j z9>k0}42-@`XF~Eq9%yMIYNHl$>nv6G_v;T6kX`zI{mD<2+^Dkr0e|z;%8}|nr6hk( zhy|47Z_XA6?08`P3BXQF{YTD$AaD!d9JmAY7+OL0`lR(6!wr4#9vqDt2|x}*n+W_k zN4pT%r?h0dvbd>&L9wQ?=P{Pv*k?BrtTv)Z*W^d}D1`^~PqS}!f1p{KRpMhTKI z=DeGp8fdeAq;Xohh#O2?p$Y#7LToW})9;)E%GeZch(z5Kw1XKkQS>^2@RY+jo5^A+ z`O!vPd1@8(&H1k_2?}&u`=BjD3N*@$O%nx?IJQDSj~CUSGE`ej*Et;Qw-!`C8GY>v zB^*gtbK@M?fLJc$EnDOxEju#1sM#lOp{C@h>P^60#!P&^yQNS!U`i1!UqBo+DYfwd z;BQ_z5p-@$)O6wL>@PZO&+1!z-bJZDC0=NcV$cJeEl)Atypn(7<5FioI`&0LmN?qB zee3m=5X&u{akKu=9c?e_VX)k+KUS-YvuA`$|_cE+yphjAC^IHL@8L^3CrgyV7gq(jC&# z-_1eD#A)Eht#imEuhLdm_d=|m?x@EkKnYQg z$e^4hkX;(bRJE$=m3htd1s}jW8+u!zc6lKR;Dgam0zHPuxGeWp`z+-fj+f;r8_$nF zQoOx_4W?)Um5A9i!;%%2M#Hf@lxy`A>7`A4EwNHs525oA`5nP$a@&q%7vyp(@-T(+ zF0Pp1a!UFV5i9eW}bs|N7jFs6NEokr$iw(P%@o_4TI zsg`rxQB0YwI7F!-;zkszLH8|)IGqpH8{w4=65f{j6uC0yu$12!;?#0arO~PX6~!l~ z>(ya3HvYB9cAT@dyJ?ZkT{q$&O4IxhQ@`ofjR+)rOg9KG*Abv>P_aCvzaDiP*-zJavp@;NLRO?cq| zio)$u^i`M}S?x~P!Oj&D7=nhosTw=n`uq-<ucvof;tOsCDKxYE7>PQ z@~H8`Q}3e_#aj5_p9@7HhZ-OW?RCq_K8F_7An+hR9TP6?osrd^U+-iuZfktgHCtd> z5N{LtB6wVaBR$Q=n$wGb3%-MQvW&#O-UyGCyn=JxHjHm|k%TOxP_!Kl>1nSvImv0J zWXt;qluS&@2i#qt*t|%J)=Y>sPEzrFR;0v--{|Z|sI* zAx(6Xx0Q$I7E+-|<@Hd&GsN|5h+;O3%j=~=#&2N4-A=i$@Uav!VUd4{BZ-gDVqreE zJo6DDn2yODb0{$d5uJB!Kz4!(8g0t^KL4}WK4~UWVqHs`B)fKW62qP9)8IEj9bu&6atE~KnM4@W_pbj*bDG@i z{0{{h6&BO~mq5tt|96-tR{;Ng-b$*=-zC%pdJZN0F_tb2!UJOI>H2&Bv#ykXn;Tda zx2*DYQ3b#(&!_>AO9>4CvEs&UsYVQRTEFZe`7Wk@^A5lL+5f|emw(@%P56^P8}~2& z?En62_wW4Jk{kW4Q0GUwg6iM@*CqgHEcSoc2NZw24`lu!EcGidIH~`qJwRWar8hM= zzwQH6H;o2~u|N8=ztt^&?<9VR@s1A7f}qp#h8L{aJw)06Ap&D06K?b)Z!Z@*Hxl_G z$2|QS`v&KV;4IA^#WeR0<&xOh9k|b&2P6>e6F8)#+zc>QBNQv25}F9Ngd_^fcBNXV zR$9y*SC!g4OmWQ_1)4TfG2Lj1ua_53n==t&;nLKwAUUKvX}E%F7;V|6m5A2XFV)!D zV7hOOT^-fLamutMvv`T1aJ%=2yWf-Tt!LxRs>K9p&Uo28lNlt@lz+@rrk~VQL5XQ{ z=`bwe4reay%{6!=Uh87)(^hbMaiO~L4$j5&1McPECpOL9w?)U>D_%38&vt$s*OE`s zL&+x^+}5Zq-W$b6*;T3t%q}sK$j+;7+oXhe5>^u{@;oY_)XnFkLlKdQZIvV?m`IBo zV^*fR^0uy_X+s~wvaLYk7_&J+bJ$XwtGMmy?aZzLgbG#N4V~64Zptz;9902C)nm?RsHSy9FW7J1b9wU2s|NS6~8)5 z5fp-fRyUx;DpjQ^R^ONICM;#XoM3hnmU>*()f1<>E+ZCZ4XRBhup732IiW5VW0uspdz1hX> zbwM!L19BVih28lKOx4?8aAGfD3P~90H_P(TEg8#uW78j347B*Sln<&i1YGCpw#XK} zp5pc_Lh!NvxK-otqAdnF8Cg(Fxf)2iUHAK5b4-kj+V62DHteZnk8heHK2FOS%C!UVY|pRg#N%zSHgqpMA2nI`HO zCz1fQ^@Z1^ZQnB3=Qt?%H7s@0>Sk7SxJTwXMD78ZqhFlXOzrk&fPl9R3<;09>v>t` z2@_G;tY*`T2>(f(xAB9hd~GrEaI+WmCHVEN^)&AUF9S|v|7LHNNVj%NC;u2=FE|OZ?QZiZ}*s#ax&&U7I-L8Vqh7*5kZN} zzan{6A%}kqpWVEi1996WM_@&^-+A&N)xQEy=%lOP_59!!#Pod#dZ>O1^k*^Nkb1}q z=21rU&tP;6SCLV>mDSCM$je>KL<(#ML!6zlP>3#bWu)bX#LOx)X+vNo-@tpIWglj# z56~!>T2MafjLzmI5(6c|OQkL!LcuvUL0X>=lh%Y|D)U?1eK1#-$$uYO$w&2GM*axI zxf-^(R{rf*`5%hgDFF`rBbcW3518Svfx6`atv5~YU)LQKjS_BXU_PuAX~U6gq!hVNs^ezO4o_-2_;Cf7<2)Tja%4Z_--FtxK>bX zo1j^6J(DBcv&%(9*|t9QmgO`H&s2~F-{Z^DOV2VmMj&FV==4m#YujW5!QH7IK~5`n zvBAPZr2rp@Bxctq=R-b1XA@*Kr!#~hzmC}B%c&Smq2Vw2wzQ`S)wXM&LZdm)P~_8F$DJ5o}CtR~pgLEM!9j_M zgmFVYdjRspASa9Vyb6g(H?))xNe{pRYI0%N7jTS1c$Zh(L-^Cn4T1C289W>(;Cl*+cPN7UV~=c!tS2rY-TEZm3b15I(?kE6BaB!F(Y7;kk&YZ{3^lb4g^bZiEU zBg9WoDIPe1#;-Ws4Oi2}$myZdf~5j8G9uOMD6^<(EdaVDedqU>jVFel>-q0( zTU9wW=D>MMOVuNP*4L7m!XT7Etahd9&n2UOz-+LS724pXGH=@4LJ?+U!^BX&DH)AZ zwf&`JRJg_Hm-hdkj3ZRs&4(PYe?5-4t^)pg9Gy%_nJM0kBPNziXO5odvz=Ecyk(wf zsE_A;R{|XuN}2n33xMpd#iz(5Ij5ye-O$?H#3!g0E2&Qrf$VNg%v^EmR&hpg$%Vp`kfLD?=x@;q2#@ssT}BpB9j5x5yg3jI_S2?N2|oQv4K za_XfRHI<*ZMEMkH)xwjeU0S$WW&7+B50WKr)9_b-`|RruQg(xZaTFlDHE@u6+M5H6 zBV|8HaDb?ziDz^kF$MfHa66+J25R7dYqtB7xo5CU*y}1h5|!zAR;`j8-Giad2hkx| zz&Lu1B}>K(&ZJPxyQPvp6sx|-Kx~(%(AuS{y7}%lnxlxYowIxDekF-Y*$TY%DKU)PA5 z97zi`D==g#P=Cr#UQK_$P*ulL)f}we5II-4WkWA~UN;I3{7^hh#;A(+5jCIuxYYQA z#Y5Ro_p5P^a-u>kb#ja*dWc_>+_PsimHs>i4R4>V#9KtE7BorFnIEj2Dfw7Cs7q(? zxJs6M&Wa~rduknA-H1|b<0n70KyXrH;@C{y-Zo(U`#C5EVFi-<&ZYMI4htaQ$pQPj zdR7X`FM+>Jdh!c}*FcAb!Y{k~vfxd?vu|hs2za`F*(TXAz6Sojtz{;-5nr}R_Nas# zms&s)_+}28r|0~e1D;)Z0B-6h0Z&dtgLKK~*eu@yfAgD2vgMPVk+f;oDGFWtnm)0A zH{kh)8`{SS1o^7Mi3+pfe4U=EqL0-!i`^;m)j)^E4dwY?4S2qg_LvlZvWS}h;-r73 z@dXI2@hR_Ka(mdULUBFy!7|4y@i4lV&OPg+_YC`W?e3g}&%AX*W@)THJNQ~L2e{OB z7kmETQu|slhw}>n_M1#2MLX|}rEc~+nZ}oP;hRj8yAa4&wYu-v9v79W(TTGIC>k(+k?D`O7QXI|!+?2~^R}Qi3Re&_ua|pmq*%1cP5(YAN zOJ#)S1Gp(O1-U%@Slz#*JXeJSxT)_c&kUp}H_OJ)D9_P9p*-KQC+}6C3|zV)rLH>tEVT|KbQ9MKJKSgoXmo0s3hqJaCud;sE5`yHwtW)B!^N$}e z3!DB-FHl0$4;Bw1R;;UB8_vD=J*NH6J+DccCyUH^FHSmV8pT0qSGK%+q3vN$6p95X z4wi9WiI>p6Ja1bcr7-Njvbl5OJM)(1KlOl7KNX$I1u<8FjDib#WsK;#XB_21`D0s* z!ZI%4jcLVD&txN{96XI9lNh@)t(5x~jH6=}&J zxxq`Z8tVh@2zm}q;_+1BbibbE*RsV;8tGP&p9Ny%z&{^N@QkJStN*0`{So~iE-9^n zME}7h<^S0q(VL!jdyww8A%7F~`cI)`S+TVuf!3TjBkX9Cr1+miy?*-!*GmLg)NKmQ z_Gs@~FuMIy-rGQ`bPasC*Kdb0!buP}d8d~Gj&_=RkrG{Wo?=O5je28+e2C*CilR&l-yFb>AkKlg#VAWrEGUmSE%|0B=sm&zKei>uqIkK)L9MB{_rUHb~^? zj-4`-M2^#>xJe<3{zB17gtzq6Yp7<^!B-8VnnjJ+2g-E-9gp=rli_m->bv}s7Uc0n zz8vEO?d!LWv#$j1w3}8;@I3%MCa6d^>ZwnwZ^w8ey%{wk7F44nrl&ty!#C@%jcFrj z$IGb_Rt6@2JBF5OA$uE~>1-f3EpTs`NsFIFt*lR^}_-Rl3zQ0nEVFbn9h(s*G?c`nAkD4bqm3Xp>2b z>_mrk;Fuh0XqKC)WweHn5u0zpmXmlB^*SH2KqzUquw<^7$!7?52eeJ5yT`y1T@Ee3 zSq3oUHp_E9IWNJturpZE_634XuPcX5UdX;44RkN6n)BRmfdh`oKoMm2&%eR{q;0+` z?b^LzHt)4%lFF2|;OO0aga4{+enrXtc1&hSGy7HB{3)^ONs-bplKsfOk|I~Vy55%c zs8&Uf{n)xI3U{4hn3eT}iC2)r_khHtaN85d0Nu(>fMI7PLv_*TBByOO0f<%a0AiJ28qm|8?qvj#7L`gShJR6V zkIGxE-I+8N+j-aLH7huJy~5|#x`(!nT^gb>4Tz{a$dA!G?I*;l6q8oEbTbp!s5I;^ z#Ht@?oAG=E<%8JPE))z4w9@wn%0hE%D2NpSZ8H+kHnCtxp%4N^F^6vX3Vfo%%JZpHFOGu^uwuayEGyn!2KQshhTU` z+QqB90esfVH1v>M2^zVJU7SWp=YL$j>=esv_oNi;6NVQJXwmnpy+my zECxxBv`{kW`1Zr>`!;k+&+Fh)_74G+EIl(wL@15mgDf|Su}f&Tny7wdp340peP*l= z`RQbk8jp?DgFHSIJ0BHj6hUfqOx`bxv@bLf+;}jYKUzKnF0y>d@Nk;nt9Ixw-hn&a0+Zuc$q`{g`&2D_9I;`8sJm5D!L_Po?>%oy(uvF|l9z!xXSg`9t6OAU7|2 z1VgQFd=wjhhdjS98xwo|v|FqUt+`EUmrDj$im!^IP0F}1V?(8rG%t&k5d>PJ4*e{r z$yzI+mcyp|Ae-#;Fenz9%N5T7OAq-0Njap%odw9%G$!5>nIgOLO`H?=3>7^0Oo29v z`q%1A!Ec^-DOWw`-|phTM4BD0#sJFUe!UgHH-aAqG6t62jJZ=NKfuDXh~GP%yBKQosYVP&0G2KjN2pO` zF?_1QKa*5=zNsrxMGBz}a{Qf5X0(QhQ>);n`Pxz(tJ;^oW5B{={E*bs*Q{!d8tL(u zU18LG1wmvc&otIzp?2lf-D8ARuuFP4O`87iBcu=JZ0 zY*kG;Y;0BQCczt06h*ou_XyN%2m@%;)=?2WW4W9G;0%m+wY5{;5t*hcAH)Nfb9CSnZQWBpqo35n3 zK>##UGe9#<$bO|{vYzkN<-DQWi25S7yr22eR+V_bi*1?hzBl2GH=5}eFylUPB0`aq zbc62!vwk>TS&A4*YpFj)$>s6X^I!Xp`_o*~!*rRT4`5AS5oK)0u?V}5!kOR3-rU%$ zN!}|c6~@0r5{LMV-gLcW==v*|@uu$>3{4}`jHtEiiPxro#V$kcGhyK_no9yFEz^RT zxa&=;kcRl+Yu_sik8D<%alz^dNWNzDC z*0O*!M{=w?Wbf!q(gW2WL*rD`_%qqDVd!2gkhK7s>Gu|%y1fvaUo1S6d!dfK8N?@@ z9YW`OVeZJ8q;Rr8qIo~uUm=qmx2sc%e?KBT_le-RAf7779p0<78%qE#fzd3R zBaIF!+G-cZ*n24^FdWOm`6<%aMK(znXVT0gF2)79LQy{V>?X{aY>b*PYxp< z35W~rboH|gRPbB|i?Tu`Len>}Ii93Nis5#v_%wpk&~eY^wNm)dvcz;ecG4T*u@SVA0 z0oDk3H5yv-Xzi{C|Gi{aQb?7TVQ?XL!=a(E^Ax#K;v=0sjiy=)6s^qF<##cHZ0unk zG$1K=Oo9qjxe6J1H|~AtR!f+~hAWA()Ew*{#YN^UVuRRx_=t3@cnCX7Ns$qy332p| zwK+J&pI+1noNc6GT;j*&ivWwL*@-$z!La+-*8v&uMSowO+e)AWJvlo!AB-XRX<=n`O>asBl75qV zdHvw^k1W=wk4mHo^p-U=q!8#}HiU5$NkwmskX!RyJD8wq&;9PbjeDULu7FUZw#FLV|YD>=Ak$SiDyBf}ZSV z$gr??KVfN+(~D!Ls@4}Ss~7$>@u5*_VVwNe?&4+}9N zm%;pznCqpkS^y}kt%7sw7=|!bS|`yh@5(!ZZQD2~?jHFS{zQRhWe=$gmd>d;dp+-= z9i-;5ts;t28lscSJM)z`uNkGmkVEJvlrAfIQfVb!I~d+CKF`T&7xgak8p{)_PMm3B zgDB9dTowqbT1^ZgZZk$Y5{?o`8S5#aGX;QyiV+nn%8*mZF#< zcs%P$L73}(ZID1lLxMAz!qMdY;J%E46V7iDxZmbw_TSQ0iSUT?j4@@I0d19@8(}G6 zK@liZt(l%f0;HJ$a2$agf(h>ZQ1j(z9V@5hM6+_hV}#NNrcVA_Wvakh7_DxxmKAUE zRa;$zN%f`w)$$6U0b&BX)(eDcw6K4#tpf8BmSOXU+G-6k4a4{Ia?Y3GmwEY5t%b-u zo4>4uKLQ#5aV`Ab^7?Hp)FyBK3CLK{e&7xkDJ19@kntbqWz|PVK8T&qE z@|fa64uOF93NrrFS_sTbMB2Hgh2S36O2~}gK*q|&5TM)~)YkRaUqQws7hAP)4ENPeo|!L@!l%<=`oa^JUaWy`}vz;qC}!mxw#B-Vxh)&TY;)Qtvid5#~Uw@uAt_ z)azZa`-`qb^n1czq{1P=*J3E`xQ{P_5oF;Z)&kHWK3UGgP{4_uhYY-+tt5tfbSi|7 zxcegVsR4pwZY%yit8aoOF-loZQ|u)T1M&rgs7Waoj9PynO(8LaJgwMgb2dLAP!RbcQxIlp5@NaU@9cV zyq>GO6+Sp>ROm~*-4?`{e)|>E$B8tGPpt2`4!04t3xs>k`g^C)4k%Kt$jLROS5+N&Q5DPag5Ux0YR?w!NOCpSnmptffFC}dPOp({VJM~hR|Bf zwNeW4Q+QJNTsp;1%v8Q;@YWvBCwlVLtGn-y>+EK{8zw?jnfy?mcvNWC zV7R8l!;b9z0RhuSWNle)tZwqC*ztXV?n(Dp5SJ-XT7SuxqA4@*S22N6&AY=uK3(qlyg_orXgbGmma;MT*jjVC(j8I6H@zR8$9`3@( zhN7Z8rexH6QL4!BXZgH%7$)M-0s>I5pF7)6_}OlQ!s~H`#XBCVvf2p2etH?N-zL#y zixFK9Lrt-UsHCnIF^v9*gv790eHkP6mGSJTHD3?kIeYvQ%4N2(++LGZjKpEW46ceZ zNo&`qK?{s4*n_owbafm-=)o&I&uja~a@oOfie|UgY6r;pIn#wRS0$h44u*zaS7y>E zuF2sN48`igXLF>4D?U9ZNxga|XvVmX7g0CD!q1hjowr zBY(T14fMUbF_Lqx(mRx!22XOw+xyVVqjb_9aU_xTxjJxaFm55=eKUz&$6bY=zGcOb z`EdpozNXbG%r+{Ccv+~Lf2hd%S?QbU1;=Xat6mQ0Co(g;SFnu-DIqTpJ`-GBJ?8>3 z?|R<8n1ffYX~FH=^}Ta3j~-UjMx(SB$l_Uv(*Zlk*s~X`CtlBX%*#Q!n~KIO@@a&( z7C|j^-z?l^i75`YN0#*-`&-9Q@gQIx8nWd!`$hXb%btWEu*Hrku<4yEJrq{~e+e zMsN#2%*R2mR-^er!}DHl`gtBdDgAs(l6FDT^!9QZe0{06u>AqnEsWItTF3zcsZ{$Q z4tdZI8gASY!3+h7)^?i)c~{bvW7nO;1_CeFk|5p#uf_xTQRjwm|GeaSgYSa(M9cQX zp+&&Pdy2>EMy`Q;I_5^HVa^c`x;%kr(SYKs(HpArikmZ}AoZr#*F^veN%>o_;dybf zx_&72KJx>quDZ!W_%J>2Va^2!uG--2`baqnXqfrv?fUSu`gRTYVEOqDfqiw&AhqC8 ztYv(k$$%cqKuO>sKh^NFQGwKs^uveqA0+ihSV4MWhT=kM;};Lof0KGsDh@=f5 zDTPaS5Qy*ei`w-|91oaK57-z5`K$&Q&4I#Y0(0X7^J@YN#{*f4PXmE#F<_mp)Cj6_ z45|U>EHy#(<3Ww5K}~qUExf^P8o}+W*a61Y_tMIrZ%+F8q>Kcb zF~&=s-+%Mh6D{5_fZ$E99e&gWu z?B&t<@BE4Hm}CI}1=Cjm#o$0gARb3cP3~79LH8UjCSq44Es*hrcCfnvQ1A?y;orh~ zpUAs1JyOZ>3xHz1gyw6FM0aK0aE?lbT*^0p;-$f}f-xl>fUzYXYc^i~Cx2pq>5aW9 zPIIQweJ}ja$nE;U!yi~HpERQI*n2OkORwX}Pn87$nRcG4ua&%S2~ipbVfs9Fn6 zo*nE8O_`oI4^5qH;KO<~aYly)9z!6&O7qehBu<}aMU2ki!dt(UIf8Be>o>H;Vg&g2 z-_V6s>+9duNYs3}Ifa^ZC`BK)B$n!_Om38QQWtEN_X0t=iUBDGCcd^-5FCItTd|{*BJkAa-o+O6%Ewqc!|*xh~Gh5q@Y5l16*j zBDFcX_{wW`@TK)R;nwU)?zO}9NwN-Un;Slc%xzbQb(6VkNu<6}eMz)-oOy{`-KNSu z#Rl*ZzJ<1{;eL5!xfuR9RpR6p1m2Cex5qwnd#WEy|%jE z^=S7%hfr4|wK_9@Mtbded07RPKaCoo)L=^$q?H6~+@7|$;P_^NYBO3e?$~Vb*6D8$TAUXditLsZt5wPe* zE&!NO0(&5$@egWF&SSM-X#HhGquqHizI@ce@>Rmjud>@8&xmhsKkrMO*Qp#)mzRB3 zY^-m2BfGr_s+z$B-l1sl4QBMlT_P|UpI^KA?|Fwgh2<|8Zj$rVtx1pJ|N0^R3j?wK zpBac+Uk2ho_YevH`w(y7Xg~AHmQnubA@=JE+9KS3ZS?A`lB@J-!`9QbSF;q?)7S6d zCSk;&<0Cw3)@Z=jPvCV&=B&`vu?(zkDXEdYJbg%;?xWQ;ME%?G0YDB%Dz@-wIG_1>sqPgARsGsvx|sny)n$2X|de-*9PLRHZQS!q?-zh-qPOw?+q#c zos5Kk=$QQX_=5jC83Etre<$OQe3w6-j5i@fU69fDtKv_}US%f#sSsjZyz5U^#pgs~ ziSfUtoM<4ke~yWj!M;P#YLMG$(%r-Mu5YjlW5{O!xYDU0PX?hIj>L2bEb6|en{`5m zqeNbt*4~NBJ-jqmXJ3QSII^^wQ9Z{!y`v5^I&M1ELhKFJ)F;r76Ht| zxzEbKh&POeuO{K+>}e*Baj$GY$nNOg80PzEIxEmK#-c#wSKXa;Esyc5NqDU#$wvaT zupYckPsqjXzLP9N(0D6Q4Q+}T1D8`iX>mKSkz)KxpOtxD5I$5V)}8c$go%4h+(#Br#j?Y}Cu|AQlb3-E0rQ-!sB)Is>ef3V@rM?V# zzo+weVW5AwDn8IlYdVzpfR8Pr5B{EY@!ORS*6{iJ>la})G@ULG5XGYTIX{+fkQH?_HikS}Gg5~v9jVKSiuEn$S1bH))W?J~_#5=!`4{?ICO z^AWPdk%lDlu4a$?p&#NKBw&ZqktJG|2h2xfnQ0owiq|2eCllFCP=y`Pv1Ub?UCFK_ z-a@3$0Y4%SG);6LGN*-5zd|t2uv6tvOA0Z}H^??MS<8$KGvTMraJMk0OO6EXuH^*k zj_2p6@+0PY=P7B?7ilK?yNAAj%^*M-qYK}NZw(EkgLECLwTLaz>ZGqwZppLoZfm#L zxX!6j<=x0~P(58qFHA(VN-k5?oUd5OCZJ84Z_z5Kj&&fQd5s0%lbyJoc&8w>fqfhV zIz|m#uNqcG+RCrQOiW7zmuoMB&6ep4Tc#|6SPGNs=onKiitF~?!O|4w$KDEap?;T= z8f;LVp_B3lCyOKm3ZN^F4h7UvK|~u6wXw1?aq?-!$%30%}=iD*?Q$V=&U-x%+BTw zPzObut)0v$RT(iRThH2Wc?#yueliqvXw9}GVw)3*(_t-)Uq*3Xu&LlXNnpGEkox-l zV>nd~g3H4thm7=l3FfWUlbGj z!Q{`{G-MRgG2=ShsQLH))TS{Wf@1fZ0a`R!8QGNaYq`@OoMD_Gi~UY%uFbMAcZpUK zDA`UnLNJ`y$6j8M zV&p<2&`R6yaknry1{z2f?g()@^{lE{)UIzoU4%6|z8lbs2ujevIQSxre zogxztGiep`B6qphb!tefBL%<@vPyzEB-WHFJPV3DEGMaYtS46?J&Fi4 zrwt_7f)tXjVMBiQT(zM}t>`g!o}M!QIPC80OqlQukcymQOf(lf3sVyer_Fe&g?^TY z;2LPVP6W&hE~2_Dlt;3eh!3pFip(mMQ7+Idi{#EK=W~Yz*$nA`mZ9)yTa4swJj8v~ zsw|n9U_IYxgy$b;wCY&eM!gwD($_DBgcq@uC77yBd{%r-5UvPCUpo3}h#Rtv%F5=A zw(6W`%o_s7?Z>wvW)`_2YAl)7V89I1Tim#HJ1EvV2QMB@7~+> ze*Emi9ynV39c2q7mO_?tYZcWfvG%@^s?wcQe1oX@j=$;qfO?1Ew9FL%4$bVkAf|2kFJH+Ajn)ck7oV}0$ly>D4jwW1%Y*t zN^SNacA>_t(76dQtx50uRj*G3EnMA3-ixpkFn=bBecNAzptIa+c zf+pUHew})h6~?oA26HFQi^04&Oz_PflTFi@^U|ieb3GUH?m&^y!^{${RYkWKHYtM3 zDT_mb(2`;@2FR1JO?buROfL7Zh1J@naSHg~mCsTko^@+q?Q=hJED8#Ez@UG107h$& zi=aOrN+LcisV^c>>OUWq$Bxb4S;a77`Ak04cr0Eoxb2nudHTBaq;ZFA*Gu_gLEGT8 z{VmyklwRu+hnsJon=5h#%S%+%x95j%gpLDlIbtX{ejZvAI(3EZXM6ACcaHOL?8vx+-62hO1b?B*`oB#rpB zsVf7f2j(h^;OO*RaFyd+_~o2b8M6sxyXu7u1ujx0Q)d_1f$o;19Mr1H3T z*Igf;5jO-MA9WeRhvPmidp`P|w+(rHtM2-m#QVbd_*#skJRbL**!8v1K(RBEWsySSg zXB~9*_w~KrpZ8}jq6WBhPlQPsN9MXjHltWt zOC!l$kc^toqL0?2a!!q(px6*m_(1cAQMX7Gn#$0wNIqN1ZH(Nm?6 zndRtk7=GRpvS5r~PR4&K#jp0^*Ou`hff%sT_pS>v=@?sCD9hKbi|a!%yk4>V{d#;U zvB8wsJ>krIS7O6paUw0u2$MMRl(?LR7zrHqK-wqDohSn-@bN|8F-;7(Adit&m(kcj(V9 z)7@T9lC@1jxtt3!p~srU;fk4AyOIN;400#H#i7Xr0S0Yxpzeb11`3?hmm)itl7*x{ zjZ8G~N_Byy_S*1NJxP^CC!qoYp`f%pqgBru> z>%$pKOy~=s!Et=VXG+9Q9WXzd*oU;0ui@y|Au4tepV`tggEG3yh@%1-ay7(p6C#3= zF)*1vzCt&p!#!V?F*cboXq!2AC)0qEE(%Lm2a$^3OcFlWGS*I}t?3XL?Z9Af2KI6= zryUWQLNcZJQ$-2?ewx0YM7{_k!!D$%yTl!xB%eBrLC^@)Yjk{bBrDM@xCH%fJ9-Xo z0@5W*S0`KEJX@_L22~)dNXka`XG=>2dBAft%E2>d*bbEw1U+&tk&c}xr`K;Kyc&;w z5kk+tLU%AFm;Fu-v?Z!iCKsd3%nk=zDd*YXqJ&-Y+&~cJmOT5^JW(206m(nb$I1>b zkP#>g!~*E2j}9&Evjk$Se`YPtmiy(5EhZXf21Be+~nR3G5f z;@?vXQg7hg?Pa_t-ZlPVXy5!kxo2l|)ncH|>vF*cC3Vhn`TDM;8W!%FK|d5Ip1J%)ce4Z{h$9;JR-4c@5v0kvXZxL z$+Gri?t%4Vout`|5I|aE%fq-Q-A}~g+FrJz@dwfRrkb>YQX|cb9OXx|1{G&{(LJ9Z zy&uuUZBQMd4D{jrikWJYAGs#c=gm1E9Bp!-eBPd4Ap`W#E< zp;zFtYe4@O87y8=V3aJjs-xqPgqgjJ~X;_?uzga2X>>y4(GYZx6 zlYukpfLk=p;=*od6SBm|M}=xdE#D?cpx}rkJOUfSuxVdZ%nKMxy;2a6$L>%By_$ZN zsxwY>dVOa+O}rT@Cd9XjoFE1B*-enc;^wWgBI|f2AQ6BJSQZb2h;ouKXD4%Gt!cK0 zll&xTvL1$V{C>aDb*nhZC4=wOH7RB`5l zL>;b4|8(%H_zwC=1y?-R5OT}4kpIIIXJ5YpD4wa@} z9EtW}^#ZyP)v(QpSBqAj1Uasmj%SQB0c7-pD2%s5US{U=vub*K`p`Rrd2Mq+9f=FY9J{+J}{j{bst znsk5dQ=$Fi<`sYOG>NgPUgz~4_C*}0AJai>i)u&|1YL9$`a_baoy4Tr43vXx?2E(X zkD@@5=(Hn3m8@kLhT|=n3)$KiHT23T*!a?Y9HQiUKv?Ug#wKE$C3;*NRvZ?M65$pg zH;77Pg=3C0@MhbDAq`?-=PI+{CHb0Ym*NP&WfA@rc+>h38adLsfnKoNCRDy8mc!}h zmJIlhNun$Ld(hhIEb}o>`?y60NzIxMkHtoLjPk>gI`XGW;%rJ0VjZLw{Wu`q%Nr!( zzN=MAT*f!Iy$*R+N#|U79Av)RU|%eM8>)I9@}Sacx8)rmNd)YRqx=#hH4w+n)iXIZVm)yQ<%J- zZP*vb_kmAid_T$Vn_0b$4q9ms6y%j<*0wpu+1e$7Fb|Xnw$achb%kIIAq#gIm=wI9 z3SE&i3}TD1kv!3TMz{`}4gA5YPK9^Mp9{#9kfPJOtPAXma%?Xufqk*OC8^qwtG>ln zMg#PH!0Zr7&3HviotEtMa2)h3ng#TQfL=-a;71n5?Ek?kuf_%Fl~AlEbg`~FPSKl# zL?JnIGh!2!2*VU837XuKNyU`N6y3G|9YwiaS*7w2QFEe9uTHt?K!tLvZ9{b@zFD?i zuE&*%e6IRTsSiU?eA%V{+W4*gfLKB6r1VtC^t8&n!@K<-gYQumM#5In;|{AIVvlLo za?&#nzamv8JC?~2m&*O2^j1LKa5e-licW-jYFtV7h3T)u_6sni*NHQ~I$WP9v?kZWcf43w?w@d-XP3rg)sQNTtOiPYgiljCz=4|C= zjAc3Oxvxu9!!9sLKV*~la zN8@gaxBP@r`Lo9NHV+@w_9_Se%YRY<^0SlN-;UVqFi3QXZ-7W@4s$x|rE;3v6Q_!S zDpu+1^Hlrasy^z5I#;((x7~MJTeH7d1Lo0ksjd4Z@~|_D^g+{ljvyx2h75hZ4Xga7 zdEwpgW3TI*k<<-FW$Ps8py_gGBVv|UFGecamtM-Y>MKfLA1qwLdCeZx9rW`DdkzuVFw0_C;k`O{W?AqtKC(`R*HNM{-9)CAm_5&%=Nst2;gy1lk_~W7F8G1H05`*nr4fFzqbywx) zAW=I`34&Kf)bXL~V5u(b!c(M1Omm$~Peib&g9QsWyk2yVd5C2_8Ol?mA((gEF7X2y z>1kZQXJX2pU;B+wfbsFCzYT?JIb`vEDS;!gRIpLH2uDeN<00?uhsRPzRs_5jjfF*m zKkxp59g+g61BaPE>s?i4=Cyh~*EGFnA6E*~#1O}DC=oZ0i?~#3zK3q%8tM5ig@~7K zCtIBl3t|%Ef)$2!7eGdRGnH#)VXez>WEV;d`0vme-Mn>V~BoZ{pLw zQ}KPW`2aqx#HarapDwUOZd4={Az4Ui;6n3yTvh+Id;R*o)G-Ou*uo11%QAja@{giS zgKnG|^%atzsV%=r&uU!{pB~!4r!Tc9XPfz-6S@&`u*^-cHPPLI+MY7j(LvnV4*7T4 zp&bu1TP5Uwsdw$W#l*B-LOwAR1p9}2*ZsaM8-otB5g4!m|_)Y19{3{v?I Dw9Je@ literal 0 HcmV?d00001 diff --git a/docs/assets/sqlflow-logo.svg b/docs/assets/sqlflow-logo.svg new file mode 100644 index 0000000..dc7b103 --- /dev/null +++ b/docs/assets/sqlflow-logo.svg @@ -0,0 +1,44 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + \ No newline at end of file diff --git a/docs/assets/train.gif b/docs/assets/train.gif new file mode 100644 index 0000000000000000000000000000000000000000..4507b6882897167d21114da6d8cc7dbe0649d26c GIT binary patch literal 70492 zcmb4~Wmp_-wyqm@O#;C?xVt-q;1Cj$;2vCqySux)ySuwP1b5eN+@TMdJ>Sgib7uCn zPyOkt>UX_TwXW4qRX?|clsGq!PAhadR4)mjvZ691BqaZeR7OIoNky2?`OPMz1oSd7 zIy$;MY2fVS{Pc1Ofj|cP2A-ZkIXOA4t*yyPNnNd-hQm4s`v*U>RZdP$uCA_jc6OGR zmwS7AXJ%$5$0m*r_DA}MUIRZsP?Y5qfMPnfI&Zc@iKS_lIf&7lVj^KiY5mR8((JFRs9Qwbrq{_Z{kgsWa=ZUJ^_oL}cK?yQuLu7~e|?MB zoVR{WZYaph^V8$Q{oU=&_0{FY`572=dUAYpc(A{>yR*Huxv{>sy0W~qxG+EWdv<1e zYI0(HY;EZK zEj1-MDKQ~FE;c4QDl#HGEHorIC@{d^&)3J>%hSW%&DF)($>a~ULI~P zP7Zc9Ru*O^Mh1F1S{iC9N(yo^(hooqVj@BUe7yI#IPc!R!N$VGKu1GGK}JGEfQN&H zfrbJAKHS5?Apl4K^nlmxPgMfIL9~Ce{4`ECDlRuWZnvVJur*(IEC;)SA)T34lJxjq zI8(od%y=^<3vH##R4W#CsYgzoJ2omEWf@Z@4r-mpELzsj8eQcpmce(e?i1igh^Lpg zkFTG9KwwaCNN8AiL}XNSOl(|yLSj;KN@`kqMrKxaPHtX)L19sGNoiSmMP*fWO>JF$ zLt|5OOKV$uM`u@ePj6rUz~IpE$mrPk#N^cU%r-mE9Nep3P)s zuqtOD0{HQ!S-3iPD27@QQmj5yoi~!etT&u0Rg*uK!fkW3GE`GAks<6-jv`%KIF%zA zL#r`dTQpN3ks*>MU03|OM7hRfHJAf$zCybvm`J9+R3){2yjWwTzHGU{%5;m?B&&S2 zSrB}*I?_;2`yOo`S+d)wpf+4e}r=Z^O0$Ggkp^@)xS2o$nd zrigS}Cak9elvw5l#2c0KV8aJfoO;vu?GK|(zxM`(oBo9M2b%#vf9kD3@_3`IAnN@0 zA^>D+0F-n}uCH(iz4PsQ824V`b~yjd!FGf&Jk3rxdp$QWG`{qz6F`rCxD%rwPO}^P z?T0Z%EWLPS8achnWOkU6Kh0j^&v@g#B%}HPQcYO#X^NPAEMP$@We6QfIl1Fx|3je*_Kk8fRc5OF(P2TR<{?p*A8U5B zW>EORLGdr~BNJ%ta3z}jT7%{yXJo+SnwPvFrbr>(P2#ZwGJmFNB^a%RxB^BDAQI9% z96%pF43Fmt$(-@WC#ks&b54-XPj^---YOd|bb@VtFe%*_Q>80=w-THVOE35@xpGGUQ~-K#g=>$-26GWO=IFG#cl1-z`?hE@yCRVk3HLg=I6c9x zYwpXGq+8j+6-}kl3FXaE$h60K;5pSUXZ(hJ=2)JTqOQRWD9ke#Ogj21cvhFzl2J4l zT#;J1&-}HH(gjL({RAkB=d(Y8N4LjU(TnHf19JI6!}Lt5#G~cb&l*td^Pz`hd%H^j z0G8y*)+#4X2^4^p?<6}DO1Sg39pedgtviT`z9vL=s|*TOHetV{A5mc~9Tr-!^KKnV z9=#Ug6icwS>HuwmS-Dx9g1`(YzBT8rY4R2X94iK(b5~W83|e+!I!^NTL5o@ecVq=q zo*alBl^$U@Dbcwr*eKDRXBnL{d1{i^!!sR_0?@lEUM~b+&A9dc?dC=r2>_v2G_omd z8(~8CtOFX6Zy)>#2<0G3(0NQgDNrmBt~G|N*=aH$1SMDowdq6zi|*8Cci=(A#ALV8;Qx&^GRTRN!K2rWBTvU*!M zp0L$(WqjqRr`DNP-SgrgK3`3Yp8a0x`a&b5M?EZrvjJ+-FVFz&q>xDNN&?D4sB8Qk z7zG~B`*%#j=rC!Xw1ITQ=n*ig{8t#2k9#OE6 zNNg@1Srao$eSohEiwXXhK>kqAAQmTO%g?+a}J|HahP3(NT@7`tBr#BnXi9?@M4Ir~38$PO+ zusS0& znQYD>_6Pk19BRb4sU`{8Acr2~_kfzK0tfr}FT*Ooz42U(J|@N1r65Pfe^jSpp(^|e zhsfy3`R?=Q4(YF;$=QIM616aVlc~(ziafS!8rN1IuWlTBvpi^6C8%se6B@Z8}=rtE4JNdBu zROi}Fn3qi$f5E=}CQDXH4(s~EaqY`t$IC-aIa~MS30h^N?IDi;MupX@i{~P5o6i7%rTz?9YLXxBm$R@X!_eRe>0Gue zC^K%$ni-a+uVpMB`ekQnZ^%#Vrk^0dr)E9@7$L?yVvpX2DudT_S%zX`vc<6=k@{k$ zN{Y@Ej}4Pm1#k6L*NGvzV=I-AO?-NmR{j^SMtaQ_eUrg;j7f(m(b|mkt;&l$<)xUzDbf)^@;M^xBcpIxH!Q)4A(4>M+668toVkY63k# zRnRVs^n23CFyIu4`D*_2xM!>jSSG0P{- z6f^)|eGmb|fZ60V>TUM~vO=x2TxnHv(u8=jJq44B#6CW1%YuJ;|JmrOGz#-t4f!=E zQbQ-N?zGC}1DlJn$*}J4sHFvS;h2gf=4TMA*UQf@jv0FcdSBkwO@>&|egf4c+J9HT zXP^Fgym!8NcndWPhtX@A&>D>a+-t)A=60&qm54h&%OKqt1F^IfZ~0)f?zPj1gLRWf{%n^(?=7vayM$^ zR0}`9MNmzR&t^AoOPt@dx)kN)tzwMD7X~Es3F2H3f?aUcUQGoI#e4s z+%k4*%UV&-b2oYEx5@S~or*<_ifc$1?LkTn4fk)f3rSzBJmvHXM-=sXF5uMN7o0WB z%)GkYiB4b}Y*Qhe~Dn!;^2@B%A& zv)<9XA0*-6_ThG7F%;r59r&s_`Wq$sn^04Hzw;GLw6kIM`zGlN1H`s3 zgojb^!|U`BDD)=xLU<$TZ<7c`pb%h!>E)1!<_8>KN#rO2)`2* z_?`t2f8h7|Aix1LP(1)D&M3$~5FQucaa9n6TNMcLUGP8I5AcvgIGT6YknoK!{1{Le zy#5QSR2i!-HYD9quUZ))Mmd;3TfKlAX1N}Y!ipcS6EJoMQ&h+avk}t4jnyR?hzJPT zLkNl*gwHk#i46>CDh%PUg3=6tqCBJT=YT%q2AsTy$=eI1Q3$h{4p}S=>!%1^lMKz( z4b4{e$7zHbI)F-Wgx>uGO#l#N_zHcY3>}dJpu~(KuVe#g%YWHMud5X zJtAUt9Yplp`G1iNfpH3fOA38%3?=V0M708#_{APHbgMdkK>iYel@cZ%4z4s$z92p)p^aUYgRjSvu& z=!+N=LKANqL~SP`{T@hdqLMWI9?AkQHn%sbqv`wlhp>(&XcwvEyhTippk(jEz$(OW zKbjy)mSD#FWNoB)g6(+ThFC<_q^8y`dBpr-;_T)S)I!m@^&E zp}_FQfW3lf9qP21p*UET$bJshMJH%Tdt*9VQ3|Ja1XoffvsEhWLpt0SnTy4s9z=h% zWOy~Ha392!42pQVPP#oQxTJ&3vmgN4WQ3Pg#*LHm9Z$%E3ifPJVX)B`jgP*q3; z?gMQ$5>!@*aU@l8wi{v^{SkbbR5097SO||Qp)`D8VGf^kHX>&RC|5HXrlH545t03rA^R^A-T zr2zOA?!k@h5YBBInxC z7CR#6m?CGs*cCS~6{b0-bs(2^;*?bLmh3u)tI?`9CYLrhmulY^P2d%8HD*jpmog9K zhXj-@6qkLK&QQAdxxX*l2+k6JNM=6NQa33d!Y{zvDWN_pHI&W`IEoewPP+;&y*bLi zLoSM<&Gl1-2jiDB1?8$am%yr3YC2bb87^DJPcxmYtO*WvdH6UlT{Ub{R;n8JI9##T zTtyg?_#U&8grJ)AxB#c50zzE`k}h)DOs-Qc!zQRn;jLMAuEr~=!g(l%J*@6GDSLBV z19sx8e^>Y(CGvzf_Y*-G1l_dgO0`^eq_9DcUx6;GVyS|bpiDxoq<^VKjjpDj3<{u{ z?O7C{lTz)!REfu1q0&-oOcy6VQg5D8t}mS@r&>q;rHIM6Au(9mX}LfnrGlL=k&~{G zn=d%mq@*FaDCjXS_v)kWaOq=nxtI&QQgB{lV*##jV?sz1@Uaewu0Fu1w$`LEE~KGg zBsj@5$Kp6g5x-Q=v<@#h_a(2crX(k1x#4oL$pF96My9@kt}>6VrOPFtXStyRzd=!^ zZit{sNhV{vxUloFRnFyOiF3=0O9t%%{I+rf^y5c_?bbB{?5X7vO0}8@mo}c`War_w z_~rJz`}Q!k{I@3U5IUC-#*j44l#YjzhN6*Tk(3r>l+M=WmiCtV6PHe+u8!r0V2>2< zO>TH4X}?+3vLn8(x2;%pEybhA4PVvT(CJHXGqcZ!1D79KJjlAqN&`0@xsadg;fk^& z1EB^{nm3lZX@f%;2zokDYB)-RC{QzTSHi{@-PgBzPLZ3wsHgF{^h&1oPNKAP;`j6u zku}p zt1k?=btzelALwWtaD6Hv*l1F8ZPPRBz+Bb>rt}A%w5ykP5@bSY(br{|!ta}QR66y> zyAF|54NyJxLkx<$t?2`^&3fZxd-BbqlST({={tU#ws}2uGNl#*T?T8+0t#1p*AV-3 zR)(WZixTihy68vMSNa=5N1EvSznk^NqEuI}jH0rPex)5NZEf-4E6#E0_fPFxGV4Ja zZF=+2F_=2)oZSAwbsX3=Hnvn|-7>PtU!rf;zhXLmVTRSCK2hR2k-5^9ZPwtRJ`6#f ztScQ~dm2JLZWd4;s3e>~8N<4wuN^>{l5LI?l`U0Wnpkw5!dRL_Iq~Zcjhp2!eXB9m zDLZ}OI#5qPsm(uT9olBnJbF|z$xMWWr%^-KJS`O1UY|TkwmJiKK2!4811Hx|A=}C4 zHY@pz0g-+laQx605js&WJJ@bErHB?vxjOcvZ2Ae)P`l=p*?az3CzJYJ`rn>u7{RX7$F8Frt-m#; zn*2u>!rU+%%_n5r=I@`T+)n2;(-swJ7BJA}buowC&|*?j=Mv?Xt}+Kh2p4~%jYmDr z#uEOnJWa@#Tjms41f68?n)92qEyJ8I@T!k0tu7KVOaaSg9L-kD(#9jx=0}O<$KB?F zpXaB`FjLxw=hK!JH3kp`mR8*`*H4EEiQwUhmX^)e_BE!F_-Ck7r#n`ghf@|lp{<|0 ztw7oY=1OJy1kA=pO=q9OFb|*SRLmEmP?l^~C;Hqr9wRqS%$8#rR(gmwu|W$E^VMTC zc!s8l_iGa`t($MwHdCKBwG?vWG&WXEw*VrWo}(MwXkBGvtIWjP)aYxpAh_byX`uQ% z4+vurt%q8!5QT2-zN}_cdbmPhX9u{Araou8vV|40tuT)9m3ZTdUp?@ zVZ#x5XN6!@w|uv*br3&nN7QX^+jaM;ZSO{SZ*y$THEd7bV&6<~Ju+=eU=2>lV)yQG z$9)a$li&gyXpuvHLqTpOF#TZJbKhqi?sRNFu4FT|9X-B%BiwzH?4{*+`S9-S5aLI4 zq$xO@JARlSjvf-;eUx@I=C+^|vMC~XSYd%yn7-pYK9XF%+(Nv}CwN>NzFJ?t&Zv3R zBZ$#Py#HNr|L*u`Ob~rSes@@Ye^m4I%<8n0VSjqUe7AUZD@c6D+i@2PzVO%DyAizgp16(CLJR+V z5Giz#+;KlMd7t=lA1rj4E%XqVvG0U&;=*)a#Dtcc0hjdhP)PE)+y0Os|Cm2V-E#YRt zPjisxLClx$3I~0V$Gr{|ZH!Wpw5z4EJJGe5p|h7?BM+YIkB|{`Xc)HWJg4lU=IXM? z?kfJ7I=Ahn_RBU1jF0GV0D#9N8nM?P#>s&BYiOQxmwNT$hlH%LA0>)3>@_Oc91@7a zvu-`x7d&ki9ov`A>c7g=Ipf0oRNPJ%T{SROW#f|aD2SkHiGHYR>Kd3>Slb_6eK5|1 z6Gg+ObS@`napeU;Pl0Ub!wzYQz()1R6V&RFCXUM)N$F3J`Bkh%*z7UuBtf-{6 z{s0{#nBtrDym@ow)4<1UdRsec#$sIX>fu3R2ukWGmc z2vMHMmMfWqrA7nnm`_$>(!9&Vl@>=mKKib%Vm4ZiM(5H6LE$^T(`>MB11pauWLEA2R zC*Q~MYyI#FQj2P@=cvOmFNs{0ii_Tmxj=1^r`lkv8c}?!HlSeS@b{aJuWsDU%ZmBB z1ed+>oNo}o$9b8b8|RIGgq~nfc)+kpeOMT7>_}T&Qt={ulpf+7UmRazeMnZ?a$(|A z3Yi+tD$YTCFs>N5c`)_q%1$KZiKq`Qt(~dnfb5Y-GzIKg%1a?g)y{f{n5C_-cu|FQ zrkh}=RM|a@WNvm76j$T3F7G>fRxjzdKDJB$bTA-pGPg1?TLi^Ck*XiI<}Y1rWQ;W4 z>vX@FIAp4S9!oAaRH;6mZ+YuOpWDaAnE0y)V|stH>q8CU?qKrwr2PEjcmxz|l7>_p zKPCdan$8TK<{T}$V|ClW?X@6Jlh8ZYEz(<2`dY1ie2r^6@gx^eH9ks;weJs~uuP_o z6K*9gj{JMufQq*S^~?9?s8%yvyt{VoVjXF>?sEonHrd7)>KR>*#~O{lP+$ZzRymJN z8@B`0=*UJrAgHztp>JLq_u~4-R#sd=myoj>QID3U!`Kg+_l_d#VV8#$8NZucDCc#f z%gpx(Xii#C{B4s^S<RjrZ<8v!SIKU$f8Puxm=dWxm-7g_u0xY=CKj$H zepm|JY;%#^gKKOukc-*h5Vn`^^k66+$#db?HG~ahcu2HELfjwHnQ5(PZPH$l55n)K z%G_^yRvu9gI=z9Iw#c*Kfl7}in2Jzuwh1Uvy;(3OtDq0)zh#!|cQqkLGOw5XsF zYuR-fBNGVoB#@oNAX6g&wCX^VdC6fV?5@~kc0K5K0ZfdC=<&^?B6vaxqJ3W2k>x8C zEcXs?bNRz!QiT^8PHlX3uoV*`PjboI7Gk8u*-{u1_gG7s2G#2nQno_#NH`^{LOVpGasx} zDju89p?}~SEj`!4`L)7%gAN@Wtc*g-i4l5wqz9++O@BBhiSoONeh{rp(8JfbH?SsJ zM9YiNnm@{C#fq(z59Vj$nX4TWPb73q7blSC3QO`#g-lJCd53S%Yy0*rYu>B#q25;Z znIDghO2rh!(-zy*C2$zmwsOpOW(IT52i;pF#$uHiE^KAJGZWiy-@xUg$p zFiy0BtjpKSGnWtJYwP5lnOq&Hw9qUZB8+0-Wh@)lU+`~R3;1)L=Z_^|KBIMwrdfK{ z_$^xW|MF_us_>PgT0i~PNcrnKgS!OErlI{^5M*@&J3cB8OGc^qoq2`4qFMJl7-&qD zkA|T#EAOYG?0Yzgw}L4ew#O7c);xrrMcM^!J9*k?xuNRFqz!FS;(~^{G7)GA`1>g4 zerqfBTd$=P;?^((M!+%>&lZp7vBWHz!vyx$r6TKLgK=bp$u+RyJ-xWKCZSagzFM5~ zE3t3Mvws5A!YzB^QA30pTCksfW+`X8|Aix3Dv~`{*prL-hE!@z=Qr z_(10oKKwhRBs@bh78fx&{JR87jnK;DXFeGG061L7P6$yGFQ2GIX_I)7SD@E-l?S{U zz#R-J=22Yb1*uJ5I2&9+bvUEu7&_c3uVmtr1cFuU2$ket0Bh)-)?tUpe2A5Ru9%+I? z#k8wPM1j+zlN>R_wrhm{(rn*l{n^59!h`tA4%cm2O3Qg>o#@&|$?ZVJ!+E}f_$DVo zpabp0mZt2^OC`IqgT44LDuRjYUnwqYd;6ZH$Kv4DSuZrOYXf z^xe~R`&?j_z8ni`{27a4;wR!qqRm^pRZQefG<@x+MeRMb54iRp(A{-#G&6BkbvA-F zalCc#lC$tlJ`iAM64ZYn)XX9*|3K89MYQ&Tcsz?ZI*S;TMRJox@{$FF&&Ee41mSIB z-2_&CNN*CxwR*ptP$gPs4nVzFMEweUr@014tot6B^y%Fe)m#>}?gtugUHoKS;(T4g zatJA1eKu`7DSiJIa9o#qe2aQmhl-SRljs9BFH93Rl)yJqr3EE&XC>AZl()#sV^`T1 z-?c&Cv)&zTK`m{tJFioAXK=&=Iju>#>~jcA^f;WiG5yJS!gJWu$#{=+c<^<3>$j<@ z$+(+y_>*<{_qO=g^yrqzu&_6gpb0NNkoXPx6Q9DrS(^_#z{B%QFbeb*p~zr!&p;9- z69kbF{mA{aMMkWdOJqSVVUSB=pDXU33-sUN4JYSMCzoE^;!6H0w5Q2dEy6A&|@VBU48Hrx<99GCo(joLq)dpELg_EwR1;=Ps>ao}zf3;15XN zcU^rkQljZo;$5hn#08|%z9pK(_50^&Zt|S!ccG}9^K9%nN~*+P`tvBqcNOM#H7qDJ zj`B1>6q*lt3O9LTCi_$eH$MJAXax3(iwOAlks1lTx2@|0k=VzO#XCbiR!^mF?F8TSh~BQYr_W0xf4l zBXk2RXMM}&y_}#tW`m#%?hW3?Oi@udEy>Kx#=H(Wo!vMpr!^{55S7)9p}ReV!ktrZ z7wN$Bh0^neRGrkw8>i3>Hs@;r7U~B5 z@P2xN-|cK|?(oeHV&v(&BBMpH=aRI3io!3lR6R_zE=q(xN=lG9w6bW`+se4aG8yY-qlg*CHnj7DUXhE*`JEG@|*`!M|m5=iR4n}?sC z*KShMU+h**TQ*mm(R}3DU0gm)Td`E^1uFJ|FR8pS@kM`qk4oo9TvAP0;>TE0!)fX- zSW+u~>@QzZ_ro+m&{XXPF>s1n%l0TLC^5IGVEfa6DE0^*{WfW_Ndx>(Ce4z1gX4gH zy0&q;pgFqswd0^Yx(<+O@C{w(i)j)%eb)CnGoX<1c0-FQpUc43juzlQ;}hlx0(l4AYEd)7>Y{=s&}XA(<)A5oY6*=DZ)? zZw#KEsLk5a_73~azK}x4md*$=Oyr+-XP+*ymMy=!d)muZ`pZ_v%U0*gR@chb_R7{k zW$QO(qtT}gdkeC=E6iI)b4Lg4lgWiNJ4>}RpHq((&@C1P&FBf&haXNCr%(3s%{H9N z$NkGkyg>)?<%jm5gJ_GRV$goI#c}n?F|qkh5(6x%#M@}OMaNQg+ao+`HIY53xGUu* zPQPTwGQEY*F@>h;nMaX(ho=?@rn_NZ#BV}7EF5vOg9#mr=k_N{@|-= zy!#*EyKKt`d+>vM#bbQM^;f2+cBbcQ@Y7Po)iC&e&+_>OQUQ6Xcqz67z%WDcoxGSZ zLE%QcgK%0s(w;qESR$oZLFb%7 zSDYbkgOLlE@4A^$CMr=f&QJ%yfbGiLrAlP56)NQH;ieK5;rurH2*y%j=K%x;ElYJC z(%SaYE}$T={c`eUz>W}&?Ct}JGXJ(exu{-1h%m=FW_%xW&hZQUr2A`gtLU!2;=KdQ zGU6FQ&pG@(Q-Q1nAq)%iY*j-p3&eUD-=ma(?;}agFVqAU#Ef6)T1>7Jbv{uZjcE>`X=N~8l7R9aC!@)P2LD;GOM1y zL%GO#UyU1c!LV*j8o@#Uwr1S62JZiQzn@Qudr2J=LIJ5}WMX3*_{C((%2d$B{%Fmv zRKtN%!)|Ir!*$8*z{Wvh!yRJ79Am=@bIBIK#>>aX^QMNnpyu^=n98RnoF)ux;THvK zNe5mz8!i%D7)f5)+ZKguPh}NMsNrp4u(Zvil6JSG3vg5totFvp`^v3 z>|pm*sZKedPKAj>H0N5G?OL^_PPN5O)zVIVp85~Z^A3B{0#2SnR=2wwAUWe)w0dgM-?wlc9rC zgoASemvc>neUF2y&8_o{gUfn@>#~Ep4}{Cfnz`lS71t)dx&FT!kue^qgRVVaE@beNn=RK zt#{8|$i$sPgJYP>U2sby{dQp?l$PziggnM4Kea|(h?GB@!&X>h=(b}xms6<6UCeDm zxX^u=OjC>&Ppsm7oSIX-UQ>c8Phx^Ygx;MG7pxvzx4a&lH5#hcz*CgtT&#<)$7AVwOUp;T7O+cB z!ed9yLvspWU57+2ooEUlr-HUiJ;_tfBRlIazDAVRMkUv_w2mWS7 z*TIn1j+oZIjMf1J*Wm=-u7H*$MSH1H%2J$z5_A0PUc`9>}uhcEg2nI0#OUHef4 zhH>2{Y@Vjbp2z8)ClTCcxFCF^Lan0{Pu*&MW9zLsIZkz--B@;>x)82%U=Zd5+7`&v zM){g&6WW#v+U6q!78=@?Gu-9}o|9VzmgWVP_ub~e0_(SJYY>4Ag!V-?v^5OD&6wxS z5x32mwk=D66}p!#D)${3!JX~e+>QWaZ9a16iFpFSU9$FFpZ1Li_X7|2!{z6d3V~gr zm%|+QBes_#ANOOfm*W)o(`CVv&+Z_#v(;@u@J#zjhx-}H%igyCImWsomJqn%`GU=3 zi4SsV+O`jAzdCL|Q1bwnxZe!8UDkJUzBOe}p`4Lm+pLf{5yKlJR>%`6kXf*?uZpEps<+^_KkGFCEiij@8qK`on9W z7h-7!4Mq}qp>lW2^ooq}(g-j~ryDfF13}{vP*ffnH{1B2a;+&4%CM^?kR_*|7W@7S z`5lW+X(VY0B+8Zhg1je|_DizaViiy}kXg@vrIw{&JPm9%N6!P^A5IW;!y&@IDjdm` z4&&4?0TY4@ zwYfe%Jfm<|Zglyg5D12HS8er$67adB@>K5(#nP#@hw{|yjivIryrA;d9!%xRqzH!b z)*byWRcmoa73p8IH&X%Zk zgbTFXoUXLEK+pwS@6LBdQiLJ|+a9h@mRme9gxa6(ua1{HB7{2rSl}Na7}}oD=#tuA z@Wgn5+TO^F+{%=wf{0oAIAR;xzBrnN3%(!z#Vt+%U;7?zZ%~E*af=6QnvE3y zzi#n}!CdQqaf|1p*#G4g7i1X^t^RLr@%)m4vA?^;WtjZ_<`(C6vyKS+;}-XhqA`yD zn_K*2qN0(+TekP_$%;tdmX(tf{!ec4l1~BD$vKo5ng4W)lOdA=R1~Jk&^BJ(;`3WH z*%JGuq~Oz6x419z&BGtJc=X2YhqbpTo3O8LaVr#&ckl$SZt-7VdV^ox;yeCUWDV{M zqzjmbEbE@pC&G|$U=X&JXS^IBP|?ZT00ZgTOwh>+8(E(8PnMp0y&*TzYiAX7?IqMD(|SUX|domzw8K9 zHHa-adwxFU`hi6e5N*S%&q(6mt?5oB?IGFB`wVkInvl<%-LLn=okr?WBXZh8{~G zbjbCg_IG!-NwEb~7*(S+T=rqiO#~r)uVA1P1@~kazPJ7-AE75WJvnw+YWXNiWjKJ! zVAVp7sq!c79Y?94jZITSWB88>l~NSwLEOImSwCd6mHbS;Yl;(NicoHj;!uxwYFHEG zdn##%Bmuje#8hsCBXm=7zol}6I*sg2qm1Ien{@*;U=CYhIzQ*=#E~JzEH!=1=iV4d zee)YQkdEo}1FFdILkB(r%Pg>cMi;|{CHDZ#nA8mM`Yw2F!tHns=UT8pF9-NI!XBH4O% zbIgSFknLH4;kIM~(*dDGjcNOYii+5cEU_|OPkK^hfm{_Yd>$Z&D*Uz=J3t+)hi=k} z1+!)@5N43H!R%xRc3r;4E5K^cDp6aPl{1$W|-l!}0F%$_b@ zFg?VMMkTbqmNv^+K3^{wRy^OVIDnsTHv*Vm?k+_?zuZG7MZ7$mG=N_o&jy$vPgnDn zkmtMo3J=V2Fc<;>d{c%%OLjuzpLrt6W!qh@0U zNcNKAp9f0HW@9G|_EPem2PuSRzb%mLqg6c*Rvyj9X&CHdbUqK!M$LIYAlVOLO+F7b zlFh-NAMEFBJ`b}B%^}>E9N=9#4|g2RA-)|P5PUd~@I=i8B1jF2;$K7t%I1>d4h@R) zUPMKP=8{uM4N0qBL?@2sQgRIq$va=fWTNI#OGpiWO}>aNl+B}68XEr5d=Xa}nn!Oa zHKMw75#KnP$LKILqWN%<(21JQ93VBSi+`CoD4WlkFf?kwdzmyDn$KP!HD;oEnY=KX z&)G0EX5oC9vVmH_Js>r1oqU;kAX~sYKQwONe3^C@TEM?AHQ~H;nSM7~Ab2}8;r?)$ z0YEDhMv$KL#=pu$lq(d)9iH^(y~@H2D-@%Wo(fjI%6>mqD9$xJ748hV$^oJkNlHjh zM<-w9Qp*)dD-BP_H(%v3hZV^hO3x(Ads}g8bnjX_!={2_3xv^%6<%!zS#!c}oOeZE zOcmPlc(02NeP|nUi=n^Cy%W{QDaMY)CoQ|^)Kt7QS{B2F$#n;o;(YI%u&Sg=P1Y{M zBLa+NZ-2)Ea@ew1`JX8mWY#>y0iP13vzCHR*8XMqD8aQ}H-ex$hWElw4Hj{OjwImqxHD+m0jlDq+%R8?! zE?aqtdDUvj&*wG4!F>6}x>FBNJG)RP^CJ2A307DyX-dHgQj0|{U@ZAKJe^?w3uBVs zUo_>L6AibOhCcA9I%-2b-Ty$HPr&5uT_aG11FF#Qpe|}RDb2y<6-#uVmnNLC%GS|8WI#4RJNZp!vMB3d?ba}m*Ulb5napnu{lL{NUoJFP1_bo1%LN6GUk9k=Fc6qIo$4vp3i9Md1lnf8(QnAG4zDK_ixCiP>eydv)T|A*s zRyfaa;G^)NYGgogd>VK3rz;?UJymA=g}U*Z`uVpXO!j>jS6`s=XP!ayZjq~p7G6!b zA7KB<;t{>7D;sE?e^FN^tp6KzwNUyGb@jhw@mz6nBmS^>j!~3Y3IAa6+@kaTuy`~D zw8-L>m8GvNo==T(Qv+K|t6jT)mt&7}pY%ZPP)Fg&=)@%7_|)H7JU$rL8T<9+AFx}y zdyV@NnBURJR)1h^eg?CqPyxO|Q6AHra$hJs$yIT6_uzQJ&R|IeijrV?>7t580mQc6 zhLC9Mj6$&>^t^p~$--RlBu4Wml7ncY7GZ`m1%D^J2@<{}3)L7PKDldT@so;>;2b-0 z^3i{j&H6RJ)1g8eCxyYyQJ7YA4>OUi@*xM&DnQIygMW(()dPXHziwxtzh~aVcigBLAEp! zrhcOwU_e%2cL3K7qK*m2*}--sor&a5lq7n0PNWp43VdV&{ozh5uJj;zoHAv#NW6N# zVJ@1QJq@Lc)*R+uq7kv9L9$tVj$VpYe}YJ=?VRy`n&aLd7Ekbw5=sAthMLU--QfYVEm_K-G8-h_rGPP6!sD#_vxi|0>!RG1rYa#U29Uwl+tiq30TQrS;? zTv|J4a$MH9SNy83ZjO%ssjjL9DNFuRSK@T1wX;7=Pa$;+1|_HUEB42y4IBP+pvIke zQ&7`EehH}gr2ZJxa@J1=ZoQf_1-ISpm4MryZjQkn00dWmc<5zmtFBt&(8^L|CREcN z^cv(}y~XmfEX6olW+wgk4ks+GgdV6SgQPCf)mg zEc{F`zgaXYD7#rQYdF1Gwi;l#U9p`vzg>0QFS}iHy*<5M_e5a4+wjG;xZ4b*D!@5Cxu-0vnDmf!ECI)Lu?GXoeOAP2b#77vGo1?3M%r468mn-RJ+=;;Sbk>V&`E))=Rq=E&$_0M9oRnaCzM56~f7R7ag5}HoK|#g8 zQ&*6er`uO`#ma{PC;@gt;ag$C@&8?2Rencu&4k4`FGixw^2XtZhG+YSx(dUmcvV-C zNUMt?uo6}jPkEK_bZOOH!76HacihlvY*l2GgymI z#DI0ClOh{SmV`S8QO`(%9JmpI*hmP3ISiz**obtH&BcBQerrHacFHO&B~A(r zL+<*ju8c*9{kxzuLqAacLtSZ+vqt?*U9oHLz7zeI=Uc|t3M>izF|9=Jk|D~Zx0P;|oT^;=c!Ng>k8N(XLgF0wBlj-AA3yXoo_!7T;SN&&}$kr`3 z>-PG$Jx}uczRgEmd_Uki0cqV!A2FWo{5XMHsb*3V;X*+8w=^`%Z|?{azF@*3vYo)! z#|ucY@8(9YcO0``yCcnir=fk2iT+7LUrC8_!z4@pp`jDwDUDQDRd;Fd%hY?4yLq8Y z7ix@V{-U9`htp&l%2!()FaJjxTDGZrcO>z@(a_DchcktW|D>VQq!Q1&^a?f;90jugW&eyzZiMLW^5oc}^Y@5U(`6z#^V+aK;G zX!|z>E9sJM=_Un71{r)XEO!b_wm_aXNZpuR989%`_cl!TI2ke|aerx&$n?fBImim6 zEI!B%gfM)_m*K*=WX}*W&@j$ z2I-P60qK_RMv-oj7HP>%cXxMpNJ~qDfJ!$Ah;)Mj&&F{^opIut`+wd4=Xs8Ksm=?I z$BDGb_?4>8HakYccaJtg=BjFIAZhGRJD%Nh&OVa)ii$*@0rd^%DiY zd+qm1$3>)edR)-FPDm2Kh&3BVAkwhbvS-=Imocz|u+yU1(Ri?o_Vgv$NbpnS)8(0` zFR%Lv_L?qscstv+ur2l)ZsBBFcQTTuly$8MAnm`xeNwXDP3X9_-$UxpaL`K;Z*kB^ zlV1W+;JU4YzfoW^GjQ7&ylX|x8(E3Fz&DDV{KxRh_laJ-RWZLjdiRiX`*>W{@$Sil zetIg$I|{|2lWE2BP>ylKH`^~}AJ=KuO*uZbbbO~aGsih$42{AuYxo-ZY{sZ~+Y!Mk zd#--P=cAj$lu^%RJwl{m+WEXBM1heZ3Y@kDn`ax+IQabr>_#DYzgGyV?EjZ2L~C0M zGC3fGqkG-{X>R?ah(W-v3c8h9pzFT+E$K>fnmtTwnJuUxa zMhot!-H8J{Z6+U-*@Zh9OMcp1Hb2wVvICqgYumzTE^FT;ODXHvU~w($TocwR>sp0O zaIahx`O3PNH9{(@N&bOCe9Tk*s~OEL2J~e{V`r;4p3rbaKAAKKzqFq+_LuXXFfHbX zp0P9xb(poS({`M8OityTGVe)!@$Q!yO_NAvNrYtB4*uJx*&V=LfF)4(+hAZ)q+{{1 z%IpuA6v$w({Z9ECCgnf=H2XFf2w{`prX)fzDIqvCs986d6dqH-+Ipx^3_5v#s4b zIx;yGLPpj`6!}1VFhn5ydZ8R0sEZKOv*%xmI=R#SX1@Yd;VXK!4>Z{H7Q!S@;PU}1 z1pwVSq_4XLp9|-9Vgc^g`*1@iiK64@h_CVZ1Y2c0qX() zT3|sK#L(OA+5`~|P%&gQfOcqJp+e8coMgFkF-{oENe0x5j^qF>heF8TJM#i3dShR- zLz6@T+vpXI6)6z}>5|rS<7Q1^{CUZabAcgY)*J8yu4sS)P-+k`nN}-fH7=oB8i^zP zo*onop?ifsAkP9IRLBv;Ojk~F*4Q3aoD?7|pWDA#!iQAS00o_+H}d#e0;-uMzv3!| zR=Boo#|SEML-c&BVRfpXJRF@Ft-X;?7!A-Y`SR8-y*a2sDj%A`f2VcF8h6|C9Yjl> zV}P9YI$)%G_d1aY;P<-{>?HPI;UpJ(b<10a?DsfFoMpn_X*u(R#S7C*?Q04>4IZF% zvOpgsoe1eRm88NG%H>jhNUI!-0;U?F1!OQFiZAH3r%KIi*%1Numg$+S(wZrO&}Rm# zV^@Q`I{8YCJZ&JZq%5fLwP8F8#3&TNI=xJ|cI`7cj%3>yHO|7RH?IEeG!fJ;M}jjX zXgt77oY@V}<$ImdEa3oB04gX<e^Fz-Pu39ED&Z|7} zfGX6VU}-#2OqKgYlwmkFIp#*MOVZq$@{n|t+mwcyyu-kWRl zIBWB^SCvJ>SX!h|U;wAjABpWB`?93qAWuBJ?evhtcW-PC!4fM6NzYOnyPE!CW8J{Zm0))}VlNS}Cr4~lH9+a~t9! zGee^r+mS7NO0NZ_YN48oF{_(z1+Cb4l`xKx&~TJosl`&&1;$o$QVE6G&k~U%D2D3j zsGz_X$!gPz4C0S>WR4rXE$~`bVrE}WG?UhTqi!4t#T6H&=$!Q-RX%m)XhYCb=%pq) zT9qBKXO0~qRB-Ajh)d(TAPWxH)4QrrH`p5k)(7Bl^%oWHlNpaq`a-cU22Y6 zBV6#Gih68%l2GGvHBh z-7Dfq_NPu$RBqAt+8eK+!t3+I^&DTwZ98$=4|#@)ezcn6f~^zr;CYmWeO!==dryHC z2OQ=&OBqZ71w!14w7L~Dh}I5B)XufZ)SrETM*2!hq0IpE+SW_ezK)fJv^+ft07ZPw zrETqPy>KjzM;{Q#K zna@mlb{y+#JV-8FD6D0P<6Bp?|EbOB^A*&<`0F5|5T$+P2FJmIic&5kUzO${^4gtcU0x7>3IeOU)D8$pzHEy4Z=GVlobI(&Z#0Y2=;3l((bj=Q<<_NTM@_K`5W|8FwS1q}?M5UI^E`5Th zarp2-fl%?Py*rlctc}F`6KWQa;A2VKHDx0mKS^a=)ifVo7Yfx~N`LCL05)u zC?>Gq)Ah$e*XLDG4R(Q-aFh>gxhZ|I00F>A8oxRps`uk2(Hr7Voq(ryGSe&~(IPav z%YmxK=5kd&WEB!oYC&2(vO{M9fEo`nQjT&(T|Bhl=ryU{qrl;S;Oj9%9w)9VX1dl4 zb^<|b&Mb%ajsT8#C@`brGff~E_PPVrn^{5*p8 zIinyR?=|U^-DoG>vs)K; zb*kO{I1{{3v*&yk>OlW0HWW%7XmxJmk_oio^hxF*PDKtddZi8i03xn`hg zahO#sSf91E310?0aES|(W-;j3v+ml6vwCA(*k!GOle+HBXOoMBCuoBYCnR}C;v@Dg zskH0eKAzdabVO=4KR>?JKAJ-VvXUyBfIz--YyZj)bA5JxqQ?=%g;I{59<_yh+P>s3 zVd-VwB;)6hoB8vqWm_PLx}KOODhYGjuq%8*P%7|3Qj7|BVbAQ{Gg4qlr>Ryog+s5m zPF>E>s9`}x6>WIzEq!g8nUy(f;$Qb%CuIf}x zP(?Q%hAq{ADVRMZiPBjuPnsQDy#bDCyN%^23N!X16ndS)qo2nglj0**k*J{yvyl3A z#VN6aO&rTAMJgRZ+n#bMTwwDPBt{tWJxaBooM^vrw4O0-oGG!95%*qk`h|MG zBTK%4JR^+APHYhjOn%37(JQmC;`oxhNsJfH>Q!|zt~{lEkDNDbG$Og7sp0)UI_ogu zi)Oi5cL%RMfm2YU*NOD1y4 z1}6s7dSYFTp)z|*;tjysFMHVOS7uOGq7i_|#TD2qzy~jac=kfMO<)$_p++QG=;?Jc z-`zuGtESbY;23~rlz~~`DiCdBXLqTlbg7X{u90r4k);ReJuq9pP1b=)rfjZtjW6jV zSFEq(WIMw{9&p6#8B1}g#wmRmqn@m(@qyRt5f^Kg&}(+8^0qKKMWZ9ihy)YP9JRtH zGoq%&_Bxz22R6$2UR)fw&@s_Zcn{1?vm;mE#(4SJxFm~eoD&z@(h!V#L>ki-c%hM? zKb5Q!A24Q&?#n;$eL!$z5m_uM_avEqfU2-hM&~3nw*k^x2F55><3 zE_At99Lz8NgwC;&-w%esm&WL%$KK$P^}XpVL%1BmKGk+}QPkMtxdWl2D96*Nml|9q zWsXbM3S)It6rE({#Y{*;2Y94Rlk@QWyZ( zZ4KbDa}wnFX(I4Mt74^+_dD3=&oz}E|D$JC0jUKt^+;>AOjS$u~;vI85@f!V-o`T zKXapw*r2?MA(c`C)hoqdLf!Ru3ew6KP1N=^<&L#*_Lu1`GeGoyL2#YgEJTZE>c_xq z_=d?Rr7sF8mQpGbWZIL<6X`75;kMN?5T^)@MM+G{Yj;5X+!Sq`ATRhpbz|!O=39NH zm?Ut6+(*%(334mVtaufz#>t+>38biv4ZS<2T|tS7lNnMIAvT;9#*O+(spHR| zcSy)fj%lK6XIt9j-LI1)SR5fiF9o&RbQGqm5O$ml_MDY~E-dISMI-Q%v^ijNYNDZT zy&MyMtfEvUIoE6NVuqIl6P#W1xtfPT#I?k*THx>X1u4iY#y1~#srFFMXMz!L3DMCFWzJ5^el0Mt_#+XMB_W()TQQc6LOu_jhL21 zW3NbpNAbl%okfL>Z*AL#?Qny+S6t1NQN{lWR*09Iogln2bx!LR?D?(cp>62bZuhjX z=GWZjH;Cp}Qs=kk=65dV_aF!j{)Hp$1>=hOle-#csS7eM7Op>QD_mVJ00kCdbQVGG zi|}cSh%Jjq^NT1~i)h447y?UJI!oB@OSoxE_$^C>^Gn26OQgihWCF_+I?Gh<%QR`r zbS=va^UI7^%MdgQi@*w-&I*V73Rl_+Psyn zMr?JpwuZiLBTsCnvp)Z9-Kho7xn&HdPo`~?=xj~5z|GEYKi}J4941;R+kB65YuSCrHezR!7;bBR=jq=FVOs+8zul2;{M+6S0T*8Ww3`Ph)QnEgTl({=TXrB;W#z zo4&7Hu}^rtPfBt?cK?7v_khadfF}KbuJwRn;ehe_fSKfw<^Ca??jeWAA(z78ar_}$ z&tYT+Gau>@?|o>|M=*lBhms4(66r^>3rF(TM~WoJ1ow*<}{cHf3 zkjZpzBf3m!txb>ZD`QKi3xrwN;`4h2IVt-{^hXeBdhH>3c(EEjP?c7g4 zX8!iYIO``3wAOv{Q8e)O8iTV$Wj-raCCWB8t}rL$rTlKt3A>eB-RLP_S{ zRxfySBt^T5-X=eUR>5Xg^dxT6+s`w+qn>GZK^WJNo6X-cuK#eeQOy2w)%}HW-T$Ka zH^%kUKe+1Zm*emdePdihuDXwj2pGR-Tz|XjvVZU`GaUHoRrkzzvOD@?5sqW_!EaaH z!STPi>V9NQxPSlZ9qdUR50~ft`){wV=1!rYp#E^x-9P|iT!}8h_d_(vK;f@Gp_%`n@wKfT9QTWhc$?=hIaisUkxR54>i5} zRiSi-SCX(vW)5Lc0ZkAiqI7ncGte|AMpn9zB2EWUDldtr!!$p{6|smcD=bK|Fc06$ ztf-`R6QV(X^wm5lzPb>v=%^L4s5B=Gw!rUabP=x}j)StSF_e|Dyv7~Z`6O05?a*#n zr|Ejj3VS)sMXX^>)zhj`rDWhwd^NmFR&6{;vdkzT5(|+I#5Zl$oh*8A_pnhFN`zlw zXM>-06S~Y)Hj=1RKI{nlnRfl`oQ+~U*-`{tCrWKF=i(2uG zhp&Q`{x>}Q@XYovJbb%O`T_Hy($}n`e3@?s@!yR1Wlg`WofCvcj$RZQBBp$AvzdY~ zwiSk9ywS2Qv@>(=`7U>ldg$y3fe@Y+@Dly1i-Mvz!Vf5RD^L(Vfy_XJ347a!4`emf zO>~gqip6q3*6vO9S3Laxvs&>x9{x862%enZ@E0EbH?8=EhyPQp_{zio=ME6|dl0P< zcmz6WU*@d5fB)Upl_jj+9~~ez_w;;(<<2~Pv95gd{lt$YNPghqf6c%OPHi`zD3#`{)A1xd_A|l;g%=9GRHmPRmq}JAL;nU@m_U+bP86}QX|AtU z&)%g-)RL{x){H9K1jT*F{$0W+~_l632q0|udm zv6?XrAXKw16VgV)Go^^#*Ym-WwrQr9_|*X-r7dWS z0VzcIsRyqRmEo^Ec+Pk9ZajFQf(Yi-{9e=6>2AU^R9!@;h;|-&oU=5mmnZY4`HOY) z;B5ZRXB9WQUWVwz-3ke;O)tJ-J21NI#@+V(Bc(*oklJ#sq$fpZ@e+d4W zKKwUliSP6Q@bC12&-@s(u zOBce2#=1=bB-IP@g?lMj!@7@K|M<3}QFG-z$)14!8?(gk_2K-ZU;qEALo8Hl_@*!8 zA2`H*(wFgX^??#|U#meXPQ$q4aeTZhf-$8*7`IfSvACCUQg}(9F~tAT0P%lNLHr;7 zvw=?_{ts@5{{sZ^e|SI)IX`uXB^XQlWhP~{o96F_oWts1FH1{fg@F?!+VgRxI#FSnMH%UaeBc*|NBfwQG;(-yLGPgmmb|H&=-&H@(AGgt=xnl`#V)#WczUq-XZ%5 zSr6peN$u#;7gLY2Ltnhrdo<)QqeTDJqdQ8xux&F><^P=z6d&U9a`++VnDh^Fj`NJ; zKe2mA#GSePu@y%V+aAgPyH=c-Kq@TOM6%R0a;#^6)rylY#6*h}{O}bear{jyjz`xQ zmlvO!Ix-pJyPH-VXK0e;MMx`7fbPUb*y`4<{|rf+_VSMRsr>oIr@bpls4E0QwQQ1h zZ%Us#Y*=?>I(>PFg?(%+;5`G)D1<#nd^!UxSdcIfxWrc|VC?9w=g|W9Wms;b+Ob8f zi+iI-^XtgBBZ#kKq+5(fW^lrF$i$F?foSd3FcRo8o?#hMk3R(DliE_)=0g?0;HgIo zl%m@*ASRQwBni@YY)uoYce*ROtU%4661(ERcVfy5s+1Hnu_jotLJo76m?RowGuCUQ z;G+=NtM669h!}#u6gkL!5JpjWV&Y8BjErhmy*WrIvh!e+0wY`RGV%k`gi~Za?Z&lWaJ!VIf?DS@VA%%h|qm<|b zIJf&QYC1tiQxqgsE?_9#>}?|iRGXX}14=&9c{mF<-xV(snl^Mzd{3Vw%c5c49a zk#^YwymSH9>sp8r?iDc_*1}Shfm{jVi#gmd%mv7Eg=mEV(}N#iip+YU;3W;-l6D9a9`>FPVvf^%w!d=H)?0VD?s zQT0pdyx=w{*W>}u-O!kgxq#MY@d_V|qO`_<@7_gHkNAQ+Q#2B*%GME2x45)BkVFNi zL`NdrRs}1pgP3~_tGfhEGK664#7J=@Mc3Pik&`d2M=@ZJJ|LIQ!&Dsd;>PURq1+a; zJQh;BHM%u97Xb8WrPLc?pZ?W}Ux#gQx_2`7)rl`KN~N@Q7vjW^k9LGO@rmn~!XQq3 za$N<86aO7Rg&U1P58}jkv~U->22t4wbiLMRK)Kj_AHe9fNh3(y=gm@O#mg!Aa+~KZ zZA6(^g|wZ+r@g(iE6*)a+}5tCG`?OAG=(!%0`zj|KYkcDpCqASve-wb%YJn1*060|d!Aeuiy#B4dE$9JyeSq$7Npb94|x z7u##Q_WWu;K15dFA|Sk|aCX6*J7hKiCd+v~zb~7`5>wZ(KGhZ=9kRTXkl^=z15SW9 zCJXz6BM{UrkNASOgZixQ)iq04<|E+{PiS2=KZd0JPzSvg=C=tzZn`@mBng@5d+T8mLN_+hpW}Z-peYwoLwJ%ArY;Wlh z=CSD3L2{0maM-SSY^eZV5&6TUt%AJUXKG$<)|knQT3#Ic?VZ+1hcKf)Ks>j-A=^LA zISz6e`%{hbx0GULSkRcDPWQq4xi4kFr*=Vah;;cUd1j(%`-zbED0?*#9Q&;^<4x(I4M08RNJ>@-kmZ zY9u_05z1cPZIky5_uU2M?FdOl6IRKrcRZqc`VZwcF+1pBJ)@Qk)=BB*%}m1jv;b;0#Q5^PE&h#Da*=@u&JV zPO1`yN{l&O{L5@q<6V^+V3LZxw?2oB{{HdeZI>znB=_@7K?a29&1J1(cePC{HUp4 z^!@v)TRQZvbEBv~{Q=T90ZFjn2lT_;Fq&ZkdX&W6g}PX`vRE>2qkC>3Bh2H9)$ z`gD-Mo&aXaqAN>%J6c$U54G$PqUKo#?r?=4=im~G%ULHWN@akA>@r63Sr?6bWzeI+ zW$dQ2SIl9RAx~si@Mq4txrQsl90ym3Kb`dmqEtor%Nm(c)b(=lu|&lWu2S)y_bG-| z#pKJrr_(&|SF=)%^C(!o>vBGzhf_RK4cwMow6>w&NFj9 z>@-}Rb~?B&@agD<$Knimrsi_g86j#K0k80+zQtiom!b_{7UoCdR0D`wkf1U zsrS&dv14Oe;F}4@bx7KWknbeN2J@>k=+=zo6PJZowL31~b%}k@_ORA`kb|m_SN_Ny z*m6p=tE>>``kDBwv>m)>2j?1w)#}tf0oV`a{Jes9|9Q^};2<{s^LyIX=L66ohlN15 z4gPfKhG+#?a%$mcEN0Y0A2W^}u0Kz?pnOVOa61t$xJ0vS{e(4@2GdLu2-%8$Cc|1d z>tyPXBOlhegG&T@6knsOS`MAVOoF1QC*wukNR)ahF#3kc({xl=vObP*b2ZTUV#tZrB9?kC499-DW8Y)A$k*IG8TzzwC#M@%pq9-suURh|iEee< zyAle9?aX6?+FRJ#8&=JOslfR@ypKbV7GE!z7uQEz(MK}hM-d3%o)CVx=el;{!vYqf z>xCxJ!>4*=F+R%Xr|-tTXD3|YI~AufApv){z|2%lNM6sKzrrs-+_%r(Pbq`1-}~VM z6Enhf8Ff8LCL>IjNmJ!BcyY`C!&-mx_kJ=N0UNba#9aYmi#%1{qJHou;wsR~)u3MJ zKsGQSO%0TW8E|G*T+6XF`46LFFDpd$@pKajmE(BMdnG6Kr#L15^(g_5nWuAqOCxFr0DpSIk=+{6QfZR!};){@193v(Fz}Kz!c`g z)E=l>t%TGXyKl9}jO#(~b6Ism$EI1w7>i}_s3uwlGRIQ)z-k7D$yLN?o|~(k$Bx*> zI)XzvnGp}IX%jDnk?52X<$gR?a6&#k9)5qU2qS80$%FlUA*VC-77Cfa z6>_>jF`N$m(?U+LbUdVxlN&YFJ~1OZMe&yZLEs8-kz}a@+oCaF-P2YIr5PioKS<5W@f+N1rQYS?hN#zFfC|z#=8tv}H z5zS%v%JIdgTK@M=r}FTJQBU8N*e{ap8P-M&xACF&efGwv?5SpkGJP=XX=T`XbF2vQ z(f_fKQ^8RD?F2O3pRA-Fgboceb~3l+j;cN!V~`pm=xJ>Z-Z1c-fG_r0(|-G~LmX zpB8dL77X=jkAT=`FrH8EnA}9rZ1Od~7INy4tNyf*Gv&?*sqjOZv!G!2wt(f|B4~y< zTibsgFnLUiBTuUrs|pNw9Cu|II1kkG>Z3%`NvZ=QJPAt_PJGX8Eu3VQ)VAhi9jEvt zg?x~9KGmrx<4O9#CASzpSeVKrbB8FXz&VI9NGH3Be}g75dK%6&H-Y8Vlf0GoTH5?2 zH^j|&axT)ff?qujag2+)C`SMFJrCjM@>b*t{=HsBjCtKe|QvaF_tf4dIQ-g6J`oUOK;+ zl1aGN>4FN>)yAqOy23g6L!HBO>cv-nI=a>2Es|(&?_Z-IOfY$L;UDRaTmoW5!hu}g zP#Ak)`3Ni^HB@UBUWBLUBp8PJtPNH^%v#Yr0~YmJBl3wL@60~wxG|Y9I!(ndh?TP_ znWekm@eG&&oB%Yumym8@hekSNs^V7d zSDlb<;qt)|NNC1jY*rayKkF7|26_4YtXueh|L*$xUFknh)9MD>Y z@;qi|AsvZH;E5gj5dRYoo=XvrbIMCNQkj-i#YOO+rD^~5uGEOnCx=&>u}?l)z6knp z%!aH~F1MRM=5&>68crj7}A~b3VM1#-F`ppEI9#+K%rk> zn`v5sQ|=#|yeBm;L%M!ov;A|IgS@2wO_R50VAxHQw`oLdLgFLTq|o@(j7)wsniS%! z8+yoHKnbMDJ0YbC(&X)wPgzs{QX8=)6#wROnb8&m)$`wL^8WVP%n0)Sq{;iImrFAy z=zr&O_y@0zEmk4#ah5!?ExUPqnQn22c8urtOrR;|%OQeFtITF^k}D!MuAL=e$ro(y z&CJg-1}({-n6dMCe~>FRd{Q*GcbGnxdLLQgtV!XZN+C<(X%zRT_uUIk7L32rL%zRE z1dPLClbWPmd%{8@YFjv4_|oM4zxvwj{tvvi?+1s<%4V%U&#tqG<8J^( zzRQa80^|Vvz{h}Zki??0-#^4fI&ud5|8-_7x0`1Qz_i1;^1;^I#u z;(zvQuOhS4|1W>-=@e=8{U;9*0+IOrLsZ2ASA7qWcr0xMfk+f?E)${u;PB!E==`94 zCO_3a^zVP)+*p9KXS-N(?4){{hju4 z^y$?MkL5kflDC9tUs9RwK*(JJs_kjUrz^&VE_MC$Eiyn<0UNIUi0*I?*|KKx!!?5R z85p>ijeo}NKuz>?Io-NF7Ds$de0gzcx6s3+aR1u#>Jmp047?K2^E}V2Ai4J1BJ9l& zKoJN~@WGfC(f2*l3efkcyIY^hP3+F407>PWECP+*_0`od10tI+Qh{V3n(fmV%-6(=rjej>{ZlF^k^h#P_1OhuPp1S&|h}Y6wwZ; z^mE}ew$qPEV7x6W?Fy zsHf#Net4i?L{|&Cl^K#U2)sRqmwvhi?_AO~O`LMf=dP4j6VO0)Wq|pa>?p9T@ZPZG zCsB3InFX~HY&i%U$2F3Lx`TuVYO>RU?&H6ncPqI7j{YlDS8q5XH$yvCD+vR_=3 z{W=-7KckcFY|fXAQr@s%ztYJ{uQ6$L{v$fsEm$7xgb%yRtxSfu4;VBkg60??zZ(WvrJu^_= zM-gim@QT(fGf4Fh?AIoiq)J14-fWz)f}gQpxx2(kaZCI>jO2-+`+V4_EW+Z;AR?vxjs5B&b}wQCS`!0STCD5MG-g3v0!o?nzZ#&jWq7bg2LwH06f7k}0BZ%C8GAZ9I`BnZ+Z zOA0YN4}vtIg(Ax>`Z6=wmO0nfG&D9fvz67q{BBGIx5c6lj&_!C5s#%qPYjGt&v78n zQ_WJYI4shkalKt8M|;l>*ron(53o&zp>!aO`icAM0Wl!6j;gJ}nzkFtH9A}r;G&#S zs#doN-;NIyV`x?rCJPPGNfA@;8HIUukMjO?j?uV+A#ym?;}i=UMOmLD#&B>b{WJU0i52nll-gNW0i2Jip0`JhHAYY3LVvXwQV>lDL@6hCr zcx#fKr!8-&qs(eD*vPDE2fg+gBNB%uJ?6Mn@T{WRp6%NUO*u zPb?osoZ9EK8kD-zaI&`P47lCL^^@=5j+~4g|FX>^;qseY!f|l7~)@9XP$U`G+$_w`dkP)1xtR~M!ie^q#%=a_%xqyGZPIw zG)3FSmqoK7@%&stb`bZ@hG78z6-|zxkTJ$m#P3*Vd+N+&v1&UK8L<>G3Gd@^gW4zq z_2Q;N{f#y&JrgbBJl9A76%NqQ=C!8l=uhVh*Y7>w`H;`-c125*p^aizz^vm=x{?{v z5w(#OD(>8$9d*Mx`?{Uc5Y767)74ozT!Ct^cc@D+x;foHhwB)=jtG@QpU9 zHZJueZSuSAOoev_V~SZPN;-r#nJX{5Ux2Mkcte})f7X-ij%1TZ7Vc`=NA)P^AWA3~ zsjHun714IkZ&pRd(acMXk?ty!;do6Z-^$`ewuJ?t)V7>+Vl{UHe4_Mw?YCshe@7q-oAV$|r zBm8EiIU*3+(C=B^mkvpkJonJYS%40MaH&-IcY>6=NE}PlV9cScgI0=PV>AcA z1TmY12P%S@LZs1vwjn53>c}_+w{PmL-0t1G8(n93<=95+`^=6$%)!EZO(3%E; z0lyd?Wyg3sF*!9oGiyISx40Ba41odruC7C>7PfbG_x2ACkB|w90W8M zbR0nMdBKy&Xcdi5coxk(6U&Pv1}@D^DQY0#^Z-YgErmcK$#CUga? z(n~W=Nk_})0UyU)DoG8-s0MjGj?1vrGAv*& zQZqItXleI?MJsq4aoQ?bxNiQR!hk=*lr=!Z{mC8B0SXL3QJ!Wz>!EN(I{0DScI|Ye z>iqolhwt7SFYxzN?tkJ1)+qmVw^ZRI`}e!0UmhYicEk;pn+iKB{!8Zxgc?t}rBk+&0ly^}MO$QTxAB*U|QOS7o8P{gZS zw8-z87Hkf-k9F2+D0j(^?p>0gU1^I@cN3C%bDd1x^0E$Y^F>P+4cI8_>EuL|(#qj7 zS|;=*2!uaMAOpV)xetFmuXIkHgC3Ws6%^H1nKzWoWxF}s_jBUoPt${!q?W(iEr~@f zLHjlQlnJY<yf=K}MUoOh%iq`M=+nq&wfOBX2k5%;TP z1#NbjWOMN&7Rp7*_bcVbiTjx5-LmXAz6CcGw?PmdRx+8#8TEGEn8BB zC}LjP0wF#!HSQG2l(#GcEh=h}WJ(yiiEqhN^*44{RF5F8o7Oblqz4(tbzEd1OxO;~ zx=PnuJM_z8Gtv!fWnNZ|g^=_hErbax-Fy%x{Pbn!Z167i1*F`s^$PUeob>%c9&9T3 zj1F7ZRk|BLzT;U>wkwhi4O#eSsXi)4FPr}4lEErU##ULWK^75V+o7=85LR-Y-4e;y z0^~ru5oz3j*REQZ+c=y=?foHT$Ic9?Qc z3FRO$D=(Fpu_`yVo7FSoWt+2`4Xq<|-bI#JaBqTgTC_ltugCWz2@_uqN^W;rG2%mM zz>QKs5q}@&;`w|{%VM|zJIy0ZEO{gA_f&3q(zYs0A(y|Pa<3q%{DjKwC^GMl>iY@< z1}ls&M^_>G4AObhLX4jS`PgN(5WKz=35gaqJUeamzP!4%h-pLdOXta!7YD{KohOnB zGe;Eg?TGWc4g;Nw$lo)p@d4bBwDbLMgF}wjANO?Dl7p~8KXx?z3%&X>IDAVe-3$)D z)2piL9|ni-zUzK^eo)0(T7_g6Ha`6A;1K2o7W}*O!wNGa zF5Hn>viaG)m5?wJ-4KQbviJR%Oom|v&ks9?({6jAgj%zidgm!%%9d)7ejEqrY7PR_ zIJVqtpMIPlYO4FWUIl9irZyhg-;y#rJovz{=dBoHwOZK z>1g^1!+IP31q;JZItJPu7~fsEsK&S7BT z03-m0@BZu$|L$fn{_ID`-X9;_9~R?38xUGhruXVyZ^A+coIhENn?QwGh)2deL}Wvb zO-g2qPD!VPIQGKebL5_8Wpbk#lu`otib$%N(OytjHc*sSHd3?-ms8X;w$q`%=wx|C zi!Sw=ZIle(k~sfC(GcYu(}{TDIkrXe2>Kq`h&MAN8)Ksby!+(K+q64!t37=b$8D1# zXKo4k9VvS8X8#_Vl4$nmsz$Gz8L&G(c< z`IH8%VA~zDgxLH*>H&nK8Tkd_EEF*tutsqsX^(ueOcI~LzrmNT<2!y<1!^ZM;usF# zfy=&CCTI8nnV!KM5q{Z($~o7b@Ewufb0JLqY8q{vLsF-QRkEmaFFBXV?l~QsiMJ-0 z?QD+LqQG?0iI*i@WKIjlQ{#8`_A&TP-wyBq5MgNe!nD&7##PAHLL$TNq|f|HyMWs> zPr8@Ce{eE>%e~R`oVKpNLgVL#{WiZs;~hzBTZ2hR`1yWWjQ=vJ&v}F?vaB4E!m>MwWxwoVg4`xrCjv$#F^&R>B>&;C~ws5Hv%iIn>i59ICdBBLihr%`5GiRq*`!CR zzM_LXxCEknKxy{8`+!OPGM%(lLqJ+^S>|S$j7ifDBI46#@UTpI z`UzRX*h`Vsf%3GgtKcVXz!@p)c01yb?+1imB7whi>^)WP>ZA1j&<9VSpYnKs(F36x zp0#JzXo!6#g<_aTv(NUm@)c4|v@mL^(Wp3ipxu}_;}Bb<{C#Ajx60RCw!;rs0_?{{ zZU6UcI~d`Dc`X}vuQYaB>Oii0`p62d;@u?SatHm<9j=ENe5m}?V+!R`$KxtqZYR0c zs52+?OhEUuzz_v@0?lH*t8??5bl~Hyi6sHVk3HetpHFzbS`qfvNqVj*aYL^^0|X6_ zAhjKYLYF8selS6Lo=7m>(DwoD@Xq>Pw`)jomsehH*vayG(_5o7nSM9VWfz1+q09%J=0%c3y9`j?|TR1-Ys#y<=asOU3AQqfhta9 zSlOhnXa#<@7@uG%lGisb0Y(debJ?u)L#_Cr30KtoxW~ShqHS=n9sH~mjqUs%2;65t zVV>`@P_xy_e1<#i&gsR47&4-bi|F$#{T# zjvhQ{tKpH9)z6(UZ}4D{>M>KdLe`1ZX@4n23-N(Hk$TO`#}e`V?~i|`^>z(E96?j* zelr^EF_8Ylk3oL)@TcNYu@=N}iAncjk|8LWSmaFEM+sEHA!;d2yU43fVy$R3j%;)|qQXRquY zouH3=jXt-)vweSuWU)!_{W%3x;aXvPT__w2dh7o~-COuowXSQ!lSxe`-J#OmQX(SK zB~mH^f|PVOBGMq;4bt7+(%k|g0#YI%-6hC7FtOHJd!N0}`M$qk@MDhgJokO&ffurc zw3@FWQZ%;;8)LihMf^ZU6v4toswd!PjsxqomNqX&*;7G&iCLw*AiAAhSWOfiq_BF$ zwaifq;U--|)fVW%@~|kOm?hYov+G`T0CDhQ8292LB=eSQogG(?4o>cKzErKo5=o`2 z3MnB3i9)Wg(Alk>t#2FWqNKE&q&`nFT|*5cZ5P#h`QpuZJPGCKv!c;fw-7$O_>*9? zsRw0C1D%4LcM+xqu#Q8_vS&BAv2K^8FOD2|G*o_ImiF}1EL!Yly)zVIj(pLi2UrFZ zu(NhI-gXe4NS}*`IKo0Xi;^)xxc0Rmgf|Xl9w0$~L}i}RDm9!FS(V$uz*1Pre=P{v zB1HUB(e%SvKkG&LRcPQyNt-$AOI{pB>c`Qh=<3Tke?S(%v!|G38_ini0>L`cn4&KZaxs z{+ks|{Rqf_I(vNC=r7gTMF6Y!r-G0Vm=^!*6-|?w2q9XckKW0t20RjoF_o5Fh<5|q zpDxDxNmo;+CzwbqW#4SFUk<26yGBvmpD(^zPT~4^&2y#i8d=?GUt0tN97`=x3yF;Q z79BysF=5XQbnzV;ZkFc+!?EfoyXp`7pA%1p1}*RFxZh|xnA1{g-g|GV@lk)ZX7eNB zV*7gQ@yfUw{Kun;MyFH50}{|_P(4g`k~4xZEErqQ6P`^L1$&EM_3=yLBzzo5FQu8KZ%V!Ew182P?aMu4?fay<;i%~*KfeA79x=TxrRRNHo zSTFvn05hu4nL=4k{TeF;;S=??!|N@+#wtG*V2UYp4)~TWrI2cg`)Ke_2U!gP;o%)( zqG(&eRNMrVG=%Obl)fRfL6p=t%_QcHrb82t=cpxKN$_x(hh-Yc(Fdx#D8yTY)7oh- zt3*p0G=46?d@SouPAllIgaY$@IA;Qe`Ul`!UpC%3+i<;K1$T*N;{Qp(ofImd{(7c@ zZipEVY`lI%l*#Er**W9AS}(r78=5tNxH!q*w80gLWAZ*7E z0ySjM`Ha)*KH{G?c?QWe%tct~GGr;-65z*((5Pw54W9Fc=BKXFk&9L9fG(c1g3$Scoz(*EfOHdXKZ z7a!QKwP>afZuVCHRVbFZ(X#*swR2O?XL;ql7>8>`7~hZ3v$^)~1+*O{xj+2E)QIQA zeFU-ne<$v{De8O1HU|5LoPmb6V$pHW6fonH=#x|3K*||3iD$0EPhyZW&@eSix>}=_ z4wWt^v$Cbtr>>2o!??S%uRrhX7xw-93r7-60KYKbB)~QE4oOXXrsNM~o-qE3dYu+|l=Eu4A;JX+EI6}3V)DvCc(o3Ds zsz(*&-O%#kG?1`G59B;3)b+^3h#UuIJPDsb$i*-DaL)=;#V^YF&|7%9dNc90X!?=W ziKmFNPTB=~UR}zR4dfdk_=aJ8n`=es-i6&!(EqXc?yu(B&()*e5gdPa|Ca&{2MRa59--( zAfb$>Vw`5!O&!G}Ugg*k6ArrkCPr^?NABDW88FUqQ|^)A5_Wj|1ZlrY;t4&K_(!=N|WuwZVoFDHgv?Smii(e58>- zU>I6I_^mb=I6j1ILKtDf?L?zujZotfREf=F64TP}qi3i2hDGhTI)liIRj#TvK-I}_L8QGs5soXWrJ`6KU z-tGuHFKh}1$GA7TJcH#~rG~lx?j~8bdXGB42*<`wu}JHsO;-el zB_EIWK%tluIC)H}&2=Ns90n%o&GEu94!P`ZncW2wX00vh$E%gO(H6 zIu^#_jvz|HOz==#NKth2l_tH&+f`wVlSLVjoTk+DtPo4cGRiPctk}2P{&FmalumsK9l(*2EFE;h!-MtbE(xAni$foV z-Dx{pi?h$x;(!Y6^J{T_M+5v`+OG+#|HylEq!RpPEk06-{dFzQ36%iWV)RLCp7Fdd z&Bd{T`Av9+Mkp80*5V$#CjD>P;NNS>{r$Hie7+E?0aCPErJkZB{u+alw>zwcF!)f*+yIVL9pFd-2o%Aj z*V-9BXp7TRGw_ziYaJmkn=!Qf-f~xydVlo^iwokmZ(t_<_nqw5->6MFE+E=X@0Up( zOgyECuYBWsWi9_zKp^6yxtUwihqe)Fb*l3*y6uk^UOhpxKebK{BvyT&9vEZqkXewg zvYbzW_;{tQSNfWH`4qFtTJ;~Bi%)}$cG>`Ku=$7@i^|^EhqfpC6;IF|jb{UIC>_ox zcsL%BWfZd%0u{g zT*%Oo+=bb&&;lG%DQP4CZSbI}Z!OSV+}SB@zZIH@l1BTrx!CU>^?q`Dmas;6Zfy#K znvyHaOZ5n&+`HI%IyY}Ev4uCFq_aMi=+TT{j%<@kXVb@#xWwPU0pp%d5Ys7!${5N! zS~$;Hl6WggJTbQJ@jQFQA^f$F?KpW>8lG3tu$o0BNW+Z6yi>2;9*vN~5;W){>`3%G zPBz5fh0DBg!Mw$V83q=Dn<;<~?RvnrlZ4@)DcrB=;={w?wQznPKTFRu2S51EKO!~kU{lq}0L^ZdGTp*Aae9|p2PB< z51YO1^|brd(FC@#9+~dG5#B-Kl(}Ks7M(os*)<6QPBIP;5r~|d)OB-~ zxezO21O*ZACb!ploVDqYE`> z6p|OHT08kRQ1bTO>6k&qo`+(>sFTVFy#wvKqydfBMBdiDy!EhR_Jc6L+wG<2H_KG> z54R;EM8x_X?<{^TOFa2t&N@BSe;WGntjGSWML9|c;hR%r`s+UK;2-!!XgK51Kz!J5f+q8GwzmiA**zqbD^@pFnLj08W<+ojGVuqEXv3C7OX(Y1-IW(z_Pz3D5l}02I zhlkB*|9*r;ILa~O-}Trt3igx-Lh|GFjAJwhx+e4+lACQ#t`1wz82kU!V?R|fI~yzDc=#|bu#RA< z0A2LM+G7}}W0fA6dc3`kXH@>Dd#N2^F|zI3$lv$8m{ zqlKDh^AkwwZzDvBgXn#?G#R@DQo(>1?5iA?OwcagCc+l8>8B@^LiDQRvr|MR0%6`O znS1ODAtvstwYeFXut-LVDA7T#I3485m`Z-8Eauq?CTxM@KnZT8Jn!+R1)2=>z)I&% zR-QlPl8VQ4{|7|D;SqQ7xGymXMn?ibi~?SK@|kBN`Kny2rJwA0u z_u!CI@5BCq;j!_~w27%<)afa#z}dH=FpD!`=*w52>x-}p0Ys12wvN1Z{r5kBU|o-B zHbSdi>9*d>jm}MDctR*7VU+3(fe6Gb!4FT8YJFu)NK<~9A1ktMGgULzt=fouG zs~(Eo;**dMllKwU4kuOhP7_l$>6D9nU;w8RMc^+%Eus^D6*ELSn@}#+M2GMP58nU! zn%?w}*7Om>KYBLKOARXQgnpW{i*FML0R8mWgq_)-^aZ(soea$@Tj$b;Ge98zl7&si zQW6ZD>sCk@L9$kZVkJaYL(JWXob8YXQh@hNh}$J8Kq_i{rS5c{GePwb{uB-IrA%!1@HIk^FNjt|6a1^1M_zUB^_r3 z1`A>De>pMWOtKf^7ZSm$)fM-pRQ_L+82|EFcRjmV0VD?eR$}}^LCL%B|J6|OAJDgN zpLLci3Ky_GWBh+qvS&ZeZq7zT1t)m^?LCXce*otE{`UUSsUkl1pKz*#>*E8Oh}dBK zsM`r~XM$5c*iUaS^elMhSHY=6d`o*rCuiGbxUT*IpG?`L7YsuquS;Gs^o^ad0}T0Z z=Q-Xj7!xnWykA{d+1TC*+1fogJQ7C_S^oeE7l8ouTcaBUn_i{kTlhjS^lvVu^KY;B z7D<1?mrLnIig@VRrIa<4@l_y?gMp#)&)ESIF3=z8w_n+T+c^*ZnjHWxrO-2W;Oc#Z zK7YYLX)7M%#H-u!1Y`)fS)ZtZlAwY^*O`pGgZme(Igt%+$ju$^9yIM`E~#fS34 zIUm)nMyg4I(Y}H|e|me#h;HH>kO`1Av1$ra`Pe1%(d9$re3^NGdL5Kr@f1O>NiW#| z9-fHC=fTnh#f@ZO(hkUYdNPQ*xc<}Y$vtIh^*dw*Dl?bIPP&WH6HXQ(8Px@l6H zKfnM+fZ?!rjel*SESN>YOD_m_8jeAm10R@3y9RjE_$^5bHm<&7LxiSmHA`_Yxl7g_Wg%uZ>e>a8;40{T z#rtv4R_rn&+0^6R4j3Gi#vXJdLu(X|m`ki&5zv(>5wm`#rm^AVWV;4^s1$w2VRoW3 zf?JBSz^i>UmA=z?^{FbIPTL128e`cyM_GAsO}(58>O*;jy_pmK`;+^E8@Y1{VA5le z>YNSxLdHl?Hj){B|Iu5J`iAWm{o^jy<_~o@1-mFzFdA+=ax*TkpKhr)?0eOPUcp!ppCIjcd2t#sO z$$N3`Mc`BSAt-Z@m` zw#$1w$r^-Ty<%K|KbZ68pequ2M|J&4W6;#rm~Fk1cmaAZvN{u^It=WZp-YyT`XrBY z1@b=$%f_4;!3DyyeHAG_g=Gh7#&H2*+5XcR1m~}49>5dA2cAD8@DsQEAL_^dP9al> zOYk3vq!JeaJZv-}XJ%5a$R7xNqi5OY=-qE-(lbv;0AcH=0rS5fN%ey#mBtJo{VIV9zs&k*vGc7SS(s)H58ilH$) z%|9{030kCiN4-iiy}5RI{p#$v$Uf#W59%)cV36HS%)$#us9260LV~FGQH_tCPbwGQ zO`&>IvxX34Hfgr|QgAmlUoLbJrI@5WP1e&Bde>%k4X+JDBKZ<`0Gmiy%3)&~mOAkz z$5EQ51ZJC3_ngv+O6uk6Uqjo zR~2UGx}s>+&v0<{Q_4@qkWXtJxe@;tV~F5pJMh&x&S(H);mm0-Jio6vxH%=_TR#pq zvqlKPz^C5H1F`#j2y~P|b&E7}!Lk_Mb0IABgpvVQ*?t;~qt(N>?df!!)qP~=qx3>X zrdjmMmArj|a^KI#$uwN;gtr*X%D}Q&`tiE3O9gkq)Y3H3-im9IyHq!G3 zKZ*fea-$!UH5U;oh-_DW?TT`WG3l#sHJ#d1d4(B(t36 zYLUigU2=JQV@4ILd*feQWd4=JB<7sVnEj4N3)dV0R@6TQna=(_OMHO6Z+1bQ=5l@{ zyk~l{TT1J9 zvNvJ7ezHIPl;z{WTx0@+)jW;cG2&tdNzc)0%@p$S7PgSnN!>)?4VI(zf}5uwo8dvQ zdcH7NL3lvI1|wD`zak(w&$lCmQzC0b%1(C06K*|X`%4_$FX>0Iun>-L!6zIXv%QGS zRq(OsL=!?VX%-g8Wp4@f2CO7;2%Wq+wnD8Yb#Z3vO|v53q!@C7la1EX%aX}{I__{p zLyKOV#le^+4k_fbv|_>oMFF0b6vR2Sq68=@fsa+dq}r?<{6rykPu;c1@#Q*1ij#r^ zP?{G=2_-mWi$heSD5$>45(U#$fS1_6SU-Q~)9XTNAeJL%J~1>CJH zSeLbdMD+t*Budu-A}%Y43dJHM?zSvDxi*6gYX+2h8|3G6I-SH=yNLU6uLb%@7m*0& zJ>;%ET*y2~9J;DtUkDaT?O5bhJGghvv>`Hgr}Ro5Qnu_<$>C%ophf0ZwtQsQa5~R! zmaKpFoeUpvL|bC&?FKqkK4rL^zS119gzG)EK|*T3@bp{U%Sgl33R2=bh{S2h3Qo>1 zNovE%Etv$wu8qlDsO3a_v}%ar!|hfHwLx^cC7_NT>(;`yhh%86q7}|RIxb}cu|DC~ z(Y(YDZb?jvqX$dQLy5!g^kh?PmxN#(Z;Y+1FyUwjKe-Yee$yu#-e`)tct~CP-`zOl!=M2r_q!QQ4ef zwk#@Yr;O9=sy438X{tYe|I)!nbR` zJbTo>{>!!B0Qw=QkheDH*Puc`jCAJFiHMZLjIxuY3X4rNB~D6ByZS6XJv%8TC!Yk> z{??Ne>N08rU}&o1i^1TlAZcR5XraNPZJ_8T>80po>gTGVY9B_Ye2LyN#QwtWHANB2 zWVs@wK=~E>Eae)_OYM2~yb<9&@*TRSjXrjR77_J*SZD4>d>C~M3TcrLzEg3yN)g!v z$W!H>)!T^fu@Ey&(lydZ3C|`(Qjya;b2FxGk_MbO;1?XzB4v@RW&sTlA!;#eCTl*4 zh!7DsJ>it#V^K{yA1y%*qX~24YzT-L

    HGf{m1Ff(9-pUVK6L1KPz=e1Ro61$!g z0-X;nb5MoT#BS65&g_Z$Ykq0kv~5m)HRKNcIU2B!I1Zx|CPiI?c#d{iAA*}t9> z!<`tyq=9dSiheur(0~5*t@1891fw3YduU69-Re$ewGWHm1Az_W12_slSGeZG!~Pfs z+E4mRa>{A;5aZEzlBDU(o|A|*?edRe$6f6bT&U7VXT0glE0cX$BfK<$1_f>i?1jH0dcG#iON%K%qV)O}aMIs{ zt-ks(amq!PvMz^to?W|VbyE(xU&3${QP&r?!)wtG)Z?j;Bm2)2^^$b~EXh0uI!CSU%o< z8;}G0Xf9t$`tV(pZvNqXoGnn^miUzU=yWmlEU-bO$@ggKK%$d}GQWOJe5IHjpr3sv z0T$c6@ndu3s8aD>&Knz-st*?V5f( zT+Ar=_(M>k6W_qzLc!@reub`j{Qb?T)6&9J81R6z7A+|l|8rmihTnJkSxV(A{Y-*I z3UMNB!rbF`f6YDvkkOnvcUOFKVy+0NpJ)5)$aoF|1*zG zJR|1W-seS>1~AW0G60WFp;6v?7Hzmmw$=w0;F z#f`e&V|qAN3uh``)TY0tw^!8)$}Q({&XiQ)Z^$hv_My)e>5ycADW@lL0YfP{!Jr17 z9mB#Ki;{)~p2VE|S21XvL0}q{pvKb@Hhxfp85QmQWl%*#mJ2wSo)ipj-~+8dMtv~& ztLBT5(W-L=)dilDs5clOyr3_1flh8gO=4h4H(-jWQ84&RVHx)YYDf}ufK>dl98)+v z5UVo+4BSQ|-BJLlKoKom&q{Oy^D%PMAsQ)T3kffVB+_c}F-+m&WcOt_ih<<~2sOGN zH3G)*(NY2V#0QJ3Y$*dK9cwYyKt<~x1zA9#b$YW(B{tC#p(FCL%Lwbtw*-ecgi=`g zFGxMHY|M4Ua2aQSJ4t&*S;@oRtbJo;i7vlx$u=iLuV<-;s5YCQRwXibw^K(U$F-XW zQqzYRQGsFg))4DjLysIBaYf9t3$n)rILXKX3|JZYYe&n4?aUumN?$oD%$G0Ze^{&i zrtl2C&VRh=ykSJU*wNomi`&H_13&qgZ^{8TPOV;er%JcrI_88EY2!Y#y9{us>cW6V z-UBH(nDg~c702uZQl}?$vKQO>5B5znPj+BQpr!NYeb*3X=rcxXBHcf%FKvH5h zXBVh)mMf~1;Q8a~Om5E)w)q5j0{Zvg1D@ZYpZ|G!-d}~_Xb}R(Reni%Rdwo_B{1fB zLvu?|)BhIW>H2@_SfU{oy7)upEH{u%B=vnXpY`}PD#c8E0`+`0kt*Fc`}{vf1JqXI$Ne)er{2*Movk2g;$|OwPIN zdq+j9&Y8`K=bm%#z)R!ikt_9sum4RWS)VLfQS!eKYA68<2FL$PIP_(%MAiS6&1hRI z@dEw&q_h*PF1ub4L_5+;f^vl3#{ui0E-ouRF zGQ>S!FtQe?D1tVs0Wn0SzUSRdjg3_PShyL{9c+5 zBK0U{qZ!~g-CD3w)YRR`CF8vO52aOlf;ZKm(r6l|q*nJ*9&!>?C>CBkbe(UB6|#6? zRrNOd7A96|)`w_vE$(Cn%{^{q3R<`zP8{l%3S?y@cXPUlXT!fj@4*u(%A}>Vnasa8E@eDE-Rzqn=Y6D%awo)3tstHI!{*`? zV3VG@b9}ZJi-l>N1^A!)_(oe!5uo2jxO;}=S0bnq_fVK(R; zqGm+ae2fNf`+O`o{Vc$r=Y^^?tqQBwGlwlQ-7j@quI9Q)Uh$i{$$nhGUcAf%z|q8# zlHUTKfxS48%?!pn@oRQP?9*(kiR{8NkDmh)CdQxP=pHU)jhw!-0RMyCOt{lhtx5@| zfU&*HzGWHeHLKoZD-un4lvX*d6>rOznG0s6;nr%3LnMFMi`TJIm>Vle9jJn#Y*FjB zk1q1I)s2O=SGvB5!!e?qXatq~J z$ldOPi|O9(=Rezvh4GC)?Zu_^;I!JEAz3jd&Tvi{ri(>QXS**D>B-ImX_@Phz+SAq zH>TD3WM9D%oiBVgkmVhGYSay#qov24b*^+DwU0hfEj=ID-MIl{`2lTy3NlR(B0Y_t zu7^?1a+Xo|0-@Sr!?DRq0@)RzKz0R{T%oIIH7UZA)JD|VOgG|EOo))UU`Kd?J2MX% zR4KIyH#^gVnFA9=>mZ46uuzDrR>RXGvl)4_@G15Y83qMq3mHIziW*^Hg+?V(v2F58 zJ7~Hnd$H4%XZk3pVB(e7B{NwR2|qYae%eGUdWsfYG{n_3FjtxPf4&7wfKBZ@Owwjj|~keAdqS$2@+1;C^uQOQ&o61qW1{c1Snrl5s; zEFAigAYr0h$kvT?-gl1jSzX5f9KHL9igAIWTd6K1qFe=?*^(|^bxS@h@mQPAfv%?v zfTOQmrDp3*>e0?7_v!nh(@@@y8SP2us_5)BvHuK5_agS1$9#sPUv>7`NYh<4fz0#F zyzjF=qPKg8qRqFetG6fx<*wGRb@!xc{-&$XxYk~19eucFas$ETxg6&@#`HDV?qV3< zwWPeqZw+Hs_8Lgb->o|u<0`K~73(BHg-%oV)whTb3K8N>-B!E7{?!^z>>VsDFUVn# zn_-$A;GZIsPHTk3RKinvx+8Enbm?B?L?Cr+@zt%J#oxc&N?TZ@_7Ou=U~I5mL;SX& z2&oT2toqe@VbGNoy@$8CvtT%hDT6jQGRRUaVQ+~FN8BjXR;=cP_rD|{SEcT9If1k% zX~lwqI$IJU>8g}9>oyu2|354>A8{>#XhOuXS0{_l` zF)m3ICJVsBP%4RrCK*eR6>fRR6^)^6#aTz&GA*k#DZ`frfWdD5(wdR>@ELWx7e;w+ z>O5|I{j_wKT}j;q!(VbNbdD3r{)v}LMhQejw%om8)!4u@b7D4qpe*T@kVOI6xt35O z5D&;2eEa)Ou$Yzlb6wiMnBL+NTm!@>?(<%UadcCMxIgu};_U6`FYX(l7DO3J0wQ;{ z$KVUX2&YAlph`&Lhz-NK5y_gt^(Z2V8ahs_V3g4ho<}8*xBh)6E zf>xS#iXw)N9_G+$^3Kae1<7pnOrRvt3)6m{j27-0+3|+ECFJwW-T1yhwr@R z6P&SpffnB=(0)(_QWPffv0zzzQFt$V*3T^ZBA0`BvI`>U#DIShCqrt_@~%=LO%JU=Z@ z>jVGb%GUI3=42hB9xWqvJTiZO5KsE$@v1~2-}%s|KpRlzZI;d4KOjqP=e)x|B1?{p zUX>dH*Y<#3p|9yJCwhzVUxLo}If3*RQRDS*JrqBsw{R_zl1%zHym7-(AP{mNSNa^E zNQR2-_@j5Ye39IMjBQ%Pk!1uTNDM@3=@AD4EcqX!2eiF|RI|Rj#pa`a?k#kT z)Q+;>xBpg z7QO)nOIZ!8-Yu|cvk3e-D?6QVIhi9+QG~C(LJXD!$6q3~zQXk%l#mK}fU#s)p+@Y~ zYlO|mxZD$nPfVC}^wq7L9g{GrvWoX*j$6uK)~Q)5Ml!86=)1`hvo~p|Q+E{7X`21Q z2Zk`Iaj(0%9>LAKP!ERYzg>7-=7s!ffYT;<@3_n}+8d)0fiv*B8yQE^3G6*VyBNV{ zqMYk)+GUsBXGSiMK8FQsXB65JvZMA@U^zbSMVzo|m))3QrQ~~yu92e7J$F}JTKMg~ z4*x357^ephP8X8{#j5d)B25O~XFIjqFTZ+@_Sm|(pHFqQygAZ-?Rd4X#ex+<;!8QJ`1IdNxo-^>F@YxZd>iHZL3kJGWQ2IV;zNV^R`9TEXi< zh3+=oWYG7CO}Jf29?b5rC|Zl2go8n-T(wC=7MVgEKo5n^0U1UoX$u+ArkA)>GS)e_ zScbR=5%aWYX!ZHPA!B5#oo1*@iW|=}8<%L7p!T z<^#WQGglG9ghhhm$NF(B@%!Q(-U^B8nKNZ3W4P7j8H)E%lU5^9Os0<1Rf6^%hMu-0 zQ6?J%^CA_7QkDBXC``C+D%B-3i8jTK;_xd2R3vxqiabEuj%<)hCw4z{k*2qVw4wO1 zClJD^xz4h^Iu_*`((J{v8j2lN5$A5r>cw}!MpzJD;I1l?wu!Jrc<-)z?cHG?QOi*L zc8l(Hx9I-*6>P-)9#r;6tg^1;*i?ujnV8TcSWhmFBr3Da2c-1KJ_@BsSk`El!wcmC zKB{a(E~yHNiyXb6^s`c++RYSsStYL^YYB~ioF%0zITSO7V*u%(J2L4SiXR*{=0l~w z#OBP9bjFgGg~CS|bff|-IdvGu9beXtEalzom|Fh3MN0?>cHtT4_h$+aJ183b`1^&U zXa9ilceT{#By4|>g*njzdA%Rw$LV&@E}UPA>;Qh4o0%@mgdIOJ=5wIGI0nEEpUJW! zVSf}uazH^?RyiC4p15LQ$jCj zR~9?hq~0lGu1lis>7`JApgf^JZR(-8EDHruz}&_`3BeF)wjsv@_sqZuKEH@%iqazv z=C8%VzrouehS4fOKqGc)c3c!fbP;E!nz!Xei0Cbf3*qi9;@fbYA+$DwQQouZ&G_%+ z<7ItzkY6~HSYz}qzy?w$(PM70`phl_$DlWr&)K^!jy}4QVX#B5N%9)qz;^g8JOwi* z66cD+BjgorH1LB1A9ZComIs7SXMDtY$qnSN{ahGPx?z`toKkkUjB6h0qclWWGIml zzJY1k-W(o$qF=;Ny90hhSiM0Os)u4r=HT5-_RTUZl>S}2k)21CM%&cc-og^MDH!e<;gQy)?f$)9s3^bT+lUy_uj$6)LE2w#hKrKwkL<( zsoQ!>b4vlZa4H^lo@F?Cy9Pf9kxP%#OGiqLF)BZti`HqV@{BV9E}T#*p!g`BeV`M_ zaO9a;z;{a~qk-kYAzDl>r0oLu;q=*r6qR|x_i1s8mD-P@Y_hvkf#M^2Iy`xvrL4@h zXn-Gv>Ko*I+Jj4vwhZzrlOGx4)XXoP@x$W=1wdE@BT`Fs@Jdk+bIA(R;+uvBgkFVl z@6)^P#QLk{LVVq;KwKC@zW8BLbCXf*+Ds$199j;i$mq%F zp#Oy*-tObwpP-fILwtkUA&BpPQ-JZpiDYLsYu?Tf$-##P0SO`FGThMzdApaTO;@e? z?mAuKNY?z5ENc-5hbczm?W(P%6T@4om&oJp%=qclFFEmq(O=k3_+A*>VGKqD_TULw z-!pzVUgll=cYgTm9_%h0So4X>INO7Nq%vCYO^-J_Xc8n>IEym@dva`(!_F`)3=dk} zE!8LfN~WJ=StLt4vys2!hyV7%`KkDbl(8dSz}vjg4ebLd1e>}E6D`-B`WPR2yRrbc zm{pD0f$X_vlqbAirzcOt42(X+6aL_@?8zD8t@O&9K8l};>_nJ|U7o6J%)qpXw|%>C zzS@)he#&tC5>`P;dMQze%-I|n*{lHurQwr3Iq3X_^V_frz^@=9(IyC<2~tvQAi(Co z2{Zc7{R+C&AQ(imcj)Tg@Jd9aFfY$!SrUB~=l0;)kO)LG5N72gc+-)=X9qnQeEXgK zjz`-GAQ4^g?A3nu(a;D8rw)v3)qpckSR^(rKfI$7jQ6Ak7`H=@nZo{spY-W)1VP@S zkeK9PNKEN7*l=B@TBR4@%HgyYsRpqpbVHV0JE<*xdLp{n-OrTofS7UVm^BQ#6Z>}3 zuC!)-l4T)SCXV^F0`gR6uoP+X-7*7NcnrB@iOd<4hgfIpTG9f9!6LD^7{_rR=*NM|WR#g+~bNW;5r}->wiXPu6-v zr)vA?Zq@AF{Ck{6LV07#jWH29nM}^=F37{sUmG7+C z?A1)87F@egGbrL(k{u$#a(`b^aY@iNdCg;08!1O6fXgOQsVvvTR#Hv9JiH)uH;?It zq)K9ZMT5Jsp)t+PQ6FZ4>iX0?%}1S+o%9vieIgsquey{s?GKts1E?SAYseg>-)pXT zo@MU~o$_o_D8|dnexQFgpxLUG5H11zwun%2EEvU{@ev*}`uR0^} zw)h23gmKD!`{BDn`_*mvh3S#KMw=EqcvO8f;WfvLI`=<}jWREkXpnWoPTcOC#w)ID zUtN4$9$9!B!NT#`%{BLkn&cxqPW6JI{ zV!@Thspn>%R}Lpoofg8V@;9E!@J6Da+OTj+-#(^3j>LOv*Kv7r)8FZ06ou^p>+Srl z@Ef~xEViYWl^$$nXtHNd@PWX4z1#60ju-gV*ufwS7!cgaqPR@;;AteYRq{PjuvB36 zaEA15_WhG3#s2D%ioV_a$0y5o#hl^LoPC9%TPu1pHDj;(_R4EcR?PxyCKdujt(1>w*WNCa{Aes&r8qbP?TfA+-}y<#(kLr<%V4PaEYb zPwpx*;K~f4VikvD<93slcM~#mk-pw)}=^ey%@g3y{AcsgJ` zwaBEgQhaJ-@zidS+P?Cs-;-2gLV~z5FVj{qJtVcg6^4J7w3O8z8K^x7%Zh3Ob{JRR26*jb+ENO zHRNruYbcm38Xki#gw8C4BQbdj#53oJ!NrKk0S2=}KZUA@>?T-CeZNPqkslNKKIuf71QR(`dDUKy zs+zvwS@m>0I4WjC}vNKHfYPdpGxR!T>2~wnGbfn*F^2e_u zovucCMo0N6#iV6jlHQ1MgMlw5q$;}_D`yrVZ2UA@TEc)t@lg^fQ8GDCGNp1dwS6*eOfr3SGUHe>^HDM@ zQ3@MR3WstEmwgI%ObSnR3h!77!cxk8XT>c1%!RJcKzf*DTg42e^8M1_5a614&Ii)BlG%PlqgQ^45ht&J6nZL9@HKA z9|}33#^>=kr_?+9KXZVK&!jKY8-J9(r2Z$oem^<1{_p?KKNd%h{hvx-{s(saGq0Z* z&jIW8&tbjam3pQHJcrxJmwr~e+ko$#Lo@!s>i>Xd#DKmf_x{s*!fP-A=0N)pH|8zW z$3eUhg-9P_)6gf+!g-?8wbJe+c;{r@K~8Li?_caDs?3s-xctjTcI1zJ5)elb7)N-Mc)2Hx%9E zE45is?XwA7xYM8j(n|CWH$yujDA^+U6g4B}T#}vQszigWc=2_9?G3jUEE>1dXJ%If zFLn_nwX~#cZcb;pa2!R~a?Z_w;<-`s*%XTveeZG#G$mIS!r#WDbHZ5JFNKRpC8sT3 z!F7Mj9ZN-N^lWoQXb?dzJed1x!F@NWyK0L)KDcc;(x!Jo;@rYbDo&EHCmxvh;MC}F zkQdCVwBlc37BIi;r5Dm1Ll3-gVXTJS6047@4FS!2Ib;D{L^mw1R(Qud%D{R8TceO( z5C4Ss#&ZXXZGCJLsxbvQ3uoXQj{8C_3tuv z%!Lw?cfx_?M>a{q6g3souEUO zRdSPg<{)i%#_IbZhMP#bWF&e~OWuTiN)NF5aWWP^WzyVBViRJTj zNLnGJ57D08AzI274Y<1m$(T)~&nf-5IF?g3pPdtVzpc6@+-;411wSRmN6bgq zF=*dWgcP@z23#yA84AhWLZ*I-2f4juZn4$XBv@;PPgf$f-N!~UxP@7w(z-psF!vX$;T^R+P6|ec&VlOqrE55xC!TK^}b{2dIKWM z@x~`Kq0HXdMyXqtJtegdnwqSw`30`-Q0cCIoZAo>5oO)w|#tC3r%X)aI98< zdxMD|p3~BDWH1n@cZA}0x;~J$h@N?R1vGt#rAw=t7HAK92A@M92YaeZj4;9eXF3)h@N=z-XhLU%uPj zvn2&_JxZLjRSrGSh?1O~uJfXNaftE-O6uo!owpo&LK9a6XoMt6EVrwa@7bI^H34y#}9HeQNgBc3R^Q zCfmjP-dEZY`fzH^;?x*upncxbd0E?VD@GYFC-q(Jy4dx17+tQfUoCYAk&HVIi@9+d zGT@A5oJc;Zo9UxIKuUT?3f_3Jk2qV_+uko3vrI1&ZCV!p&gmVvi!T=&SS4KreTIWM zN%e1MK9WaDA5LM1=-&*taYxDKPv^&S#-^9Og(%*bt~T%<0+n?`Qgq3%;5XoSLIj^! z&7XPGnI1cA$sNwGA(Ow59-7KCM8*9QQRR}LVi6Hyu}wIdn!k&}?Lq~@sBk!A?Z*2e z7$a=MxbSzb=MW5Ldsfbp!Z`Les_sfc>Y8W^-9-#FhjbK!_V_`99xG7FZg5w&a}JM( z(Zh9mXE?n*SlP4{T^ZIf926V45L{!u<{0Jfn1+&IT$j6AzRG=r5pcP5#SX_2#3OE_iBtR?t-iR2GOlq14#DCbcM-@#9tn<#8pB9;YDdT#q@FQ9y5|haGTBzS-H=&i1yu{p}Nb0h-~SiYtyG`3IS;1yZm47|?@u9%Q}-<+pkT|So3`WfiHYml!B7E z@A&b}u!`h6eEFty{=zR`yW+<{_GXZ_HmDDK)+|2`gt9&-{`5fUPy73$)c*GO%mQ?X z-zNDEKlyKqa2hB%;6$pY5*n%uLf&>_TA?OY- z=xQW*gEPo023Ud>NDnF)fKVgD!X!KiGnRvxHSrgg{s^PKYg2NQFX} zgZom2S$Ks>C^A~Og-8g4Me~LDMuug`g Date: Mon, 15 Apr 2019 16:41:34 +0800 Subject: [PATCH 3/9] update to using git submodule to maintain docs --- docs/Gemfile | 4 + docs/sqlflow/Dockerfile | 24 + docs/sqlflow/Dockerfile.dev | 45 ++ docs/sqlflow/LICENSE | 201 ++++++++ docs/sqlflow/README.md | 24 + docs/sqlflow/cmd/demo/README.md | 82 ++++ .../cmd/demo/slides/201812/sqlflow.slide | 40 ++ docs/sqlflow/doc/build-tensorflow.html | 429 +++++++++++++++++ docs/sqlflow/doc/build-tensorflow.md | 135 ++++++ docs/sqlflow/doc/build.html | 406 ++++++++++++++++ docs/sqlflow/doc/build.md | 128 +++++ .../doc/close_producer_from_consumer.html | 447 +++++++++++++++++ .../doc/close_producer_from_consumer.md | 161 +++++++ .../doc/database_abstraction_layer.html | 383 +++++++++++++++ .../sqlflow/doc/database_abstraction_layer.md | 92 ++++ docs/sqlflow/doc/figures/sqlflow-arch.graffle | Bin 0 -> 3985 bytes docs/sqlflow/doc/figures/sqlflow-arch.png | Bin 0 -> 99876 bytes docs/sqlflow/doc/figures/sqlflow-arch2.dot | 50 ++ docs/sqlflow/doc/figures/sqlflow-arch2.png | Bin 0 -> 96987 bytes docs/sqlflow/doc/mysql-setup.html | 433 +++++++++++++++++ docs/sqlflow/doc/mysql-setup.md | 138 ++++++ docs/sqlflow/doc/pipe.html | 374 +++++++++++++++ docs/sqlflow/doc/pipe.md | 91 ++++ docs/sqlflow/doc/quickstart.html | 387 +++++++++++++++ docs/sqlflow/doc/quickstart.md | 93 ++++ docs/sqlflow/doc/sql_parser.html | 355 ++++++++++++++ docs/sqlflow/doc/sql_parser.md | 65 +++ docs/sqlflow/doc/submitter.html | 377 +++++++++++++++ docs/sqlflow/doc/submitter.md | 95 ++++ docs/sqlflow/doc/syntax.html | 448 ++++++++++++++++++ docs/sqlflow/doc/syntax.md | 160 +++++++ docs/sqlflow/doc/walkthrough.html | 362 ++++++++++++++ docs/sqlflow/doc/walkthrough.md | 71 +++ docs/sqlflow/example/creditcard/README.md | 149 ++++++ docs/sqlflow/example/creditcard/index.html | 427 +++++++++++++++++ .../initialize_dataset_and_database.bash | 127 +++++ docs/sqlflow/example/datasets/Dockerfile | 4 + docs/sqlflow/example/datasets/README.md | 75 +++ docs/sqlflow/example/datasets/data_to_sql.awk | 20 + docs/sqlflow/example/datasets/index.html | 358 ++++++++++++++ .../sqlflow/example/fraud_detection/README.md | 101 ++++ .../fraud_detection/creditcard.csv.test | 10 + .../fraud_detection/creditcard.csv.train | 10 + .../fraud_detection/creditcard_data.py | 59 +++ .../example/fraud_detection/index.html | 388 +++++++++++++++ .../fraud_detection/premade_estimator.py | 115 +++++ docs/sqlflow/example/jupyter/example.ipynb | 106 +++++ docs/sqlflow/server/README.md | 14 + docs/sqlflow/sql/README.md | 13 + docs/sqlflow/sql/python/README.md | 13 + docs/sqlflow/sql/python/index.html | 305 ++++++++++++ docs/sqlflow/sql/python/sql_data.py | 57 +++ docs/sqlflow/sql/python/test_sql_data.py | 32 ++ docs/sqlflow/sql/test | 0 docs/sqlflow/sqlfs/README.md | 56 +++ 55 files changed, 8539 insertions(+) create mode 100644 docs/Gemfile create mode 100644 docs/sqlflow/Dockerfile create mode 100644 docs/sqlflow/Dockerfile.dev create mode 100644 docs/sqlflow/LICENSE create mode 100644 docs/sqlflow/README.md create mode 100644 docs/sqlflow/cmd/demo/README.md create mode 100644 docs/sqlflow/cmd/demo/slides/201812/sqlflow.slide create mode 100644 docs/sqlflow/doc/build-tensorflow.html create mode 100644 docs/sqlflow/doc/build-tensorflow.md create mode 100644 docs/sqlflow/doc/build.html create mode 100644 docs/sqlflow/doc/build.md create mode 100644 docs/sqlflow/doc/close_producer_from_consumer.html create mode 100644 docs/sqlflow/doc/close_producer_from_consumer.md create mode 100644 docs/sqlflow/doc/database_abstraction_layer.html create mode 100644 docs/sqlflow/doc/database_abstraction_layer.md create mode 100644 docs/sqlflow/doc/figures/sqlflow-arch.graffle create mode 100644 docs/sqlflow/doc/figures/sqlflow-arch.png create mode 100644 docs/sqlflow/doc/figures/sqlflow-arch2.dot create mode 100644 docs/sqlflow/doc/figures/sqlflow-arch2.png create mode 100644 docs/sqlflow/doc/mysql-setup.html create mode 100644 docs/sqlflow/doc/mysql-setup.md create mode 100644 docs/sqlflow/doc/pipe.html create mode 100644 docs/sqlflow/doc/pipe.md create mode 100644 docs/sqlflow/doc/quickstart.html create mode 100644 docs/sqlflow/doc/quickstart.md create mode 100644 docs/sqlflow/doc/sql_parser.html create mode 100644 docs/sqlflow/doc/sql_parser.md create mode 100644 docs/sqlflow/doc/submitter.html create mode 100644 docs/sqlflow/doc/submitter.md create mode 100644 docs/sqlflow/doc/syntax.html create mode 100644 docs/sqlflow/doc/syntax.md create mode 100644 docs/sqlflow/doc/walkthrough.html create mode 100644 docs/sqlflow/doc/walkthrough.md create mode 100644 docs/sqlflow/example/creditcard/README.md create mode 100644 docs/sqlflow/example/creditcard/index.html create mode 100755 docs/sqlflow/example/creditcard/initialize_dataset_and_database.bash create mode 100644 docs/sqlflow/example/datasets/Dockerfile create mode 100644 docs/sqlflow/example/datasets/README.md create mode 100644 docs/sqlflow/example/datasets/data_to_sql.awk create mode 100644 docs/sqlflow/example/datasets/index.html create mode 100644 docs/sqlflow/example/fraud_detection/README.md create mode 100644 docs/sqlflow/example/fraud_detection/creditcard.csv.test create mode 100644 docs/sqlflow/example/fraud_detection/creditcard.csv.train create mode 100644 docs/sqlflow/example/fraud_detection/creditcard_data.py create mode 100644 docs/sqlflow/example/fraud_detection/index.html create mode 100644 docs/sqlflow/example/fraud_detection/premade_estimator.py create mode 100644 docs/sqlflow/example/jupyter/example.ipynb create mode 100644 docs/sqlflow/server/README.md create mode 100644 docs/sqlflow/sql/README.md create mode 100644 docs/sqlflow/sql/python/README.md create mode 100644 docs/sqlflow/sql/python/index.html create mode 100644 docs/sqlflow/sql/python/sql_data.py create mode 100644 docs/sqlflow/sql/python/test_sql_data.py create mode 100644 docs/sqlflow/sql/test create mode 100644 docs/sqlflow/sqlfs/README.md diff --git a/docs/Gemfile b/docs/Gemfile new file mode 100644 index 0000000..bde5004 --- /dev/null +++ b/docs/Gemfile @@ -0,0 +1,4 @@ +source 'https://rubygems.org' + +gem "github-pages", group: :jekyll_plugins +gem "minimal-mistakes-jekyll" diff --git a/docs/sqlflow/Dockerfile b/docs/sqlflow/Dockerfile new file mode 100644 index 0000000..eec46da --- /dev/null +++ b/docs/sqlflow/Dockerfile @@ -0,0 +1,24 @@ +FROM ubuntu:16.04 + +RUN apt-get update +RUN apt-get install -y python3-pip +RUN pip3 install --upgrade pip +RUN pip3 install tensorflow mysql-connector-python jupyter sqlflow +# Fix jupyter server "connecting to kernel" problem +# https://github.com/jupyter/notebook/issues/2664#issuecomment-468954423 +RUN pip3 install tornado==4.5.3 + +# Load sqlflow Jupyter magic command automatically. c.f. https://stackoverflow.com/a/32683001. +ENV IPYTHON_STARTUP /root/.ipython/profile_default/startup/ +RUN mkdir -p $IPYTHON_STARTUP +RUN echo 'get_ipython().magic(u"%reload_ext sqlflow.magic")' >> $IPYTHON_STARTUP/00-first.py +RUN echo 'get_ipython().magic(u"%autoreload 2")' >> $IPYTHON_STARTUP/00-first.py + +ADD demo /usr/bin/demo +ADD sqlflowserver /usr/bin/sqlflowserver + +# Make python3 callable through python +RUN ln -s /usr/bin/python3 /usr/bin/python +RUN chmod +x /usr/bin/python + +CMD ["/usr/bin/demo"] diff --git a/docs/sqlflow/Dockerfile.dev b/docs/sqlflow/Dockerfile.dev new file mode 100644 index 0000000..ccf1776 --- /dev/null +++ b/docs/sqlflow/Dockerfile.dev @@ -0,0 +1,45 @@ +FROM ubuntu:16.04 + +RUN apt-get update +RUN apt-get install -y git +RUN apt-get install -y wget curl +RUN apt-get install -y python3-pip + +RUN pip3 install --upgrade pip +RUN pip3 install grpcio-tools +RUN pip3 install tensorflow +RUN pip3 install mysql-connector-python + +RUN wget --quiet https://dl.google.com/go/go1.11.5.linux-amd64.tar.gz +RUN tar -C /usr/local -xzf go1.11.5.linux-amd64.tar.gz +RUN rm go1.11.5.linux-amd64.tar.gz +ENV PATH $PATH:/usr/local/go/bin + +RUN wget --quiet https://github.com/protocolbuffers/protobuf/releases/download/v3.6.1/protoc-3.6.1-linux-x86_64.zip +RUN apt-get install -y unzip +RUN unzip -qq protoc-3.6.1-linux-x86_64.zip -d /usr/local +RUN rm protoc-3.6.1-linux-x86_64.zip + +RUN mkdir -p /go +ENV GOPATH /go +RUN go get github.com/golang/protobuf/protoc-gen-go +RUN mv /go/bin/protoc-gen-go /usr/local/bin/ + +RUN apt-get install -y sqlite3 libsqlite3-dev + +# install mysql without a password prompt +RUN ["/bin/bash", "-c", "debconf-set-selections <<< 'mysql-server mysql-server/root_password password root'"] +RUN ["/bin/bash", "-c", "debconf-set-selections <<< 'mysql-server mysql-server/root_password_again password root'"] +RUN apt-get -y install mysql-server +VOLUME /var/lib/mysql + +# Make python3 callable through python +RUN ln -s /usr/bin/python3 /usr/bin/python +RUN chmod +x /usr/bin/python + +RUN echo "#!/bin/bash" > /build.sh +RUN echo "go generate -v ./..." >> /build.sh +RUN echo "go install -v ./..." >> /build.sh +RUN chmod +x /build.sh + +CMD ["/build.sh"] diff --git a/docs/sqlflow/LICENSE b/docs/sqlflow/LICENSE new file mode 100644 index 0000000..261eeb9 --- /dev/null +++ b/docs/sqlflow/LICENSE @@ -0,0 +1,201 @@ + Apache License + Version 2.0, January 2004 + http://www.apache.org/licenses/ + + TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION + + 1. Definitions. + + "License" shall mean the terms and conditions for use, reproduction, + and distribution as defined by Sections 1 through 9 of this document. + + "Licensor" shall mean the copyright owner or entity authorized by + the copyright owner that is granting the License. + + "Legal Entity" shall mean the union of the acting entity and all + other entities that control, are controlled by, or are under common + control with that entity. For the purposes of this definition, + "control" means (i) the power, direct or indirect, to cause the + direction or management of such entity, whether by contract or + otherwise, or (ii) ownership of fifty percent (50%) or more of the + outstanding shares, or (iii) beneficial ownership of such entity. + + "You" (or "Your") shall mean an individual or Legal Entity + exercising permissions granted by this License. + + "Source" form shall mean the preferred form for making modifications, + including but not limited to software source code, documentation + source, and configuration files. + + "Object" form shall mean any form resulting from mechanical + transformation or translation of a Source form, including but + not limited to compiled object code, generated documentation, + and conversions to other media types. + + "Work" shall mean the work of authorship, whether in Source or + Object form, made available under the License, as indicated by a + copyright notice that is included in or attached to the work + (an example is provided in the Appendix below). + + "Derivative Works" shall mean any work, whether in Source or Object + form, that is based on (or derived from) the Work and for which the + editorial revisions, annotations, elaborations, or other modifications + represent, as a whole, an original work of authorship. For the purposes + of this License, Derivative Works shall not include works that remain + separable from, or merely link (or bind by name) to the interfaces of, + the Work and Derivative Works thereof. + + "Contribution" shall mean any work of authorship, including + the original version of the Work and any modifications or additions + to that Work or Derivative Works thereof, that is intentionally + submitted to Licensor for inclusion in the Work by the copyright owner + or by an individual or Legal Entity authorized to submit on behalf of + the copyright owner. For the purposes of this definition, "submitted" + means any form of electronic, verbal, or written communication sent + to the Licensor or its representatives, including but not limited to + communication on electronic mailing lists, source code control systems, + and issue tracking systems that are managed by, or on behalf of, the + Licensor for the purpose of discussing and improving the Work, but + excluding communication that is conspicuously marked or otherwise + designated in writing by the copyright owner as "Not a Contribution." + + "Contributor" shall mean Licensor and any individual or Legal Entity + on behalf of whom a Contribution has been received by Licensor and + subsequently incorporated within the Work. + + 2. Grant of Copyright License. Subject to the terms and conditions of + this License, each Contributor hereby grants to You a perpetual, + worldwide, non-exclusive, no-charge, royalty-free, irrevocable + copyright license to reproduce, prepare Derivative Works of, + publicly display, publicly perform, sublicense, and distribute the + Work and such Derivative Works in Source or Object form. + + 3. Grant of Patent License. Subject to the terms and conditions of + this License, each Contributor hereby grants to You a perpetual, + worldwide, non-exclusive, no-charge, royalty-free, irrevocable + (except as stated in this section) patent license to make, have made, + use, offer to sell, sell, import, and otherwise transfer the Work, + where such license applies only to those patent claims licensable + by such Contributor that are necessarily infringed by their + Contribution(s) alone or by combination of their Contribution(s) + with the Work to which such Contribution(s) was submitted. If You + institute patent litigation against any entity (including a + cross-claim or counterclaim in a lawsuit) alleging that the Work + or a Contribution incorporated within the Work constitutes direct + or contributory patent infringement, then any patent licenses + granted to You under this License for that Work shall terminate + as of the date such litigation is filed. + + 4. Redistribution. You may reproduce and distribute copies of the + Work or Derivative Works thereof in any medium, with or without + modifications, and in Source or Object form, provided that You + meet the following conditions: + + (a) You must give any other recipients of the Work or + Derivative Works a copy of this License; and + + (b) You must cause any modified files to carry prominent notices + stating that You changed the files; and + + (c) You must retain, in the Source form of any Derivative Works + that You distribute, all copyright, patent, trademark, and + attribution notices from the Source form of the Work, + excluding those notices that do not pertain to any part of + the Derivative Works; and + + (d) If the Work includes a "NOTICE" text file as part of its + distribution, then any Derivative Works that You distribute must + include a readable copy of the attribution notices contained + within such NOTICE file, excluding those notices that do not + pertain to any part of the Derivative Works, in at least one + of the following places: within a NOTICE text file distributed + as part of the Derivative Works; within the Source form or + documentation, if provided along with the Derivative Works; or, + within a display generated by the Derivative Works, if and + wherever such third-party notices normally appear. The contents + of the NOTICE file are for informational purposes only and + do not modify the License. You may add Your own attribution + notices within Derivative Works that You distribute, alongside + or as an addendum to the NOTICE text from the Work, provided + that such additional attribution notices cannot be construed + as modifying the License. + + You may add Your own copyright statement to Your modifications and + may provide additional or different license terms and conditions + for use, reproduction, or distribution of Your modifications, or + for any such Derivative Works as a whole, provided Your use, + reproduction, and distribution of the Work otherwise complies with + the conditions stated in this License. + + 5. Submission of Contributions. Unless You explicitly state otherwise, + any Contribution intentionally submitted for inclusion in the Work + by You to the Licensor shall be under the terms and conditions of + this License, without any additional terms or conditions. + Notwithstanding the above, nothing herein shall supersede or modify + the terms of any separate license agreement you may have executed + with Licensor regarding such Contributions. + + 6. Trademarks. This License does not grant permission to use the trade + names, trademarks, service marks, or product names of the Licensor, + except as required for reasonable and customary use in describing the + origin of the Work and reproducing the content of the NOTICE file. + + 7. Disclaimer of Warranty. Unless required by applicable law or + agreed to in writing, Licensor provides the Work (and each + Contributor provides its Contributions) on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or + implied, including, without limitation, any warranties or conditions + of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A + PARTICULAR PURPOSE. You are solely responsible for determining the + appropriateness of using or redistributing the Work and assume any + risks associated with Your exercise of permissions under this License. + + 8. Limitation of Liability. In no event and under no legal theory, + whether in tort (including negligence), contract, or otherwise, + unless required by applicable law (such as deliberate and grossly + negligent acts) or agreed to in writing, shall any Contributor be + liable to You for damages, including any direct, indirect, special, + incidental, or consequential damages of any character arising as a + result of this License or out of the use or inability to use the + Work (including but not limited to damages for loss of goodwill, + work stoppage, computer failure or malfunction, or any and all + other commercial damages or losses), even if such Contributor + has been advised of the possibility of such damages. + + 9. Accepting Warranty or Additional Liability. While redistributing + the Work or Derivative Works thereof, You may choose to offer, + and charge a fee for, acceptance of support, warranty, indemnity, + or other liability obligations and/or rights consistent with this + License. However, in accepting such obligations, You may act only + on Your own behalf and on Your sole responsibility, not on behalf + of any other Contributor, and only if You agree to indemnify, + defend, and hold each Contributor harmless for any liability + incurred by, or claims asserted against, such Contributor by reason + of your accepting any such warranty or additional liability. + + END OF TERMS AND CONDITIONS + + APPENDIX: How to apply the Apache License to your work. + + To apply the Apache License to your work, attach the following + boilerplate notice, with the fields enclosed by brackets "[]" + replaced with your own identifying information. (Don't include + the brackets!) The text should be enclosed in the appropriate + comment syntax for the file format. We also recommend that a + file or class name and description of purpose be included on the + same "printed page" as the copyright notice for easier + identification within third-party archives. + + Copyright [yyyy] [name of copyright owner] + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. diff --git a/docs/sqlflow/README.md b/docs/sqlflow/README.md new file mode 100644 index 0000000..5d6407c --- /dev/null +++ b/docs/sqlflow/README.md @@ -0,0 +1,24 @@ +# SQLFlow + +[![Build Status](https://travis-ci.org/sql-machine-learning/sqlflow.svg?branch=develop)](https://travis-ci.org/sql-machine-learning/sqlflow) [![GoDoc](https://godoc.org/github.com/sql-machine-learning/sqlflow?status.svg)](https://godoc.org/github.com/sql-machine-learning/sqlflow) [![Release](https://img.shields.io/github/release/sql-machine-learning/sqlflow.svg)](https://github.com/sql-machine-learning/sqlflow/releases) [![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE) + +SQLFlow enriches SQL systems, e.g., MySQL, Hive, SparkSQL, with the capability of machine learning, using TensorFlow. We slightly extend the syntax of the SELECT statement of SQL to support model training and prediction. + +## How to Use + +- [Quick start](doc/quickstart.md) +- [Extended SQL syntax](doc/syntax.md) + +## How to Contribute + +- [Build from source code](doc/build.md). +- [The walkthrough of the source code](doc/walkthrough.md) +- [The choice of parser generator](doc/sql_parser.md) + +## Questions and Feedback + +Your feedback is our motivation to move on. Please let us know your questions, concerns, and issues by filing Github Issues. + +## License + +[Apache License 2.0](https://github.com/sql-machine-learning/sqlflow/LICENSE) diff --git a/docs/sqlflow/cmd/demo/README.md b/docs/sqlflow/cmd/demo/README.md new file mode 100644 index 0000000..02fd487 --- /dev/null +++ b/docs/sqlflow/cmd/demo/README.md @@ -0,0 +1,82 @@ +# SQLFlow Demo + +Before you start, you need to set up test MySQL server as described [here](../../example/datasets) + +Start SQLFlow as +```golang +go get -d ./... && go run main.go -logdir="/path/to/logs/" -loglevel="info" +``` + +#### Regular SQL Statements + +```sql +sqlflow> SHOW DATABASES; ++--------------------+ +| Database | ++--------------------+ +| churn | +| information_schema | +| iris | +| mysql | +| performance_schema | +| sqlflow_models | +| sqlfs | +| sys | ++--------------------+ +sqlflow> SELECT * FROM iris.train LIMIT 1; ++--------------+-------------+--------------+-------------+-------+ +| sepal_length | sepal_width | petal_length | petal_width | class | ++--------------+-------------+--------------+-------------+-------+ +| 6.4 | 2.8 | 5.6 | 2.2 | 2 | ++--------------+-------------+--------------+-------------+-------+ +``` + +#### Train + +```sql +sqlflow> SELECT * +FROM iris.train +TRAIN DNNClassifier +WITH + n_classes = 3, + hidden_units = [10, 20] +COLUMN sepal_length, sepal_width, petal_length, petal_width +LABEL class +INTO sqlflow_models.my_dnn_model; +----------------------------- +2018/12/16 15:03:54 tensorflowCmd: run in Docker container +Job success +``` + +#### Predict + +```sql +sqlflow> SELECT * +FROM iris.test +PREDICT iris.predict.class +USING sqlflow_models.my_dnn_model; +----------------------------- +2018/12/16 15:05:58 tensorflowCmd: run in Docker container +Job success +``` + +The prediction is stored at `iris.predict`, you can look them up at + +```sql +sqlflow> SELECT * FROM iris.predict LIMIT 10; +----------------------------- ++--------------+-------------+--------------+-------------+-------+ +| sepal_length | sepal_width | petal_length | petal_width | class | ++--------------+-------------+--------------+-------------+-------+ +| 6.4 | 2.8 | 5.6 | 2.2 | 2 | +| 5 | 2.3 | 3.3 | 1 | 1 | +| 4.9 | 2.5 | 4.5 | 1.7 | 2 | +| 4.9 | 3.1 | 1.5 | 0.1 | 0 | +| 5.7 | 3.8 | 1.7 | 0.3 | 0 | +| 4.4 | 3.2 | 1.3 | 0.2 | 0 | +| 5.4 | 3.4 | 1.5 | 0.4 | 0 | +| 6.9 | 3.1 | 5.1 | 2.3 | 2 | +| 6.7 | 3.1 | 4.4 | 1.4 | 1 | +| 5.1 | 3.7 | 1.5 | 0.4 | 0 | ++--------------+-------------+--------------+-------------+-------+ +``` diff --git a/docs/sqlflow/cmd/demo/slides/201812/sqlflow.slide b/docs/sqlflow/cmd/demo/slides/201812/sqlflow.slide new file mode 100644 index 0000000..944beef --- /dev/null +++ b/docs/sqlflow/cmd/demo/slides/201812/sqlflow.slide @@ -0,0 +1,40 @@ +SQLFlow + +* Overview +SQLFlow is a bridge that connects a SQL engine + +- MySQL, ODPS, Hive, SparkSQL, Oracle, SQL Server, ... + +with a Machine Learning toolkits + +- TensorFlow, PyTorch, XGboost, ... + +It extends the SQL syntax to enable model training and inference. + +* Syntax: Regular SQL +.code regular.sql + +* Syntax: Train +.code train.sql + +* Syntax: Prediction +.code pred.sql + +* Demo + +* Next Step: More DB +- ODPS +- Upon request + +* Next Step: More Models +- Currently only supports TF Estimator: `DNNClassifier`, `DNNRegressor` +- More TF models to come: `BoostedTreesClassifier`, `BoostedTreesRegressor` +- Upon request + +* Next Step: Data Preprocessing +Introduce proprocessing pipeline such as `norm`, `bucket`, `hash`. Also upon request. +.code preprocess.sql + +* Next Step: AutoML +.code automl.sql + diff --git a/docs/sqlflow/doc/build-tensorflow.html b/docs/sqlflow/doc/build-tensorflow.html new file mode 100644 index 0000000..9632551 --- /dev/null +++ b/docs/sqlflow/doc/build-tensorflow.html @@ -0,0 +1,429 @@ + + + + + + + + + Build TensorFlow from Source Code using Docker - Sqlflow + + + + + + + + + + + + + + + + + +
    +
    + Sqlflow + +
    + + +
    +
    +

    This site uses Just the Docs, a documentation theme for Jekyll.

    +
    +
    +
    +
    +
    + +
    +
    + + Search +
    +
    +
    + + + + +
    +
    +
    + + + +
    +

    Build TensorFlow from Source Code using Docker

    + +

    To contribute to TensorFlow, we need to build TensorFlow from source code. The official guide is great. However, it interleaves the native building process and that using Docker and makes it confusing because packages needed by the former are not by the latter. Also, we found some useful tricks to start Docker containers in practices that are not in the official guide. Hence this document.

    + +
    + +

    Build the Pip Package in TensorFlow Development Container

    + +

    On either Mac or Linux, or any other OS, we don’t have to install and configure the building tools; instead, we can use a Docker image where all tools have been installed and properly configured.

    + +
      +
    1. +

      Get the Docker image containing all the building tools:

      + + 
      docker pull tensorflow/tensorflow:latest-devel
+
      +
      2. +
    2. +

      Then, let’s get the source code. On any OS, please install git using the native package manager. For example, on Ubuntu, please

      + + 
      sudo apt-get install git
+
      + +

      or, on Mac,

      + + 
      brew install git
+
      + +

      Then, use the git just installed, let’s clone tensorflow source code:

      + + 
      git clone --recursive https://github.com/tensorflow/tensorflow
+cd tensorflow
+
      + +

      By default, we will be on the master branch. Feel free to do you change in your feature branches, or switch to a release branch, for example:

      + + 
      git checkout v1.11.0
+git checkout -b v1.11.0
+
      +
      3. +
    3. +

      Then, let us start a Docker container running the tensorflow/tensorflow:latest-devel image:

      + + 
      docker run --rm -it \
+   -w /tensorflow \
+   -v $PWD:/tensorflow \
+   -v $HOME/.cache:/root/.cache \
+   -e "HOST_PERMS=$(id -u):$(id -g)" \
+   tensorflow/tensorflow:latest-devel \
+   /bin/bash
+
      + +
        +
      • -w /tensorflow brings us to the /tensorflow directory in the container once after we start it.
        • +
      • -v $PWD:/tensorflow maps the current directory, which is the just cloned TensorFlow source directory on the host, to /tensorflow in the container.
        • +
      • -v $HOME/.cache:/root/.cache maps the Bazel temporary directory on the host into the container, so the intermediate files generated by Bazel running in the container are actually saved on the host. This allows us to interrupt the container during the build and restart it later to resume the building.
        • +
      • e "HOST_PERMS=$(id -u):$(id -g)" passes the user identification on the host into the container as an environment variable. We can reset the mode of files generated in the container to this user identity.
        • +
      +
      4. +
    4. +

      From now on, we will be working in the container. Let us first configure the building:

      + + 
      ./configure
+
      + +

      Usually, I would simply choose all the default options by hitting enter all the way down.

      +
      5. +
    5. +

      Build the pip package builder:

      + + 
      bazel build --config=opt //tensorflow/tools/pip_package:build_pip_package
+
      +
      6. +
    6. +

      Build the pip package and save it into /tensorflow:

      + + 
      ./bazel-bin/tensorflow/tools/pip_package/build_pip_package /tensorflow
+
      +
      7. +
    7. +

      Change the generated wheel file into the usual mode as on the host:

      + + 
      chown $HOST_PERMS /mnt/tensorflow-*.whl
+
      +
      8. +
    + +

    Install the Pip Pacakge in TensorFlow Development Container

    + +
      +
    1. +

      Let’s now try the new pip package. First, we need to uninstall the current tensorflow pip package, then we install the newly built one:

      + + 
      pip uninstall tensorflow
+pip install /tensorflow/tensorflow-*.whl
+
      +
      2. +
    2. +

      Now, we can verify if the new package works. First, we need to switch to a directory out from /tensorflow, so we don’t import from the source directory:

      + + 
      cd /tmp  # other directories also work
+
      + +

      then, we can try import the newly installed tensorflow package and verify it:

      + + 
      python
+>>> import tensorflow
+>>> print(tensorflow.__version__)
+
      +
      3. +
    3. +

      Now, let us quit from Python and from the Docker container. We should see the tensorflow-*.whl file on the host in the current directory.

      +
      4. +
    + +

    Install the Pip Package in a Clean Python Package

    + +

    After we quit from the development container, we should see the wheel file in the TensorFlow source directory on the host. Now, we can start a Python container and install the Pip package in it.

    + +
      +
    1. +

      Start the Python container

      + + 
      docker run --rm -it -v $PWD:/tensorflow python:2.7 bash
+
      +
      2. +
    2. +

      Install the pip pacakge in the container

      + + 
      pip install /tensorflow/tensorflow*.whl
+
      +
      3. +
    3. +

      Try TensorFlow by starting Python in the container

      + + 
      python
+>>> import tensorflow as tf
+>>> print(tf.__version__)
+
      +
      4. +
    + + + + +
    +
    +
    +
    + + + diff --git a/docs/sqlflow/doc/build-tensorflow.md b/docs/sqlflow/doc/build-tensorflow.md new file mode 100644 index 0000000..5360f1d --- /dev/null +++ b/docs/sqlflow/doc/build-tensorflow.md @@ -0,0 +1,135 @@ +# Build TensorFlow from Source Code using Docker + +To contribute to TensorFlow, we need to build TensorFlow from source code. The [official guide](https://www.tensorflow.org/install/source) is great. However, it interleaves the native building process and that using Docker and makes it confusing because packages needed by the former are not by the latter. Also, we found some useful tricks to start Docker containers in practices that are not in the official guide. Hence this document. + +---- + +## Build the Pip Package in TensorFlow Development Container + +On either Mac or Linux, or any other OS, we don't have to install and configure the building tools; instead, we can use a Docker image where all tools have been installed and properly configured. + +1. Get the Docker image containing all the building tools: + + ```bash + docker pull tensorflow/tensorflow:latest-devel + ``` + +1. Then, let's get the source code. On any OS, please install git using the native package manager. For example, on Ubuntu, please + + ```bash + sudo apt-get install git + ``` + + or, on Mac, + + ```bash + brew install git + ``` + + Then, use the git just installed, let's clone tensorflow source code: + + ```bash + git clone --recursive https://github.com/tensorflow/tensorflow + cd tensorflow + ``` + + By default, we will be on the `master` branch. Feel free to do you change in your feature branches, or switch to a release branch, for example: + + ```bash + git checkout v1.11.0 + git checkout -b v1.11.0 + ``` + +1. Then, let us start a Docker container running the `tensorflow/tensorflow:latest-devel` image: + + ```bash + docker run --rm -it \ + -w /tensorflow \ + -v $PWD:/tensorflow \ + -v $HOME/.cache:/root/.cache \ + -e "HOST_PERMS=$(id -u):$(id -g)" \ + tensorflow/tensorflow:latest-devel \ + /bin/bash + ``` + + - `-w /tensorflow` brings us to the `/tensorflow` directory in the container once after we start it. + - `-v $PWD:/tensorflow` maps the current directory, which is the just cloned TensorFlow source directory on the host, to `/tensorflow` in the container. + - `-v $HOME/.cache:/root/.cache` maps the Bazel temporary directory on the host into the container, so the intermediate files generated by Bazel running in the container are actually saved on the host. This allows us to interrupt the container during the build and restart it later to resume the building. + - `e "HOST_PERMS=$(id -u):$(id -g)"` passes the user identification on the host into the container as an environment variable. We can reset the mode of files generated in the container to this user identity. + + +1. From now on, we will be working in the container. Let us first configure the building: + + ```bash + ./configure + ``` + + Usually, I would simply choose all the default options by hitting enter all the way down. + +1. Build the pip package builder: + + ```bash + bazel build --config=opt //tensorflow/tools/pip_package:build_pip_package + ``` + +1. Build the pip package and save it into `/tensorflow`: + + ```bash + ./bazel-bin/tensorflow/tools/pip_package/build_pip_package /tensorflow + ``` + +1. Change the generated wheel file into the usual mode as on the host: + + ```bash + chown $HOST_PERMS /mnt/tensorflow-*.whl + ``` + +# Install the Pip Pacakge in TensorFlow Development Container + +1. Let's now try the new pip package. First, we need to uninstall the current tensorflow pip package, then we install the newly built one: + + ```bash + pip uninstall tensorflow + pip install /tensorflow/tensorflow-*.whl + ``` + +1. Now, we can verify if the new package works. First, we need to switch to a directory out from `/tensorflow`, so we don't import from the source directory: + + ```bash + cd /tmp # other directories also work + ``` + + then, we can try import the newly installed tensorflow package and verify it: + + ```bash + python + >>> import tensorflow + >>> print(tensorflow.__version__) + ``` + +1. Now, let us quit from Python and from the Docker container. We should see the `tensorflow-*.whl` file on the host in the current directory. + + +## Install the Pip Package in a Clean Python Package + +After we quit from the development container, we should see the wheel file in the TensorFlow source directory on the host. Now, we can start a Python container and install the Pip package in it. + +1. Start the Python container + + ```bash + docker run --rm -it -v $PWD:/tensorflow python:2.7 bash + ``` + +1. Install the pip pacakge in the container + + ```bash + pip install /tensorflow/tensorflow*.whl + ``` + +1. Try TensorFlow by starting Python in the container + + ```bash + python + >>> import tensorflow as tf + >>> print(tf.__version__) + ``` diff --git a/docs/sqlflow/doc/build.html b/docs/sqlflow/doc/build.html new file mode 100644 index 0000000..e185dec --- /dev/null +++ b/docs/sqlflow/doc/build.html @@ -0,0 +1,406 @@ + + + + + + + + + Canonical Development Environment - Sqlflow + + + + + + + + + + + + + + + + + +
    +
    + Sqlflow + +
    + + +
    +
    +

    This site uses Just the Docs, a documentation theme for Jekyll.

    +
    +
    +
    +
    +
    + +
    +
    + + Search +
    +
    +
    + + + + +
    +
    +
    + + + +
    +

    Canonical Development Environment

    + +

    Referring to this example, +we create a canonical development environment for Go and Python programmers using Docker.

    + +

    Editing on Host

    + +

    When we use this Docker image for daily development work, the source code relies +on the host computer instead of the container. The source code includes this repo +and all its dependencies, for example, the Go package google.golang.org/grpc. +Code-on-the-host allows us to run our favorite editors (Emacs, VIM, Eclipse, and more) +on the host. Please free to rely on editors add-ons to analyze the source code +for auto-completion.

    + +

    Building in Container

    + +

    We build a Docker image that contains development tools:

    + +
      +
    1. The Python interpreter
      2. +
    2. The Go compiler
      3. +
    3. The protobuf compiler
      4. +
    4. The protobuf to Go compiler extension
      5. +
    5. The protobuf to Python compiler extension
      6. +
    + +

    Because this repo contains Go code, please make sure that you have the directory +structure required by Go. On my laptop computer, I have

    + +
    export GOPATH=$HOME/go
+
    + +

    You could have your $GOPATH pointing to any directory you like.

    + +

    Given $GOPATH$ set, we could git clone the source code of our project by running:

    + +
    go get github.com/sql-machine-learning/sqlflow
+
    + +

    Change the directory to our project root, and we can use go get to retrieve +and update Go dependencies.

    + +
    cd $GOPATH/src/github.com/sql-machine-learning/sqlflow
+go get -u -t ./...
+
    + +

    Note -t instructs get to also download the packages required to build +the tests for the specified packages.

    + +

    As all Git users would do, we run git pull from time to time to sync up with +others’ work. If somebody added new dependencies, we might need to run go -u ./... +after git pull to update dependencies.

    + +

    To build this project, we need the protobuf compiler, Go compiler, Python interpreter, +gRPC extension to the protobuf compiler. To ease the installation and configuration +of these tools, we provided a Dockerfile to install them into a Docker image. +To build the Docker image:

    + +
    docker build -t sqlflow:dev -f Dockerfile.dev .
+
    + +

    Development

    + +

    Build and Test

    + +

    We build and test the project inside the docker container.

    + +

    To run the container, we need to map the $GOPATH directory on the host into the +/go directory in the container, because the Dockerfile configures /go as +the $GOPATH in the container:

    + +
    docker run --rm -it -v $GOPATH:/go \
+    -w /go/src/github.com/sql-machine-learning/sqlflow \
+    sqlflow:dev bash
+
    + +

    Inside the Docker container, start a MySQL server in the background

    + +
    service mysql start&
+
    + +

    run all the tests as

    + +
    go generate ./...
+go install ./...
+go test -v ./...
+
    + +

    where go generate invokes the protoc command to translate server/sqlflow.proto +into server/sqlflow.pb.go and go test -v builds and run unit tests.

    + +

    Release

    + +

    The above build process currently generates two binary files in +$GOPATH/bin on the host. To package them into a Docker image, +please run

    + +
    docker build -t sqlflow -f ./Dockerfile $GOPATH/bin
+
    + +

    To publish the released Docker image to our official DockerHub

    +
    docker tag sqlflow sqlflow/sqlflow:latest
+docker push sqlflow/sqlflow:latest
+
    + +

    Demo: Command line Prompt

    + +

    The demo requires a MySQL server instance with populated data. If we don’t, we could +follow example/datasets/README.md to start one on the host. +After setting up MySQL, run the following inside the Docker container

    + +
    go run cmd/demo/demo.go --db_user root --db_password root --db_address host.docker.internal:3306
+
    + +

    You should be able to see the following prompt

    + +
    sqlflow>
+
    + + + + +
    +
    +
    +
    + + + diff --git a/docs/sqlflow/doc/build.md b/docs/sqlflow/doc/build.md new file mode 100644 index 0000000..774dc91 --- /dev/null +++ b/docs/sqlflow/doc/build.md @@ -0,0 +1,128 @@ +# Canonical Development Environment + +Referring to [this example](https://github.com/sql-machine-learning/canonicalize-go-python-grpc-dev-env), +we create a canonical development environment for Go and Python programmers using Docker. + +## Editing on Host + +When we use this Docker image for daily development work, the source code relies +on the host computer instead of the container. The source code includes this repo +and all its dependencies, for example, the Go package `google.golang.org/grpc`. +Code-on-the-host allows us to run our favorite editors (Emacs, VIM, Eclipse, and more) +on the host. Please free to rely on editors add-ons to analyze the source code +for auto-completion. + +## Building in Container + +We build a Docker image that contains development tools: + +1. The Python interpreter +1. The Go compiler +1. The protobuf compiler +1. The protobuf to Go compiler extension +1. The protobuf to Python compiler extension + +Because this repo contains Go code, please make sure that you have the directory +structure required by Go. On my laptop computer, I have + +```bash +export GOPATH=$HOME/go +``` + +You could have your `$GOPATH` pointing to any directory you like. + +Given `$GOPATH$` set, we could git clone the source code of our project by running: + +```bash +go get github.com/sql-machine-learning/sqlflow +``` + +Change the directory to our project root, and we can use `go get` to retrieve +and update Go dependencies. + +```bash +cd $GOPATH/src/github.com/sql-machine-learning/sqlflow +go get -u -t ./... +``` + +Note `-t` instructs get to also download the packages required to build +the tests for the specified packages. + +As all Git users would do, we run `git pull` from time to time to sync up with +others' work. If somebody added new dependencies, we might need to run `go -u ./...` +after `git pull` to update dependencies. + +To build this project, we need the protobuf compiler, Go compiler, Python interpreter, +gRPC extension to the protobuf compiler. To ease the installation and configuration +of these tools, we provided a Dockerfile to install them into a Docker image. +To build the Docker image: + +```bash +docker build -t sqlflow:dev -f Dockerfile.dev . +``` + +## Development + +### Build and Test + +We build and test the project inside the docker container. + +To run the container, we need to map the `$GOPATH` directory on the host into the +`/go` directory in the container, because the Dockerfile configures `/go` as +the `$GOPATH` in the container: + +```bash +docker run --rm -it -v $GOPATH:/go \ + -w /go/src/github.com/sql-machine-learning/sqlflow \ + sqlflow:dev bash +``` + +Inside the Docker container, start a MySQL server in the background + +``` +service mysql start& +``` + +run all the tests as + +``` +go generate ./... +go install ./... +go test -v ./... +``` + +where `go generate` invokes the `protoc` command to translate `server/sqlflow.proto` +into `server/sqlflow.pb.go` and `go test -v` builds and run unit tests. + + +### Release + +The above build process currently generates two binary files in +`$GOPATH/bin` on the host. To package them into a Docker image, +please run + +```bash +docker build -t sqlflow -f ./Dockerfile $GOPATH/bin +``` + +To publish the released Docker image to our official DockerHub +```bash +docker tag sqlflow sqlflow/sqlflow:latest +docker push sqlflow/sqlflow:latest +``` + +## Demo: Command line Prompt + +The demo requires a MySQL server instance with populated data. If we don't, we could +follow [example/datasets/README.md](/example/datasets/README.md) to start one on the host. +After setting up MySQL, run the following inside the Docker container + +```bash +go run cmd/demo/demo.go --db_user root --db_password root --db_address host.docker.internal:3306 +``` + +You should be able to see the following prompt + +``` +sqlflow> +``` diff --git a/docs/sqlflow/doc/close_producer_from_consumer.html b/docs/sqlflow/doc/close_producer_from_consumer.html new file mode 100644 index 0000000..a418f6e --- /dev/null +++ b/docs/sqlflow/doc/close_producer_from_consumer.html @@ -0,0 +1,447 @@ + + + + + + + + + Closing the producer goroutine from the consumer - Sqlflow + + + + + + + + + + + + + + + + + +
    +
    + Sqlflow + +
    + + +
    +
    +

    This site uses Just the Docs, a documentation theme for Jekyll.

    +
    +
    +
    +
    +
    + +
    +
    + + Search +
    +
    +
    + + + + +
    +
    +
    + + + +
    +

    Closing the producer goroutine from the consumer

    + +

    The producer-and-consumer pattern is well used in Go concurrent programming. When +the consumer stops, we want to gracefully stop the producer as well.

    + +

    Problem

    + +

    When a gRPC server receives a streaming request, it usually calls a +function that returns a channel, +reads the result from that channel and send the result to the client one by one.

    + +

    Take the following code for instance: upon receiving a request, the main goroutine +Service calls launchJob. launchJob starts a separate goroutine as an anonymous +function call and returns a channel. In the anonymous function, items will be sent to +channel. And Service on the otherside of the channel will reads from it.

    + +
    func Service(req *Request, stream *StreamResponse) error {
+  result := launchJob(req.Content)
+  for r := range result {
+    if e := stream.Send(result); e != nil {
+      // should we signal the running goroutine so it will stop sending?
+      return e
+    }
+  }
+}
+
+func launchJob(content string) chan Item {
+  c := make(chan Item)
+  
+  go func() {
+    defer close(c)
+
+    acquireScarceResources()
+    defer releaseScarceResources()
+
+    ...
+    // if stream.Send(result) returns an error and the Service returns, this will be blocked
+    c <- Item{}
+    ...
+  }()
+  
+  return c
+}
+
    + +

    There is a major problem in this implementation. As pointed out by the comment, +if the Send in Service returns an error, the Service function will return, +leaving the anonymous function being blocked on c <- Item{} forever.

    + +

    This problem is important because the leaking goroutine usually owns scarce system +resources such as network connection and memory.

    + +

    Solution: pipeline explicit cancellation

    + +

    Inspired by this blog post section +Explicit cancellation, we can signal the cancellation via closing on a separate +channel. And we can follow the terminology as io.Pipe.

    + +
    package sql
+
+import (
+	"errors"
+)
+
+var ErrClosedPipe = errors.New("pipe: write on closed pipe")
+
+// pipe follows the design at https://blog.golang.org/pipelines
+// - wrCh: chan for piping data
+// - done: chan for signaling Close from Reader to Writer
+type pipe struct {
+	wrCh chan interface{}
+	done chan struct{}
+}
+
+// PipeReader reads real data
+type PipeReader struct {
+	p *pipe
+}
+
+// PipeWriter writes real data
+type PipeWriter struct {
+	p *pipe
+}
+
+// Pipe creates a synchronous in-memory pipe.
+//
+// It is safe to call Read and Write in parallel with each other or with Close.
+// Parallel calls to Read and parallel calls to Write are also safe:
+// the individual calls will be gated sequentially.
+func Pipe() (*PipeReader, *PipeWriter) {
+	p := &pipe{
+		wrCh: make(chan interface{}),
+		done: make(chan struct{})}
+	return &PipeReader{p}, &PipeWriter{p}
+}
+
+// Close closes the reader; subsequent writes to the
+func (r *PipeReader) Close() {
+	close(r.p.done)
+}
+
+// ReadAll returns the data chan. The caller should
+// use it as `for r := range pr.ReadAll()`
+func (r *PipeReader) ReadAll() chan interface{} {
+	return r.p.wrCh
+}
+
+// Close closes the writer; subsequent ReadAll from the
+// read half of the pipe will return a closed channel.
+func (w *PipeWriter) Close() {
+	close(w.p.wrCh)
+}
+
+// Write writes the item to the underlying data stream.
+// It returns ErrClosedPipe when the data stream is closed.
+func (w *PipeWriter) Write(item interface{}) error {
+	select {
+	case w.p.wrCh <- item:
+		return nil
+	case <-w.p.done:
+		return ErrClosedPipe
+	}
+}
+
    + +

    And the consumer and producer be can implemented as

    + +
    func Service(req *Request, stream *StreamResponse) error {
+  pr := launchJob(req.Content)
+  defer pr.Close()
+  for r := range pr.ReadAll() {
+    if e := stream.Send(r); e != nil {
+      return e
+    }
+  }
+}
+
+func launchJob(content string) PipeReader {
+	pr, pw := Pipe()
+	go func() {
+		defer pw.Close()
+		
+		if err := pw.Write(Item{}); err != nil {
+			return
+		}
+	}
+	return pr
+}
+
    + +

    Further Reading

    + +
      +
    1. Google Form: Channel send timeout
      2. +
    2. Go by Example: Timeouts
      3. +
    3. Google I/O 2013 - Advanced Go Concurrency Patterns
      4. +
    4. Go Concurrency Patterns Talk
      5. +
    5. Go Concurrency Patterns: Pipelines and cancellation
      6. +
    + + + + +
    +
    +
    +
    + + + diff --git a/docs/sqlflow/doc/close_producer_from_consumer.md b/docs/sqlflow/doc/close_producer_from_consumer.md new file mode 100644 index 0000000..0b506e9 --- /dev/null +++ b/docs/sqlflow/doc/close_producer_from_consumer.md @@ -0,0 +1,161 @@ +# Closing the producer goroutine from the consumer + +The producer-and-consumer pattern is well used in Go concurrent programming. When +the consumer stops, we want to gracefully stop the producer as well. + +## Problem + +When a gRPC server receives a streaming request, it usually calls a +[function that returns a channel](https://talks.golang.org/2012/concurrency.slide#25), +reads the result from that channel and send the result to the client one by one. + +Take the following code for instance: upon receiving a request, the main goroutine +`Service` calls `launchJob`. `launchJob` starts a separate goroutine as an anonymous +function call and returns a channel. In the anonymous function, items will be sent to +channel. And `Service` on the otherside of the channel will reads from it. + +```go +func Service(req *Request, stream *StreamResponse) error { + result := launchJob(req.Content) + for r := range result { + if e := stream.Send(result); e != nil { + // should we signal the running goroutine so it will stop sending? + return e + } + } +} + +func launchJob(content string) chan Item { + c := make(chan Item) + + go func() { + defer close(c) + + acquireScarceResources() + defer releaseScarceResources() + + ... + // if stream.Send(result) returns an error and the Service returns, this will be blocked + c <- Item{} + ... + }() + + return c +} +``` + +There is a major problem in this implementation. As pointed out by the comment, +if the `Send` in `Service` returns an error, the `Service` function will return, +leaving the anonymous function being blocked on `c <- Item{}` forever. + +This problem is important because the leaking goroutine usually owns scarce system +resources such as network connection and memory. + + +## Solution: pipeline explicit cancellation + +Inspired by this [blog post](https://blog.golang.org/pipelines) section +*Explicit cancellation*, we can signal the cancellation via closing on a separate +channel. And we can follow the terminology as `io.Pipe`. + +```go +package sql + +import ( + "errors" +) + +var ErrClosedPipe = errors.New("pipe: write on closed pipe") + +// pipe follows the design at https://blog.golang.org/pipelines +// - wrCh: chan for piping data +// - done: chan for signaling Close from Reader to Writer +type pipe struct { + wrCh chan interface{} + done chan struct{} +} + +// PipeReader reads real data +type PipeReader struct { + p *pipe +} + +// PipeWriter writes real data +type PipeWriter struct { + p *pipe +} + +// Pipe creates a synchronous in-memory pipe. +// +// It is safe to call Read and Write in parallel with each other or with Close. +// Parallel calls to Read and parallel calls to Write are also safe: +// the individual calls will be gated sequentially. +func Pipe() (*PipeReader, *PipeWriter) { + p := &pipe{ + wrCh: make(chan interface{}), + done: make(chan struct{})} + return &PipeReader{p}, &PipeWriter{p} +} + +// Close closes the reader; subsequent writes to the +func (r *PipeReader) Close() { + close(r.p.done) +} + +// ReadAll returns the data chan. The caller should +// use it as `for r := range pr.ReadAll()` +func (r *PipeReader) ReadAll() chan interface{} { + return r.p.wrCh +} + +// Close closes the writer; subsequent ReadAll from the +// read half of the pipe will return a closed channel. +func (w *PipeWriter) Close() { + close(w.p.wrCh) +} + +// Write writes the item to the underlying data stream. +// It returns ErrClosedPipe when the data stream is closed. +func (w *PipeWriter) Write(item interface{}) error { + select { + case w.p.wrCh <- item: + return nil + case <-w.p.done: + return ErrClosedPipe + } +} +``` + +And the consumer and producer be can implemented as + +```go +func Service(req *Request, stream *StreamResponse) error { + pr := launchJob(req.Content) + defer pr.Close() + for r := range pr.ReadAll() { + if e := stream.Send(r); e != nil { + return e + } + } +} + +func launchJob(content string) PipeReader { + pr, pw := Pipe() + go func() { + defer pw.Close() + + if err := pw.Write(Item{}); err != nil { + return + } + } + return pr +} +``` + +## Further Reading + +1. [Google Form: Channel send timeout](https://groups.google.com/forum/#!topic/golang-nuts/Oth9CmJPoqo) +2. [Go by Example: Timeouts](https://gobyexample.com/timeouts) +3. [Google I/O 2013 - Advanced Go Concurrency Patterns](https://www.youtube.com/watch?v=QDDwwePbDtw&t=111s) +4. [Go Concurrency Patterns Talk](https://talks.golang.org/2012/concurrency.slide) +5. [Go Concurrency Patterns: Pipelines and cancellation](https://blog.golang.org/pipelines) \ No newline at end of file diff --git a/docs/sqlflow/doc/database_abstraction_layer.html b/docs/sqlflow/doc/database_abstraction_layer.html new file mode 100644 index 0000000..bba96a1 --- /dev/null +++ b/docs/sqlflow/doc/database_abstraction_layer.html @@ -0,0 +1,383 @@ + + + + + + + + + Compatibility with Various SQL Engines - Sqlflow + + + + + + + + + + + + + + + + + +
    +
    + Sqlflow + +
    + + +
    +
    +

    This site uses Just the Docs, a documentation theme for Jekyll.

    +
    +
    +
    +
    +
    + +
    +
    + + Search +
    +
    +
    + + + + +
    +
    +
    + + + +
    +

    Compatibility with Various SQL Engines

    + +

    SQLFlow interacts with SQL engines like MySQL and Hive, while different SQL engines use variants of SQL syntax, it is important for SQLFlow to have an abstraction layer that hides such differences.

    + +

    SQLFlow calls Go’s standard database API. The submitter programs generated by SQLFlow call Python’s database API. Both APIs abstract the interface to various SQL engines; however, they are insufficient for SQLFlow to work. In this document, we examine all interactions between SQLFlow and the SQL engine so to identify what SQLFlow authors have to abstract in addition to calling Go’s and Python’s database APIs.

    + +

    Data Operations in Go

    + +

    Data Retrieval

    + +

    The basic idea of SQLFlow is to extend the SELECT statement of SQL to have the TRAIN and PREDICT clauses. For more discussion, please refer to the syntax design. SQLFlow translates such “extended SQL statements” into submitter programs, which forward the part from SELECT to TRAIN or PREDICT, which we call the “standard part”, to the SQL engine. SQLFlow also accepts the SELECT statement without TRAIN or PREDICt clauses and would forward such “standard statements” to the engine. It is noticeable that the “standard part” or “standard statements” are not standardized. For example, various engines use different syntax for joining.

    + +
      +
    • MySQL: SELECT pet.name, comment FROM pet, event WHERE pet.name =event.name; with keyword WHERE .
      • +
    • Hive: SELECT pet.name, comment FROM pet JOIN event ON (pet.name =event.name) with keyword JOIN and ON.
      • +
    • ODPS and SQLite use either INNER JOIN or OUTER JOIN.
      • +
    + +

    Fortunately, as SQLFlow forwards the above parts to the engine, it doesn’t have to care much about the differences above.

    + +

    Metadata Retrieval

    + +

    To verify the semantics of users’ inputs, SQLFlow needs to retrieve the schema of tables. For example, the input might be

    + +
    SELECT name, age, income FROM employee TRAIN DNNRegressor WITH hidden_layers=[10,50,10] COLUMN name, agee LABEL income;
+
    + +

    In the above example, the user misspelled the field name age in the COLUMN clause as “agee”. SQLFlow must be able to find that out.

    + +

    To do that, SQLFlow needs to query the field names from the SQL engine. However, different engines use various syntax. For example:

    + +
      +
    • MySQL: DESCRIBE/DESC employee;
      • +
    • Hive: DESCRIBE FORMATTED employee;
      • +
    • ODPS: DESC employee;
      • +
    • SQLite: PRAGMA table_info([employee]);
      • +
    + +

    The returned data format varies too. Our solution to avoid such differences is not-to-use-them; instead, SQLFlow retrieves the table schema by running a query like SELECT * FROM employee LIMIT 1; and inferring field types using the mechanism called DatabaseTypeName provided by SQL engines drivers beneath the Go’s standard database API.

    + +

    Prepare Prediction Table

    + +

    A SQLFlow prediction job writes its prediction results into a table. It prepares the prediction table by

    + +
      +
    1. Dropping previous prediction table DROP TABLE IF EXISTS my_table;
      2. +
    2. Creating table with schema CREATE TABLE my_table (name1, type1, name2 type2);
      3. +
    + +

    Most SQL engines, including MySQL, Hive, ODPS, SQLite, support both statements.

    + +

    Translate Database Column Type to TensorFlow Feature Column Type

    + +

    After retrieving database column type name through DatabaseTypeName, we can derive TensorFlow’s feature column type via a mapping such as {"FLOAT", "DOUBLE"} -> tf.numeric_column.

    + +

    Save Model

    + +

    SQLFlow saves trained ML model by dumping the serialized the model directory into a table. It first creates a table by CREATE TABLE IF NOT EXISTS %s (id INT AUTO_INCREMENT, block BLOB, PRIMARY KEY (id)) and insert blobs by INSERT INTO %s (block) VALUES(?).

    + +

    Note that Hive and ODPS doesn’t have BLOB type, we need to use BINARY (docs at ODPS, Hive) instead.

    + +

    Also, note that Hive and ODPS doesn’t support AUTO_INCREMENT, we need to implemented auto increment logic in sqlfs.

    + +

    Load Model

    + +

    SQLFlow loads trained ML model by reading rows in a table and deserializing the blob to a model directory.

    + +

    It reads rows by running SELECT block FROM %s ORDER BY id, which is supported by most databases.

    + +

    Data Operations in Python

    + +

    Connect to SQL Engines

    + +

    Thanks to the Python database API, connecting to different databases follows a similar API.

    + +
    conn = mysql.connector.connect(user='scott', password='password',
+                               host='127.0.0.1',
+                               database='employees')
+conn = sqlite3.connect('path/to/your/sqlite/file')
+conn = pyhive.connect('localhost')
+
+cursor = conn.cursor()
+cursor.execute('select * from my_table;')
+
    + +

    Insert Prediction Result into Prediction Table

    + +

    Python database API provides execute_many(sql, value) to insert multiple values at once. So one can prepare the following insertion statement. Please be aware that MySQL and SQLite use INSERT INTO to insert rows while Hive and ODPS use INSERT INTO TABLE.

    + +
    -- MySQL, SQLite
+INSERT INTO table_name VALUES (value1, value2, value3, ...);
+-- Hive, ODPS
+INSERT INTO TABLE table_name VALUES (value1, value2, value3, ...);
+
    + + + + + +
    +
    +
    +
    + + + diff --git a/docs/sqlflow/doc/database_abstraction_layer.md b/docs/sqlflow/doc/database_abstraction_layer.md new file mode 100644 index 0000000..c5d9821 --- /dev/null +++ b/docs/sqlflow/doc/database_abstraction_layer.md @@ -0,0 +1,92 @@ +# Compatibility with Various SQL Engines + +SQLFlow interacts with SQL engines like MySQL and Hive, while different SQL engines use variants of SQL syntax, it is important for SQLFlow to have an abstraction layer that hides such differences. + +SQLFlow calls Go's [standard database API](https://golang.org/pkg/database/sql/). The submitter programs generated by SQLFlow call Python's [database API](https://www.python.org/dev/peps/pep-0249/). Both APIs abstract the interface to various SQL engines; however, they are insufficient for SQLFlow to work. In this document, we examine all interactions between SQLFlow and the SQL engine so to identify what SQLFlow authors have to abstract in addition to calling Go's and Python's database APIs. + +## Data Operations in Go + +### Data Retrieval + +The basic idea of SQLFlow is to extend the SELECT statement of SQL to have the TRAIN and PREDICT clauses. For more discussion, please refer to the [syntax design](/doc/syntax.md). SQLFlow translates such "extended SQL statements" into submitter programs, which forward the part from SELECT to TRAIN or PREDICT, which we call the "standard part", to the SQL engine. SQLFlow also accepts the SELECT statement without TRAIN or PREDICt clauses and would forward such "standard statements" to the engine. It is noticeable that the "standard part" or "standard statements" are not standardized. For example, various engines use different syntax for joining. + +- MySQL: `SELECT pet.name, comment FROM pet, event WHERE pet.name =event.name;` with keyword `WHERE` . +- Hive: `SELECT pet.name, comment FROM pet JOIN event ON (pet.name =event.name)` with keyword `JOIN` and `ON`. +- ODPS and SQLite use either `INNER JOIN` or `OUTER JOIN`. + +Fortunately, as SQLFlow forwards the above parts to the engine, it doesn't have to care much about the differences above. + +### Metadata Retrieval + +To verify the semantics of users' inputs, SQLFlow needs to retrieve the schema of tables. For example, the input might be + +```SQL +SELECT name, age, income FROM employee TRAIN DNNRegressor WITH hidden_layers=[10,50,10] COLUMN name, agee LABEL income; +``` + +In the above example, the user misspelled the field name `age` in the COLUMN clause as "agee". SQLFlow must be able to find that out. + +To do that, SQLFlow needs to query the field names from the SQL engine. However, different engines use various syntax. For example: + +- MySQL: `DESCRIBE/DESC employee;` +- Hive: `DESCRIBE FORMATTED employee;` +- ODPS: `DESC employee;` +- SQLite: `PRAGMA table_info([employee]);` + +The returned data format varies too. Our solution to avoid such differences is not-to-use-them; instead, SQLFlow retrieves the table schema by running a query like `SELECT * FROM employee LIMIT 1;` and inferring field types using the mechanism called [DatabaseTypeName](https://golang.org/pkg/database/sql/#ColumnType.DatabaseTypeName) provided by SQL engines drivers beneath the Go's standard database API. + +### Prepare Prediction Table + +A SQLFlow prediction job writes its prediction results into a table. It prepares the prediction table by + +1. Dropping previous prediction table `DROP TABLE IF EXISTS my_table;` +2. Creating table with schema `CREATE TABLE my_table (name1, type1, name2 type2);` + +Most SQL engines, including MySQL, Hive, ODPS, SQLite, support both statements. + +### Translate Database Column Type to TensorFlow Feature Column Type + +After retrieving database column type name through [DatabaseTypeName](https://golang.org/pkg/database/sql/#ColumnType.DatabaseTypeName), we can derive TensorFlow's [feature column](https://www.tensorflow.org/guide/feature_columns) type via a mapping such as `{"FLOAT", "DOUBLE"} -> tf.numeric_column`. + +### Save Model + +SQLFlow saves trained ML model by dumping the serialized the model directory into a table. It first creates a table by `CREATE TABLE IF NOT EXISTS %s (id INT AUTO_INCREMENT, block BLOB, PRIMARY KEY (id))` and insert blobs by `INSERT INTO %s (block) VALUES(?)`. + +Note that Hive and ODPS doesn't have `BLOB` type, we need to use `BINARY` (docs at [ODPS](https://help.aliyun.com/document_detail/27821.html?spm=a2c4g.11186623.6.577.768231deoru03E), [Hive](https://cwiki.apache.org/confluence/display/Hive/LanguageManual+Types#LanguageManualTypes-MiscTypes)) instead. + +Also, note that Hive and ODPS doesn't support `AUTO_INCREMENT`, we need to implemented auto increment logic in `sqlfs`. + +### Load Model + +SQLFlow loads trained ML model by reading rows in a table and deserializing the blob to a model directory. + +It reads rows by running `SELECT block FROM %s ORDER BY id`, which is supported by most databases. + +## Data Operations in Python + +### Connect to SQL Engines + +Thanks to the Python database API, connecting to different databases follows a similar API. + +```python +conn = mysql.connector.connect(user='scott', password='password', + host='127.0.0.1', + database='employees') +conn = sqlite3.connect('path/to/your/sqlite/file') +conn = pyhive.connect('localhost') + +cursor = conn.cursor() +cursor.execute('select * from my_table;') +``` + +### Insert Prediction Result into Prediction Table + +Python database API provides `execute_many(sql, value)` to insert multiple values at once. So one can prepare the following insertion statement. Please be aware that MySQL and SQLite use `INSERT INTO` to insert rows while Hive and ODPS use `INSERT INTO TABLE`. + +```sql +-- MySQL, SQLite +INSERT INTO table_name VALUES (value1, value2, value3, ...); +-- Hive, ODPS +INSERT INTO TABLE table_name VALUES (value1, value2, value3, ...); +``` + diff --git a/docs/sqlflow/doc/figures/sqlflow-arch.graffle b/docs/sqlflow/doc/figures/sqlflow-arch.graffle new file mode 100644 index 0000000000000000000000000000000000000000..1a9b5418a59a55332c7765e59ae259df44f9b059 GIT binary patch literal 3985 zcmV;C4{q=uiwFP!000030PS6CciOlX{XF>3YKiib|>wuyxOFbHff;kZ5M0Z zT-S!*e`>#ZVXpu6dTsMxyGJ|iug81Jz;nY$IsUYLc(9|azi2kM1_O^Zo4f5@ z<@oTR)mB6v&F0>Rb!GiLiUu#6&CAQnhApic4)06Duz4Kt0Slt5Ly_4Fk)Y9yy6d8t zQSO=Qh;O^D6TM#hWAlPty_Utfjva~XpGRh!P5HgJYzKiYKm4&7MuDjRYf)T-_kFh) z@Zq3wBtE?j?6WhE(amOJlCnnBYG6&&1sjx7oHE&Lj+4=CZ-3CYdu)df zMVUddHigO@y8gwi+>4cYZ})r_-Nrr?)o7hFktpju&54OXz**vFH}x}msp6L!F7M}2 zMr^^3{=g%)_Ha`|!ur1u~B3!^${3AN>&-lbN0bx?l);5nK-hXcgq6TA| z6a*((pLyJl(r}wx(NPA4%_)3GCG7LS{l$IT6ZLlexF_ZnBsfbtZb?DgThXuX{yXpM z7o@WPZhPMj_W73nar@&zYwvvP{OFTP(CL?Vp3~bA|Lz`rLf_6$_dk1E;`+`WX}9sa z@87}|Ea3!67Ty~~}q^Q(Wz&Z4_C4>ITHe{2f|ZMnbLyzImHdP(lTDMAfP zybOX>U8ja&Y$%3>uWu++y2S~T1kHr^35I!e{c_^=&X?LTZ9Z5=_M6T0FnGf3?vd|Z z-HhZinu~KD13DR3oyh&c4(%%z9Qa-Kb5i*!O=rf8aWXLiZFT<{hEXa(6bxB&YO`%S z7jg*pyYblMIw{jhB@$C<6;7{D{q?uHp{Umz%5Ow;YePXuHLf$xn&hzK*L?& zYhXJfqf{C(f_bksA>rZ=6Z2g+9C-Fs%dx$Z$|1$|l~;;i%_c3RMIJGi+!AT zvq?chxv)|QM!FWCi5$l zfk&#}G%JUYlH9XtCrJq@B#o7yWT4UnxYYibi!ZZ1k?OaUcN$b}kPU?p^SaP@V;c)m zX=7mZx&Xp-uBALs(rn?m-Q}0rrb`>m;)2;^kZs>td{b7+4i9`5oY-A=m@jPPB~rVl z+$7~2TS1V!i%A8=%*113gKSNc{$xGdxymIzv%N4GAtr`x+x6xYq?Tv@#qCDtWw(>{ z`J7u8d{h=*4FOo2hk(py;E>W5)NyqV(HJN-+~%6MkqI4b>!; zMkysk(=8(w(O6v43kv!MUTU3-IcIksGr3Wl{fx@eW}gX7Q+V0V8k?W6!qoK!D!}(b z1(JHJbs|Ce49DaWx|uoh)W6h*ZV}D0utl+{85mi4kBg^zd^E{UKLE@Yt=>$akBn%l zmr{buyhj%5l((2-s1rnIsAKz~8~zt{bGv5{O z&bb|gEK-%pSo?tu*`L?K*5Muxgz=x9fORn;onFA0FF%}mLnhv1tABUc&w(xG;oXV7 z_`Wl+gKj7KNvJCRB81~NYRWf^_4v(H<(sC(CYlwS=z45I)YwF%mTwfrCR7YmalUTg z+tDzvz0P+>ei6y%B+NgvornkO+FE>fp@x-?bEYZ-cfgdcd{`VP!O-ulwcjX%fcFBs ze?4dZnh>sFkTTHcyB5^b0_sVtuAb0p)zfTcRIo8k4C)Eg(>=@|Dyt`xkY}Wx)SA_k zZh?9N^#tk()YEF!lUi9l6&Y)}>S^|~>4FWYCs0rKvw~AuJy9Js41_Fd8k%O9hFloL zw4ov)s|doDX=$n^wUl8#F)5{+bxSFuij<;ALk2cf%hHiabx=~Eq#lEk+HbdyJ8M3V z*eT~1Af8q!p70&SldNGA4|fXNsXEYJr%GA*n7T%UJSfC20hIJ-~sTS^`J+HxuIyQAM-Q=pa;-< zfYAHGPL%}c+U4kr*`?H5iUzzKxaE(DXx z3&BOkT1oVaxhD`38#sXgE5QpRKwrvAuo#Y|Cr)6&Tj1~tqlxpPj_>XuI!cFuu% zf}L|9m>%xVIStn&m|*7|=%xG9OK3@Yi3&foP{Y)SCH^BSHY^-3ZBpy0sU=jmT2d>h zC8BDEhABprS_ZYS1#$`G(qoWI-&nxkf9$L|QQ$RQ4ti;&dZ~JNRZXo$FO`U)f{kWT z5KJJL?qSAohvm6A*oscnhGiN;E2)Mq=vC833v^HI2@|$z)=b!}@`Oo@SSBTgNmYXo zSf2yQ^r$wldfc%+Mf~px^~5~!Q8WefBal!ll~C1VtYWhk2~{GG7KvgH5(*?#g-OI+ z*XStH8yc}pim6GksuA)e*XXcOw}PTo6qJnMhD_CV#RLvvz#)w1d(t-sUL7ZWu@1lk;63Y; zzA^R)#;eC3#B>7eJw)t@S*?63rF7Qb+dbFW>35e0JaI3p2Oh+90^mJF;2k7#Q;d8* z^Ogy^)ng6;<^Xfg@R)rJIAoctM;t^?1H?T<#Jw91t|BqFTjHD!8Ad9$g+ZQXPzzM| zo;hSz4r-yPBzaL-@`WW|2x68Y@0?fHm{RpcaV@1hs(m--QvdoDFlx zva&%fsA)XS&2p&j&2os|A)tkxs3+Jh2hs^5T7X)be;B;7SVFWWv2>$Wf`AsFl@|Wd z#id9ktm>(F79*APq9ZC^5cdn@5(KdT`E<(_iR$tR z*CL)F3FH-5$JJx(B&KrG#UwoqLxnN~gFE8yp{;iSbgj$RStu8!yxCg2Wm z_pF0g#L%lFcm@1SJwWdn$F7KRS4Zp$__cb#-7^hckzj7+kb@9u0J-NFxgy3}9g!;_ zE*b#tnFg*Px++(Da1J~M5QhM}30?HL%y8#i=pG%!U5lAB)d8OZx@Ja=%5of>IR{fK6=%-@8Zn$yqkGOASFjI8LN@mY1U+Nw1XH9MdwgoK(H*ar|t$ys; z4(qe*CX$)mRsDxO5q`|W1zlMn&sFg%RwplVdtSXUbx<)i@G9^s@ap2cYRQ>1r6V?# z=OZb4pcHCoPmMxBngWFah2|-=&HRuDayj~;bEhFBDK=v2Gz5)Ws`XGGR=~a!poS+v z4Ttt1;sZ$s{lPGLU~Hfh>(NLl9cH628wI`uDfLju(m7=@(js9yM%EwfJ}V8+A&XSBvLwgii`#??YI&~1 z!o{of*=3gtWj~kv@3R>&G=ppg^jMmqz!IYVd>GB=i7nsl+Yu{4d?DbCv~b+O;qP64 z*A3IRpTr2sjAm2&789&$=XN-bg{N9bnmlVB1+Ewlaw%sipsC{>`ApL79PT{fmpLgg z-Gy3`t1xd3Y)3R{DnVRpEge)o8Ce*M+HREL?xa6AWh=?Q!rJ%P&g#MJjkb?nR?VTjZo!{k`X;@A`IT~MvK_F@!xAUY0W(=Tlg1zMN%l}5rpK7<%P%(-`8KkK$ z^`lYLl)tVSD;%s+P5#5k#qgaBBpE`Ps;$ZOlcAS-M3^@a&3S6ybU7ObaOx>)$=$d4 zpyk=8`%FFtnx+|fuQv7GgjYlN)MF>Cmxf$X<{_fXGv-JwYlMunNvJPZjx_mfd&5Fs r6vwW4QoNXZ*;g6jpPev*K7X=@wBJ|Ru2IHP{6b- literal 0 HcmV?d00001 diff --git a/docs/sqlflow/doc/figures/sqlflow-arch.png b/docs/sqlflow/doc/figures/sqlflow-arch.png new file mode 100644 index 0000000000000000000000000000000000000000..0a5a20236049dee1546fa2f1357d6f3320edaf8b GIT binary patch literal 99876 zcmeFZWmuG3-!}|6bcaYYlpxX|-AFfzbax{SQUgdxry!k*l!SDr!T<_L2na|?iXdI@ z8r^m6=epl7_wgRb`{CIK8{uST&huQe*8l&jHBlOB^0-(OSV%}nxQYt*wUCe?QbLGFdXIyTgNt4q3krpbx>;HaYu%Uo)f|;w{Wv_^{{hsh9X|q%-qG(LyVpt zaiYKe`+c1rcGiEN$=Usn+X6SpiTDdAHwPEzU#|_GDvCHNtYziy;^>KZd2MGq4{;vR zUl0EG;lI!J`&lwBPOffN?(X1e;`jb|%0G|&`|;{-c2?m25T_C6{^RlgJofL$YuLJY zxPY;Cv$Ifi_ONmTC;Q{se_r7K{D?nZOOz8aT>lup-|x+@qu|XE#}eiI>x~x23LmRv zL_(58QoJwq&H<-+Iw!S(TM9Ov-7kM<4QKKH<@b`V=e+Hn}B1my?h% zi%4O_YJy8XT_LE)q%0bX9W7ovXAO5Qb=|j>EA-#g{l{ca2!yJ(5Um%%n*M{66$VzTgQ+~e`zrz8+m63f-$2Kk zvu8cDR?>-i%8X0|{rfAPc(LP{fU?o{W5MOo_RInN+0DB}ftP1={pMUmP?!hwjr}K9 zyU%qS`GxZFcs$2Zi5nU=lZg<8;J?4L5|L7&`ET`a)m~V2Z+N}hu!uWG$uGG%SwShvD3buj^B(dvCeLy79Gdj^I<&6AWF$;* z^p7veB%(YbMunbqE3PSzt;tvy&rsI_0nCKXYm81?9cB(WN&G#)z=;Ncu3TDA0x$B)NKTb zt&7C5i>$njFDq9|-ien4WCy@^2tP`g z(JH?_zk{7D4Y@Fm&1aEX8_oUBXFu_dTUjs5|8!r-8Z-7_(!ix@WB$qawzwI^Zr|Oe z^{mft9Yj#ENL)vR)~Rg1_C`^xNnD@CyC2QiILv+Ri{3izeu_PR5;>J8bt6`%)>?JF z{e0Iwh8^Wl!fhooJZS$tb2g2bpWy1yb8n>-mLxi6hK~aU-BdEpN|8eZm>t)S-!)BrV`edhZ@43A=t>ZbrUq!tvh8yy}h^)%JthsW2V; zqw14G_DtgQbf&KFx`YZ1;-p`U;`0?3X;A{Jx=fW4-a1SPqhgazNX#^R@hjtXU6foe zN>+1=sX$r>r=Q}>N8+9lv}Vn7Km4R4ay;iAKf(cXOx`FGR$J*)j0)aLPkA;f-IYkoAI(uN=l%nh;_kRxdVGXhl*AZL28u) zFMh}iz~`FW@4<%!7Pgy^x|l;YvNP>hZ7m+=NzwU)A`7@p@xJLYK>jl1p4xq)Y; z6jnP69UbXb{FPIo_~NhDu}Rrn=VJj|P+D5+3xd&Amo*>GIS+tzXsS$>+U(UqP+w z<DnN{@ASFoC4ZKuytu}*lYZWQgNlapSf6$?!+1WS3|^Z!L(fmz zc3QJ56(!t9cWM3ir1nV@m>7_W6qHb`@ATiBZ}vRc#0e|1+^y1>Z}H|d|NK=yD6n+5 zUp$S9{T=wWv$(3|(-iO9G+&@8z$#TB^--OoA1zY!`r=@$)<#{g&F}5@LWfCzEah{i zqH6r4zG1c~%?7PvC95Byp`N2u$MF(Z1hqJFF}MT*Uq-S;=G~vvJUMn2z}-qyTX2{x zNA$kC*{`!wJY^~K*nf2h0zBkeoSz5F<2}@Jkc}5vT!x78qkchFU?|JJ>w2%m5IRn zl=o)iW8-5B0sW~;qfhul4@=atS#$mOqsSOuYBhV<;ofMW)5tTut(eL0w9ovD5J!VQrYvf?lc5~9kWVx`oqhK0acp&z}(pN6LoUEhe{ zbpQPJQKn%;)^B4j#N_S9y4Ed&WTDzb%b!M_eP@w0)WXtf2d+ssw zgoh0+J#{V%U%L34>&vH0yvK?Zoj-5*ppd4lKo>vNpPc*gnt&b)BO_3Vqa;%{8zAMj z{Tjyneo~bPrGz;z*H<$sj4Rqq__@w``mL(WSD&J)dO{0(=jnV{B&-Jgx<_$T zLIujMrkM~leF2v*F+o@6Dct-Dp98Ne4pnE*m>=}DeI|YNobZcDa%pWfn=w7X5BC{m zf~0&7n)a#ccqSQiUMo{t<)rXES8ma{0ns^op(_1v*Y)l>I-iS!sm(zv&dUE{eLHgT z9b4bCo5WO0SR}VUz1GI2s4(w)Ce_3B7Xe%u{ZDOYS$&_4`}$0aY+c)!TfMrVK43Q z=Zm&+H4#+`en*z^5>;%qZU(wm`w04*nFD1`U5CY`)0BE^B=n0L>uy+YLw!z+2gK7D zsS^X(6e+rU`u1FD4_wPuTkW$jlib5qVE@OpA}$vdTrQCrLxv7ye@Km;Ed#>rhqjFt30K+s;-|aLka2V7(COT9$$?4ZS-Z?-1?maHkT&7bs z=_7IZGkI}z-^hEO34pIO@qn{;l}0Tq;grq^pif&XP}-dgf+BTyYpUuw0F6Tc3)6vd zPenn+`sSN2{dA?O<0=W8vsI%+C8NM<_e)y~JX=I(sL^ixw(92j_YR$gvdU(q1R4_V zls$}KZW22UxzoXm(}Pc!(VXoGZBLGyjU(`=$J!yX3cGVnM#tM~2}j_>H3uakzeflU zj8M)EpBM@v4QFkajZhjV~McEV|{@$uO#%fmS?>;8!53_XU* z*>b(w2`5lKHzzBNL@verPh16JG(!aY+aoMexZ_THwx{c2yL)j%lXZ2g?xA2Zx}Cdno`K^;h|T)y`ETS&M-V;4 zU(V-?e@)9V-fG-7_4E_1Ibk}WDr}^eg955Y+@r_U)1ol3BvzY{`yUbzJ-P7WgC50E zg=owsJcCz$4~^Dkv6`fx<(5bI(_xe$IRrT{k~5#*3m8`wXp$&;z3O zHbbrD=ep6W6h*|eg5X^{vklnR_V<4^SSb67)m7M5weNv-scv;a>-E*Sb4|L~{&wx? zrWSN6k(rLTlZdsDWjxT_#u0%g0lN_C;vwDZb7Q#s&7FWPGsGVNcwv#3fO%^b%HcMx z=RD#0i61soYZH&?#aj8&s76VieK&P1btl8PeiKC49wt~T;=KJB@kr;oKCNWJ{$R~4 z@mk<~Cjb>Lq<-#=S!n<3M-}?v)6Cf(NuPfPXX6$cq`cNHv)&tsA6J(!%6s@G$bTuk z_|BuNdX4gW!E#DRc}voZ>ONX4lWHsy196iMjnV3dHpjd3V0~&yD>` zAH05e$Md+svJZayV`IN=DdiM=Tcbh8*jIM`4ORaIo!8~gd%%?Vva033lFdT>c3Y^> zl)c_SdN>B8S^@dc5N{-xd8HQz%AaO!w;P#7X18^w+d|3lZP|=iE$LYd8#%i;uex zGR7@v4?Jk2hl!^dO0Bca7ZacL)lj^(|1`GpQav|^j&l2z+pJSVRP{!lNSSE_sRb=q zn&cE*>CYNYg>I5b_w7@fWbiw^pb)Omt1Y3>f0}zryp4AHb6mZz{MtQeme^gU?xj{o zU;qsHAzZwR2b)Sa&SE?0-Md7Fc~{Z*Lda|pT@UX&P~GH}c=H4w((e{KrbNnN$O5Y4R#Yeh0*S!EH`zinggsw#Z0upmzx@bLL&>yLiQ=IG%e&W3!C*s{8G3`!gcSw1wYV z&=TRIn5d{R4z)O#wR;)daz2phf$rtL4`1kjk_2(WY@wbmAjOIppn0OQ^Q05EUnTz} zxD)>wI*ZJn98E~W;rI6QN`$6Zmhxs&XJyWB{b7u_f<;y*jO>a0uGRgMCP;O`e zLZ|H~3eW(G6=n)GXtPzVmq)YYW4kl8FL7u@b^VTa+0&cMy^23deBqkJDtwXPceoif zbMH1QuP&cdU?b#1xCXy!j;P)&Di%Km&Tcn;tqh+n^ZsqWi3)Nq$D*-a{mmY@Teik!AKtcbpR;5wwB%ufUO*kFHIQVZ{-UMA= z;q10c!1sx42I8pO)r5p2R0%9nw8y!Ep4|&7bp4Ted$7T4N1s_Fs3Y;pi!UfWo86!u zf|ZmMSTcHBwP}4cH~y`HS|YtHRjs>5xSJTjE}J$u>%xF{x?jGzS(9<+YiED-&1o3- zGw!#F7EU2#GHr1#9elh>2t}yWh=P*y0X>_mz3vDMH$h4lMUPTSP# zE0g=YJm+NRx_sx!9d*+KqO$|$2N3Jmb~2p{m-PA0?Mew@q>Qq>0%8JoO3se`ugp=b zD7Z`|!`&*3+w@bRdzoBjAzL=MoApk!k5Xxt2jX=?U-v(XwWxRY!Y!oFnZfPnt_`B9 zH5}xw4drW#7e9NadiNvq|Kf@shOFRpY%~ceLkFv?XR!xYS{M?B3jihAjG&GvH${t_ zA)30QSrB$czWiDZEYBt_Y%0?w38e&qRGxt;@$fu40YC z%;>(9ROHy$#HgrwrGrB@7)DJ{t7@;6dRH6k#yR`@$&$pH*~Z6(pHH}yt!q@x57vhD z7c93qjqR})&?MVp_O3RupRO|!h@mJR8^+by8v}`>`NyYzY3u!|hl$cPzq1cOgZJAq4N> z2G>apDD}N1g7M?M0Ji%m8{}*{evdNW{)hJbrw*+k05E2eL3C~VzcgLMuRzIv1yLLi zw|V>D7{@$oj?AihUpgLSkV zB2>W(q|L3hc;(fZfv+JM89re^NT;z=c*IZ*0XS|m!^xXiIjIT3Xe2KXoj+)8BaPu) zxB0El#nzBGY0w&++Y=L9MYGdD(tqJSKuVGVQaP6?Hwz|&VGpQ#%Gtu6MW$ULZhvwB zxtV&WOg`HWlHzCI7<|7*F&6>wz~lGR_E5eRs7dysjmO^?Oz>#M%L5zc+@cLVKFYdn zPjd@YARz5{gV`d?6~Lk?E=qpK1RxnG{rK8fED1x=YVmT)5PV#4HWOdfex{^fXGj0U z$E!>vE{&>;#cVo*k>fN~zPb3zgyT`a?vS{?aK8kg+nHIX(r~lh=ZCo`?;ycTjeKdf zD&w}F-j`>G)P)zWXGc3*q8EETc>B^bjwe<_X+i+1KJZ?#Bf0$o0oVBgy{xoHSps_t z*b8)9Ba@}u;8MUQh{T0ORt;~{@6+T^XLRnC__p~~VH^Iy@Sho=54J9N6aRE1Hh|6J zQ9GU3++QWc+DG+Ytt5b1OGTKJ7T1G@#gh zZW8!-f93shdIk+13HhTuxMU*`?1T|a=D@5cyi|aMS;IDc=Jtq1yjY1d_AQPHuJ0Y! z?X@lo;$q+Dnrc-?X6o#c(=2);2{$KSX#%0BSZO3T=xQUJoX0Y<-=N*UX`!YWxzL9w zx<1nU2fZ5BArHdNf$~G>+xcsw-#zj1vP5agTsz!s<8bX`*t4RIMtTv|Elw&da8&(4 z?=<|UD*~{jmzLlnFKmUa(?EkWjp|VGY2dwrK6|F&MZ-0>0)ne_tGl(?>YV34<65K* zy}Aa{{d)6~B^dq9z~g!EG7;-kK9?_sqv&{4-_n_Q>d4SK%&n2(}<$R#@n!4-SERDNFiGKbZ= z=^2!!S9E$JFfv1v9x#!_+o~^KsS?(v$eQB%UtOGX#C z2p_VDv|sEi+s!prvr6ZE;8HeQ0gO~bHfp#Czy=%Msu_( z4TW$gTfAL>F0Pu^1BX<dECZ17k(` zda6W`H()_)%0J-55S830KQS&;!i~VaE_Cv-exfv>%m{Z}9%Z)l_Ui6<^>j+!`KlLM zGP(Tpma)e1^$_2Q}OMeb*1C zOuWUl&-Hy}Ie)xuzyK&a+VxpQp?|>} zp(fGG{#CI#&a}L^-(^St*!2b|ILg~WOEUw~>-}Jb-d#wb0s6Nu3h$>ny2U5idviA+ zuC>wii(LO0 zc}T{mPd}VkD@JdLyR9gAKG`gNX(*7zreF6?XgP*=0^iPDAOjFc;dCjKF193_9N&Hf zJPC)5^-beEPUA3oeKnJ+Emn9!_`fSNfEVQ~Ci?Z(P&BHUXNZr-@gf%z1d}paH@$GO zu>uYaip5Lo^Vx(080rz8bV-)apx*j0@Q>aniq8s$c-KgvHoDrzXBS{*++R zt`*bgM+uQrICq<(Y|9%JPC?|Vs~>pA2P@c1pX&GaXhT=F?z3qM6Ct3A{`~(0RbpCt zifmeHC;uB7@j{@HmfJQ@{~H<+L*Rg%oe};2f&(N!69P1{^CIi!f5AYopNxQ^_|=Oz z{08~|8M3jH2LPhZ!R=A{e~~fAk^x@XnAt7A_kRE&V8~M=vr3IszSlGRFCw#oIs$h6 ze|V!nsB937lOe#o9n3FjfLP-SP%0GWEZ6SGe;Gt|vj4C4&<`c$8X7uc!HuHil_`@T zTp`gL{o$P;TT!f%!Bef6RSIyQPnskU0FOYoKUwP&YX!*a39(7ix932@I8$Lz1d=e3 zeGD-kGUsu%t6-6ZE*O#OPvfm4_Z%iufL~i8^t4xp7FZ_;As|6g5VF5MdQJR&X^9|O zRkuzGhJt!UnK?n6$$GA9emXlhGr)!!IPfg^`MGX+x-XLvt5oY)Qn~y)2uH_-TZVXm z-@7GpE`G?Dz;majug~t9CksAc0_chjX4K-PnhgXy@zFh7F_6HX6ShJ@C!W$sDqs3@7NNq1Q?B@nqy}b9(6h|IqiIBxS?XIIfaH0^Se&bff zhsFp2`-yMyK$LvQW9O$y*C$R|F+yY}ylp`kVo{fXStLb^ogpiXwxW{Y;dsbb#e3xu zPsS|`0u@ZU*a}gc$XpWV+qG|TTVEkXajLG#q1Y*!=AIq|7}Q!1!c^G-iZF;Dq9$A0Gxd<6jQP+yoT z`D6G4yKW350?6;#r{v-OwiW4M0zkSFjy&GO*pMZ|J>C=c+@xaFulqKCJ4%2-<+gSK zFlzFOfv1clHe91pE1qUuF|;X{jFi@+A*!?(EAYh?5z<-jr*!eK9)7Xv_@#A5D3Gd! z^L7=P2nAl;LxMMu2X9a<-k`ZxMK>Wo#x>&m-mMZ1sbT)m#}+%|l=Aw?gH@Uww&RY| zgLM(PsFhaWCC`w)Kc>O!H+6bJKhHOPXKFypGYcdqF2uvPZcedRnkd;qd}xQRkY9|5 z8`jda9zK4Gir+_YM;VDy{br#3HU2TM2vzA;gJlmj%(8% za~CRit4B)&CoQo)LAdA?Jz!ZehDZh_Z#-BNw2v!;|7aItDOK;pBRHW_x~#9eXX~K$ z=B*ix`=!#}O%sPPJ<{$Z0t5HmfAU2*8MrGU zm_!QV1>XZK`ais&BDCTC#lqS*%Q)&aTgN!sVJSs1l%sf_fz#3{ulin_^yI!Z!S_@y znald_AM86|l5{Xzg2iEOL0ML*5MJ|0K?Ftpb;C*t7+u^Hpa2zMuJEGzj8av=t7oWU z;niJKDHHUdJJ`R~SsTl-r%FpNED#dslNWKn_$?fyV5~gT5>^TSM{|I_1w;Qx-re6G z?*jRJnrihoFTzJQbZOjmy)?UxQ7O+Vi37pHypL!fAB*1C{o{Rmi%G@T=VB381=k#= z;u_Q)`}NjM6LorqZ{bc+S9aVThHWJx1RX0BLp)+CmghYXn_hy)jGe1ozXlu}q-$J> zdelo&?EOB8sY=Y_o5owY0zM0a|U{vhj;`ovro!^^wrYp}96d1QOpHeeQi82OM1|Fhzck6V=?`{E6odiA=RBv2&>S}a8ErjX%S;jx@R}%QWI&E1PO8?fN!tJg30RxA5W02BodPzzTmD zTqNsWo3upKNn6|GzOGbl-uDxK7Q}|w4DAUyWW$+ZK^Ti?U7iqBjLN%)CJ%!zdLnER z2|M@apJLx4<+6CeeH(RJpwLnV;2+1%4;jB^IMfN_XE-i_3PUlNLnq8g=Yk*_10&%u z)HMz`)ZUn?%DoQu04*0o326$ylEF~DvC=g_Qvpf=84*2DVXN+b*!Jb4!f2VD2IkjK zX;y8&C%!bjsTZuN#Gt|1--3vU9Zksw<(9Eee(U*eOA?4VAku6rFmxIP8>(^<%=`C| zrf@514dokiT#dj>zkrxdQ;W>W@~T!9%Q8hy>8Bhb(u&|3qC0_`0M?7;WJ=|B?D5&WJ0G$Po%(u(x~Rz*6GTQ zxG@rXVPnjn<#W{6-+)FK{hs+7pgDR%iEMrahnqp3W@;0%5BE5&gU)7d&L9RoNlhfPn=vg-O4}`mbs>*#Keo;f{l#)CP!@fo!bg3jr67=SE%`z$%qL2->P> zSh>FXeqAK}6svg|sYGAKZ29g?@g+#w*->;{ot61t{21_SxBW;Tdj`B;B_M1S0htF8 z2|}zo@ILjck`CtoA=r6FY3zJ{>-oBYGMM>q-jQ%0K{pEeO*MZHQc`^9+Hj^ea!4^3 zi1MWXnb3-51;A`8u-IyT3H}R#=M~Vnn`~uGTa_&krI3Fa>7SuIwjZDyvkg_IVw~N; zCO4>T*%kI(juqg8gJdJ8`}#=gFQZ(XgcZ03;}`kR42huj^lax9pj8#V(v1@!Tsk+&1tW;ofd4+0jhz?Sv_E*aPd97o!T{)vAo4U;C*tu9ioiM`$@%W` zc%kaCMw+_Bvj#%*;l zj{EED$9rb=k08t$W_s7wv1t$z3 zJ$&`*L+l^(8wdG-*~CV)b*~#6$W92u5Hqd+K7G0#j2!w1gj?wl(k_Qy7fw7#tAAsA zhVY54M{_FBy&Sq6@!|9^Gi?m7VqXxQWdb1;{yUfbczi(4^_=HuKQYgCLW$5A_hPsX z04*>VV=#)Gu3jab{6V$3Y_?qpIb-?^Pr#W+_NR~;X4lXq-32|CjPcc9vM*T7;7fr7 znLmwD);UG&^=ykv5e$kD;xho?A!(oPvmjA8+NtsA7Fll5I*y9g~KnuX`OkncV zrLRIaoo_d6Hhf2)JKCM!+G6auzIqbAaDvc%|7Rc3Ulrg-hH%6_1<<_lLZook1UF6918chY-Pjy%mp_j0Zm)IKEcEq%e0Z z!Ptb&+ySEgEtbaskuU*0606z+NAB-^7e5f;_zPX$mwe1zmhT8Ii4TBqb|)p9sJaKK^h(hmRcaDsWrsQT?mA&ce~vd zJCP9)rhD5m_+pGM?WZNJK-L|5>!8!|W9bUus6k)>vw*C5;UYyF_=|RRh?7@!1qTUR zqlxLQbcCrI1Ji`H;Ttw2D!sB}z-RG}sn!8Gujo17_3RKPsdU^>zpf*6uiP$Y!RfeL zwUS&ts3J2eFZdJ@O4K1}z1e(@(_CX;d5oVCVH^rYI6dc;c+PVzz7$3bk+(jEcRa?O z@ejBJ!aIY;v{~sSP*M|#?mLMIszsJJaeaZ+Oy+hP6P}@y%;d+^=FJ`Ls9%j#TM33O zvjz2T$f3Mf1)UpQ?6y4SCX;N}croX_5U?5e^PA==2%13naZMK%dx_OkK&q)f zwucyyoNL0)ys2E%C;E;zu_?&_%_gZ0edg?R4nk(H_isE4129qtp$|~spXJ+&CM$#y ztLvjppLODp%-wZv-ngwj3B3fp%PxCVc#CZqPuawD@utCw^xnW*#pK(5>RRZ5MYu4H zT=zyf*EWb>xY?Es3;BW zjlJ5L$?i+unJSKfg)uM<^e*6eYRlw3QF<5Up{L0OX8sKeb%E1swP1c9y*!U)=Nx{z zi3s?EVp<@BWZ33GtPYAoi`v{77znRx#szo=mc-D7zwmd(q~{hPTaT>35bl{^F#>Na{(7)?m^{jZ9L#T?98w^T^=GWaNfjU#v^Go$Q%))o1^?PX#!3Y^8d2wuz6TN1OKyo-c>(Wt{dy zB}^zp0n4B#IPuhWOHD1NLslnf1YL)B8uLwR_t~{BX3U4KS9Fk2>1hZokxK!z`NK@w zKUy&XE=rWoNd6pRqkuQl83-PFDsO?k3fh0Thp}>>z9z}hpZHp31$%KCHoB&Hbx$*= zeZ3i2ZP{C=2+ctFu__f^HASc}0RNup6=oGzf?Z1^UrE^`vPHa`<#3%|a8+Vilyb!( z2sWbqz-$Q>dfoA?A~YA^V<3N$!DS|Gyji^w?I& z=c!4Z4;M9&f*0U~j8Cwu{c{+&OgoWEO*#V8{LE(SbrnIT*+W?^Q``~O84?K1gz&Q= zrY8bK0R<9qhyEZ_I=-Xy7*FKh&AX30yr07iUMu{*ujeWN1`du*O!u*-DG^ld4WY#0 zF)nI?bNSSn>9G^-F$L?-S2>R{?!`L_>2Jx>Cg807D%-u_2TXWYXFa_yB_7w@le0rVjwcv;Zi zR&WJ_4wA)Zmjpt7iV6JxXj2N%GXM~y#2=I>pde$e46a_+42_7HnrDb!7B7vmKj6Fi zeR3%572&(xrsV9FVZn5XZd5GdccW{GSPL_s7(|X;H&-}ugPfO>6*+d}8q4o#Vc1 zo#$_nc83@tRl~@Lxj0vT%W%TT?&@D1F31+r3$t1m`o#h-8V^A269i4g6Yl3SKt}9^ z2E0NlVGo5K6OcHI<8FPmZvEu`I;fOh#=fhmICh8_?Fi~E!^P%|iqd;o+-lok(ZHL4 zoR$rGj@%{XJ^Pgub_3FY0q*|fY>*}W5PJ+lXSPCEq-@^+ZJo{#^{v^mv5ptSrQmb* z+MZql)?eQp?bj)_0q0&*xqpMo4cTCekDUy%*HuL~2J$)o=;J;)k_+tteJo|-hXmOWIFXO`>3~v~-+@L`Xr1zK+y&QxQu}icvy%omZn92A8 zxpx&X4BX5Qz?EM4_OXx$v`hq_=pRrcjThaIyZ}*LbL&Kf>#IvE5X=XHGs*l{cfg`} z%zH5q+l;22>LPH+=@D^ML|Vx->(d|*6jni!o_IR}A?5vsgy8$0dax3~wc$6V-=MG` z!X8Mypg-97> z1)goZj;Ti8nkZ%9D!?j^NR6QJi^L+p4*eszg;+-kxM1$rAj{bNt_eSxg2~AeKyR=L zb{$h>#*9-a0@!bVEz<=nV0mzBp;!C&VW*i2WwnVj;ym?Gf2VAzYI=7$d(Z%da{B?&1oliC=TE zjTE74kUeJmv!5F^Af%jdAPf*zJ9~fTiVtSVtSR3S*F$qaQ!7lmEcd4Y@`1}eIj4fpKqCaqv? zMTO&B<45iEiAo~_kZCTPI-mDmsAw6H%s>`e&vO4c%lg1n6oFIB>xWa7IhyX%1fH02 z8ypM*cTE(^eR#b5t#WJVn2dBV1Lr1LTrBlWe)g5IrKV6DqgKW*qn|Xk`W{Qa zD!PbJMh*P-7%ib02t-Mx;x?7j<3tJzfzq}ts6sI@!9kI3M^n1{CnHfAS zJR zaMsil9sr|T^%@ik@1E%}u#>rX0~&B3Kg8{lkPY@b4*6|)vzWnY?+#p)x-Vxme9->_ zN}@2Z9K7az_wkYI9Kwz%<&<$?K@_8-GeA$Fu7&r_bOd$KF2ty<;ns)ud1hrV3iH>? z{tPs=_O`sm7A^91iN*~W=xl)=pxqO34b@Tcp1<$9`~tu429IF~h|Ac?;QiSlIhugL zIWOee{ZFYZEaXRTaTvbp!}*ydgH(v)5u<6o4ieh#a2(bFHi_}|xD27x{1>qOW86p& zIGjUufQ-cpPNY6?{FxU^K49&$;FF{6k2dw({M(?acMC^c3_q40>Br39$DXK!_3MER z3dQ4K)9yNGU=!8pnvYtRGhF-iVEw-JwUDjC#z`*g6VSFCN!Fv7b3SbK!S8#_`$(fK z5lMC^!4sY}`8}0v42WA2Q_BN1TI7OxFFd4&|=|8V(XL1Pt={uUt3&Kg3qJ`s=W40h>t~o%*R$~%+TO=B?6qt z#CScCilDE2bb*$1NF&Bg$R&zmHFb2<3earKMe?WTWH}ND8Ot5)7HAwLuqe{`2ze4- zZrC)&&0F~MH1Za*;5DFoKS!T>K4wXooBc1%8yhwSgUX=RwId)GWX~E1~wWG|NTETL-UYHWm&p@z5mP%Y2K0 zxX-jPu}n_O8UuP={k{rp)AvHSGE$*nt6^grdcfE;OojmHpZAfDg+M9n#;XK ze(znwDcl>17D1A-oJAC*O;D`~h znOhnw^xY%OkbNLJtPXO-F^oYLy5U1eSaX<1%8ch0;g9VlV~48@nAAUaC}{bHHiuJ zkHnJ>u}bA3aVvlLsm^niCC5s|yyBQe5M%OcXVWF?=~2bQ(j*IY zJ9-f=a1`L5{21)ltJ>8?LxvGSMjJr{N_QR|pQd~_G0-(@uY;+z`s*$G7Hb-td)y(tKZSy0{;FWOw=DRI+Q05lU-K`V_GbWM;SC{EKsBMm619R|Zq1W?ZB5 z)mSA^x)2*nDEi?O1ci?8YeG@$SHseg^TLSM5Zh|@02NtTjRC#*UqE-2m_?&HG(FvssXk^zowdDegbsZ z+Qk!qB#HohO%n0mWearel%YNUJetAF9NnQ^*J-gnlC5`}+NxYnSH*#17vx;S>{44m4^2H?1+tnIitI9YZRJ7tIXSp!9m-`ZAYQCNH57 zfIhI9qyO^KJQ@7~6ReINi3#1>4B;H$Q3)G&B2JXyHq(ny9 zWH(-7UV^ptD7PP4P8Rxxe7|Ao3|WbrNq;xoppBOKgq4Vd@wi2 z!&{*~N2kv&oi3NqpGHAeEVT_{NPA-Gj5ES(ma^w{o+~v#lzlOsK+o2sVdC5ZR*&gNDiy)a5dM z>hhQT`1bR-X#ESq-n->yZwggjnGxXmo;xQxOj%kqAUz90v88p_(25AY4llZkS#$^4 zfq3O*Zs4FKQcIFvCe+6m^~ng7XIrm}BQisKDcuSpmFpg9fm_(EHUOv&bs^42qX{^> ztoMDHbd7_7M~9J94K5}&Us`=ffSWeq+H~5p;hgLM!v97XjwWX4gtcLO(MCzvPcbgu zKkmMEY`8D1hVj${Oe&Wwt8=f}dM8h~b(*-HNMWvdw7UZ_B;fW3!MiDsjadRFK0~Sv zR=T0A+XH6XR`V#w{AWqQK`7sPK+r(47jkjT)?C64lf<`Kptdmf{~J`{!hAO9QMmE&nQy<9Lx))hWJ1UkVmyl{@|i3 z6j6h=M2Huzh4`cx9MXLm%w#>x5S~00YsL7`YzG!S%!_%Bwdf>zNDV)|kpeUpb6UdZ zDh1ju(jT^u(}l+g{L;zwgbcdUhd0F536PWsSk&@dyK4LA^I{Sto6PGwsVK-wc-jorN@bV?um+pxe zqI_Fe%d?b$E}M@!q7*mro*BMiFgI6JBJ5upgX_s&2RDR_xC%BdBFGC|1m(T3_j7sawJO#;Ts5fZR;vcG8tjV z9!a<4Uw(+QRj<4IOXKLKbCni^jq{aTFv@E$G`Qr(P>Vc5csqzsl>jo~ATTi8Q*JYW z9Hk{=!d)GXJ4VfQ6XkjjfS1Go_6?-b%RaNjA$e&MG&-?#c~7miFhMo$`++9Qjbtbc z2O{Y8E;v*|0lEoEvYR6~9OUk*V@Alh90>?>*Ch*M<7j{mp#y z{{PNA4|5M5PF?%jd#ztB2+!%HP3A$qfVemClK1>)Ai2U2?y2)oxFkXtk7hU_4e7qh zOk${BJDyFJo{7IMfLQF>xhnBs9iP3&-cBUa+TL+4t`9k#&3;7F3Mhf)fvep5V-o1w z-te)VhruSHL`->Vv-WCmi2jVLN7HLckIQJQTgyo4m(sCvQWGRe;gL zxQIBU{dXp)p1pglJc4z!RUDX)I%m@8_sK~_Hz*y0Ug*vQVH6sb#*9vd)DoroHXq=s zze3TKjyfk7_4ooF)>#ek!#)ATwsMVM&#-kqM5FQL0`>HOJVSNl?R$y{R z89)6-D;m6Z#T>Ae%tIiTUCU3mdL)0Y)Fiz_yf66-%j)K7ak(IiyKe-J_ZHHTL=dXTVRg=~c& zPqZ((M{->wh>16b4jp_2HF+M0H^NU;f@rX?(}X5IJ_nkub66mb%&EJ4mGqJ8@pp9w zSr8|`F8v6K9{}4t{>{*FXtyo_wz_nppY0n~Gvjj5Ho7j|{|g7b(3=N>KhA=10q$;` zccnx;6&=S?Ii{e+)5EOC_X{*2?Zo$;X&efPL${#{dw@V+39QPbCeQgVVAh06eZpK0 z_Pk&1GEUWfYnSl-oj`=l_f#SOfPgUMD)kG;9W$*jPo*2kRL(n8=KMlq|@V-SE*dZ=rAv z1(Js}Alw+l0OjOa(M61~hQ?w} zmeXG}zXPq|)o0y9-0@TSf@;mX*aywM_?|8&tk-b7)_(bV5zyT^MlwkRfFzJ5VkJ8p zJVG)zkgwfyM+T3K|1{3OAXN<%=%A7Sdo}jjAFH%~kss{6Sb8NQ;iZb z4&u~lKHq3O41m!NZ&4u@?vL3oXIK<=dw$%Nj!~JgHjM=EEOq23eshkf^djh_K zIO1%3GsEh)KiSi4D12i?@*jXa?F3MW3u-yCCF%_@rPW?dyJQt{^w^7m)IC9I25VE^ zr5u$ujo&y;I619$yzu`pYp+$H8k#7(y=7+tIw@rh&*wz)O}t{H*uO<{4J69H}w?$P9YlqAr>dlfMHqO zh1@v>a5QZp@^`!XGQe3x8xge`zeN4ZmTtO4%s}tcjQGj$LG()!Pg>m7NdU1JZlVm_ zcn`qlWX&%6_8OolnmE<6hq(NdxU*EN8zEFRKk;iqc!fiSj;}Lh!6?ODCyUYCdS8U0xFb5Ps9> zJ8(k73ruscrXE27I)AOnlQKe;5dSg2Ki3ZH(jF(?uAXJeGL!Z`yEItuX!s{xUy||krcwIE~ z^)o-XQhVXKPz_xQJo1xB*FcOomO#IR&r;7NJ7=$fEgle>pkEbN8 zjW?HEa}Q5ps^GP$Hn@Cl)x5EYYZX>(sz_)Fu;gC!C{r?PNa0cm5%187Egu&joF!bg6O2Sjqr68eQ?~4y^_;ve1@Y>C7wfj>V8wWd0zlij&jfk`i zwcSs$WMsQzehR0)<*=-qN10#kn>t(uGq?gUj?Y+Bkq*I%tf0Mk^swhYY>U%^zA~_HNa{aD{hnB_{ zZgpNL5lgzk=UFm?vukwgy?`4y7rKGyvw~IP25=}|BFiY=TVm?0-ThmO29ox7FVwjw z3^L4A0=KtVHiHD)lv#K_WJT`p!p*Q`${~Pr;ep7l zxxQ_O9k{oic>+hvto`-QV~f|jjU}KIj-f@jm3PF@abA1*Ty7A8ah$W}BB^)eBc29b zjB2ycqaV?3A-MMrRYc9ps-NKV1Vgg%@M9TD-31$NLH|#CKaL%j=F&lzrOl-W)l>MG zaI!2%_6VGHONNGsNlFUo-*EEDMNwxcTbJCy5v$rI6IUsz$OS&B45ZdXcZDu7h`PIk z&O{Tpnx&rzwQNMaSc!vRrcU?>#o-_Lh$wwz<`K)uPk=TCwRRk`Z1k=8(}9c&2`9u!SMj=fE}85&M@>!mv>P(%{ilz5w9gqLN$J&UxPK zbWoui8?zWrLio@}ubOK(Jnerr{o0qR9=R8KBMU?$RF_M{zH(6$m_8DwwruyTzK#>h ztp<2B4-zRELow;QD&gPxU}CpTs6j{cB;8WSNtsP*12aP_uN4^yQ(VfrP1e1unwMh8 z(uIErBoY{pu;R`rCq5D*3c-aPg((ejQ>axy`e)wa)l>AsOnl~h(B01`d5$Cu(Q>tU zJ(KizAMIrD8qWwlvosV^xOk~)y*-q2o0$(I7>7aG7?SK#q(fs4he{`@oIOfBo)$Nw zX)9KN^z+Iq1#`SUlH#=62xRwWa9A~ZXs&=e^l)1>_WL?>5T+AsdRxsRzqhvLI;^$3 zbb=zL?&N{ZrIwYTnY{###p=FTiS1;`{fm^%BN)7wN}|sHD5otKUps&1qkEJP zi`yxP#PMWDFA?f4p7OQO$VGFMyM0ySFsOPNBSEqPUD;5DnTmf$9J8Ifz%3dF(fosG zsU?l`W_4y!884%&K$(=0KO(u8%95-X>6zIlQB9$=W5&e0!hIH$Chnf`MRJ| z(^SmtYO?_bA?g4Rhbg%#f=?H0?CBLHpGzgEmn6~3X>96mJ4Iom^`g9RtL5G43&kyC zOo59U@l;iPj^K-TGP3SJwH+-a;~j~Gt&}-sgT-Ypy4M8idVNR27>*3GAkAi%;;%j0 zAE`H17sGdkqpm0X)En9PX<0 zet~1GxcR0Lc#6nAuqUG4+!`lQ@}UmLYGsCQ*i@h9m)S{TpQ7n9 z2w+z@pt;y~+mtYAjpU9iKlNjLc}QUF=5~3FN%_6R@l(b9TQToqersvsu}a1VY58{a zh!65Nh0d%|4LwsOxbwRE6txIQAFa1o-y=#7-rH49$sQ>&;r%UVl<$42a^%url0Kbi zQ%Bj!uZD4INk8=Q+s6;RMm^8pKCW-0fJG;@Xf9Qf>>(r?!EJG)!p)@}D$wP+*0v^F zOTCF?r{vvi<-4#{NiP*JPUfNP&dN94lB!rA=3p1~!+b*6Vf5H-7=)D&iIFsL1v7-| zLUH{DzIDhR|EYF7Nxn?KF4rrG!Iwm!3yM2V+EX!8e?ZKiBjD{BEQ1*Bs%>7AH!S6n zU0G~&0P!q~c2mG%GhXg1Ucwq8po5RSaEpFJH>~pu6VXb!{PL97wl039Rjw=Xq#u(7 z6kYxd1a=x_Qe3;DiA(>{w(v;^br?>hGpX6s_Zb86Z%M>h<|3ZEpxnnA8!)eayo7aG z8a(q^b#111sk(mQ0&F#w7N;_E=bb?G_ZsF~lMw$R!irqB5HQ7jJ26#qvb~hP1B;hO z=8h6ZQkNMPWxpPzK+)Ophj(LQvdg3=hcY*TrcNLgF00A+u=4`Ry{ad;uL}Rq$-M2j*Lq_`g>8k{E?dfUCSMx((UKY~^ zye3Tsk(CB!LeLRLPCppzzXDP;T%~|Moy_`2mSYL7IJB@@2c`NuM9(Q=eiMzd1^r2! zuxD^y+HEB-qhM96KqO?jGJW|YuUIv>Af+XDr$d+O4{JgDyRRPu={+Bc11lK7-y-w= zl+s;G5^Cu%B9c$aiiY-2Al?-zqF+dVdh;+uXtP$bzX*2z8e4>DCMa=MI=3L>uprsfU8{?F&@_Av&aGYX05LhoE=|5 z>yreH<}Oa}4epj!%X_8)rc8G)H9>*=xWW8r9|-XWP;0e|Gu|Xg#uz$)UGB0A(;rM0 zH?0Y=@2uERs`${)y*eD!Gnp^<0w;QzWhbeG3G|tylMwQ@3ArkWcqw16IPJrwLy-NH z^|T{x6Ho7Y8&W$$ma%6zA0&q7<@Q5PN4{oOxckzGCJ6JNG7|=1(N0<7InbBb<8IV} z`U1%y+@LyV=L9-HBC8nEFT%|7yVT%XV!}W`-{$)d^?)h!@QSj2=6?;k{$xi%2qk(h zzR|%VYH=Lv0KNqpA*Uzf-ZRm(c}fpCwd%cHC5jbTh2i zDIZL1^qGE~g`*p4TCgw2Lww?*59D0x<_CQje~afmQQ6Rxmw~O1 zVQRMf(Z}0uv#B$8l1+)kG-JqG<=)}~-pWi0TMKhY#7(~O0Ytj3d%jtLb35bX1&vFl zf+r`B5JQ_e<1Uz0nhwB~l!P>w2nIU;j-;OAHdueQVh=&j?GfGywx1UfIN4*6Q*#{8 z3G(+n#~-2~Nt#70!~Jhb8D_CVZIfw2E@i%?CYSC+A)fBp+Qp&=aCa}9R-#U2Y?P0t z&AEJyPwXo!pCQbr#Zn*tNKJ#V%1E0dLjkLsD?cc#o=gC#(Rt-`jkHeh)2@hGf>8a5 z3?J8ta!J#2!jFNtOBYSQVZUXkX1wk*9efgrJM4W1UX#I6!nau7MbPmpBAK?{ohX?C z5_YsHE?v29hD{4$GkkOG<}bmfAyj~Y2q9wv`NV?$s#8Un!G|J}@hG$=`D{cED_^4k zcyTUp_`_ysM?x5i`vOCniBOmcT-bxF?-=i$l(Bsn&H58u&Wo5iZy{wfqx(82wie}D zE)Y&tn3+|zrAQ7{l6+@rr;jj zX%R;+RRpxYEQ`PMc!`eY;~^yD&PS1md5+Y2yl`TPeZ+GOpPW7YGJ47RO^d^6cENPn zmJy6{)e_4#{8rPUj+KeJuZSSe<*m>)%rdW4V;ZRd@#@NIjh}z$Ru&W_y|Va;L`Jc4 zUT2HU&(Yx7jd&UmVa3^~1umG*RH0(9Tec04N>q32r1@enhnYkuJp;r zyu(>@GQ_LbRltYkl@!J2+1z&{P+2GAO$n0H5`6ZJo?kmLbzSo}%C{ViA*{kAQFN^| zQ_bP`jTQ*#%#$&X2OA)8>=Q)qoriUcgz*F(q4`?)x$?eD(0-*{ntZa0kHe8XuUTfP zXkxun55~0H#91#qdp_SzAyZ)iS#+AHnPXuwI3qWz=B+<49(nI$8JrU=Mw|JaG)Nu{ zsWngG%I*5HtOZa~{JFm;wgM9tiq9}2hp(g#bBzAl6(6U4$p?<&eltFefxiMOiM}ny z+?u+PdX%*OJm6i_WUTBKn=}Pr4~Y&G%^D9EUIo1F#A;Q+CzAzY^a034U;FcPX8b{e*Pje4mFeFLZ4I|s zn3Ej8v92|2Uu9dMDB>daO&ilFQ|MNGB4}P*4>!k-z^7K&p?GH#mZAIRHMAT`raFUq ze3NME3-5`n{cgeDuy~1KL^POL&%S3u$T96tkO#U37#h4s*9luK*dD=oT|h*V>IZJJ zEyt5l>8*?gqb<2_5rCv^;I-SpX^f3vJA%UWd!m zv&)rm^CachgH_J1kij|Bl;TWCpM?4R|nN*IXi7U*ZR$|mse+p{_CN``Z`txj?tx@)Q8G9ezqDeKxr#Z82gOCz1HbkRTKK=}6Nb7EnzwB(qT3{fdNu6g2s=)!<)%Gn`)46=|~oP@5Y zg&NuX1qsop43MsJXh662Q-PP^63DA5*~pLd!W_mJA9W_F1@0{4Q(I zjV_HK)7xB8NmhB5=uh6|sH=E4vn<*8YH&gwMlyNC@K{wmU1#QFgp@v4*@|~RQuDdG zuaEbJCvEv?XY1|0eL9XJPg4ny-@1NDZwat6{iUKdA4jqBf=^g0OwQAAXK6G`tku*n z*ox-}ORTneMIHcWRE+dlnyKAx-V~>GKV#YQabp#(Ps3M*qef04ByJ}#Y`x9gt~qbCq$WeQM)*?hZVmnWMxQ;`?0Utmia;4^?_*%ryV~(a`8j3`mY2d` zxDkK5lfk`Zf`3c1I~9Xw?#wz+B}IUNrw(CTEyf@fUdz<7^NenRPCOzvLyWKxTIZ>v z)CZ!HSKyejNDw6ZT?IiQbdNijl3TxK1}BOU7*J+JnP0cE7r4EGwNhaFYi752HYvg; z?wgiugk1b{+^bHRZmf?lsT(~eV;#7f8)@H|UCWuH6)xK)A8iDwMiB-2lNf^RVLJ%Pi!v}b%LFk$o^{f|| z_c<|3LxIW@ed-GD?;9e2a{~)EB=Kr-;;O{1O59uBNnIZA&YGcG7PzQ^m&bamj(KLAMqG1#JuoBrji^kI|ttK(R+E_%(-DJ zY5sCSOmd7SR2bEEb=k#!HWk^XmW1q zH!q&(QSYd@&@KaiaYW99yODQwGEKM0a1@LrRur5%E_xdcWJsouk>7gT5OhvnZAZ_u zVj@xne}Hm+6mUyt=viZzp%Pt*on?FG*C!@BJ_F9!S-B^3vGbsB|G4}MI4469u&z|V zvg!6HgO;|x!*}p=E#P%3_R4dyUuR;(c*jv#a9U75m8TB}V7aFck?5p$$ zFFtI`$-rO!{mf!E%T5`$$U{ zxin1XUrm4O-Nhl8l(W+i_{lbJl*+_U^PKu-q}QZRdyMcGTro(JJIivCqF%4ruzd~$ zt@|I=?W1wcv&r0^5uN8P<-Cz{qy~01P$S6ZlI_UwvCP^bN<3!j7AS+RIFA#P!r{F+s02?G!M`H0wae$ z7SiReQaJX9C5&Wa+{;}LfK$uqQ*PrU#CbaGbsbt@Fi>?E-FXGL$V0P%#Sw=Y~q%d@a_`NA7}l z3~_VWVeSth7Uj#CK@MHh*?9RguJ!>ooP&@|U|+XX7CFYaA1mj@u8_$*uyVN5cypO!4gfC9_zxqVZvWxAkmx>vVZ)iBC@m(V#cLe4>pNghX~UKp z59j_hSFck<>CSqn9u$CjhJ{6fX8MJS!s-m9ll@={#bedrHG@&XF^8G(0=tx`@U+{C z{iNg_*I_D;Kd5ndVnrEi6J}TsKfmN4gLnF>x-;PDr?rqMTAQw-@eYc^xUU0)L67AM z{}ov&{R(Xc1`osf?B=_aE9Wr2+W%R~5fQX? z?VHP<;mW^Upab!dR4@B7y}R~7fkP#RFZ+|iQhP&glf8uEaEAL`fq06w+Lhxww>qfN zByoTWkO`r2^V~g#Rk)Js{hEpV*m5i=fe4vH3|@>(e|?87=z=}_owxKm34-SK6|={H@s@q23evX@0|-=8Y{m60A%M4VqfWAsY%vA(bD z>zfwZoBAZQT?K>lifsfP%F<_@ua~(MFbK4^re4vhwA1y%z%H$iKa-M@sb=NSr@*MH zxzRTZ97{h$;`Wcteo|#YS9nHA3Z={#8EWm%muJM3eP}0m_uS8DzW*Vb*Jr0zTs!!9wYuZC~xU9GVrM|9gdpMYVnXnl|9(;2ZnjbQIW zip=S->NANeXL1SH^sQB-cL8Osi$A*{#6Dlg-iu9cWcm1_MR+%v|HU>(+@DgLWLJ(b z)A!)!r3;q&)wT7WBWzW6K)TmREOB5|L+iWDi-L9@mx4B44=Ng?&f%`r+KnuVPYu~p zb{;3GxMjbwdiYhi6bupd>Yu7DEO2^Qklm)~LFj+c-ZCV>eb5HjBd6ULUW8|~$!Z{J z)o2V6xD;k&j-BC}b^)a3)rq-+s-aeDFhkwFS# zMY;7hwTX}T_hhn!6V0h)2R|{@UT%h*m$QpqbHtZxS&QTXZE^xtWBcaLPU!2kZ+vL+ z-R$~&`uOWIL8U}wo~FgWv$*qWk4R~OO_pu7vT>E7S);^I+XTQu>dsu-hkHC)O|O+} z7lBNuwaJV>tDmJ3OzGG^r?DkL&bPittS_uv?KvFs+}Kc69z40*q1A`>g2+` zrB88(EWF3QmM;6%PFjb^XRuIb>tt2Y^z+c{vVr`*K}Kw2{^hf{im!wE9sAuE3O#!Y z_Vx2*f^nUNT{kbf530US_83&H>Q$8HG%_CGaijNwA_Abvl->-A)u`yCjCRv&XhvLgqKaPrst@ZAq+72ZL zmaPTw2kI0uH5u^vZI+2tZ|OF*ri$jTo`{~)kdleaej%0KrnDuI%~ZQbzQ=OF_-pqr zL4rMg&K`DcbZvlq!Ly73D58bBX1chF-HRJ`*y93VHb_k*2Y$qJ!-?~H-}B;Q_awD{ zNe6GiY@eYmsroj9iD3NMXAXfTJ?aXFp&H}AA_rnM_MVI$(I+0g^JYV_67*21I_Yc{ z7ti{PrZQuy8fo+mdcL1P#wbqA5fzhV`R`@z^vW}OV3I_0!bvJ0>2x~_-yK`7y)Pho zT%)fd92rGPDIVF7SuEA%v1Rh~Y$0Ywb9=VpgYk<3pU{f^!tR8t1qovvQXB*w<8_y= zji**z3Uu{yy&n)Onn=6;fZZ1V;Sl4xB-=a|+P!|nXXbvFDcy0M9)P|1G;{s|z9IX( zd03+b635-RpG75nINWq(1;0%$Itrc~E47fXPP>a@Qd*E_CkX^8VJ${&)%?BvyiAG%;0%l$ zv~6sDfX#6=oW;H(b|{{sB60o~SLedBZcnzow3n=YuXtfGG`_@U)9pb$KJb6;(Zb)S zlchl6uAaz$R*QvHfVQV@@p`5ot7AXAtpBgME2@-N7=j-$zi9~v;M~u{eH@hdJO5fcd2_{?ji0w3FzN!pLT1jOIYdS`2s0m+YtM%6B~*&b9n}j zL`O)KUbz>6c}Lm=s8#T$vPr{#Oj3eviPZ#21=S6i{J}-b{7W3OGfhOgYz| z7mH*R2trO)9WX(LE;EqZBFG!G4y4K`G_BSE(p3WeIYxdCvD_{>2kxhe?Lwhwl(&iO z7)Fp`k~<75v1x4k*{_4V4!rTeUx!sNi0TDI0zLeOK%yLO;>O4uhR$RfY(JH7vkMjC z4@0Oc6m54aCS|^5qBhAQZm;3Pwd<(DsP1@Q<$}r31o2$kAWtIiq1peX0}rg}9~B^Wqy^#r5Xx^tYl?`$|CRJqj|>nkaB9qn@n z$^t5>DMLsk0xhr9flI1@eUg#qsMQb*%d`iP`%@sWszBlK=sT$Wsz5zd`@PE-s?OT) zNVeHAz5(91l%iexh_IuPNeFuz1N+y}FoS`%Z6Z*Pi0-5uq(3)6>*@e}V-*Yn9ZBRB zwY6)9k!K4Z^a7Wpph=CNppDP0^@DoI6R4?_g(cO&dzScX4`4uNO+15LEQ~knz4_sZ z?WbH09^`mOTu+Ew$^c0AiHrO{iBrmU%0F+<*UZF}0|EGH^Rbo|TDBz) zWUW;y?OQP)e`6j;*!-DLTCBQHl|w8;Ikcj$Osj`^duTXEidOF+)hG5Sg~Kp80k188 zW~0oc&v%AL>W4VkGjKcH6f)P*W~&SYvF0``msj6moa^LzN zbe*0)kZLuTE6Uzfv|)cIC|u(RG2)5Ee4Dka2o9kDH-~FWIP|Zz`KJqVQ9Pn)C|dlI zxrJyH&+Q_qCxaZlr5?Ih=0Pg!#XZ;e9gajljrK*}z@L#~0oRW-nH2gB5d`a$#Acjh<>0j#bp@4`-?VW&xNuL-Kt!Kx414FAqNM(LZHe zf?ia3&XOKOP&`Y z8&?i0*X%|TLR64xWc~;kQ5!G!;dxdb(<9LUB_!bDDbVb${0rVrBPA`4dTSR%m*JKw zni%xyd(KpN5z$uGU9h9e>7yTRi2sAN@uZ_$?%7#*n=VSym$xTQYiU z0$NwJwKwm+pu*5R6P>BCkt;FxN>msgQVh_YvzM;IM>5t!%)%EJ7CK%U)eSw&;X*4- zGLGJ#%bgBX;^rNVfF*Ihn8haPz0~+Y(3qjK@CA&t{BgFp2iVq7OkG*MkMVEH?$A{jvEVjj zub$_&DLBpA#`<+<ro1k`CYWSf6g$TyuWd*^Jp7IK-Oey2Mis4~ z=^_^^t+q)PGG04Cgo$;)35N-NWt8YCrP5#w|5uB##yK#)txn!q0(Zx}TM9qLrsla9 zq2mr^_vXgjscFl65F_xQGngwT>&Jj7V)cz2(=!yQCC%Z|QL0sf%CK2o_G6(=QjFn=v0-FIJ61+ zmAT%tRw7QBWg|!v#S~|Nx$0{|MVI+t1ta$vs9s*ZFZCyI)Cn@u)~)wwswk{kbzfFa zB=_#UryFfNU?bNs#Qr8htITb{t0#;+Iy_1hKgL^;MV+GO)a}e?B%74E0npLPBQU-l z#rrJXzSi7c#XxHNLK2T_^#<>3)g0Zmtfg6a_!H7A6!jkP>t2nFvz08CJ3IHJVq3xX z!Ek5q@@56GqC=l9gL8qZd;57olQ6yPfFjWD(EUcM2ej7frhh2QWtOz2tho|Li@DS|fm^f8jnHoe7AaZ}G-i%B|* zyCIUj^bqH(E{3r*mNVA)jkWR= z2^to7*o-5VNqoj$uRS;GOc{H5`V6h-%PrANF|RVsASJ1V5j+07A#}!l_nulZ=)Qk3 zr0&4E=xPz$ZjThV!Q=Q#AT%#RPk3DQI{hPIWuRW;k1eAH#Jxf30e(ffQ-l;L($u3HjoK5rJsMSO2>Dh+q(=gw1&PWZ7a_+(YG zQp^tAYW`53Poc2jphcx@q+`o^gc<*^65zh7cinyXRN;I^cBOO$ZULpnpm`oLiO!(G z=tYk2i^e>+F8x6xMFK0%^h~pICH?EFULn1VV<}fMvR^3ndNI4wY46HQlRPb^ywTrk zTW~k`w#}W_H>{sh%tz(8Ja4wvL|ePqYKQiEj4b$duSME%DP*jAPTW_)#O97G5)gvorG9{5EJ547mF5OUWy-raAB4 z_VpOX{kgBbhGs_oeOb9n>?4_`(cvFBeSX%rkIX=D-tTUPlwj`E44~&9`8Q?|6;idN zp4tS1ob)MH3A=bLqqrXPMWtGo%*Jq`0-bgao~waUi#*)5b*YWR6yCi?rUU9d_N>vO zy=@O+@NAD!dhf*NOl{7uAd48=Qg?Tl_-3T2eM>%y`m@t8oxgb6@z>21M-=>}iK1Ln zAJX%?J9=hs-L!a-qZYIhqh;8)?%sV})#sjZ$0gAwQGVJF1@R|p_a5W}byC3a;#DG%cw<=Vi#P|K(3?y;%S2l{B93BmVoeO%uCdatg&4Ty-`D~BDa zqf=}90uP9iCoFOfeI@JiUiZFc*U*}hrqbM$c!9nn^y%`DqSuXHTC%n+Zu1{`8~Ow< z3Ji_J*D0?edbn|S{Z$XncEv<_`?P#x{HR@yuN(WV>Ra10uM#(ZolX7z?mJUzExp(n zYV+y?f__Ey-%at4p9FQY3vr>m0>Aul+3`9NU?$c4KU1&-Q zK}0#qdyn+@H=rN`y&3KG8}F2)m?LZGX@lbkNU-pJo@9y|bOyC+1GFG!FKdezUJ7a$ z#Qpx0wFel7Sb-Hd$d41&!Dtd_`LRlM^R=~td5mgDApSk!o0;`CHaMvqz}fjeKsxdF z2SU?AliJtZWqxfB@5Y<2J-Bxa;gfYpVVJ!nc)eW!6l(P4Z#38H67xOyPR%>YQ#L1?5Qbl(&atG{nuynt)7 z86Bbysg9Q_G5=sDQusi?s=t6Av0^!A!7LoYWM=QWKzEuNMt>v1f23rSxV;Hk{d+f{ z!OYc6t<(tB%9*I$V2< zad>A*2p4|PX`@iMZwPjGcpG+ay&q2EZh9t^r(1OtWDJrn{Z-~$M~DK~8b(-FgYs?+ zc3AQmx^9M8~QS$Aq zD(~<3t*FP7hVr1NTAH|3q?mmjXYvBBIB?$@b-u@)I*gmB8`NmyCM`PO>0cqV{NWvi-Dgk7IIFh5Q-A8iV{P1I4F9`90ybw}$Vf3?-hH(s zWce*h1RoWM-$QlSf428BrJQPZtKdJ@KkgH_GnxfBD1Kk>zZev9&dS2=M4Dd$_1=Hp zwD1TN8=B;ti>gxC|NI`fp?G*USxyIpF8{B83~t7QOK^I9mtzf1V5zixi$s@&A3C|Ld!R4wI!Dg%bHbHb5fD-txrNPLuT_cXJ*Z z4!{%%zS=hrj}N(fP0iUi{!gZfOKQn7{pdF`K6>CfVR2~HD|pwpM&@m)?Gr?2pon6{v}z~|fd z3b8M=rBX_D>S%RH`2kBEqHr#u7>jn+-RN8T6G*Ru&b?;}8G1%zJg;UGYLuY6c*TQp3GmxR zuy%|eIXJNE6+L+m8j673&nJZnSZbpnIeer-N)#hA{oaRi@Xo-QCvFdx?pbb-u9S^= zLHzXyK#q<8*BM(fs{j4i0?5Y}i6=ll_Ct(*GdEUMkqlCAApDmg{lcpy7Nmx<3q>3p zFk7Yy^a|GjSt|s@WgD*$Z=Sb=<^7euJyozx#a3JAVg(=i4S^?yUK%-CPDZ4S`(&T%9(x` z&w-3DW^jFTTF_-0jbLmH?mg1+VG2^a(d#!0&05+YJQ`y z&;wK)3mgzL{8wY&kmE+BUq9D3GQMHhll9eBv*#gtC5e>})K4kvM$a z@y!vQHk?r*+_YCopH`$7pcqQWvVOq0j&}_@#-x94=|p2FV&<2D9)l&v0ZZ<2!mKc^ z!)ald#%h=@(}IW1kl7D;8rEVAS6+q!$!KZnhLA+sDB0~_*nVgx= z2NL>9xEp+Lj1MG3lxL0Y<@o48al}QF+Lg%^l^_>~OzOas_kB}zdZV2=Lj0u(cz(9! zD#%AW3H}@_3OC>#b?I0lDyFjY&8^RGWPNQmBkOUrrtqGFx!?0VT&0TIde@4LpRRr2 zhAuxJ5NIa`;)ZK+vIHkVWR@U$rUGz9xvKg5a?flePDda0gcRO>2{Cg9 z^0F+NnGa$z<^^dzG6xsQSg|z{&4&pU6CXML+3Mo1!`m>YszP|0SCppb6pGq7c;isC zju&ADfc;2GhC!?h%LVg(HL+z&co9#647-a1bBwAU>so;7*5^5`>on{FZhLmDo(&2N z_m+ghs!v^GZ(rgQ#k7f%3>zW z{n*%8zxQVk1Oii?ee@nZvhs_N-Bc;X4aEhsoK`k6`9aRt!5%Y$Uv1m}wy-|~Tghw8 zQD`SC{T_-EEc>?{Ja_cHLVKTD0z~{R_pSG1MAIYZf()b`)0x$bd~8>Eiou|Tw2&-; z{i#bU9#3<|MkUFM1i7?1VnIazR+k#JAOyL_f5mrjxJ$sEG4B&khwK?uuxFHUzgs`S zn|_)JUkRKiLlrA`1I!j@&JlAX=-ZOP*RbPp12Pu zBSkZ)z3ri~EC!*Gt5feF7~em^t>cjQ#7${dLj2DO6+8gILxm=o za?Oe1g=x4~_vi?|y<<#~QC)4A3WY`wE+Om02ZYP1{37Sh9xG>MS) z-<|FjVnlhqZ(Ue~&g+sE^s>?-3%GRSo+K6;FrlI=1FW(kNzXzo=&6$&7f+ z>}fok_pvJ30`@~S^W44FJ*r=8^x1X2qVu7C<{a(+@B^NH1gNd=xzjdS_xqABS}cL; zts0?p>OQoUunEGv~I19ROzeJ`{$MBv7iLGOaVR8_PhV7jiH`hdHv6R4)-2Zph5iS z2LDB;{w*1TXW_~v?3Ymeqf|kig;IryNKxSLL;v*%rdRN6vyw|!)6=os*|LzDAHtal z7V1TA{Mm3i8jPL=>Kbps@ts@3x^;IbH=5&TnQq0Nv5@y>JRz*9-ueL%Cs`JAVOo zsw!A2<3O$o`2tk{Rgz~FOjt<;EuL>tr36Chz$_Fmdp|2%yM@q90$|b%Y$rNR9>tMe2zbAb_+TRazgmjP}(-HuXGlZ+xIygBX7>j~K zs|aBtoeE>1P5ub8&K==~5tJx~%{jHT3=m-hL>2f0F4Y(W<}DRgPnH48H-zk7C^r1` zb4n@rSr`hyOi!%9_SO$$O%~FcfQ9vJ#ymL!$v$SGI7DSEF~Y4#{t=Y}%x=)kd|mpt z5prXP@9ORdImj$CC{0CGo~J(R`mQt$+Vm1&-NuL|HZ@YA>!d5q;F6^MdeLcISjb$4 zMFb|>eW{v%l@LOu1XAZJ*rlgZ1RWQQ+mFg!K3}1_@Cx3yh1aX$M}}_qVT{}d?%IN_ zD~X7fISwyw6{-aW2@NTBK=`Al(r72SFG z?jy$(`045Lg<1rymL5pBVp5H?&(btDdYWf~kc+*%#eEuqUV z%Rjh9X51Fzw__2sT~KjV)HWj)TjKL-e*^_3+O{V>*x)*?`PXp8t)% z!&^Otpd>+R@wiMR`j|slc#i$r>xTI5fZk8s{|d8A`6sq=VAk69?o4dmmar2r#r;NR zi52Q_sGco=!k=@U{1LaoO_27NLU1y_?Ie)&eMs1`*j#$(ID;KX6{zaK(Kx>BQ#pzI zS5C6Q1`9E8J246r6=Z-O+fxiA~&<5^%VsXHe^&KUi!s6@9?o zy^b)CUc)rpMJaZDSDx3>sX{1{mcMeJUx6}nHYh<}VVmD_E!tx!VO`%Bpc;A1!2(*K z#~evnOlx_7Z%A?D|D-v*vpd8EYLTVK-^->*3Q8Bfv?frg0VWhm;R54^)}`(!@0#3 zmU=D;%iW!wnz(8zxMnjrD2#VaPN4APpLucf?iLz@Uk}0fK^=$xO#j@_)Wi z*stMJ!hU@VurGvS+q+qRBKQe`1CaYe5Gh&XR;jxG2VZ6)F@`evv!kyaG_do45Gzha zc`?ON1q0gI_Au4P!0`OP4*qFu(k(nvM*?vy3Z(l{B7qNxBB?vQ7KsK6|4S$s;{OkM zZyru{+rE#oTFOvarYOmrA#>(&8AIkNGldG7hmy!L7nw3-iX;*WWlS;^8jPicS{0>~ zS)s76+tB+wZ@+zfzklpM_CEG;^g28**7~f^=f1D|y3X^w&a1D-$Bpxks%(yKVLMtX zMF>no-8&;TxnKh2Ki6~UP^W5F<)w8P7eaKmKJg%09{Kwl?F`lmeLHfsJ5&3XhZuR< zv^EcGTe7@gTYI`@&l=IAa(h81H?*rWs!y4Oz=Z)VRXG4C_xGM%m#L%0C=Va)DmETP z%ptvsXO1#fWlO}vkVaa!<7?_i zcq=x0_$)Yd9UZU`xqcq?Ape8YLILp>NhGW)T~YvN=EF*K1v6A?S|JDmc&L8xKV=4k zD%7>`iwwcP=w)`(!s*K1j*YsOlRy%o3)QV3=`M}uvZ7=Q7WS}a0R)ypCU{7GjQTVN z%Z`S*y{%8r+0N8%R8Wypj$s$uz-0``(djeDnwMihzDgy12pfuaE3uMKLfg=OsK(15 zVJcZNKBvItUuW9J@$jKs0I5}#$lzIB} z>9;qCa>G4DGQTr)RpEI25q~7F0`j36I(Wl#AFtdz2FQ4J0C?{~ymjdZ^!_A%on zTmR#PP<)V3kyYj68;&5zep~YCKc15)`svNNC4YJW(wB6SBtP#I_?mk2=;? zS~!3Xp-PGk?N^!X>Vp+}4R9u%Tq>Sg}!I zJ_Z%mrw?dlT*$Uy?9O@z#gNm3y_D-=l8ptGrzf3+h2b6#cspX7h2XP&%L&2tp9eGA z0;`I|nFG21GdDxG5hOPD2;A{j*lia$k8b(v2UtUh1vsk|**_a}Z7HHhgfMz(+RdH5 z_Sn{+-6jR}|H-#8I7>iP46F@~ZA_qdOcAzgXxERRWI&Vl{6R%Fm?7cH5pekI40I8` zh9q)Nhd&yH@lnX~9&g1Zo{if>w1Y>r;>O&$9R#+V2JG^tHwwrd?d+jCDqt0ct<&*9 z5k#b=sCu0R81oo3cI6L)!%5Y#N0fmO>>hm6VnqfUu3$Q+n=Zzkr0-O;ee;ZgqAoZT z6?Kqk6A?uh%xtX6N}dJXny8*=Pe&ls*O$d7Ad#%h30~^BWmdXppp9DCl=$7j2H3x0l_GG-1=_-RCCz?gJ()+)ReKbnGH!RDOBVQPg7)&uk5%9|5kh?*41R2K zRmT(1W%NU<*nRNox4sHntJa&Iq-TjcqmC{xOY`WT6Tbvdsdfmfu*jp)Mb%6fot?n$ zJOA5@8G6Bg)q?XW2dCUl*)cXKFfy`zd$kJ4^xi>pQ@fUPX#&y@19*BS6AV$UJS#zU zI}rd~)*Oq-v@sBLl;0uuz6Zcbvm0pEJ`Z2V9DqkuWCIST?sIKQE7R1BQ{78ksCf_a zdG$D-7lwxz;QIFexoMBlP0LE&f`Gh7lx6!h%f-Lp*?cgR=UNM3Qux3YYI&mMfrM)` zdy(@HnRLVQ534{xv%iDEXNn{`oj-Rc7Tq0bzHR6{zsLaRxxfxN$c#cJ#(JUY@^mBgi&;Q} znj2`dFVS~{FJ84qJ2Xjs2Vr0tjIFzWzu_dkrWPpu{P~mAjVVR_l$)CdWH|3AL~t>^C!}zi5styLr$5}X6YG7z6IcUF(83L4iZUL<5aiq%y6Byrt>1OZ zl&eQg8yT>VEJtJ2#`JDtIAJMnw{W?~Cer~cH5`<`Y{N1lO%g7VID>5n0lCMD)Q zTemaWgAtPFX@U%Rlg7L-IUODVz0%)ML#ETyREB{gLorlBHH~u(2qzfcaU7Z%Cptyy zN+&>C;SMZ}gO`Vs+(yV=(1ef>zbDKXGMiA~!;QM1w{`r|#gBWb-RXFrZI|cvwqdM| zPEWe&feA@IXgz!G>-0I9-L`f{>PaM#qIf#}mY8w`&iLv^7vRuQ?YWLOfZ@q~*zRr+ z4utNfy$bZ0bEpXFS^UFaa8r#1O7xs|&2zgDr=i5YPMZ?uuos2)C%g@8NeXu)E&FgHTk(q z`HSl#xfTx8ZE@J=^c4h`vGHImar0neTLJbv8SXw{q5LW~?f$U;V0!RoYHMH)3xKlt zfwUFqnyUc;{ynUiWuu9OwM2UQ=M{vHphEdw>zQ^Zk{-6FJoz9Vvdheg{=D~TT$`m%BecU3wWbgNo@@hW>%a~f48~~}EGfd5s6V-F zE3BGmhW8Tia1G9q7PKV4jXRBY`{v=he$bM_sm+)B_=5?RF|oXDpHFp5AkcV07Niez za)f)N`PSc4jEnel^60}HyT}Ye3kqpal!V_bFLDXS`pVf(@+h51s5hA$n$z4Y1%@aW zmIGf;eohJYx4EP@KMDzf7IsLorrK7ln6Sw4XH1Azy2K(U={*H? z(J&Y-3m040R`|h$MVNGk8z?k=k6gmgKzSC}n*Qy8SAk6nF-{a33FM%^yc8-Evm4qB zOU=tUW^5I-{J(MR4p>hLp^8EH($D=wxB|9D*i4wA8k6Ta80Z)%MbKLn#Pv7O z-9D)|16uoP;CNnreNp59RvqOpi;oNfy68d2qv!hzkz{L}S{0O)QTTgv}VCT;%10Ds(d9rQ)zX6u{0UN#UGlC?kVtMtUX0yr72c@cz=M4084Itk4?I3j7S5HD7;+@ly z6vq-248(y)v#g`5gV5@ok&0o{JAxP9-5FRSIrt|y;YW312bTq%+0!tvfm07fK3(&TYr0K~NI=f-*vRy4RT@Y_+? z2oIO1pLW1-UT))V2Y}GZp{e1}Mc09f8wxOK;0JKZ_uW$XO)J(2KRiczmHZc&v%y7@ z_Z@Y&pgx|n88hWJn++l)gnD2>+nE)}P%40yXh~C%`wEtxq{!xe6_txTtdD-^EIlm- zM$g{yfJbUIU;YJsKtn%Z4hjHEtp(ZsZ;ILrW=2=DMpM6*$LtMR@9Tj&$p?}#8JDLz?C7Can1gs#@NykDpuX8^ls)9D6xz)~QyP3XJK zrwH4tW3UmmcsUZDPtKg$ zcVClxy>HM00Jqs868iZ6imt$sqPxU=Li=x6O&)UY4xcRs^8XIV@CqXu5YyeW_&}fX zU$E1R0Mu$-JMh#0is9%@6PSUduJtOQ|L;G+wcmx+DlbLy=x=b#9kHd4(~bV?T_7-a zL13KufBXYII)nj~tQyk9Ut_V&0sJmoD#>1z!;! z%BWKyKhv5mJNOF&lb|bK!sv`&WQH4vhBGX8%`FFFEWYqK(xj1HnIZLqA>r$7|w!b5wmxUe!wMF+x%Nlsj^}f7Ob)|7;+HL)(&?VH5~|E z!$X7v3hb!=*tMvPPnk*H2iRtzzpU=xHF)^%s;$f6Yf`17xSP00ZKFfjV`MJuiWBHHW}%q2E!?5aEK_7+(a}EAwhjP>J#weD0l@n)SoA16-rb{6Zu+Fx!P1sIXU#nk zsw7--7k8{JNJYp?5|+0Dv=ro^4dLPp>FUZ2#`Cuq-&BnLoc(m~Nk3~hOsmrxS8(ji zX@qM&1><5Q0TO`_+I0#nBA`dH;CHq`h_;YjL;V-CFiU4Us^`vE{Td1y(J9!w{LuPM z!2)OW$*i`q?DoLrF@PJBcgX`rN?} z3Z$({Y0Ht%vP2Il6ZJt3eBP4l?BnK}7fjZ>G5(cI5fD*Pv;T(5-+L2D$v7CQ{)Rt@ zumQ2RwTf_m#Q-jZwQ5GV{q&vxT6_7q;#nbYy6_j8B z&pds8zbAK1fo0S^P)WQ2NfSdr7j&(<0F4k|-#L1@ zZC~KCm*dci`P`{IHOAZFDHF(;B!CJooX;0QhYx^Bkb|9LZVQe&(SVH^kAeS+6l+t1 z9EYLV*wkguR^%5p0D34CTTwc8WE5rA2q^~TUq7^x{(=lqS2;1M7brtVB%VUev@mGa z<_DzM8NoA&Cc=|EX`_KoUq!9>sb_`^9o~kyfC0b$W>d zQ0i;XL}|$5p{#1|>%VsE3&7?Mh>iY*+U5k!%Nq4$M(KY+y6Fbx@wJNFijA9QU;;dh zA+;^qH%DPd(iZgZ46vaF@|33xW}(~%VtY+005P9~|KsCJK7(E0<%`-cha13d@Wl`4 zUgcP%H%A<+=}m(sz^I)7=(A2-2HYewQxbrtL37AO*V<%3QXrTF#;={_cROXDqGAz+ zVK*3KP7zNRpem}5oWP`>7eexG27uW(w221+_M`3%O!q9p?IAP#K16=ydYtSN(gBG6(4z4^xq_2O#*Z*TINNfVS;V|Pv~ols=}r|ihztV4z~O4N*4q6; ztN{cJk|2ySKB3hV1|?gwZYCOPZk>wI2a^|Q<1q};Sqnn;4Wu!jQnf@JXrn1$4^b%GP-(EY1HKP->|+c14D z>X{}}O+GYu@6&}O4A1If*ZC6I-zd$s1RU!dINFw@F_Z4+wN{b5P)KvGoipAXXt?3% zOw>CKUG$#|*q!oXlR62@>#G1H>%o}i@QUEbGfGgLN0PXP(rd_hr8-mAL!DG$$fAu- zV1NYPASXB}ZRMCuI78oEs_#C_y65G!dkBmL>oU~Zbr+eJf&KjWTPfeM?)D5ep^vHO z#h*0YKbD_M2;H{FG_QvV&7ZbOP|q%U_d*!wXWfi@=1pkM34Ddk>id{CHuk zUpxEam&DgcvqrcR0J2Gt1sZ9gnM{JV^{*!yez}9eR^O0Sr{b?us2W?VDa%Wo^1b(s ztY@-prjSm|(HSk5B!JQ;6mhO++_W!Mk;{W~!3579RZg+Fs5-FZK=RtmAJYJ%-cQ_a zIc!Z;fOK>{9x;Kpzi>dishq@;vKCbtn&IP)jV^eI)EdSjD#{7)Y0??7TK;<<@L2SY zQRznXd>w<|U-a9Sg*24jNI8&pXnq=XfwDz)D)$8dK<$#T(->yrSlvB-$@%4PKnf3J z7DojdH7nh^{PsN9R;znYKmmo+xpmHp2>nWb8ol?*2H@_94+(CTzn+sopr(z2Ko*La z)P(1cwbHmm&P>@e5~pUqLo2U)a^z3Tc*_=*>wMyuR9%C|{A-ZV?mIl+o`DZn51gUN z_C0LNU23udb$Z=rptGN4po%hz9oz zy*znP$ogKlua|NOkJV`5CCQ^_5t*Lh3N*={+gIvJo1Mq<)KX*4D_~vpi?@IR5(={M z6VX4R;C&AfpKqw@hv!UPGHG)_3TD|2fHP@J%E0+1D z*JZ_7_P@Lb7OjjF6K5v-UT~DgTy0X`^K4TY*7VwPrA|SUYbq< ziD8oo*+JiGM(7ZI-~L zFi@wD6khu79)2R3k}vmQtn+~7s%rHBxGdjxX75ER`C-MX{CN!{zLuxgADR0xriSg5 zZAYFBmLru=T~CjZTIN5Nlog_U=8L zZVRN)`ow0yk3hbQm!!zl@z<5^Nw$O3k+qubRzVDJqljz2z|ZT&${k77lF$# zfR|q_c1|bjE@1d9lx!K-`$#hqUtVPEjV(Z2WzuImxTl{bqt(si!N4-AVP33%W=FE8tS)NFY4 z6LQK-eN(Ny%VN_D01m&w0bq*0_EYld?tYZQU?bGELH$;(J(+=(Sg&&+m3V-!9b&Zb z^rnCvQdfV|Poqm{{n?{QSJ$VA5i& z13+|(s;aIJH+tb-;10$uP98d-tHfRvhgCkNe#S-#EhaUvgZ_*#AHMCV+;g_tbV=+yF3+LWPVg=r6bp zy087D?upFn-~TKBB@YMPf@U;x&aD$i@TXmC$MRyV+0ite`;Wc>mY6ZDVk})ZfjJ1> z9(~%=>3Ya^VewjA`fbLx^>mf350Y&$!6(xwYkUD$o&QNq2-~}>ThLp|eQ4}$q1D~9 zO3tjjEgs!AUiV%PfKxOB9F0Xf0hQ!9;-MHQ&JU;_N!E zz6P$Ze<6q=sGE+}W9O|94L9)Jc#P2HKF^mh3u_QUHF&cAu*!deykt=$L={@3kWn3n zxt^8p&aWy|C=DaY`aPJYX$jOuDR5N%0C+(NvrKG3BLNW@Z4LxaDUiSYh5D?>(U3C$ z#!&8+gK^*BdRz^%@~XFIU0q+^N&&wWAE7SDqun6at`7``48d|P&!gk(ELiH9BjR*s zq{APWJ}Xesl>pa*#0^&E$0C?b^OlhU0uC7ZXr5AMa1uQ4Mw zR2WX>1vi}?huJ;BVdf2p|Bnxzx_c=bFcGu+{%*{W4g}tA;bSU?)N=rU#_v#uR{>4h zg60SQ8hp26cCq&vYISE!KrJsNKtuY&c<*kH82<_K7w4+@5}W*48LUyz~V4Hu;0La}ex} zTmh=g`VQYn!c4#W#84`0Jerj!G-d>fmOdXc?+v~Mu%cjMnQV%mw++0!lmxnq4M7^*-}taaoRGwssJ{mO%00k5ZtxLOw~zBCj#q zTubEGQz~lTNFw#5`DuQUD9k7mdGPM=p6~C$T2q(va58k0eBbA=isnHcF6lz7$lHS4 z7)FwwNOXLm=8Qjm59DF6_*N^ueIJ4W36tXD3mO6u#D)}lRz{b%Qvhez%2EWXB=tD3 zRFTUF79Vv4+I3Afngh6g{Q{uy$D~pgES)4FQzetnJ2(mX)Wj%!t`8d5#lbJ@Y<1vm zdCRVr`*fE=*?Nl3Ev`O4Ydp2zu$Mg9WZZbDoQ`F)No9ROm|dFq3CWR5%K>2F3Q3f4 z=c}krq#Rn4-ljA_%QDA_YkpGTy}G5^{FZ-~*W=xvY!;(CZ_==Q*`M;o?^fQsaT~Wc z!2b+?f)*no^!8+?kQ+(4a0{tYFxT%q>^yF4?Rg{o*ua`kNeEsU3Dn=5;+_v2PKel^|*=u=$$6u37>PCljgFmX0P=WiHD*{OEnhZMX&9Z^=C# ztE#s=?bw^Ig=+TGgKuFzP@#2!yIXO-T+f(Uop$=I_z7}OMnQr5$cp9E%0}g&2~l7B z9a_}+s}wKD2+a&2HGf~la^fmCsbHwA!{hCYFPJ1;lC4}-2jgTuG06d{YEyL9y4pS4 z`5}fDQe<(zi=g+t72ao5?4naZI0 z$|cA1V;wXzI0={i;g|%otJ|_t=?boKshdr0=gISz1NX^b#RO6Ik@tU}eYmdEBL591 zVUg|}zbDD%zd$R%(hv}jECV-kivRp2DUc$x36c4G{_}rre3yXw_q|zv{LNG2QC5@B zUbOWu;Eap$X2LnCK$@vS{C@YKPhA-)=gjJyGJ9cYq;{vyMz*X1*)rTBa-|<8yMeo+ zht9Q&29AHhW`t%WdwBi5ZQyUVoV>9G$1#7zBQ zy0H7m?J8&1%~piw1Av2}&JgI-34#UvC3AWwc7~(-j@`o=Z#uV_r>$Ez9tC+lK2K3O z&y5h38ww8_jTU7{lA^Iln%PNKik)P?$^&EPBcUJE)cm5Gu-6(^R{(HdkB-bXhcACL*6sUR&6A(~{qm8`Ch>1Jh1- zJ&o3XYe5AhLBbh?%kX-vz*tUuee`r8=b{}0$Tsi!f9w!qgE6ca^jS9g%61So`JEq- z1+V!@K%G?q$ZtgMr-3_pyPT01C7{mF>ioaEAS@DOeU1<9fi_yZC90@2z@`&L=ICg1 z%iv*D2%(-=o&pjjs%KY3Ti`%D4xPz&fhHdSD8LO`*UJFVpS?`F#o~!cm6#cu+NX^h zO{om|q5m&ch}`eblft<;g=M0kt(Ux^!ABVj{(8`fz7r`qzC+6-Uqj5PFXTXcbpk9o z9*o*#0-+@w@oH86-z}NNPa4QZae$NoMLQV=pu2*ADHxC{z+!RHNagF%oQfkKd(sJA znr=W^0#KWcd-0mp^b|n32R#O>N+Ik&2iER;Jl#iu^~W2HYORRGmcltY?`T~^R91Zv zVR!#sW@h6P2)YeRt1^31(ggW)+y?S8=vhr7p%^QrONOH3<>Z|5G2E| zr4%{Pz)AWBpmiz1>+L&jM|5)GZ9#&sU(M#$NqK{9eZd$3$eVyMO1%pRy_iVUAvV-U z2rk#;4)dS8ay%fzr*e21WcnF~q2HIAS@@bSy&L@@p&J1HskTF~FgCqmAS+$tYD3(^ zie(D(e^cJv@vlp4Vued=V*TD|!G=?V#jbqH^lP$-JwVSB?%Y}j%MXrYGG^}s30ab+ zikg|=#K`A_PsUREZ4}ho3RZEk_}38O!k6iA%}XAjeDysnX%tNM;(z9J|5V89HybV= z5olgq>kcXZlq3C7Q1fi;r~mvg8gYAvnm2d)e}3?XI{ts-+nA6ayK$2aQ%i3SD43v4 z-3usbtz3KaFTj|O+SebXJl>u{qEX~4ibh=&8xU4Q9~~>-f;J!tfdsXz;5m)RK!uSE z%n9aYmdXA!3>s{L#LZBZo&pfDGF8yZ5*Ye51_O2pgG>t2^M{abLG?yIOv@YuBWmmJ z5m4^>lO~Tnqm26o2y9O#0Hyk{LFiV-wP4;rkPm0OL&!kE`$h2{-I zPXlitRicD8SDw0sw)Vm`T^h;(ev2fxm7hM|g>tjm%W^)S-==p3JZN~AUi-7QYbN;W zuV)88U;Uhw$aw>b`wHm6Ls_6au+vF`R+0grsGf(}+y0u+-!lq$e-d;?YlElH2)M9c z2C-!aj3BuIpSt4A^syq_2XzHen9z3fp3-Z7+=Av2=z& zf(9}o$5PYK?Jy3O0l7-1AYc8bC9v0TX*z?E?cDe)eWbirVC={Q)%8;t)ySx2RKULZ zre%pFuP7v&j~jBK3HRKj?ok(LnJ!-@AqP8yL&^(Qmh=bAbdWl$Ns`gr$DVZx#@3^r zx+%kuEA&ShdT-mvSpXto8F{6RZh8rD49%Yt!#5l|l9FJG5t3XzyQ(l~^#b~GueFu9 z)qz(rnzuRArL?+`ABKtAoYlq!BbQ!LSo2`_?UvVuc}iHA^(lRg(R|x0nvj><&UvF5 znTQS>rgY$^w*QX$PW{f-!CY`o}TN|r`lV^AZ%;Bm3p5*6g6uqY^tx^fK-+@sIZ6n zH=geRaAy8B&^)8bVKZf%e0pg0e6lsN107Bs2EA8aWY2#EgNd zl(l)xA#0;XZ_{D$@7dACd7`k8Gf>H%cUvO_nd_v|HLZ^Dsd0 zG|GauYU8fVa_pT1K;q8e*<38)A)R^D=7_iuEQXFKOBfGtv_8Vv2}mW)0AjGy_=nmunEWo&ZEva4`zz$FyYVz`UB%cZ_?12LCFPuba$vY61?z zgN`2V*`h(`AiLwaY2e?;*bc*2u3yBRN?U}-_);pm)@{G)ae3HJ>;H1GHF`w++GF|* z(-m85gIY3VuHd0V`{mEARSy@Qg1UDer!u;W-$WB~n*Gqu)OS<#xaGp95aV8O?YJq1 zH9UrzTW5WT?4BsxMaJTt1~C5ifMyu3UI}I!M~Z(Uz=t1C z-E}HLnZ>>~Ny60(nIzvh*^`GErw?yeUxvA+PPY4@7e57pB~`NYC&{bO!*0m=2(K0D zCN@ngH<@x~E-?SQ;&-aZS=ao=q#U8I0g8*J$9EWA3dx>5daFF58)b?f(hHay?`0K2~R*08KXtag0Y5M=qP&+|7LKIKy4BV|t|lF5<^q?N328lU_JEGFvT74;|4`T>PL=uj*{mz)i-s zj;u0tJqtHJWr^B!`YZN83%#p3^j$WLsx_^C*oVN-?FGVZ<}Q7zc>4fLzm{z z&^O1ojC;4yoTj6bO+fcvj=NqrXly(nbu(;0|pn*<9u zAskcL?(l4$WNRC2BOrQ`k!nPFsY{h7ap4pQ(7ohLZX>M{k+uX!uv!kqoj<>;sfDrT z$#<--K<6I~E|*&qn{b6IhyrADsx3WlLH()B&XzCG+=C=>Co|0T;GCDsm;~05=H=)% z>V3&DZ}l*wQplGNjrMkJ%>XLlDg5vHlLNi2Ad9H(p9kjW(j098=f;)3LJzq@XhfD> zOiu@f2ij9Zfc*CUnUS?N5po4`YA$G%ht`5P$CG3#&%sgSSH{*aANXXvM-q%~XmlHyU zWq=F{ZdL>WG(VrwbN$fa$+r(4ysGkeXAA3Xzw|2DynoI4$@LCWm-T!F!d8JEVo<$yPiBV7AXFvY-d4Tq77Lo*NM)o zBw;&6XQ;T27>2rupu`BSa)}rJEQa;$7<=VT85SvP00KGTx06s_)%_R}o~M4VYm0)V5HpMaM=Yf-auzp3S+-EnzMaNy3rL} zJ$#>T7ZAzK9eexW{$6QMJUm@aH2RRu<5W;n6~n0$1n(^~(4F=Kx~*PNlI4K2mV-0o z@*`b2c2Gj|8{|F9QL^j98-_0JRXA$`<13;@c5^{1uNY9jlhfd>WDBaFBC8X*WkoQ~ zyMFpHGb|_u5ZuS#JK3YMoCPs;7I>gi+d+^LrOO9Ypy9e`W;UTK$=>NK)ko%y{iS&n zwvA_YLUT)|E+-+G6VI*r#E*|TnH)1!BnmnSG&Tuko-rW|?E@!rN-0UqJ+D;weAYAF z-YCPDOo`*A${jNYG&lGvfNU>j*iJ71?ZACMp%^h8 zraG`AeWaHUHcI)*Oux!z?~ItSrW+k zia9Vn$4W2}k8Y&d@&lV_Mz*zc8Nix;F2zaKz;B|J(xTMIVq#702M8y{%cMog+H?>R zG++9#QVX`~bFOn6qe%((xElRJ$zM@8%gha&r1fMWtOgI)-2Pf3Lh<@A3Hym_z9ACi zD^OA02tN4Rm6VL)+{wo3ZF|$3_c#`pKgY%$(k0B3|8`%XIv0+m<&|NZ;y%V;9^J_I z1=ax@)KxFP%S4;g@tOz<$Y{heu}_iF7B@a2n}K>`!IAXDt<9;dnkV2kMa!OTV}cNy zJ&CQ##%Afmjr#nP#&rybF;D}KHyzva%YJ?$Z(vA*@*Mqn3n+~F*^z#%7Wi9J3k)|| z^S6vA6w>Sw4#vD%It1?zP$ozzE!A^8*-p($sI9xfBbynzrrg?vSsZHf$xiTZ$=R`=^axF^gUJy;#sitB zFd}~0fIgXVcAGI}g*0Qtz_a8FRJpSFN0RjI;Nrvinlj9@7Mxj+6yBODBna|E?{kQ} z^RhK~6-KV|B7?xCaeVkLHP_uT-uXyFjMMMXjr?BUVgw>S=o;sW#~J-H7mP$8zjD4S0spy^bKmcc;zdWyvk*61*a$9&Vakm z7s*s+_ZpB;JHoojC{52Adf1q^#mb(Iyq3x-hHvU+IpAwMT3q`{Kj&H#j7`m> z&6k6&yh^`w+~U)T@#3OfWoBH**3mqnvdCJHTaCg=*hT!)dnFvcq6*{Q|UmnG*DD(YVV8Pv|wJX4r?;Wh!dSeC}cuNI&ZwRZ0FN;vr62{!$^E zqtM=5_Or2mq)fWIA-nRQmwR*mEPMN+L`GA>&^^Ph*^W}7!}{tN|IVnQcHygT7304` z$HS){q(1xAeA^3m6=R%$pTpMd+VMPAQCVg7y#;M_I!RVEM@CE$TSy8S{nj%!FuvCQ zKnI|SqYKj?F%iu`$cYb?O31tAHj5Uq=NM4pthe3J1|SVi5Pq*CA#CSCl))c6pt$$? zpR-s^1J2@hX;B6g{#r9ha!ew5_WUQY%*q{5^8a?loiAN6sE%LVewZ}o5-9B2AsvyR z)Rmn}&i;8ZdpXX3ck|*)g)r%_0dEy_9dw6PAZhFt)JZbcmwPB2TA4T=e!hciw!yPm znp{%xOcu782BnSc|D85w&E1G}6NDaUlj2J-vUI#Z-S;cI+$4kCuWUQ{+su#PuHANJ zKk)4gx@&P!i6fM67K3PR7g_UgdBoISeym*M)jZ2Z)DLkrJkQaP^M&bU+BT)ckWwsB z7<+zm{X)($#RD-(i6(eIDteyL+YJ|qP%o6{Vziy@lvoPtoLtgdMzu$1035Vt$)64LI@SP0mX+dM=y&~ z<2!X@Tjt0jHhw=#Eo2;fVs7gziq5fZ{h?T+M3VD_{9|`2blE9TR^dM}v|DV4{{E|t zfOBzyo$LkSXzq=Rz|%ur8;vo;T=vdRz}Gy>!1iXrjc`^Z=`CNpaX2j%y3TJn^s&4= zK1aF~x(TT0alkV&5`;H)m zO@aiW1BlkX^2NB5vylE&Qg<_Mt_M$AIPjC+F~cHyNNv7_R?`XYe7^T&g&Qe_Hm1@ zP8uGDWH^EsTJJq)<1cS2iv`Tf@E~z|zc5t`O-yvZYn5?tcj`3lie;8?;i6Vt$PH2W z^B8s5GW?pMmr@vdtI~k~-v2;L&ypu(TTO^Q+k07d9K${+J~YJ%2Wf5alp`L`x~|^L z0JX!3b`ZJb)kR*oe256~MGRzG7T{?08(@rz)04I`@ZU51$tv7AkZXq;giYO}KIHTI zM0Saw$fFp>YTc#?nMPdaor&E~VaHB5?{WRQ3u7%OU0n#D&ib9Sb2jmvw%n^7D^9Zq zDI8N#$w%Q0Q%|WoH6BQJCUM80yibq1VTxOXj6zMB#O?`>{|!gfR`uf@N!+8*c$H|E zd?ESwhYoE0i1nU%`J615iWesP8!N-)0vY?`z;evAxUKkhoi;RXa>0$)b#chLrzd@O z?xtlZlw?L1>G|?{7N@@YmIn?KefO2*{37=h+*8*DX`Xn_mCg?xfK=Xhs7qx`gpioi z2x#9I7MhC zuOBMY_*BQ9H{rxV>nRA4g0|Za(6Izik?kjyId~ZMu{k+0xHU*?4Us9NCBA7YQlP7` zMU@6b*_)=HQ2KS5M?V!Z4JFSH;!2c4#4UPd$H=6nd@4MLi6d{14r{Wl2a=ZUGL7sn z!5;YFUW6PIMsP@CB@SN3G7)!8P!x#{A$hfv2Jfv&tCltnhJ;HrVSV}w3{Mj|qf}i! zKXVD8mL@R%zC&P@mNl($_jKthv*p6E!ft4dx~?yz5cne*R7xh`#yIAa#R{?H8_yJ| z&%Z23|4Lfh#d_TE2#fxm^DIuG8~#c$(lg!xct zB$}X=RC!{h_!U&bFFHQMw!Q%*t{6x|mNJ*9_0=YcIaEW9)O$fD$4~r9|Gut?Poe2- z0{$3<0Yws9gXeGP7UOm9$%}BgPiM0IF661?DkQ1qKsP3X5wtl=IhoeumbJHHLT^=Y zK+FVdsE6=3@E=Kq(PB@Zr60Uu-xRvLcN%{NL);onB)>|eZ0Oi^n>lHCi$D(c8)qY~ z_PxakeqWAz!^^*ZrokU;1B-Fv*rSU}H#(+j-!~mpx5wyF1~Jw|cF0`x^gk22<;3x+ zjxa&zvi9b9%KH&VJHG9|`ejB;qU%JR?{PdP7+d2%wIyjJzAcXBGI$J){nRF4$VNaX zOFMgCy2F&pyxu}en%%D=HyFHMXw;51$u0tTodS?;TNS^Y^RT_ubqj+xI5>8Fj~G+X z!d^Rm*6SP;^$7V!5v5e+$Hedfpd=~n#Ec=5tLYFyGeaMKMJ=36nyGbV$9Mb z_)~It`qj_by|{i(KiCNgvOyiG(w9Rn_wVTBYPGCQN833tUOS=toM&px$EnyW%^H|O z9*ONb%!i!}dvHWXeervH6H{mFIqL{1Y@{b>8n+x!6K(3hgjf)eba8+Uq`w*WmKTrR z#<2@qi;mp6`Ww1$z&x}1=Bx=XFN&DuYp|1Cr^vO%DS>|+f%puX#;~mmw?Z(e(yx2o zyMq!Z_X_lGhuQI#^?ZLGlb=;MQ`!x@IT-uqAUgy>IAUx(E_Q|C~S)qGQON6yOasl!QO^%yl`u{xEU(w*r=Ml7UJsH;4p|7v*Ox8)()HV}$OhnU$ z1>?`JVb*MK(LRwf{B-EwZw!@f8q_=12pS2PeJ*ta8YOSLs`8sK4K=YYhpgZquK)Q} zEECIqd0mYioX_!BVVqBJS?iDkOrKMzXMDK z#I-M2WoGb?W2Lp!efOz-A((1Hzt!*??h+}A*LoqYJ>BijXhn7X^JeCY5T-x26=SeU zd!T2B=t8E+??=}>ny1t)Y+!ILXYJU-QDg35DfAQd$y*K{qhzMYe?qs9IdQ<2zH+jP zv+EQg>+`|GB9Y@@4Fr>taWo&m>+&+V={*6!*P;DF+Ge^jLjkOkI^^KOD7XT&X$wLP zJ^(NDwI4zLd?`DX?Vrm8pG5&80o5m|f(HKV)GALNK zOg4kW31;qjZM6aS_rnmG*+Z-NZo4rlxu?f1veGzNwk+&x`MJB{)zq?CfzIxxpDd38 zuqP>K8}wtnCMj|*@}b2-6eo3gXCp-~-8J~y$>DLV;fW|{=}JN#t>$7e9E4SJm}7VZ zx_%CTU93S{CUoX>pXXWZOvDno7n(zcPttX9(1R}BL zDDb3Ftlbb!`dD9rVtH1#Q9q6wwX&F6vWHKACKN{W?vE=ya6kjYRt6XwA}uX5(fm$H zUh9}h4P9=)$EBY>@;alTqYPLjd!z6^s_m!iN98?Km72DZS2d!jk`p^%z!k$r*p5|P z9HDcD885$qjM?LcH9W*S!9l`P6#dBuhDm-oXLQF_;j~0ZrLG(X5W56$#q2-=xXgIm zZFN=ZgG*S7=TRsAq_ZFJB#A6jFLo6)(bH;fB>-t?4HC+0lsIXDd<@s_2MLQ}Jsx99 zycOL!V@D_{q|+PGK?Cl_8X^yEqU6DXMMUT(D9T#(vy zc%~tctmOML4@jS&HGdfVK-_)E913IM+R;=i$L4Sf?>8nx=?uq9n`!Z@*I0YtD(8fk z(p4~e+#!=n=ctXc5XQZp2&1D2kKxm`haPOmB;O~yb25LH$AIemfi-kx&A_!E2KkI5 z07!4kYC4MrUFmm_hCfw5pJ&(eaO}jK6t@pHa{KPQpY0Hlg2Ys~xYbNBD8u=cMfHP} z-63(`PPc3U1X+VQt`(SuiMI$SnJv?&XVuP8_50MkNgy5uOMs*tw;BTcm7h znQq2)Vq9>nb9t!!Pi*N9av&qj$P4sXnc&v^Xm&fwTP(>lG)R+ z=p@9w0Ad({BWhlT_3M_I=6Stcyj^8B+=(kr_76sUcj`@CKlAF6HhB9v1+U+M^{-mo zWkJ9Emz85ZpGEJR0h0&v=37Qa@gz$O%Q3+k=T&J!R|;FBNP=rlPYhWAp?qVfvt7U_ z1OL5`VX3uTiBN}Vg?Huk#o$F|RW7SQtoG9%hw`GJAWVdH`wyVWuI7F&c?C!YLUfb1 zJHygCR|F2KU^yN3;@2CH;Imck7>d&YBZvNllkJc#OEyX8wP%kf_L zj=iMj5ko)~pyUMl&EjS0C6~W6G9E?<{fE{IiV?_k#)?l0JgyQN+7*{3g%ixu7faKq8)hIxdp^e4_<@O3Hx9BqcIn( zuBiqAXol>q5#=C&c7s7$T~Phivxe%1CD=&afRpi<-&wl1BAt<$= z_5;KwY(_vUTN~p;zAI&hLJgjPBIQSw7t(zSMwA1Xx~O`Ow6lMIQ+ZJM{eg{}Sd&Wf zV@L|PgK9OzsNhn_m12k3ta%)7r*(IYdHg>4DE>%n7!-@zE58>Hrg>%9?x0G59#&mw ziYwT~2EvLNcET)7EurD%Edei{htUs}Xm|y>sm(ngJ2ENI%jQROHf`--f|02Jo)9rF z@T67tCqR^G;7G0m22eK!t_$#*mj#3NmKBMRumQr1s!wp)l zUe{AG@N)sJQx2vx)Ru?j<->D_KxULyV80v>)?Lh2c2gs65nS9 zG%l!gi$fQ1RzUkN%iGp6`XCjk&v`DzdTxqf>kyuLka}7|58K`4es!+5Y{Khnz;U{Z zBokDW6kJIm)10ji<}zPTGuuuN=r$u-^Xd1GY>M>9x*qL~f>l!c_VAVaJ3imtEhe7T z^K}tYhM~_{*F>h6QfbaUE6eB>UpU3PyZ*~FiJQ|RdhrjyR6Ap5o2~_b@t>jh;|CMv z*7))V9afAER&nXAvvAYLy*RBaeqK4Mkk&JW9?}*P5 zG2qLXhOTsOpa%+Hlv*9JPsvIbd{NfUYgJW}MlT=*O`x~TSfr!B52BgLmTu^Tnu^jC z1b83?_|aE^EF}WCYmZU4kN+h5oLp5O+c6whZ~cSaOdj^7m||!1%ZRrf50h=;$A-$H z7TTqTBX?{=1tp;Q$Arq&{BfgE#D`Oup`9PUtsMQX$#xWLe{2C_$k@uWaE9jCjRTCW z3A>Y3u|#oe8Ci4Lxnwx#%qi)46NpB}N`X>$=(<;#w!Q8Q@gilVpW=C+T571rWb+y$ zc6RP$koi_6??|TzezjwZg!o^*c4a#rn|RtYKLb4oJ4E82`UOvDX4!swU};Y|mtXWd zpq(FJb?053U$Hw+R2TK{MJjq0NRkZf_{dj(f%fC89n{0VDw!C1?S@rw`4j`u)X6+J`cWvC;eZyP<*did=$1E`v*bcB$Y3J@II3H z#c7hOy&sA^-Zw`Kzoz%duGO2$Yk$EUyCpUXu`zAi9#dTX)N+BFTb?w$IfB?AR+SuG zKC0F-q9PO7-P(k8)jK~>?UA#}ZhCg)dFGLL)eBxzMCJJ}A9A|uj!X9aDiqDG%xJqL z$ot9ag1E;d*<)toqF%`J5a?T4IH!=zk^pW{>Z?TI5g2Wg1O?&~X!^butU^2;|G^s5 zbg0Z{{8>g~NHGQRuyn#@A?Zun$;7AIY->v~g~pz#Pqf}>M@+6iqahU%YUUd#v$ z^oH2+>CohXEAdRC&4JGWC36JviZ?jRy6ic1NqH6j{-}E&F;J*ocDE$KsYC2wd$gf?YV?OD|S`)q%^x} z%pP-DBl5{gQz{zV)0`AQz>IacoJywY16v);Gs@x>8xk z@J1Pw*$?)@C=iLWI>%qNC62hcU(Nu68fkxs4%naU=kl2Jj?NR_LvAvfe&T8(5U(*b zCU^$UGLl+of*u88ED>0CeSmfHzl&x)S)sbL%5s>f)7Br$05DcGGpYPEzFl~hnkv+w zAlB=^mE8c~pri~^ezxRO(8xc3K{3nz!#1HGpCKbj_&Sb_rH-YGWi;ou|82k5FqE#$ zT~(Q%m-$*k20;_B`!u6V50eK`k!>l_OK=Ap*18?rI<=O^^U|4DLs{AaX02=L6RY4?_gF512UeeF zqx9CpUr^D;q70&AQ#tW?6X zae$(a)#!|QJx*;$cIFsJl=b}lDGD#q(~=(|`H1Z_1T}K*$88PDXaf51SKP#suZ)NL zX^I!Nr(RYp4U!x_3YQ&_aHl zl7=A2O+Fak4se5l<|$oq_!H(Yu?F?GxbE6+8QqaAw-Dw>^*iy(F3#h!r}~=u zz3)+dR}*J8w=(|2vT6RKXtl@TTPd$(eQaw7IOU}sNmBwoA4{~uz_9U;k7eWRABh0I zsVRJ-js$Bze{^Y@-1E(6Tr&1Oc13zBjr@Ydq(MU@%vsgvspjP2D|X`nQwMzkjFN;eYym~ z47smBhWv2qu8&(kZu?07k>cY;DAp5fv7MTCK0SV6+2S((P~@=aXk!dIhAuWbPmX5U z4fl;_eRdl=FvR#Zk0md4*Z~;p^hdVzdJOw>(Mx;rt4VpLdu4%TMX|54# zGv}|2cLS)9To0ge;;A#6+j!#-QPp0CX-eIpcDeC#ASFd>49R1<#|)1dAMbw548;!H zJVL;hf}ZXAPoW+%P9DpuWZRm&!cv9*Q-hbgTEqa?15O}*n{DxAJrTtQ%K&qc}yMfE@qqKtTV1FbW0SPl-rHq@?MiT(%a&?tn7Ch3Dt z3qg0Nx_iNQ{Qkw`GR~SLAE`cW5A}+g{_HQ~^&vbg#hi}d%n$ZTl zep#S7-s(HS^?$MV-r-pH?;kK1viD5($S5LvZ;2E_ifq{-Xz^LWaUw6G> zYms@3bfE%g=E-Lfm(JzvGnGZqUrtv4{*M+R|Jk5hD96IZ^1Ab<QE~r99E;rV>dk-wD_-RqZ_s)FV(1}^%NyTdFci+lXF+3b<|<;}01A>U z4D-iC;YZ5x_Zi8t1LJW@!B|?Sjsw~|>Lc&C#$s?ENL?Y!3hJn$AvEucpxeHNV;mg3D-vZ34-yJ+f z(b3@{=}S-teU$z16T%@LlALn0s9-PfnKsrhgPqMsQS(3_z*}Qf_!;z1PyGO7vq)u z!W-OH(XxZ%RaU+q=h^^+5PO%Wy_@}vsy(Xplgo!?;X%npp|iLqa?_vnygvi_w#&jL`GX;Ij6hehLgm2xi%Edt`T3r$_Cm601dx@<06 zqo4YRT(M<>sr--saIG{JwewM%e9@(v)qDH=+e35RzF`UJTv$ z7**aX(7-GCCa2E!M)P350%?iT#JE=EK{F8J8qBC3F&Nj_v4ar`ecPnkpg;T^<7A&u zc+!`*f4%%l^}V^^pF!X~0tRDcWE^jKq(eO)9>G9PeK|3Uvm9tBo}+{eajtf8cESwL z5iBD26~oSV6%(d>(lv4=Q=~4!&-C%z4yT= zy7(T;uO>!1N=qX}v;zJ^9EFGQsBTLcz{uk`db=i1mGE}cU}z`)$U3-2B!O3yX4C%C zIXP!1i_IHFjuvZ5rL*kjmPR$W)>g6v`d`^K>TG)H8sCEu1 zmkRhgQHm%Ic&Q&g2K`Z)@B$@4=yvFh+^(O(9$!zA!mp1T& zE?2RT-jXkWeJ2R%=*Nn~KI$HCrO8XavK{7&ZR3TW;+HkSk+F)nFD&@(Q z6K7J56t*W8o$1Wh5YFT)xnhr|l}N%*4;7xk>+ca-9CrWE)Ws3zUG(s$`fqw!B^6&+ zk~^!eEZcd<;rOmMb7R7n-on6tiJkN?>Yyz@^dG*&CI?oOs_IpdBhs{SsWs^`6xdUEwFq(YL z!pIOvL9_TO0#EyipWLXj>}W^N@Pq}m%a&5!qwqeT8IUz=J<4;IW01{K3krnR_9=Cq`1E3m7t7FITd zy?(qOE?U_sE}-?~y~!aye^ghd8ZzeI#&?`Yp;ApqU4|GHSf~qT-(RI5A~%aj<)yq% zI@^|O9k3TZ{z{xOQ4H5iyo#u4q|y`5*qFqT=F84U5RnP@(*-XQ+e8`&MRI+ zqmz%O<~6}M*(g8o?c=VbK{hEtaYdLprb3L(VE`Ou$~0Y;nJr z1zZgHRVnW8!jd$`HO>hT_u*P`60?NSBjd)1#c?`6!kWKv8Qv5$8K z{YF$xez9IM(WV{;tiROu?uw9{Wy(?xat`0DNSVE&suGqpaNTUih<00~Ey7JE`swwJ zUGsZKx}R0p(6+y>L2htg{#4t&PEvJ7Tio!d&SC50`v@-k*IOHYcA$}n<+tH7CXTa2 zyE*#2Tb1lYd%a*Ff91m=(Bj?fc|qc*iqobLUhGEyBkCW)ZTe1I>I*bJonDRvOd%X2 zg0U98HJfbLnOEC{{e@a4M=CE5X^8F#VqS;&Q^eF;-hUHY;G+^bM;ad`a3M@jb4%GT zVlMbYXPI%`sMV#$ji9j7ZL{_CDT}@cVb@9rUP;+;^7c+N`0Z5hj*%qveL2i6;ld@qKStU@4k;)szDb$ z$v~Q0n|8${rz{_ow_Xv$_tPZP}p(o#&QOf2E+pZq8W$Ge-*k)0Di_p9%W`VXzhD6IQreGyU z{^PH*F{RCb%9?sQ6susdagBUfpE#aXaAO>()oj?EYl7V7SWC`N%I05<+g#1INj(;` z@nASit5!{+UYV+^z$}oR_XQ0P<$Pqo!ABwG=Ou)9VZHQh=1nK0RzXuq(5W_5P~A(4 zXbs`s);T$qTj@)S<^aQKv+okuXwc?K^l(T2Zz=x%tt{+ z9hJD0N4N2sE}_&StBMGBSRBod@Den7>rA7ZMfXJTjNe$57;`xXol>cqMXPNbhp#IU z5!Qz)ifT#}X;Y(wCW7~Kgd^Vh3lWO$sJ-B^zYqqRb2LSU9B>GxYr?3{i_He#W}bU- z&QIkxmDPjaMO8~NyG1tPVlIsmP3TWS?bqTS!nk?t&>D5eQ8ifj{QcuLvAypuEaPxi zx$tB@_l%e)+tNtFBma_w!M(m4x~U9xG={nRnH#scWW^T4Bg}j{kBB=yc51?{a+BN5 zG8>NvOOOKXM4Fws%k#z8x=1J1rqc~MyA+1COMP!(Qhq{eM!CR6egQk$AfGKT0C#3} zGMbgFOhon?K6UgnKqX47_(0jxIkADq0a}KS#ey64%rDto+L&}ms8wRZ?|=UkEq1|_ z9v`)0tRmUm8Ca-&+cmomTQpOx^&u(QW;AQ7c+z{0Ohv5N_I{ovG!apF)adit>%~?a zYanLeSUs=cB6(GeMV62w9MeiT+T5v9v2ZI|%#{Y)$X)`F4}W}yrYlY*=YS3(?EZ({ z^}vlOn-bxc7!hnlb)njrU8Dx5Ms7f9^8_2`%*u{p+m#6I!dI!*ln~olsUC~*(oKEB ziOW+P4amt;g72T-+$?{y#27;n{U{*;IRRAWAb=Gb=V^)Wp5>}XxR6s#kC_Gnoewn`5ZkU7$_l}vrH|8n^V z)3Mhn*@bR~aW?LJjxq`?V(3H0nW4okBcezkzY#E*w-LP;A|%UM1Py#8D1##K!epi& zWY@jPxfVqs870nP6FK`#!=Eju=`W{p(x@m2>^AzRH4dLNr~OZTf7~((X#aLiXxh`t zkLw99%JqOaioz%iOpNrSH>xf7{IvAl`o;orqdP)&6f$!;uu!#NM1Qn(RW|@QY7YZ+ z{&08xuPe!umovj@4W`)#DrXN@tLml+>{CP1D>bMhVzB%+bn**PM0;<}(HXY|>{kRW1|hFxbgvCEyDE6h3idRb-Di3^G*VoA~6E z6a{0exw!$WPydGhH86y@;FBN#)*v5?8HnixM_lu+7t^mPq982I^-VoFprJ*(ct6Pj zvkh=ZM&p@H=Gq!;94H|d8Rt1V-G}En{zc~vRHafJpZ}Y-V*^-YkWgOI0wst9yuX)m z&&y+u0}fHU_&=$UPZB{sDca9m02#Rb>kq~g7jZfCo(-iJCF%q_&u2{(CSwGKskl_C zSWQZatd_r);&{XoJ57shY0U`0;O7E?Z-*gH4 zGkv{+YfkEJvn<1^T~*Zh>xc|_{Q}bw|`4ioFkYy z!T1Ip;*GM^`y)h~ZbXLjYrNX}Ki4007(mbI1*5rVCnZ48$+b|t`{%Ho&LbDlY^EA- zm;5Iu9E;)ug3S0#%ew|IC_nX3?-H@Xw4SFK6izpq4u8rZooJ@CC;!hyAr&eOdK9_= z0jx`O@`8YPX!36!&ejc?^2MWcngd~4;A;*_$(|o*jKB?|(r9NII27f*pG}~QiFK<7 zYH3&>X=_gSDK|x&+xypcD?=WJE5ouG1Sn*YfXS2=h)N5=Qdh6aQjJe17mL%63*|+$ zbh{Dl=KKjbt9q55RT6|fJ@8=vkw+k88$=rimS?<(OSRlh5G`U8xRQ}$g@%6&!T_b{ z!s)L6>2rk{L|+XZT%7USdAacCTfRM>%HP{Qmb(!|A$8a4q(4w_5qWjHsl9311Kc?| zvYTW&OpL1B8cb~Ji5-Z(A4EE@AVdsCg?geXmj&SB5uiSx=|Q%33Y@M>2xS9e#npD! zzg~iEm(Y#-M!ezzu4MR?fyaC|qcAWxDKu;Ldx3RU($(CAmW_>eP%3lR-XTU##@K31 zG6aPvwUQyp==%Cg@E*GbR+Ub+H*Y5g6SliEN!O=cEo1C|-8L_gPm_<$gEE28=Xb ze4GbC!3%vBNQ@CMh_oSF_T@+LzlHe~Eb}w$5VNdtmnGipJ<=cU&Pe%cyb#SPk8?R8IGo zSq(zW&qZfZ^)qFc@A8ciMzY37GKG)R;ouRJDk0QgP>lVu*U3Hn$b+p=KpJS*$qsu_ zINr0ikR|H6>J2PXRyNm34fa4gf{~J&S7=!`{t8sM**Az*W7IT98GRUgc`MFz9A|!Z zh=96wfSG^Q%S1>z0(-Qp2FfpszR3lW7rSGtt!-SwO0)Qu|Q z7A{m$WNv1&C#}9=hork%eD$nuW^*Ti42aG@sZmCj1JRuNtWa`_pW9LA25NeWVcWKt zI(Id_GOfTh2u^{3Ls(aX%@CE@X2spL_F35+SR)U9ujsz{Q9HE&bMIsrLi7Sd?1dlk z2D1ocT^ERPe?$&oSYs3`zToN|{I>-&sE zxg7)(d&&Wg1fh6nZ^Z#3CUB>QVoyNQ`A$~SlB#t0^EG&jg5=sR$Inyj(*G1O3da0p z4!qpo$3WqG?WPkkBQIcBnkpA#!xzghuzhq(^|Q`*6n@G!wE0+S z8=L@qvhxvP;sfx>d$bkoMNbq5p!ejJRgj@wr>!9j@{I*2O$Me;+*zZcqr|aqh*pN$I8c@hXtn z&JPZs+x9rIvrK>`VNBtf{IS^ok?IPTLIW9#EJG!htZx&Hxkq@g^PO}vwLm|zyrW$p z7F-C^a6IbI&+E-*MG*UpCsL^H*Die~Uxw!#B-MiwW)rjohY08io`oJaP+SWGc;aSw z9HLw^0T*vQ;J~bomSteLXiyhOz6&>k?6v^tp5Kh;Ab8*-Ud(x+6J{VwA*f7@kY1CI z_qH78C~zX#XT zl1B+uH$Ku7tbxXMUs%8tCfgZ#S%I-O8tpi+hhF8xNSTDVqd$h3jR)l|pn)dWR@C;_ zDeO6|MOHxzR0K8W0Y!ZC+Ilq*_TN@ty6d^TRrl`_BkLYwH32WX5d9Sv8P4(hiPwYb zHxb0Ct{c71sx|#60Z}z4Tn}*HYu5A{n7zvSSm^??XY{~ioawi0Q({#%;wVIO$yC-( zI@|ROXlpWu$g4(B$ov14T0#t*=R zcZFN%brw(&_+7vNy&(BQM7SyFpA8}u=SjcP+x>ye;JI|@4)Q8W#8`}KOMEHN6sS43 zp7^6wFl(>6jEh15dtf%R?HRL(*K_iFcqo5sBjOsT3B552ud%4b4n$-(UN5!y0<=rt zR0a_R9jpQ#{l|c&*%Y%}d;yJc#;hbcREDb8LHHA2{Ys4cqNq7#ptBGn2@fk*L6>r< zJ+_lG|6$PsEVgx!3Nz3E7{C8Ky9v?*M!umGQfJYgFK68z0=Va}ABd~tlnLK%2vxr% zbOIVcEuLABg{vUWDqEF^j`1bdCX47aqTfe=ryK<*&rNz>>&kw-9=e;@L84%}_o<+m zxIFlQEr5y-NygwRtJn2|bCYJAl2Xv6$B?qTvg~nS8-r`*n{%CH+>pu@=g&pRlD`i2 zQJJ%mm-AaNK{SperRO#%Lgxfgqpt(E-@UyPKmEGnpPlH`E@3=7_Amw31qvB#zqcI& zf_b-qC|l96O#T?*E7g{!CMfv(ZIw0MDZ`oZp*7Z=(I6Z>OR;yx8?0;%swzP{&Sz3z zIi{&foq0WYw>bRl+ih9MT>&c`rZMhQLjq&H86O(W0H~n}cb4bY;rFG0v&AFk@UWNiwIystk{1Oi z&^g|t@V1Rtol#P8RdggT$sxx+?tv2~$U8s~8!2+8d0Kf@R{ zv<$yC-8^5Fi_-3ih36*dNnQp$1M}K%KbQw}EzU#*IMYuR&9!R;P2WeAoxg`5wV&);EA5w3{!T!ROXh~T&KeF zk7y^U>C2ZR$0%d#qPhP4B|6^2Umh5FrQ~kz>8Oq_MDuijJrF;0?e~k9GLG>XdoD?S-+P!jP>4|~b){qmEI$@RQX2L1Q|PCO<@;=}27%e;aRGSL)dl5w1c3^Ss!X|mM$-e{nQhb};qKGdzh||1 zPo-Sy%D5DV_MbyR+#FCT@omVu<=IU+X^Lb`&wum%P zU?osxz9^q&ybn%hj~wv7?Gz~fy3#^uDan6)=_ zq^S60HYJSYkIJ)*5~B~(3|aHa+W131YN0RDnk=3| zV{DU-Fhl>w*4ODMeHXr|E2!s$<`vwv7`X0XW};UaXmzAp>Uu${iRY_?hzSE-aV1^C zIg`e>R>fW18@at?LwmF41(?{g<9Gr+&be$_FE2F~A@JWH!|1ZPu=HO(#s~^7ASj3zIeP0J5wQ7)aQyn0 z@`wq0U}~0N!d?FEl5?!YqfQ9Bl$>s*oZClJX{2n@-MDvuAWD-$VwYu&g&I4grGrIL z^8sP*KN?+vD*;s&+7^4Yi_Z5lfg(=xX5soMMBbv1OU zn!`GWs3cO$7u7al#GDF!N|aexUnKCKYzSSaz0}8wdK>#UJ2&s~V8=?x*yCTiaCfP0 z2KVC{31|a;t^<3Wpp_}j9}&!gTaQ+=-!aurQX};3-xx%@wQtTd|zU*39fcyfw`&{Gh+4KD@0~Z&~P^*`x$G@IZnMZu)io^98 z0w%pNQmu#uUik$*UpLU;km1uW$lmGr;@=TT&9eX+i>vpgqW-uT?ZPwi^F-d7_}_!4 zk5MoWB3oGp1wao%ZMvC=Al9%YbHQTw6ww>P$N3Iy{YxJXlPg!Uztj~f>Haf5q98;H zuMj<|kKkx}9d-zAL;?bYoQIA=ZgZ-LMIcfsn};0(3*tR+t>)$8B!K0dIX@! z70F6p$qe&Pn@)l0wg6JY-}}8F{k0yC1p9c0vs{sG1rBzOcD zufvkU4ce`E5K~@(F7)MCt2ghjAk^aPG1Ek-&eTjnVB&f2jYZoTg8&~0NDk_dFMqa2 z6_D20HLt6c4SoE|RVS6nCf~Q$vZM_lDWK$74-{ixiJ$F-9Q8Nz3R>yGWbPc?qdM#d znNc&1%(l6Gf?U|Kbx+Xf%LBNuP?97~yh%rH`OhX49+BOTTptYiG6=M;#xRN@MXYII z8{qalBDMqiYFD6LT?9+EL|7b%F5*6>6L>ME@@5k*F~Qt4Oy>Kh=1z~-{a;{#J)Ae@ z!e#5djag)Bce~O&5N2E|aWNpmMK-#?Uu`ra#6)zGo`twPt`L(&bkNO@P*ObBr|*_2 zU4DjJwbZ=}cV&PB9{=0Nl-kacE>~2d>_xN!<^`&8y?sUD*z@x`YDzSbg5qU0+Gk)yzXYQ*hmIjow^yKWmp7~)K_j9E>H!~pVOqBF`#2ZKp&pe* zGXOJxSKN2*W zn8pRt8W#mMl(oVZ?%&li>zJp+_-+fw)+K#6>=J)G-(PjN?J;)*+!#XCxi}{ad+b1z zCGd<7nn03#;BXTBx(i&61Z2Fve^6lBt?QEnRV7~DUsa18NuV)%^ptOW&awyQ7cZtR zNRUQI9h%|;T5XG(q{vLwFjWW>V*LwPVydP<@Mr_%jU{k#&Z+THff}bnp3;S(o*ZdR3P^w(j?e?em(1`F=z`J+2Kv{oBYZ zy3;T*eOZ!+sfWys-*SGKJBYyn(q_U$Cyq@S8+Vl1FwhhKq9mAFewQa?r0iPd2^wNgY@iu?vA_sv!fctZ3}(1?!O|!hQ3qAXL70j)wC)a+ zmiON&F)rMNHV*M-ElY`d?X*)y&gjP$m zSI!u6bmOj2b9FOQl%k!~0z=PaCUQI=_G3_)S;Kj>f33w<1Kd>f1LI}>y8q<>ne`BJ z#4B##Fe=1u-g%+VCl@BWp|ab}lamCamFs2rGm^AjfFb^xz% zQR;HvpbzOaETV95D(HDc74JcD_KczU#@C=9T1AR)EOv%K3!Se?{c?`*m^kVQWboX| zdXYYs#|rk#mvrv8c$ZH|=-gN8C2!{3*t8_-?1lcUah|BF%qC-X?_Z7ez-*P9dym2Y zeMh+x;i>H|@Nf7IYx~X4Smu{G=AW>apH+SiI6cG5JcEPFZZJEHLw0Sm?d)UXz3qT% zJ>Ea=Rqswfb*CQ{2s$gfR1+hrMrhZ*apr^8?(ChYGLgM_(hE3yPXb=y=6P2YqDFDX zvjbUwjQe40oeAmv0^{E>A2gdJwv%^YLMy*eVmQYfNHc%w24OBm@M++L($vb_#+!|2 zuly`DsD!OoX|818fH{pi8`Jsyf`3=(JwH%DWb7OICZH8INMXRbxG$hV-{GN4%l$qk-5rk-2S3M(bgbn6iG%h$q<$)987YTq5 zUG_K*?D*aTwbtM)4mZigv$+=oJqde4Q!B7V7F(X+=6(*S9)?Qdu%NB>{7s72__J}x z9lv{n%Cz=2@d9lKE~=UnqSP2JhF5ME>EtOt^aImb#vUUDqZ11s_G_2h@6>V=eg(F6 z2pepBxS2FE9o{k>(j&9|vuJlTexcE}83lNdcO*dU;jnisX~bGd_G{*A3XZ%UsN0d?ErVQoqUN$@|G9>cG1>U>*mstyk;Te zgT_KVmsbMuAfn}>(j?QuDa6YU51~5B-TtfLEWdB1`#rnQ=F8r)vVn+jkD_)o3DMde zSEFTF42)8{j?MD=g0{(Px!AuEmx@oL=SgD(5%MyCLjkAUjCg0r47#GvJ{xNZdJZLs zgAy+O8-N88fKjzN%EPN%+EYf7wuwkYlwno4K`>f>ug6=aTr%x81B%JT0&CO51S*UU zS^($y;7Kx>bStfWL}1<<k(o3 zQ;xv$w(W5uyssRU<1Jo{s(Gsu)z1&O0~7W4$Pi1zR6pgu%l)?&wEh(B9_KOMGFgwl zp47TGP7KOsX)U0&mZxo}v{9#FCqfFFVief(bkS{in<~>vA9=I3Y24dvd}96i{sX@luP5g(`Hp)}4^*Ygk_-q; z3>-?B(nz{=MJ+V%t6n}c=b59fXXs;`vj16f`tqPq0u#sCb3?Ih8o#0o?t2x@qN@|V zI$v>|6)}mrw{TBRS&QNVX%i22ICNSJM>gu_ftF!k`>)A^!T~8$*<_KT|E#du5dmDl zIBs{M8&DJgYVIl?=-=ie`8M{yX<6VF{PV{eDRk1*SF%Ye@m*o1*EL`m(gL~--6yeY z*Z!zTDEwhKFh7{|$B*UDj}@@O-T7ENNJ)x&UE~?1jh-o872=tG+Ng5_!^#j8x~UDzpRUV zqjBG-hr_8csl%iEKssd0>p}f<#}X!&)yb`HdxdsHAH@bOm9I+vdm%`9UG z7~Mt0^9ZcJ?N>%%AlH>K20H9Qm`dusd0<28i#YM;Xovu^*MxY`0o?F>73=|?uY#N- zPge(0mdcjFmO|iyVd;lmWC*$xXEY01^+b~g-Q0Klpw_d2T&(F=F7>l#5MOIQvRkp< zl(OnyVaIsTL3`^x*cP8h(mRi!7+A6=-8?|*dEgfb)_r~uH9GO=V#MD`ZG@8tuMp7eH96u&H2?QTUr5*ox?R( zVwvJWw0|{Q)*JEhzDLP!_7zC_{a=|bu~Tnx5f-pR5%n{xEx5lUvu=1py*96#$wgf8 z9;GR~XQnhzXJHe<42O;GkW3x|hkk+7Py@2h7l6^82rp}i-U1YANw6jADX3rvshv%t zg*fu3U=5=M&Xs%!1%UOv|IubRVn+Rx5OGL=F;y=p-xfOP&5}T@s|hW|b{CMuGZsNt z`f2w)`6wQ*2(WFC;G-8dm<*RDJ^*EhSiuZ^8Nt%2;Iw{dN3rfR#DWv)H@ab85(I%XHShWI#$TGuJWH>kJo_{+lyT)$ z>BHC)kUaDvXcvP!NIHhBW>04Ev-jx*1qKrprKx6^{KEVs-c+ssxwK!#GX^R|x-zy7 zd7nDG-%Qqg*ImmM1TR!%E~7>ERGmu-YH1EM64gzt{531{;^NF-{ULfTU^C*q-qSmmMN+Glv1Y__60py ze9i)5>;pB?0gf^G!7;HR5WQjIWx$PZS^*lBcfX*lOQsXJQ~U~FVSxRs?`{^qR`$bh zdaw!w22{kHl09fXv7O> zRk6PUb7w9uDH`%8KzcAT>-X2uoN<9|`QiI-oUPYzqDYU5LxZ($f?tYD)2CZTlp$2S zrqDq?o6<*2%?;YYMwe}b`=Ow|6#DW*FziMEZjl8XwE#ZEYyGdcVcr)F%u)o5i>~E1 zK2`^*2oFIdp@MMbv{0F12^(WppAYBAchh7 zrW~xxSa5SrzdWGFidiFfoA&ahZ+L;tuB78Qg%?w7(6`HuiLAc8nYL>K4WrW;9R$=~xPFVr0veaxr&7} zw&aPJG)K@Lc@UH5w-&s@oRg@1k%|hX$h`Tf!p!VQ;2{{TpGSNU(nb#N3md#JfyEm} z`S5&kKfbu-DFZug!g%ob5>|K`K`;EDV;bhE>&Ow~FCB;e4kLPoE>^_Tk6p6U7uv0x?~NNFx!Vu1l#gaYh6o3`oEOj0g9!4D6TyEl`Yd~>3cVWJr< z-Rpr9_~AO|=MO0QG0Lf18wIRQ9F+q3#&!1*+m0AAoI-ru3J`DJ>ZLd$Q1^%jv(;XJ zA-}~;&oz^FK5lQo!_&P}8FR*v8}aAo6#^yQ$Y8|zU^~udSK>zBvlCPh=TI6(7BDS@ z`)tfJQ(vl`WV7r>ygE%P#F{}6IxGLs4qnM-hA*b?%l97c` zn)#G|spV^;%UhT#7$v8({=rRypkb$QHSu}8RU0KU7;3PLcDH!`roM;CMewJrp9PF-F80XJW(4OC|d=c18 zD>($Fy?bX`M$-Wj+Baj_v~}tT!sy}cg9zGC8G(>i5ROlFT4Kda6VBEFv+3)b zWL`#FlG9A1?0lxov2`@-=HizwVt-A2VUhUU2GypmyrNSbuhT+)W3MxpD+lZR3Ho=* zU4cLm{gBB}?ruDKlUzC#^f#tUEj#p16@0+r$@ENq_fS`6oml4ZxF$bx%dzo{{oEFz z%;1!iA@x_2-KXZu$LGkUGwbG+h9u%NIu#p#4r}!MD!cnY!L4zkOCh0(@`q$J*NwwF z3{70IbsN!CSsV$%Z9vR3t=y%*vC@FJufa-Q*@{W#K6Tc@PU!R?Afgo+vAuS7gX3Rx zLsnE~ov-VLN)YV0+zdk1yX=_FrQl#z6U~4j3#6eE|4-@v?;hZZ;{Wm;`p@^U2RX_)a35%b@SOr-UYhI^&j6;<*d!svhbHvC>qiWD!Vm)id{1b+;c{+^6eFvIee zKi@I(_kaJ_?|~aEH@LxS$;5s_oMZpjpNLpEs-aCwhyUYyY}hGC#77}>m&4m|4Ad(H z0RASx5UD#|?4>@SrZ6R57!H7`AmXNkxYr?Mma{_U0ZcMp*M^UFrd$x4Ngz5_$?h*C zLhYaO9wB~pAsIf2gHTNV-h)l+H@ap&7SfA(6#S6m3-sWEj*|J2oVW85T8rh-<6_#p z0P~|P=N%q65ySyEP?EYHRBel(hlVkD!yolmPYh($+`Wd^3EM{j)cmi&acLF8a`>;B zqG$0C3#*#*kjtI|mWl6SjTS&WM`sWfVp-R{zSP;9Q-sv#-vc2Eam}RCYJyCPjqW(a zFj9k!8F|dROmCQGNKGen$un!gVRf0>W*ie{&iV)Bh)`@WfNY`$e?W4cByj0zf)5Iy zyA-cqEPceDAWTuwpr&VNrx~BKh$immH}z#~NP0i0 zepx)f%$>_y-L&lkpVYMSR%GqE@^W&0b$s;Mnhvls3--VfPznl|wV;Z?N6_gKn=T|5 z2#GO40yMqZZu7W|-cDv|pt3#LPa(32Ulx1@Cqpq}aS&>y##~7TSvPQv&FvtZ@7L#7)R#{z<~U41|^InUZ-(!?xz$ zq!@^L@WpeCu&EEpdV0-;D9($J7>K6Mgo+#NM}OW@d*rMcy&h6!z*(#A*&B&=88doG zXJiIKkcG8kD-IWh0%~JR7e;>uUI4FMfu5Xuyfz3tzd@28A-pla3C}PhrMQjI!32RET#*Ls;Nu67 zC-@?$09xBuk9P0>39`uhc3SrrkGUR=El!si8MagiCO?QHAEmB~JyfjQ5_OLLnVljY zt>uSQyTRACa};XYsx%+euxaJvBT$DwjK{0xs)vW=MC9ar zj}EboQMIj*f*jw90iPcu$x< z_iL;-PI&3VcH{SqTuUA7)0Wd1{lRLiV7BZMjR?gP6yQ;p9X&IBeswR=zDW-rSV3|{ zkt9XyVjR{6@rZ$?51~4sI=_g|?D53JYr^K{pw%rs(Yy_-X-nbLYn_ocQd6%o89iZAFr164xX7>s;}u)Z`L)bxC>FYV~4d7 z?D6F?$CLRP5hqJ8I=^EaTlLB7(t7HWw>T?}oZfdhJqwpS^wT(W)U8Kh@5I}U$BiR? zJ-vhR&h{-?zjpFZIOl&ta}@(nBtf_l{bjA}xsFaNqKdV@i0sZS+*vPZES&#Iz(#Q% z`{0Lbjk?}}^IQy?ST6wnPVJjsUHzJgm*5$u^u+qL#OpFjJAh0EB+3J5qA?(yD9F#} zNX>AghuF#*tYa_9^hliYN+O2m9ADg9! zxz?#0fp{H_8~Ch>w0z#1)~!&=UnHOFt|+hgdbeq5LIf$Rp~3DM^s(n{D1lN+?)d6OlCy8QB~hu^LyR9HzlrYWU0~ofBO6 zfDEkT+6*20L7p-=Bh7O+?!Te_`;6@2jP53FIWR1?-t@vHqAe~hMc_RxpdQ!&r|Kj0 z0m{{L4YIf%J27PPcRfEDatqy@j_+Em!&3fSQi9S9|O zx62+6n3+i!8W`Bn<$v-t`Yp_r721ZSF;rt6pvZx@Qm6Ix`|qjOo$bbDY9%`nb$4#h zm5t`>KWaBFv3{1~fV855f^WMB}l*Dt!M=@>b zYx@_!1rUIf*pJG!a4X|GU0?K$Exj7qc`PsFGUwbUd9gUwZt7>Tkn+sk;fzwIv+~2v z*E!G0$?bd93XFbFbx0(o$Y@@bB;fK-%H==*M5Ms-w&`%gA^1P<-QQo4!P_T^fmpr&`9pKu7&eW@$;8P&d8PmL zpMYTaG%_Ku-ufRu6o@7iBA}|+DG>i(ulWochAjezBpeCu{cpdC%@hux{iR$9>HkW; zmLw1&nEU@v=l|vId=N7IlqY_bX8iu!A5;cEzJ2?q`f1E`f?F$jyxE2o8<&G$v6*=& z>_YCi#Y$ad9g0yfKP7~HI$=jjz<&GnsU3X%ebM6UTPzW!M&o#@|9)5CCUIPZ`yX;H zF8_Z;g$MkFedhn`KXW+@1)hZUz!EltG5_mVBlkgfvUIZkZa2GP*9Ec=U~nMSv~=|z zJgJZ?yi(WpfvB4gXEQmomP|!EvQDT5Wup-U&-qn5n{O+;A0X6m0+>RG6Gp7Prf+^j z!{n|p5`ZM`G(PjF$vOgFrzGP5zqg+So|GO^yL3COKNw3!U-*mI^|XLM^qQtx2@h%l zmrU-r4#J9ml{LqJJnxDGu#~Bc^9l=6ZPLwml3IL4oNMD@lHUuIg+- zxR}up=)8>N{H)eaKXgP4jFEsRx=W@`T2=PDu`)lOfab9q=Gua9zrTvggy^PMkWEqJ z0Ub(0FqyPQ?YJee`>EOSw_8&9?g%_N0o_XBt5vJ9Ky_^y=B5)n0hwRkfKQmD^Q0B{Udm_Znf^1zu?bdKKY}}i@F~(=buqa|BUAi*|C)86do1li)HNGe3|~G3{fV%@&!ac$ z_>{ln_K{SLNSyosN%DGuvq}P=$?)eJ}MhM7g*r(0 zmR9sN4~v@Z4j?2_;c#Mbp{k2TWbxfbEAb*|SQ25@(hUK+3rOTDaD7bSPTHQ4YdZR+ zg&-T+w}GW)gUCo$q&d@!YduKMiHq9oLoVIN1jN$`*m#byWhQSv77SrO0b$#f>r#s} zx^YN$>z^Ig@uT3nn|)dZ=7p(|vz>BOr}! zv3hv`n*ec|_#L+g?6sRn!{QCmc>Nb+0aBUX{CR)uGwh?kfP$RSh(xIU(Q*MuMRVL{ z+B{hM1&|Fr0|Rat{#UG679$|t*_m;#DQE2m$BTPFEJ=NE@++m8<`^$8m_~XdPvpAC zj0!%#$6Me6^bm@gozd*39-huutqfh$G8i!Fr!;e4p>qyWa^gXzea| z3!aMFP3b2f<3oV7UJ`jb`JEq7NIiuSX6lxJacxbG=2YGD;R!xdiL91iKp=Y$$_K8~ zy)K158|mIqQf%CW3p1nmwux?U!~EWsGam1pr>+|>xxh2$MY?HtD+bLutsgL0=;<#m zsmB7#UeEl%T^yAU`br_%6BTqVI}e}HGQXtUr+m-5;@~! z^&sugj+-(|3CZ8wvWxWmaG>_ATF0X^QfwwaSozKlAcN0)P5}i>n_vOtP&AmukOhqG zB5zBM9!#&k9P>yUn_5GV(2TQ}%^(x(-sPq%Sb?~N_1i5Iyr0YSzLvjKZ+_R^F5va8 zCo-)pb*&oj@3Y-AxwE?*#JNa4+K`!X(SrD(C*MS46hn4`fD$rA0el1Z}}xnckh zfyANBy)%7=#?CgetJk6wj#qEoGwZFG0#ha1rh6w{Ck#>L+H*Er>4GAc?y`0?t*$+u zfI4~tm3Exr@WEx+CFR@mAoRSua*5>y<0dGolLZ7N=^7Oe%&c3!=E!j9=wbV+a=Q;> zM?<`z4V~1Hc4A z4t)NQ#wnelDv|ZY;;Tnkg#6`a)iMoUKOC;FM$;cR>t12@Oj_Y6Y{0Cthcj`EV%k88 zdei=g_iA*!_tsLHN<(UCO21Y^-(|<9hBJ}A!(BnpOTf|Z_zflGew`LIzt-%m7UFLf zDIw3ZN3Z=;iqfS$<5AKR4g&0xI!B70&(EAt2Z3&Yq?R16apWIJYAlsT>IbDW*ExE) zD%m=(1cb}C=AG7l^?OuZSuE&~?mX=1p-pjUC3%2<3QlS1n&tK$!yU#yGPd7kkEJn7 z8|B-4uC!3$pFj*8Rl-Pwi7W2lAO2poI-~s)oVBaN-op#&bP=e$_6aM-4@!Tn-Rhr@ z;EWa8e)a1;Q$b~{t>5eCj*~SdQpY)`M?6oj;k$W!p%I4~4x^1JBOma3HzoFa%Jb7o zokUuF$Mw5X@;u|G;c;@YuWsM&`o5ZQBo&Bu}&=Q{Sipro5Qk5Wt&zy3e% zeR(w1ZQC}YNSO*HnL@}=h75&mo+3latYkdG6awSU)jp678M$8pxD&6~&q*RAFGd+6m} zx7J59ki*v9-;D#KL7XySDj*$M#(7pv11+p&K+AQp`&N=!dPw?Io~zq?J#SHE8yh^y z%bv9m^Q{m0lV^OkJVb6dgQ6yhEMukg)qp)V*|WPStNXP%j0qg4e+MPZ*^oC=Bsqlo z#646JwatR7`q^eM{yTG=>342;{mbCF>0`H7q`dP&1>?K`yGpt;o+T)FRfQGT(Z@q5 zWBmP5AvLMzq3x;;i|K_hI5}tO zTIil2ZY}!#D>kwtq!hYIcgX}#yW+K%d+5D66~(>ze}W$KjN^rl`GTFKYu?@09eSbn zG~9x3Cl1m{SDxe=k1azuhUdq}k`arkAiGLX(OZeFkUF6tew@5xes#5?XNSE*yy}^I zzbzVfCxU9IJ(3HxdvJciJ~!Gm_sOj+GJ1b}vRa=2p}IRmoZ|JCN>PX5NdosVZ~DS& zzB>RVT4R)LZ1D$>iY2Ru;uTOL3}1@|<}deLIh9QH3!mqQPRzuL)w!Vo{^8v8oeHZg zY~k&5HWwRYNXYB-;6W%OP*3Oay){*+F+FxFE@`z}D!R*OK){{<n(ezi=S#J>q{s*=g@U)5#%ugqqZ$Wc zGOJ2S!pqi#_A9khA2RyGq>hZdp?3)_?0peFfZunkC`pF8!eK~CgnZOBw3=)DOJeVr zDGW+yQpBG2sC)qt|6o32lvw0vxq_0`_I9Ox1^!3kHO{B@G`OMmwn%N%^lC+lNWR)j zD*@LpWY)ib8LqkPr1DRiS6zL0LR0g`*xbbA$KUom1bf2ly%WtEtaB?+>CrlFK$+(y z<=!g4gSP|$fs=O-V(GsL&*HRzlzG;pA>@J8?XYV!WU7|;Yz;g+IdBS)3 zLf*m{ljDs3@)p?e{eFqr_N#%gTDdAg2aXV#OUUKE-ZIT zQom@(y0wIvFJ{yN0~j}LysGb@wUhr)^G=@?V}mHTCun4_#-iFYfZ_Q!Dx>p#>nDbz zhPLH;Q#>e*o^5g|`88aN>g??L5?w2^IdcLMR8ziLIHq^W?q~|tD#MZK*iu07MD4;* z8IZHm1MrP~zG*cFTt0b>Hl*@=iUcjYph)@ABa-u#%Tvq~yZ3=P4babfiQDm-Bz4CkX?CHXv+L?n zMByJfb4qjU`p8ggxer_rtU5E)T)zA7QwOUDONgLR-*&dwmymZ{%sdbvPTQt-4mMb!u`euvUsT?GYCjdOVsW_y76=x z^*tnkK8+k6#(x2F{?l&{RfIsbHG!2EI zF%QzP80eiOiL!dHM3&cNNy2dDvHX%&I)P~r8vTT9_H3^=gp4KdJ^lcCSM<}$5RG*f51($C+_#XX!h=oM3SgqS6e zS2^77kkIq{djW~lxcw6R#rTchsFn@7e17v|&C#{~Ar%{V;5IV%5Ql)I90P9Pk|OEy z)KeEi0X1v}3;Qvcj1k$~tiJJ6$d^~@`p1en@tTcLe3~H=OdqX#3qnHNflz1(nW9=Q zdL7M#LF|Y=>^3BO^9lY(sWSe3E?5H5zaZ2Z@gslrNNv zHu8KoGb2r)7*z@uC6^$_r=K{P;$hKWb4W{qT}*Yv>FLvsYX>cFFxi>X9{*K*HhJX5 z9<~ZJiNShrkj%D4%^cJ$XD3!Gp!Vy@H+KPId-h(;N=ED1rook0vA}@G9q_|TCJUs6 zkKm=_`3)`h>aUz(8gd;EE}g6mqWW$y0Hxg3r6V8%cWZ39l~Y)#^-j<&>Nq;JT42Ym z49GoV1EiPM@t)v8;k|F;#UbbLxXnRNVn_y+p&g4=lb3LVoaw&0>dC7C4iz+crIY>a zJ!P(wx+Qj*C1DT^I}SX?yP-kIlF!QhrXf@1I(RH==jHSdF|f@jdmRb-szwmbobugX zPDKDBw>KC<=uQ+Qw-^c$m`1=qI8jVJU)1Xf$`Q!(h<=~_?H=(;4KAt9m=Hqt%Y?~N z(Mjke@wSn3)QAKL*Wq^my_;(|@ox!^vqe3)vZ9cPek}0+nxM(VMqtlNOZ)JQ0Bo9B zHA}GMceA$oy4+J&krgOdksDr9302m3WE?#3vI4A7FhfDZkEux>>!Gfmt2160SF1j; z7&q2)(R%18shagWy=h@e=ut=(kXA4RIuLs3?Y@*M8HJ_6sTOB!C>)Yml-%7rJjL;n;57(TB@^TX zfTofwvicqcZ^3fKoy+7NVdQ(Lv~ux*@Y@; zMaxzS|Kr`;v62q%WpEi9rS|=G-}C!63>>B!Sm^_M5-J%6i$jo zO5%lx^vXe@ie#^;obe9EC?87i0sBeVea0X<0~)tW z0P!S}nY%qlV2q)V8($uGS({~(*O`5QfBO58xvUyGD82aRerWVw7SSSiSgLK;<2H(? zw0eWfcyv%Dp2~THx8CBg>?d@!x#oGUII z#h9@5zT#{5@g(_#@gtA~5$ccP8VJLSqkdZWB)k1XuC#`-K52>T2Jl0Nn5^E>m0)@}VWFIZaJa+N-2x%c_1J&xE6_467KQ%Efm%m9kcQ;r;CS;t! zS~Fr$=9$uSOA;xA4j`RW?!~-0?Uh_@%6*$iv2Dod<3Sf#K{S{U2Ys~4N&Ya%gf1Y% zT{UKLZ-pm?loqTDH-G`#zaGYTn*<&_RX+$jButWu`kce^T}LQ*_AkcyDQ7+Ag@ag4 zr}#d`i9|NEmEfaroD^AbGrKghTYpC)98tS))Np_OBkUZ(M^;!7JvAJ#XJX>0RHTEk zPHz-+)#WO=GBb0sxoTbQH9lmZK5U-N+8^3{2XxYoK@cxA5un*m!uY=E|APtx;a(BL3crwPh_| zli18RXWsJuc0a(>ah?A0Bz6|N&>48{_cJ<2cW}knq<7R(HydrAGzx3D_`GXzoO2vy zuTvS=NeR1vv9%HYf5YbM6#e8P5=1$>|BmMq`40l1|KGg^`>`&AJ=!GHMLPV7;B9Xz zZSkH2-Yb+&=A<+{mm%#kkSlD1wCl4<$eU05B?>h4ZdG5aN@>fyW`7jg=>SfZyGjJ^ zrdR*`+bfibp4C)S`g`94!-g6k|BqjD<497XHZ7?IpSGbPufD#%+f6Je^fJR>5SFR7USwJ3z;VeV*!A%*+IcJl}nOco6a97F6O|NNg>z_)%jfb zQjF>X@&^ijKIfQcBVu%E9Mx zV)R-136Yo)EQM_E@g8_uXI87h@2Pg^E)O3*#4p}pDXKpHLj(ngG$DOsV~%I1iNSzi zI0I;5Vc0LX6CdrK`fSX*oz1|Em;etAb&bIElsaRHet+G=C!uZN>|+v&VIwJ5?z){! znc*}@isHocFic^!WN*3DQD-7zYLapbzFtXc!)Q`}sog*Ce&1ghI!L1`iW60wQSQClQJ< zAO&`ltSv|vb0sQ@8g~1~q#t;p1JmjrY?l*+eOq=*Vp^&4VTmbEtT<|)NJj4We@%({ z6Jh6<8ZSw@nuq^+v3hW^&t~a*b!i=)?faa1LOdvK-f8nWl5COv#06g>e;{95JzC~1 zd-!2G&2FE~jdaGl)VoGYo{H`JeLv~?nN1l>O$PEV+TwunY(9cn4vfu+*8i8LraQ^*s_Hf1v9Zxu$)V@}WGO zk<9R)kK0P_@%q}_k^WFFPg z7Whyn{oA9i=j8s7naWd1;n_BT`~Nd7L@)|;5zUZKWL8up`pO)6nVFbdz%-2sN-RN- z$;18aj`3^jhkmLBK%)^*6nrQF21+36{t{U2H-m6D zPr6;x=ps{CuR|rdr4nk4HC~SS5ooV@ zV`F2@1W%xb=~e5<4#$10)Jh23ZE!YB%S zNgT4rg7_Lp91+EP&L;o9mel6>R9+E-J>?&8o}bTv=F2w|(LIib95X{tVh4l8#o1do zc_ZFaM#)BbFWl>0xYs3d zXT(_ei?VTUvnr~q#)?sOSZA+XRPT6j(IN~-LPVtCJ3QjGpxm$@Zq%D|aq@+nxdtM; zL_g2}$k}jOJ^BY-O}Cpya1Cb7ZdqXMeL@MS-7 zmn+lvHiK%UjoI_jhpTifq2ik8U^c4Tz9B+org~%QMAH8Nlu=um!J5?Am@?~2LJZ0U# zdEY@K|Me#~b!9*WiaZ-L3z0oo6Y^Ndx&BA=JbV&UWBQEs%EU=b<?Lj$3Z@gYDy4Tr7a2P^@6V2pr~4V~1uO|mnAS@@l z^Ws57FA!T@QDL;g^2f2ZP2TRf`9oE3m^4V-Q7dB!QkCU~NR!bM5cBT!MZwS}N&%@03lxXp?u#s@sbHIfZ zi#vE+T*;_peqFVN7Bw&$710nJfOxV8?=9nO>2q^)g&W#nNb<6N4G8fjaNDql^lU*> zrzE0a2}N7-Hf)^ur*K5e}guz2{wrRyD?qnf?v6poGjj2|YL6tso{hfZpV?y%fjkSk$mgs9#!R;` zR9AjOEG<SVg4L16d!Nqr3O6T<~P%?y>A14mq`W4Vymu>4m))7O7l(MU0%^ipgp0S(0Af8 zhiQo7{ELv1esQ=6+ue(ZVuNz#!DXt8$Ahs8Zf#O9=5L79hoB$fYpGAUO) z8w@CMo#Hd4%m6lSXYsY!@Z>DXm39TIsjeI(DMU7_IAe2HhM<8mFl)@_>QFIq-GsQ; z($5AMc_t~&ZYa$;OPD_1^jW@o6)gh_1?$>IOuK>vFa`EY>D^z2*(_98toEbN*1s>;$$h&Sr zjGD>DZG)hrCGRdwg^Q4yh`Go|3@*@q7|z`NUx@Kl-YBvG4J0kE?$xv4*}X1B!NtU_vfs z;p^Rm@i2Udrf-j585e@X#Ic_KAtjCD!vSY2S4gRYwTyK-6ZYe;3s?9w00$Rmd=+JI zZxW}DT|-b|IS`6}0?d5z+{6kF;d6)spg<*AS9wOcm${?CVFf3O!uuC1=3Q8R0Y0l) z;$Y5d3L3|1ov6rt#G|-9bZQ(zaHfnUMeKi|Q)`s^(XKC_?2;Q1+OM|plRtzBM?QmQ z?`Np^@aduwO^vIY^%e68XRf#8-C-0o;eiBU+&KySbl&Gnb^F^;`q6>|c#3MS#klYF*;9sOos9I2iJXi3>{1k83Np#{nzlF4 zer*+X-^Ke`2 z7V(4$ORtYiqk)^CwM!f6#`3iOQ=fsn7}?3nIzMySY_qwa7YiyUznp~#Z*tI zJwi`|F2wHWeqfu3FJ5L@&lryN(75^eWYO?>s_4qI+4byn3BWz^yx$T)MmYvgZEUKt zy}~8=#{noBUj=4xnO(8@O-OIMK8!gsvJKBL%`_N7(dq19I>q2cY&xF~HBcw$<2!Qf z{|E>ps0DFnv#!9k&UnyBba0h@_`{G)Osgzxw5<}(lcY-w?N|DyWdEpMZb0}@NlzZt zRUZ`)6gV2{MdR3U1}dS0vS8bZ4hZCC0w&$>(!npj$bfTC?;1h%@p*eWHF^GjCNiaH zYyd#s2(=-7Mem|D%uxB+0xbj7!zqVtb$imO)e#;>N1dh5T(A#1N0>?G;1nBmP~1;C z)~-Mu8!kg2r`WB_g3z61ys*%{Fe<=(@Mz>s8b?}Np2tUx0DOH=X3JxcKnkN2cu^z+ z#2DFC+pIxN$yf7x*3k-anm}-^W@QDFoqxt>eci|GHn+N9~9dn3)#_7||QimZUfRA`!?y%gHk!i;! zuq9$a7%zvyD{^m=#G(XU#QKS$pZ+rN=MAQ$WIva~#Jfv|UlESlEC3sGd-I{sKn)KU z;NLrGP0}YJpS>MUr#nV>m0MNtcq{utbS==bEuppX8m8?bIIXW&E)ZlbMq$!x2D_qU zX^Z9AMc~@yoCk>gz*7XzSd7PkhL?2J#!mLck!$%|**6 z6aliH(~*B{f)E|_@k^2-GMEU_3v)NDr?D?Y#9Xrisa*KYDh-eu zr^u-v{`*)?Poi)SofYq0F3)|o@vAsrncpjVND^WgIEK);z$8&yi@nOO-{rADVCZjI zdkYY=)!y1rLm)c3d3U8n*R-bW&TKfwP_M=S4BSnIbFg6!jNU7Jf{%LX-U zVHyiaU{<~fu{-1%97fcgfCHAO01+Ne^AXxBmo^*_Q*(=?0-qyBl=iS)MT5l)&)%?v zZ=p(l+EISn2HHo;e5NqeU6Ju*w;Z+Imj)YG_RE`D;R6PH*t|^kiFvKyud(>P>YT(h z!QER!-y%UPeygv|$9m4L22HNBvo6Hv$FvlDb+I0ASVzlh0L%L6e9F^}O)mb~Ls%CY zC@dq+Xu`tE)f;!sLp{=E9I+IyiD7;C%5imi!puU2Rjp$df_;g3T>lO3P zKG!GnX6pKB$U#a$yQ1b=MPv8nt`8U`341JDoz|YpqYASecr0L3c3yI~btC8JW86}q zckM<9XN*5uizm1LwidFNNV>bb!+yT3(aGf2)&%e}mvmrlET)6W%ANh1 z-rdAuz@^Ek#?DX6RZRdJ;ndF7gsh%z{_$+NB#Y0sKijUrA;7jDiZ7iTnZIHpJwh%T zE5G2bGd^q-EF7o6w!Tv#s1p44B*__iJeY^%JS{ww7LsYSvqGDAB23}~zx7Rh(_WgU zLDvOKM11)O5mUtO0p5sHD@jJM1UfU(ShpHr@C@f4a{-jqx+)(Q7#DLNEiftB<5}uS z?1`^i|EpP#K5z&YL{qs9F{`kFM2nV^g{>YN zws#vn>z;R!Vt8~2T}>m|)*M`o!ti%1K`86T_o{xK4aPnYljz3rLOj=E(c&mptMzhK ztTEa=$Rw)zqvOZg^@@l|=&_8!XB}t0L+FK`G8}c$udX*F_Kr<)!00~|qc z>pt-!1w{7{|2sylc1mS6uhelqz0|;K(5Gd7JHqGPg;4WH14d_*EA5TeTjq2gjR*Y} zn62)qt`gECl6H_)-eK4*ejeaS-j#eiwc!(mn$faE^0E}Q*}?oP8D5$fD>^BVMeNp% zFBa-Ai>ge)F$@18tB=L|)kJ|UOdR4w-Bf&j6gc3t#MD5GWX-a0lc444{Y-hEa;3vY z%on?qHCMT96Vkpp*;3x$c~!4IR?YXf?K_$`lErs%GGJ@NE%%ztZ}#71(H0{ORexp> z`W~8xiud41zPp0*cl9zi# zMC6#SN<6{vlARs6RU%bXyXqxtiYbjKfWj?pXQ}TouZ_$)y8t>(08S?;5N3<94!@dJ zDjybUj+$_c3bYnz485i(K7I2neb{cY_|L^WW(Lq&x_8`}^3I3aJ2uchiJC1fcg{Pt zibBdL3>MwcTmTswFaZY=nR(m8ZSot({;x3|b2c3Z>1MMC3`l@7p+x>e zhXXXK5eby`E@B6zSY zRzj-J!Q@65CTa&1{^ZypMK0!bxYRG#Zv894WyHJSmp-6}c`-GU zQT=d-+x3yXgF%2qdWF)h1+MaJvAho75G&yRb<}vh-><^d>ysjLp823vY=F-|)WVKk z0>S?7Y~{*GCP9*Gct&7k-pAkvE+1)PLNNJXowl=W%NEw0|he2R=|J(J)9~QUX!i^YLFBZN7q( z4E*GHhrzwXq%$Ohl5E-gx5y9kVOY&xGWW7;SN?G&qXkXx4gDls%SJ9Zk(m8B{gzH? zSqtQRUFhzjfGW@Z{GBrp&xP= zK!{(j1i!`zhpS!>sk0-0ug#~S5*jFaLk*c-K0Kg6#g|Zg1YFZS8Xz-KTf&eJ>y80P zuQwyo?=qYZoOC$+`tYtqtI_DmD5o-8Dq66*2TI@Rxv2J>oOonNx`Z0i*2Hg)TW|** zOpkd-Xy}D}osZzlYj7+hrgJP5Q{cCq4{FZs88R^<-Me?zMmaI8EZIF7f;zN~Sn7!h zQ9glV1uk@f9XE^K#-05p@g(oQ`Kkp?osb&!#sO9tlex-eE!1+xtFpk}h|&69wLTmT zU&b(J+@d2@8chG`pHmB6j+mxh5Y)cG+S-R`I@ z%8j_KH(#`(ApuL+IX48haA;G{fMVcv_hsWn!r+awyqQ$Ne$~PkRGQr0^O*YDKu1Lw zrYY^k$al&(^X~QbFfoeqdZ1a`hE}4u)WNk`fmu1bVAMoc_Q&Z~txsK{VseA4H%E*6 zrS(&ISIWDAWOfk|*sUMt8@!SS)HK0HkQ1;9jpw_2M5PP62Y=lLFe8D16Q|Dy+Lo$Y zR*w(0eF;>_vgf^_NUGiuxO|%_vYuM!Oz=kL@Mvq-~xp=#x#Z;xC2Z{!V>OQ>$_02bs z1eRxvibBLHR=Wy(nAe~I@C{LMCB;vqLyPI~pRvbOXz zwJg2tBs%SPa|IfD;@`*gcP@9=7}_3CP-&1H>~1pAA!tiR7@KnDr^ zfI1_;wzN-TLeSaW6e}vXyoGx+c^$M*z2)Q7sSILrjZYG{j+c}8H}3B`aL=KuH+RZd z$yEoJ_;~Qja<~|VV<63Kzm}|d;DbGXT`&W&Ovs~5E}ngV-NL`-Jbz1p{`1%WEYtt$ z12R8xiiwD5&rLN&xl8|9+W*z1{m*j#XF2~?k47^5^0mCv%!1LS@*#p>YD!v)h4NPe F{tJI7qU!(v literal 0 HcmV?d00001 diff --git a/docs/sqlflow/doc/figures/sqlflow-arch2.dot b/docs/sqlflow/doc/figures/sqlflow-arch2.dot new file mode 100644 index 0000000..db16cea --- /dev/null +++ b/docs/sqlflow/doc/figures/sqlflow-arch2.dot @@ -0,0 +1,50 @@ + +digraph G { + graph [compound=true]; + subgraph cluster_1 { + label = "SQLFlow Server"; + node [style=filled,color=grey,shape=box]; + parser [label="extendedSQL parser"] + verifier [label="syntax and semantics verifier"] + parser -> verifier; + verifier -> {codegen submitter} + + subgraph cluster_10 { + label = "run locally"; + node [style=filled, shape=box]; + graph[style=dotted]; + codegen -> executor [label="TF/Pytorch code"]; + } + + subgraph cluster_11 { + label = "submit to cluster"; + node [style=filled, shape=box]; + graph[style=dotted]; + submitter; + } + } + + subgraph cluster_2 { + label = "database" + style=filled; + graph[style=dotted]; + color=blue; + database [label="MySQL, Hive, SparkSQL, ODPS...",shape=box]; + } + + subgraph cluster_3 { + label = "computing cluster" + style=filled; + graph[style=dotted]; + color=blue; + compute_cluster [label="Computation Service",shape=box]; + } + + parser -> {python_client} [color=blue, dir="both", label="SQL | table or training info"] + parser -> database [dir="both", label="SQL | table"]; + verifier -> database [dir="both", label="SQL | table schema"]; + submitter -> compute_cluster [dir="both", label="job description | training info"] + + database -> compute_cluster [color=red, style=bold, dir="both", label="load data | save model"] + database -> executor [color=red, style=bold, dir="both", label="load data | save model"] +} diff --git a/docs/sqlflow/doc/figures/sqlflow-arch2.png b/docs/sqlflow/doc/figures/sqlflow-arch2.png new file mode 100644 index 0000000000000000000000000000000000000000..92d017cd212691212a137c1ce6913bcb4131c1e1 GIT binary patch literal 96987 zcmd>mRa{ls`!^y;36dfr4bouHNC`*a(A^;2UD6;30!m1CdFbv&L8QC8yQTYG#~FV! zbNOEXFZ|57!FBdtdp+@u=LE`1i(;ZbMn^zEz!Vn~kw-wds|fxCpxy&l$h#1y!GG@9 z%Zt87DC~W*fq)>0ATA=L`0mbj5=x>%>xI8SE?p2?&ijvV?|fy@${&~8)>gzH$kVfa zSF$)_e!4kVI9_u49{v3|a>-r2I~(cm$r;Gs+uR)X&XyUuaU>YXd@=THviM@Z+8yWe zl*#4vay5xT2>X>GKB^!BBDOaI5;?-1-~UU;>BAr`i@eepgz@LnpPzoqZT|1?f*;^g zd8a}@5vvIO_b249TaEvF#al2P>7y(SG-T^P%lahJdZ%`thk`x$`v!i0nvVz?c{VOh zo8!N4&0DY+Pf&K{N)!R%zn9aIuoVNpEb;v3S;*JG5_UIK2nheZEQg4iE7D<0^1mmB ze_+@lSc3Z&2jQQM_!?hJ~nvsDj%tun<6IB_nX{kRD_I*F8o0e$x#g zF_PgsR^%>BW%zTSIZtKNP@{O-y!LaFDJKKK%5O6T-pc*iwvVKDMr9$^((5>X{=x4( z3ZnoIM)X3?0q;ML0Dc=u>1frD^v~`5wfRJ^5rm>B@F1KN|9KhfID-E_OZ%WtWr|vh z9c8~l8qHVJ&kzl*G#$!}w<53Mr2qAxesiDNyY(rhc zVJ|O-QTL;*xu&L_y1m8^d9a=9wG=Na*2c5+0@ae{TE{ITaJ}v`F=uZG|HXRwu@CXT zo?4OR4y9sLSCjl+quB@pt86*@?zozqbM9p`4ntDf?WXz)4Lao}az0Lj=d0jU zoGcaUG&Pdge7nP8*oog~G+wN`v3>9K{Xa)E+X7JtIcRE_v)}hi98zKJW^St4NbdLd z$QbDN=YQbI-0$T(oknaUeK+s(%|qGOZMP=tn{7iJr?vLHA7nlqg52wIb-Mx77O!%A@XIa?n3AMf&Tgg*hOKWHV?CC@)?|m1d$w$e-$Uv5#pg zkZV~lG=Fd!RWm+6YQ&S~qNvREDp0Jq@8xdO_Bd9mJDqhq?Bg9+&>1V#;=NThWo^D_ z9m&zCD%7&;pt(^z-gJq}RmxSQFEr~6?MddFN;G)^6ZjduQ#sXpbKd8rY_`l)DNyhK z&o+tifzW^@+ZyZ@uO++153Ysqov6CqoUd6P_HdqhC_PR=T;w#N}tsLco_zwk6Ei1*tQe!#b^!1#dwZ2k(r^H=3wdg#NT!D6yk z!vCZy>Nnn~N_x|!i@9D`OJterUnF|U3bpE7?SpLIPFXj3oDF()j3p=+sIQPQUZsP; z)fTHhYNEs`YIf|py!nnoc0`#|zNPs%m5>qlWmd9lZYZf~L(C{ojz#r?zxK16EuGua zC`J|a7Jtl!I|1EX4w6Psbp7&;`;xV*#eU$cP}}wL?!CI)Dr&l_WsM0kne)1BC}O}e z<7IlW|A^b}tEp5RrRrT2tfT88mq4n(@t1wChN!#CJ*7P21iTve`XGc_2G;cl{-O7C zJ(GAG^No7qSLN9bTNNpH>V`$wMScGD;_vwcQR93W(!p^Zs6NR`@kG~LZYvVEOUe4i zy+)v>up$y8a0?brf(#N8rQ)vWb7CizkJhX=6 zRv6z!KJ@JGCu#Qjd zGs$jZgCoyR`}j^1z4p@0&mFavpN*Sjd`65>kM-(?+`g}|U%NOZuoP&L(>6f5<2h&a zgKa?kccE(+0%7+h4f?YB^PA6)&lO+7k}NAHj5s{6-jPWST;vw}nWsi_sHmg$Sr>o5 z2UB7^oCfD}o$6&!FI~J{D}hR+>*v9Z$i}R8d5a12q6YWLX^+#HwFfGB$~{?Bxi@_R zx7(rgn;DBpQC_1DN@SsL)gZecYKtz@x5gB?BhZ`tJFiI(Vsc5I6Y;xOheGM?PL@Ii zCY3cTR}|2da@5L<8h3qANY_)6o*ngFq+QrayDx`}UOk7gZKgR0M+H^x5ig+pQ^N#5 zpceb*zCJ5mTciqoHFz?v|3KmKydN4~J!b8XsG!DWbFD9<2vLffk?qk;YKt;e;KY74 zLn9fBrz9j{JeYj;L24^iuS=SDXL&AnPj}#g1&wW|0d_3S@@oZ?)1ZalepNDJ{Y&xv zq1)yysB1Jy{b(BB`Eo?2qX7$aW$deIL1ggPIC%u?rt6b0ET`sQ9pp{@K6Hif?wYv5 zH%42N@8Aa>vY$3&iq&0U_FsV(nIbq~ZM$Oa%lIPfTOzmpRc3jK*4QK&#EhDue86g5 zZ*6eToxMPTA!og-f$}9tA7wgQ3i1ujZG^1a7P$H_0<m3q0{E~&9@fuNnFJ-oM6?;#NVPv`{?$i42aUq>XLcK~9}$#nA6h?Y9xUr+ z;b!h8x&P{zeI!IFV&&Bu1)_GAtZ}=vD~MNX@j0wR{E_0{jQ^pp@!HE zP~o?Ahu%fzW|<`3r&zwA%@U3i5qspcJ6$7q-E3`8kCTb>GueKt?@FU0YcNNF*4o>4 zOw0NFbm0No2=8t+mAl4fnO{7I=BQ?e?Xzg57n7{VtBHR9gf(3R!uwoFkqI>+dQ0_` z42N1(Ihe^u`So&svEFwKrQ@DQ1EK}COFx|wCCnovv@OD`!$zu+#tQ3p z55c=0KZ%UbDuprjo>8t!5~8cH?(^e2_HKVCizyBx|C0JeSX*`0ZI5ryUxjOtdSR)X zvfHxXy!Tm)3Sj`5=h^q=aTR4b=+Xy5?O5v?(oq#?^2_jcnXZUOUj(D5QjxOcA4HL6 z-+U~Bw%dWOK=&nbGEHw{_y=b_CJWJ&PbV1DpE&=u_bX4oJPp5iz+jOpz1*#H zn{q(Azs9k4DuPZWya=u;GS6@)YG#cfasaV$$LF>l22cD4tTOpM1B=&ml_By3W zb1TK*Q5w7S3)3+2i<2p9k0UWkNmFNNg+hE<&~5==3_c%&W~OAEUW|@sjX?g#!BGwC zol{eJe6gK89ndp1Ru%(QEULPKFmzVdN1#HKVfh z=~>4{lM0(FF*+wF>t@0@{I=mrwV&vZuJHr{Wp9j1E80}j(f8W?K?7{QR-uy zre(?x%Rgy3dH1i%Oef&OV8_99Yv`i##Pe+7=EKdVbjK^ip!Rk)A&47s$9bXURHb%o zBuD&n{hzTM(IZc5Ca;t!DLDuV%FPu#r1hjs-6nPNiV*@Es(xf3iEqDY#dXtFWT~`T z7z-*i-A=Ho%^~~3EItBR}qJ$@5BtVv}Fi;&s7b?VYnjx z1%N)%c&7%zzHe2d(IrXK8watFcpMFQO3BvuzO$Y43&Qpm4a>+0QQtztG0GAbpZlBp;%bpaF76Jz`uB43@S@(? zKltq57s*~M0^w1@*Gm&N&Bw^*6KAiyH| zu8y80#yx&7u>UHe^oUmvk`;!MW_+(#6Ra&j&o;p)_9i9@sn%gb&vA3OqqF2W-Cw{` zuo91ag>}l9E8r&~k`UJC=(N}WbXy`m2=5ut%LK{(;50a_MDY1U4Nz~pX~xjMk^nea z_^9OB>CZ5Z?)>L{wTl6~`6-cu2lKBJ`g@5xsQ;&z9(h?h=I|Ji8~S#(-i5<-klK7G z^Ycz|3l^FJT~1QmOXEp^W~{xg4gxFzVgzW5GvS3{5InkCXSYC`TT#?>GdZm7dk-_; zaclG@{xQ4R@|aGoLyiY%&CmCa)1grel^_oja{VO!AAb-Q2q+B0#r%P_MC$$P)43@L z)&>LT8OIoA?RqZVdo&d+qa;rFGeq|02ma$Xw}O>1HOFX-jG?*og= zovJ>8K8$~Eqyq$Y+$*;uz86T;Z{)S&IjyZwp4fOdNdfSI+o)x`ldo3x)2i{@h>V?Q zIZQZTzb(XRGbhUxN3p_8Ndx@3xxTqNlwqxJyxiqw-F2EUBs3rV8VEnwEDcBUTO)Zf zNiM5#Cm_XNm^(E5h$dxl+b#)UHydtmTK%UxYj*%AQ!+*9LT3XhvX=uOm+~Xw96JVF zdII_nQ|F+yc8{M6B49*kt80CzF%E#Wq^X;3&0gcB#m&{Z)3j~dxry?-SMJJ+c}_c% zP9sVx6X5wXPl;b)XPn9ln?uK+z&n! z08%z(>q!WdR@h(MpqHAcoU-D@c>hg@(|nZ4E2+1wP>wO!T6|}9E|&05 zQZZom3A&EsH7|KqtVvE@;s6fX6gc;rL&Nx{k6*nj?n&Yuqb8^RL`_=QaH5_v__Tx# zFe$gl+ImpKrO%v|1(lPX)K1jUX4XBkEr9&kMuuVzgIOE7yqE{~nr?S9IM4r zn_2Q0G}20u(eH-kMAbl6w(L(8@?-+wIbPi0`fMrG={quE?5qUa?7OAUo}Q#0M{-^} zfUPlC*LYZnJakzc`LinDsfFWzO$fOgUY95vwi7jq|{^?YO zVKALt2v-8Y`|8jzhmFBjjD%uSw5NfX9)V-r4P%6tTZMJKRavKX+oi$uzm6A!SuY$R zFFg8V+z252Z$3TJ1)#Gw382JEi-{Ks<-#j14+M@*NjNNG>7}H%B2VVLyu4Vh9E-Z0 z%GpTaqOL^^^}5>Rzgr$}!4r51yIhOKi58sC$x1E-i1Dd_-B+`q7p-qTs#lm9s^fkA z_+gZ2BiP-Rh68yIAjW+;oSK6Vkk@cT$B&+l-}FpMN@o$UdzZ#^6sfp8XMDffsrXxE z5NVk%N{eu3S(Dvbhor}EwB2@A6lSa%j+NcE3aYmKiLLh2SElXzj!*B835ZWeD&rH# z2Uf_i7I>_W4ei;UtjP07HuO=_>E&N>?U*IyNsgqm9&8D0j`CKp>boVBKcz$9j5Fe z1P3?J{|i*@G^mQ?GpA-SzFSfMhP3$)#mMF?b}d{pT;0|vm&4VG<&KE!Mn!d%ht}tS zLTzAtTYqf0Od$uu=Z``mefRB@5#LtVAUFSH*#F6GxskLS1hRtP;12TB)nsi@uc6xM z+W#8nI4Qt#PaKNwF$xOixI?9eVobPedJ?&ZvVsdVYhDP{QqXc2qVgzM*8D1#Mdc43;xJsUZ;FQ&Y(^>0}zafMh<~Rvd|FR)@07HWzN{G+&t9K>VH}3~#lyw)bdbw%OIw*2_mi2*#%Ek{ zb!QqMSbG%HF#G6udVAEUqBu456I8NuRt7 zzT2xq{?a4B)Ssb-i>$VM zag1LQrSpqoS!jx1{6*G*graG0tHt~`pV|6!w#JKYQ+oxh=>8Cf{B z6gS|)`l}hw){pFfr_$ zxdqpPakmdeN-D}*(6OO2>)MT{WbEjDv|DP;Qm@5vUe=H^7``06epVlfRB?e{ixU*} zQ|U?3{$suX&DNT)H4MSl7H%sZ%s4Ez+UnrWO&z#U{=V~9h+nUiNZody;2=Ux6@Jjl zJ%{OveZ+sLz*G29QCz3rkv}=zB>5j?vW8a{6bhw!(P&@@)E!bHH zvz*>)dN9i`9L71>AkU8wTRmu1L^eiiQz{e^006ud?B-ylbt~-B1kw3q^UXmqiSVTif8VkJk6nBgywnDMw(QLm6c^m4ZFQd zMuGaVY>cwn2sg{v>@dVo*RQW$mfpAKZ!; zrhl!Z((z6&X{TaLT_?wEl9r%DBghLT#cHeHooAGDD@v?@$=Ug+)hUgS6}O>(P#7VWj1 zG|)$$y*YT-Z9<%uWNCXtrc8$a`R%)P3j3{+@|f5f%;+Ac>4AK=)A}}!sW8ejP9g2T zg}yJMAmhJ^E=K}KklYG!(&63y{CCeGfb1J5V7slWyA3;&wn6RwwQcOX3(c|*&7mc; z4Jl^A0Vn(E8{_&j8Y<=YBpz!pW6;vgHR2aPg(J5Z1?L7x@Nw=Xz*qug|3#$l^B`y5hKXe(6A@6pXM z!)DDl8fP5Oci=bC{Ncvw>Te=KFclw24p5(#oNh`%KZdf^;Mp6gREZcC78;pwx-u1Q zm(~!TO*)MlJHs@#Z^bveyAZ9IM$SICOd3Gynq6TsK=~?e|Im3ZPAb`g!&H9Oew}>0 z)G&E`bBc3M&1t9fUDlt&A13O3`|z9^{@0J#KA>Fn8Fk0%Hn^QSHN&cP2CS>Iha@UL z+sB=qHr-mnr8S9(Ekn*3hX`iXV)BJ>i;5_23S*~)!{)Oy{jz64j*EdtUD0t`@6udU zzPlcGbK}HoxxglP_m&I#pl6wysDI_Gc0|P1YwINB7Clt>7QcuQEHmo)jWywbJ~$WN zr#^T?ndfzTbJ9)~Iu4rfWi3Sg))63eSvb#mR55GSvfajuXj_zrPZZYGj%Zk&kE+I) z(#gTHqXzF1bFkH|k-l_hOnEw9lrljR6Z@IWnx!z?LC0Zd2^Hw39f(3P*@lOm))rg zN!_-e6S6s0_*w`hB$N#OC3oR6*5j?Yz(|Zt2lVi4eaTI zjhb$gbCdg7K&FT10T*zDI`qXQu-nlQ)w42KC+$Kq3~28Jmcn&4NZ`|Pz!_P9yyyhP z7_~g*f~!VlJc8FyVbi%;$pq9;qKGxhY{{P)p&a%;Y`n|ZW%n?HlftYJP8Ko)P$%vQDAB`L>T)fzo_S3;t3dXUv^m8fjaQ$nJu8tt#gNtFcW z;{}e~9mE=l;CnM*M5q9A=gR0ed-w9vw`6Q4VGXx8mnY|IDHzT4v&tdq-{Hp(KJ2Ju zRT{e;6;WLP;=n+a$zYhYJ)9Hq`7wK^XaRLDevLfOyqqY>3dgCG*REC+gHbPS>rR*9 z$!?udFku4UzW0Pt!V09kd|HMmcP%-=c&gD#tFyoVZ^k|q6;xIn>L=cUR(P_0=t5Cb zDLytwm|C&xfK#XhYVRcA8?A@3WSn!HlPm#sWb3qU{b6MAMzg7r+t&ExY{_L2YF?CB z2$Elx~BYEgbU>`^O z1XRpaYPf&D5(9eh2+`6^E$;;>k;Sn3l%bkbTrU_;(IQ2k#(1 z_RYOJKFio)A5M5qcOoL2miN0?mGP5Y4_XP6TF;FuSE4m00aLYoTH^`~9#k(aU^g`$ z!41rH=|Cql%5?aC*!wa(Rl2^3F zLa;nG_C!ZoBu`~mf*UtvP=?o5v15$K8r+pmD zGD@(vavFfFCf@J4IAI5-5=xbg1f1R9orfEs_e+ggh|rK2Bn7v!3?dp9{E3yTnIE?^ zBZBa6z@wMpbvu+vyqqV&AsQM-*O_l_A8)2itQ?caTByT}B z)M6ARwp9vZ*R;<*hS~?fEaOsl4C46w^ZOWPjfWEp^mdNxX-0joB4`T}73x__-;TNF z>xZ((bec%hy>UWyxc@t%7=SF2i<@dMfhrP(;3f-kG@8QkY9i%tSal}>4H1M;LxCzSVkDb zN<0>ji174NW9ZM}kj+hnWuNdQ!JdOHg8gz3X1b1A^}F5i5yJ&$I6L)qe564c!3|Fa z)y%9EkHTqNJi?Qjn~N=vAGv{3__;bwo-F7N<#1s@AKRj#Vgj@y9B!4|LMB!Z3SVUj zP88R{^iDSOTrRZ@yS>-xen*gEEZCK1R8s`R2_djF+(-oIB%?O(M-T%vvGPz7-cjp& z`E(7Xk59w6fDBqRZ(T)X?C1Q;7I0n)Jf@L5tZ}Wos#%_FZjP{>{3> zX1$4D3YWS5$|MXRlVtWjx%L68kU zd%BV5r;{~}aOG%|O-}lSjWjF`K^1l}FL+N5<;!IK-A;u4&&jDKB1l}^>4O>2#1ub# z{|S^m>ouIHO>zGfIEP=IXYYR2bbHzKNkc>FQA4`Ii&EL)&6>6D(fke#^O4Gen)Sqq zUn?Jq=K21*Q(SPTPkSDEd=Lym?M@MxQ!di}+@ce7HfV8z45OiRoy44Zt;VQwZEGC}OAZe7y|6=Md>v zE&hT`cVL$}*l9tq<}Cn~Oh#y8H1z=QksO+A1$-&M{}|i3Eib>?nGYB7e*-qnm`N5GxN-NA8^#v zBmv55jp*tp$7a2ApMH@B($Gm>eDITa>h6PA^JGh>ha6~DBc0n9TSK(@qg-pBE~g7yjguUsXMiA( zTt_*fKBcUwwwuVF?96{~>$FR(etMFMQIzAk@atl$XbL!$wC4{%;d@x9yy!+0y$$p+ znq-;fA3(&~ZY9(y2b5O*a973X5$Np80bQCl>^fm1uR#Q)ERq${PK&C9V}M?*F$h_h zPAdYn!%TEqE=}zZ7-zImM9DnLZ=d+Jx%dmZfI`qI99)atE|0h^6a~Xe_|URBZ?{lx z9K~^GRzH+a5+~1s*$$=teD#&HphrNA!hEzw{EUrn4PTu+r`Zn5C@4X07aLiYptB!M z6}r3ESlz=hh0mCs?6OK^K315t5xgZ&Q)FjuHFP&zK|T0Pmr0f=^YM?Ye&1fhNDM zCx?vXluqFKNP$c|nZr`Orbz1Oy~t(mrFg2&-i>^kqTaRK!AGhPx_uX^av9ASU!?cw zku6xfO#Qw9j2f`qIbEU8mWM&-Z)y{3eA|7J4=*d{sd@8m=Ytpv_fxb^4Y|(1tzsy* zH#!{&9AmUp4*u&+*rc8}0E98*2L!+y+gPY00O%{-y-vRCdnkT=Ui1*$H?-SfvY|}4< zg4~n9$?DF9yGI7(tFOl`XzhZxoYeBs>%Q?n<-;3!P~>jiL}o_z@MEBfz{`|p0SCS) z5H?|Ws6w))S=r@p^wRcuej!bh-Pwk>a3BeMPj>$$PDXSNT&7fWdqEEU=Re2U$CfSe zw@b-DV044Vy>8=a8K8s2x{!;a&_14I0^xTKJ+b=djOIqPuAF*|JR$)*6ANY`45I{& z?0PMhe*0xxx-9UxZL8&hKlw_(!?na}hKKArp4Zu`CO$InE5Cwa)MoD`f3mA#Ma1?Y zpyM;=jL6k?%&a)3ZCd=|BfiV=NI))f9Ta*Rf>tYlp0lydbJ|G>syoNf%YX)40xFC#({tSjRt9LHAB7-QItfko5m0e6`w zsd&iDMY7xOMnD?7d5O91G9rTye)Lr0UY*D`fMrm8qBaue#G*oWh8z|a=!Qb&)9TNS zk)H1WIGB1Ma24jI;qfN&XC}M0`ks>fIZ(nARU5jKfkgi$7oWS(?Pg8j_FyYS4`Li1 z^c*FUH0yI7P{YiC5~~@GnIAR#s8HVyhts&o%d!M323G;-1gY0ImwTSwaF(e_!gTBw z#A+tTh%wqq^HWxqPGcx6nE9d#b_2DLaZbqGm>kFuy*$S|Etgyu~Wl{DmFiekFu{QGZApzHd+_npe15QRN z06~q(csh;AFTWq`%{lAo|=e@Fn1LIg6!? zvK#j%rYhwr`&_M z;YCp0u5AP4-Fqyr@v7TJJ(Cu-6T!>*4s?G~ckm8EUKFld!+lWg=`Y_i74n zLiG;qo%Bi&huk{MCm3Guw9fJmz(@+E`&NoX?G4zk1M_9is{>ITR)bYa_VY1dMEi~m zu-O&xa{s!v9jLx?)v9yC)EjGt+s?<(kfRuCO=1Nqf%tt76r1BQXsW3NOZ@8G)|^#= zdPRNgAiUkxa8ZU$E?;SH>s!UBn@h*tz{}JRNqsAtJOc z%#9EkF;4oMhxoCnBvdV8Dv_zyKBwH;n#1VRgobrojm7h=$q|eFxAvxPD_Ni#Q1!|8 zI^O_sB2DO`Zd6I5aPM*#5&9cg@ED8JZy}qiU_ls7-b-`-FIE zv4`N_gyGmQjjjY-7*UAFdUNfiQxL!IK=xbNi{we0@VId^WbL`_eB!DZCygmEnQ<7} zrP~W>*PL9XIyWC{K7IvVQy5@}MK$ZQY+S%S^!$c>kkoF7FVnOqd9y9kHT1nKQ$7}>zS-9GHk;2C`9IfR5i#_G_lSx%0>E)K+ZKxtYwbwV5!@O1h%ZDoS^2DPsQV0| zXJsiK-Xpa#`P3KDR_Uwg>(_?h4IcJcX@z>R>;`BKAJHgX`qRIEWIeF=G_XWOLf6)| zgC^z-HR=<|Q5?eyx^hr!q-=4daUacbR{_SaOiVlKPs|IRnmj3D^xzpM z`=2mQW;gs~so#HXPaj94BsFDw?C}Ys2%2cVS4s_D*d&NS((qQkH?cor8M~h_qi(mhOPc@OKp`k1zs|mduFL8N>dR||y^*f% zRNroc#kKcrtq%iftH_%^#@cS#|MHUYqgU2r>~g3=iMnl(w8bt^+l=(`1DTKs+;jm;+G~|h?6>|hPcp+S*4@<_fHY;8( z?B8k=984L;Xh~Xq2rSMvxs`t|K(K>1CYl21(+hi%2NYp15OEdNQPLb9p(B-sd)s4` z=_giIr{m($N_YRH_{skC8gGp3GAW^2V&>)6_&_3ZEjNYYz8y@Z22-ntJB0vSDX7_G zCtTHkOQD?f4Gt14l>YJh^c@Vg)lU@e4zlmNh;Wxv?1^tK#>@dc6L{Fz-w+j@RoQaP zu~73aIf^6m1@Hxu5!pXQJ2$t%KMoW80;_$=Y`fL%d9Y@{S(skIQ?j_Yk(F}7vE94= zBW3IXK;cN5QhZVFxcq^NEreqQ7gqY@LGW%K|CZ(>J`9(i`Xq(6 zyb)*#d;rw13C^{_&;CTmR_J}`(*1&osK!Tc$OuS=9~roiM%(C=j=KyU=Mw)y%?RQl zeP}y!-aNM9PAB3hZBBN48@?)RPgST}r>oJ*M>{#z&!9c6CsBWh=Z0BCk@~>A8F?(n zZ@`uUiBd>#w3+3Y3fChI4~92?YPsPK<(%G3U@_v@2Wx1V)Pt?;TQi+P zJL&rKmNs4;7utTrn(LZ&!+jvYzf%#Q^{aQRdim_231ddRFh`!8j+Vy z1uO3vuzZc%M}KY%Y2{C}6~$@$iTwp{4((Lwirr~OV)vuniM`$n3m!Ui&tHtc5D{;w zWFdXKccWx&Z`-dO;iH?c$Lzr*rfdz;GAWD=gWKh=9)Tc%>>4&R!I;2-r9@4xStMVg z{PNO_`wp0zGRk-pcf4oKBCp$<|K0tF^3zc}rp&;-1BPBPifHfAZy!qrJ0hQEMHW1p z#MNLODLSrrGZ5Vz?hN7A!^Fa$Kpi4%;RimY1BTX;SCBD!p0D_V*V3P}vQzBTl$xeV z%^(-JGQ=g0_;Q1_h2zhECWm?ye2l$#!o!}+8qaA=LZ-oNS;S5#{mnH?1hR~esc?(W zH~93M9eLz}B#%Vt8f}utn)U{5*qTn<&83|v&=U}mB7ZsiDHke`hpiw*@@Gb}!jfy& z0|}MXStGk3vnbAUmQLFxtMXUf1pzPYXEA&6$yqI>EvT6m2?*6AGRmRX(k=jexj#-F z>{vX1pVIBp@0#AM2i+mu71ezXs4C<86OQar=AiThAe_FMI+Uy8@>Ytnj(N7*1zjc%#O8auICy z;#N4cdwv}iDTpGq-6#F2dVSG){z*exY|V~m_c=xEA%o2-rcn~_T0sTWPYU??wKP(A zs6OGW%73e;IrU03t~=q8|KWq^vxr(9`Np33(G>6{9iAo+8o#-8?TCH72+$H8pxYoW zPJoe=7ZdDNfgT9T!mwXRiH3M<4wpaDal`45`t(Iqa6*w#^X;l;Z_?*V?gV>#_g@^S zEUZ*0zw*v^qKK9U1^8-3iBAknoh{WPegz~7BmIr6gIH%1gSLkH!-C4>ihiy zNkpi){bk=f<0=Ya|420f6-GKJyN9dx*G^_zTXXepw0=$4_i8w^h7WjRhE6{!>~=In z=sDHvC_IH8atR9ZybV!OmmkwV3s*UmnPEm&DY*)?eR@E@D+BiHU}bKLzl1Iu@agXN z$-{h+maA8sYDHyN=BHmSrz4f$_JHZc($DyEbo#w-0B}YCUpRUU1Gkekm-`2K>zby z00b}t#0qnuTjWg;TgQ4_+0OcZv%C8hEQQ^!1hBehlFQGeA*6E8T>q(eMiMU63P2ag z({8BcJDd0AaNJY`-p?s=TofNIXZE(PX?_Kx%ROtSL2)BST~F99=j?jW%R6E4Mbc#ugb1#=(*v-@cI`amDvoo{YMpwfo>Z6Dd7Z?mVMi+a+W%e%*aQi7&eNR1LO#Im zJgZ!&8EEJ_*fIE)1E(68p0&`Y9EkZ`9XK!0t7Qbh#F3Q8qh&i3`4;3B0y?i;TBE>1~?}$f;d0Mp3$I#v2-U3>LUAMiw5fQU1xq z!l`zs^ZP2?(>3Nsxc$J+>3lY>A8NRYKsmg{pw!*_YG}KIw#TTKY4SOdLZV~z#CzmF zdlw0pmS6E6VItC#d$;JIrw$s7;*RSDGg|-$xKj`N7#A&1{yGECfY+5Shn`4Bp{Kga z%KP`f1C|H^LQQT6VoV|X9MOEbDyqh6h0`COV5xni1~OlN*joml<$fxi7^`tt=DW$t zMb^5aI8h`IzwY}9-v^oVt88d+0UZLKf*w(#4ErOc4pJ6Wz{iz;UxY?ED)#; ztM8Kq$yeYUemVqW%BBTV!IRYr!N-Eq28euWz@@njOatwix-ds3E24P7+0Obm0jM`x z?=rg+jHeQ$2L{O}GD0&ef~VoKoZxGuTv`VLLhu1Z;s^mv@R1{uJBV-@iYkPwx$gRN zSRNt`1)|->)(GPs1;(+A0K40XT3H(zXTnQbR=WzPojIu#yPyeDw^(+otJ9K2?Q5XF zp8!60s?@M+`*hB$ihxlW-z8tOW)MDL&b)8vt23?NFh!o{db|m)u;1p7m4rT)z8)#m zn&JQ+emyeW8BAi1av)x6jO1SR)K=FlJ_29@lc`gW`d7PfWm}^#ImT1H_b0%>wPvmEXGBVRd`Vz0jHIqDzft_jU!o^PSy0W_ zNu3FXRn}UgkfKQwYTQM4EHjyW!2HRyz*Yr^dKSoq|1%m>$g+S+w;@4+jED_p9veL_ z)q#O%njb1N*&a^SHHV;rPfJ^>&4BvW6@9pjPy%1GV%5-DAR{J6<9o!80ro=+lWs^X*T ztis!`^x#08qpb(;qy>Uf^FVlW2g zJ^S@k_Z;*J=Z>yG%1W?)vje^@6kCYY0*0=mocsJO@GT#;jN zhS_NTMyrqCcOc(vBS+q}Qw^kM%ZtNwn!fZJtlGlvZc<>i62RCuolhjIOkv_1!!;TU9w9OG^hxKw^&UVJmz5%Wz< z5+$i!77=Z&QIOT;@@xG{YRJ`4he4vL0Hn8Uu<@NDImU715t z2E+Yhrg1Wx;C_}k(u?}mhmbjn%}`v+gf?*V9Z5{e%mCNqIZvv z7l|6ZT*tU}9Mf7$ymD0wz#_qx5x?>7g2;Y4MhzB;ru2airs`y{12n~ILdjXIbyHW&pK9J#`bm25NuGw@ezkncja!B(X zK1w^p=Eb{R?2Gy6**7fyv$yFgatR_A5eG=))D%luVRzs?Ly#;*_M$$78J-9>6koUu zH04;Zya?_m`~^C~vvvRUCxE-lU5cbvn10$IaPCp#k3|v`8`O^?G`gp*0=pgrDF^db zEec!t0%mbbz^%Bs0tk-PF=I&fmIE>(5xiw9Q4npTz!tdU@f=vv9%6YNP-Q$tf>M*m zu)}WXdl-m=r&o^*HR;`he)a(OK1>msl-==X}Oz3>PzJfV0Q1t)&s zGX)BF0avbLZI>ocD((*V6J?%KP#g!7sSR!51~uAI^f^+NF}iRSAM70Uq8~gOoeiL@wQuoF+|03QBDki4;2Fa&Sm2v0&;iG`}42Z6T#|d!YxVzTT8<6WD2k zt_eT|(SR#U7w{>`ce)_wThsx$aAJ>+m%s4;@%EKbRjysMvIV7-1`(0&l8{niQ%Zw0 zN=k`ziKqx0K~g%Tk?ux9N))9L1SA!Z76C~C!Mk4L=y!kKA9oDL_|7=T@7-@a&wAFH zYtFe^L#2<&!IKnR+I%ZxPc}=p3}n`A4jxcw^ay5|iCzZZ zPF%m~!ND@k873bz3-Q8uZNj(AQ*Dv=nFd8^j41!2hO@>$HDv{)e)*)^VXXMI`kdh7XPH*7S zDO|JBU>qo_sY|~U61@#=fh!~6^)mNf_XDUBBnJ74kL5-WVpldUL96WkgrR`^@h8$6ipITFQD2woCr_ z{tj^hqVs^JLJATjSAbnIc7LN=(2~$q$P*ipSt5=wU)tGHH?7mIJ>;rJL!(dlAuy7n=fW$zR@Fhttp_QSP7m+Hd}R~_)NP&Vm=XUw zE0;iD$CnL&&}4=N6h7t`IM1VRY4tYnh z7wVMP5ZPPf?5ZQl(M^S^w~jrrL;(|ZJ8>*~_t|XS1f2K|2!Y_lX`wcU{;^DjC}%@8 z3*1+vknLSs21VM}F~?2NXf;B?IQ$N*gL;5eGrK$>xMskR=Gp{|K>b8vJ7D{p6g@KKRMagD7q= zsh8b=SQM%Khx>Qx zx!FAcy&~%N!xZK#Eh~Qu6Ld=F4NhXZ5b%6?H)Dqw>@iMySum!fut^!!pWn9Ihp0eWCmch? zAB8)G^h&aAPJ&+Oal=4oW1ceZLrAfY5Uqx704 z>khDq{n~GGt_XvPne`h$6%i{YCu12oTyrK&z1No!QGwUMaJvHm_yn~8_G#Ggk$ARX zFB4H2_Fxioyq;wbPWa>*lDP5l7C(Z9{V|T5q-hbq7@DMXCB@@WcYvHBQa_e^jkx+D zk;Jg18njTaJKGp@Z|NqUPGf}aouI1RD7@H%g^n6%X6i6ei?D>syDj-RnBIlJZnCyOkSB5 z47F0?wnHfSD&R#j@cMY*EIiR|gH3tcA7jTU2G;>t9j1Q~C3JFg` zdOZ7>W%Erk&~xtaXIa*{ui8V?vd9iv94Qv;D3;fi5w^C>S8m8in}JFgdbZU_gtZ0l zk==$};2#vKrAp{zTt1O%&sO0Je7)Hc-931V%Lwt4Q7K(>p97d6?KHBvsiZcu90Hx{ zUd*h#&zA-G3ZYlc1t|CJvTpXhY|mzMVsaV<$X*|>b|OwdwA3u2NYclbG!QOpTx~zG zTv*Yw{<+=!7Q;T4^OW5 z@fA5-0I9a8uTqOO@ZXGF<>m9**ry@eUC$=DrE~xRm*{)#od@z~tTV#? zQrU8Q-Vdm^*7Fibb)BuWu4ZxwJqW);q87+q^bY?$ zuN&yKg)EVmu=-A)nRKf*15K|BA;{Tnj@n%nsJI%9jfbPZ-+^)XR%^nI!<`UWIq<#P za6-DU0<=-B&w{pJYA`F3P=0xwvkrlDS3cN_Y$Ic>jwv~E7!tK79YE`xPqn!5G~nt~ zbOp$=JA-)wFH+#0tSXY%sClh$l{;p4V}5`B&Vx4m0|?PO(gKBi7$&JKr&t|wxgehD zuOJuQ%}og5X1=&}rF1wz|5RSoTf*Qy!dhF>b zd>p$|6r-^12`Z}~iQV0~rm3xgr@|*4p{vm_{7I2-fUba?B8X>pB1PIdFeo+|sP(f( z8<`KNR$TPQZ|o;@y@v{llhi}Wa=#J>lS@V~O|eCt?nRt?1VA78oq{81UO)6PQuPkq zscv&Fvr~%xu)pbwkcd~UdpC`7PUOR#3{TJ^N~BW+Wqem$k@N7A0;1HxtSP?9aO|tB zAua32R|Ptu^0b@nh@mhNVV1$85EEq0h>p81u#(rr8n?W9!F( zh90#sLq_%U=Dj6aJuQo|Uyal*;hG*vPvaymDqG^J>P>yJw^O`S;$GLzdC z_(>-CCndi-8oxPs!R{Dw?qevKEJn}`H%Xo#>5_*dyqIJ5$+0CR_^QL$R!L5enu-DY z#JF@2An8uC*wH;R)##YdJ6U8>7?))N1M@aXgM5oLGtQ-I!B=Zv#HT1~$h%D7O% z3RObWz(G#rO)>*ehh_$N3x`He6d#t?wzDQ?F7OH^@}j;ss!%O{Tit-;FmWym_2cnt zuL|s2oJFep;_VMERC24O#nHQYk$ph)y0U8Ghup1`--7!{_@)GN zhLoIthjW3F|4oJ0dK!I?Y*%In*mq;bdow3kUFZWD!KDI`%qoDlO3DG2T-<>Fv$ z*(Qf}hXKnlM>2gXp&+f=^QsCo+ZkaJx4_e*PR+A^(uZIf$dHvi$rq?@8gn=;J@g&S zy_!)7#4enkE3=P%>OcKuEA7UH2ok{fV;HdpSnWa>`fRN@TF(|nGex*tJ^MZY`?WzZb!f-; z9-rTqN2oByD#hb9;`6xb!#@8SPS4BFUv8d6#h{OjZ7O>A@GykmLV+dXb1}Bu%=`L^ zZpPMhC}MdcEWBjIVUB={ngcn`%~_wG=cnLM#Jm$vSbJK9Zl$oB41U5Hsw@=7alQgF znh!q4Hrbn3#9wI%%?BB2BgR~_y;zsL(fxSr6KZ}co>8H(#~8w)|KQaR(n}M?K^Nhf zp==eJCXoS1=LH|w5~fXKC$&ERD5-4Ij{GmdGQfw;_tU3uS~CR07cal#Onq&I^sOn&A#$EN))%x7V7oZoJPe<0g^zyA-)=< zh8?4^TLndhpufEhI;s&6?g-qEb~R39RfC++VY~2gSd~}_b~K@4<%-T#PL0;xBy`$$ zsEB8+z<(~_A{@BHS`n$N;a@6Pu1LbSwUY4H zO`nHN{`9o4t)9kMCIY!t(GO)_B0rXB5s$5N&T-jZ&LI% zWI5Xmbs=WEJNuDS0cRoCE>@e>!cRRkerwqae#E`!xkF&FLqAZF!|1d>dJpeD)O6WR z@}y4{CGV}^Vr*IHqY{x^N>b=GSPHk88}XaAU|)c2)E0U*LkRE=M5scCkBvhGx_l0u zJ>oqBUX14KPgSC3$G;)C<~(5DdY@iDZkTLh{Cm#adwHvqAXWBn-sM}*+`V)zGKK%%obRkr^*=+%Zqg&Th zeF6$!v8ml7d;n@EaIQ)n0mGRkzvv-!yGc6OW*BvW-AdJm?Id^XKuf%@fiYp_b*m;i2H#T`{ZKIVE;S9!iv1(wU^hsaiD z7lHljGk_G>W`%nq@~LnY<)Xg&rE_rxmCBneGK}5MB35Kx(4S`>FVEsOZX6Hu1Xtlh{WDM_ zJikwpB_u)dSf7<)lunVAlr$Ym>&>wn<^%+ltm2y?Zr}&n#x}0utkC^&$Y_7D^)Ff0x!awBQ*`dJu27}f!>z}^30qI zHD7VjF8YxMD>@CxNG}4_m3O-gWZUmf>jttFxts{z!ia=Q=CNW0StJJIL*zqZJI z87R;#@m`2B>0g~Yw?uh)GP)6Q$rS*&ma}3$?1N0&5v&6%N%s{4LJoZOsS^N$1Y$EQ zCJ5Ut(M>#e=X1jiB2_Q|33H|QBp%b1sSP00UzO!<#b*z!IUzD$s#nH>#mb-Erpx{o zUy@5adqE%a=s@lmakC^}^lhY*`2M~?!Wp)-0qO%53jd}vKj_ch$DKKaC6`{=qe%2q z>fuU8&o46Zs+JBRO)NeDf9J2kb((i*68>?E{6>d`O+kl)Yfb|CJa#@~4zAz2Q)n?B z!3+THA}fX8-5Wg$boF-XAXsO}n}pJYKKv}a=uO8tUSfFytnD;K0!#b9|8*V0{QA{3 z6e?cL`%~QsK-ROLzGaA*6g2IU zo>(43x|PV2qhEo}gi)Bm^*Ab!SP#6)VOa#Lc#>h>od5Y+kcUh|A<(Axn=Il0d<>L_ zV}MOKeGBgJeM|J`7M&l#La_pkkEI~;h>(#8_X?OjMBX}FfIbK=a5afy4;$l58qOex zSgSoa*qg17T>9@vj-V*Osv{rwQn4Re7S}|cBqA`4=FfkfIRnRbkBZpQp3A?m)lXdl ztbcwxT@@OpKMT8nhaePe#R}y7^Z!HBK`%d=r1$$@S$f#WwNHrExL-!I?@ z_8~Ywv|hnNQLCDOp%x`J7KazLM3^gN8_2pig6U$0m)GTy@t(Mi?Jx4U9`+UW`)?pm zcnN-IajMIZ9vF?k&_alXf)+bzrhje|t~3qHC}_UrDy}&p9EgtJI-e}2XZ*R>|A-as zmk`0vJ(r$G^ z9%*-@4fpr1A#O8(2kA<_a5v=>Ht=Bh06r3aL=IeDlW%`s2-xmJuC_kELL3m>+!4cN zw>(@XY}IoHanx_RTmshb!D&Ik^f7-f>5aaHtEHw=QHc+?6@`X{{Kfm|$u(3yAY((o z{z9-huN1I&ECS=hoWt`s-_pldXYu^^mX2TqaEYIC9wp9UAjUx%6@cM}TG;3Aescl*ndgJFRV&vU4jI16Z?()sVVn@LPg zXvf|stjzli0-;Ej4AWCtB+yxHP=Wli=i-Jkbis;v^vg3Vmy$YxH;pkmO zsy~~V9VC$2Pvjpp8aLdng8mdRXDt&yV%t47?*X?kud#!Ni&o%vZD6<#}HuxTx7Bc zxPPYLFq!_eCi%Hc=r#f|m#0LKgiygdnep zehI$*)l`PP-|kSqe@iwN;zzlr%$q;!@8`b#`dtNP!%kATEA=>P9cr`;X6Ape6N z!UfPK$^nQQM>r2`srQ}9>!P1YAi$DNg{2COsH4f~O8hbnk|ZLqz04!RGHBgX+&lyh zaVZ`d>oBmdY`=_F4OKfB+d!ipBVLa|>m##KK1rIr%T}p?Ch)i3tKdS6;Pp^OlrYBe zK}P%G{)UR?!N&*t-H-^HAnYQ<9|S!ChrmQN$-PoAb#?N(EV(;FMgva!h%l{8MQdRT zq^kGt?IqtCv51Qq9{Rq!@ddoK__wB@7U3y7%Z9vYb8%Qf@e&U=C3&zBKk({t=U`am z{5(6s{VK>14n#s1U8l9ucpnsqKy@n_0Rww{gXh4Lq2a&b`-AZwG%J9Sqx~hxrKb!G z(VwC(eL$xPq4pjs?W|0E0apDpV5sVWzwY<>TZwkh$^ZhRsNIaWN-2Zx!w3vyu>t#< z>SX5*j%N=J?tXq>l8V@5W^cEEtd}bBqDMK>I}8<@JB9etR@sdSUo@QCK(^)P$LmW! z@Ys?};O*WC+=*Mbi*!^wxJoL4(h#c`_KQh~0lup3lh%MRDX&1p!5JxTNmi}fClJNL zGQ7w2PwcfibEzEm#Vhq2s(eBM9+M0E#hxJ-8j829qnL z!8kOxof0$raJ?X$FfMc0FM+2|@`PC55>Q-zV9^-?NjuB4iQ9;n2L* zhbo{O%u82^+DF)i^b6(L-JN`05bo!#6Y9QUcJMmIptq46+3p=Ji!F=B=8X5|){43B zrZhPYjbanTl?giR1*Ewu2l?JxK23@-TJLXxzlKOK&ojM|)Z1F|)4qtzxMJM7b5K;I z5Fg%5VVR1gEIloj!N<~bQ*~6arY|22I=)93D_zj*ULM~sU6aiM?7c@iNLzH7@qMMW zW}IsvOKiYdjn$9iJ9%y>sgdR!?a{pfI%dceO57 z?kAjD8G2#W)!L-4T$>W`9qR{5j8R5$4a?JT^%bzsSt=&W=$B_bITeCz()Z7XA%+N6 z0IEmsK~sD z-F56rO}8`kl)azxB-HLrDD8ys4?e?MATd-VYPMVyET*yBYd%V(Jf`Z|xX0~-bU)q}$3Ne@OTygxU;I54r#Zk)tJ0dYA+X?3Iyt<;8W6+35{<1Lxp; zc5%V=IJ*`CL8ze{pQ~`P{EarHyUYqj$^@)nWSXK6 zX&~9E-%arYN-Pp$j5Ek7UVfQHH?_H$V{bwOO9b(})(rR_z}`W!a)!##HDY7qO1$*K zlW9RUtC*8B(1`B_aAs~U+u;7!*5Zt?>w^kx$upO4EoKUq9C=~q7dR@xH+MSZ6o1QG zk|`EXfT~2}xcmFX=37;FU2iN#uhB@}UsuXe&gv9VP3E986*%ugolAfrT)=*MQ9aBx zGKk|%8vW$T%kRfe^;_2gz-M@yoef-cvOb3+D&R?JU+3mC!4HA7>vroAA@%f9^C6kq z#XgRVM&8Aq#GaGIwytyF)_HQ?=OH6vpPYag(pf|-U*ku0*E3x%ae7QZ5cCmjvkC1L zT^UVf#8*E=hsuc3HGZ0YRXdHU>fkJ3zMVlh0ZbOvt7c3=q+dd8UNC#znZ|Ez6Q_JD z*dzM#40xc2Qri&>MX{GHJQO?QovA9zg z8O|Mlc;;dh^UpnAjGd_-^Esx#s-cX2ah(3LM%?5R;ysOI!P9a%C>+0&uK?84FyNQ*@oru+eF0xk6?nt&u&B-Ld|gV0VGS8*)GEZ);Nkd`YN2ZCUbl zn}zLnJLo20ciqT1D96PY39sD6Y1w1W)j?Tfn92L>oxT}OtRx?N%r;YmV6GJ8)8_W+ zyZUnJHHP+LirA-!b+|r{Vfv4Qjbe-*$JN})m$-uS0n(B18B^EX#ch)`@1TAF1H*8H z{S$5n{(BQp#JhgMiDlPi|D-&SbHE^rWEg2Q&s#X3%?!3XHJ9M&xjRqF_kUzpp^Ii4oIk6?XW%z21z`WyZF4t$Bp{yMq$hT zclW||aIQ<=UwpUkS$jC+5p?`WOGUzbr2FI9#_8QRp6!=j9XY-lQhJWw9W;_PUj}0K z=d%;-osaTQxlXMTUR?WNnig;`REp&T1G_CL2{9H4yDlp4AiHBLL~B(K`;J2TDSl_Q zoTwiD$n5txUP2Bja($|USKO$nEwc*xPFO@xZ{?6X0dnb=-==?{1=HwEp zkMHo_pW}@0U}@kRBU1dfPoEI?K$jP5MPUU4YVtw-a5|3}5)t;7Y#q4M^XREt398>8 zc=hIh|EXu+voM;%622iOgL?-JbDcA_t(O8FFZSV6Gc#(5Kjd#WpJ|J*PxHkh;hTA} zq0Wec4Kb^T#@DVjsy#D6^S=J@VE@8)YvFyS);t_iAKF+ z7`gl6&NJfV=Lmf~vP(gyq}C__)Y7o<3De7RiBK(ZI=9}y}d zoKz3_D`%X1m4cqLQfiQMbKAGSdOv2|Yi?aWL78K)-gRYUWJr5a>EV%HmGrHtV{P>0 z52e_(aUTj04$1Q9P9+Gg`NR;sa5iTDYBwDBrun3eiD}Fz@Kut}i`!D{EViUVvKXHj zwP@(|a{rqu`PFTH5kg(5K->UtTrGTzALp#B>)8>q=Gnkn{u*oIEPSepHC$$1_mKMi zwHiI4dKS*H{&YT@@pOA&h0SO5cAgHF)`H5YrS&dPP2k(7JKKA0l@*ukew-m4aasYi6;OV`GefYo}h*mO}U0*yQTkeHq&W>WmkrPnSIQdu=V zF?1G%#iR|Zz0eY)I(4+#>{x%~rHl-VtH$I0PCTlthYwnKbbo&K7+Lu2EQa0OJ$`{N zWGz~<2nzylA-Ga^<3Q>3Y#TxGt*)atLNV^vc&j^$Pi8{0 ze5fu2$qx^9=HPj~?iRVPT)x^U7_Mq6f7?JjJ47hMSN4LD9{R}-ExPtW54Ptz?Mt(} z=l0H0&IyvpGST`w?k%`?T&cCXA*pyVa-_w&XU+#-{h#}(8sofQ&nvIu4 zU6jfALh7!ydU8J11>u7wmRxyDx)NtDJxSp!mN9~R^oCP(%FUF+Myr!)RFuuE(pKBunQ_?&M?Qs1Skj$D7!ilQ`X-r)GVrc&yu_s6-WUtQ{Ss zUNDN&PZb18zZvHoYQ2FSvpF;Mo`(TfT1%AIu+eqfNw7=HiZxD<%A4v^MmGCN*PH6X zPF7RBSE>o#Gtg>Y4&|WJ74D%fAWwSo{^~{x zlcgtEa4ke9Jfa@mi|HYv^(rRt&^TVqbU)Df%Hfoa+>CgkdXnT;#$%@snP#fAihQ(; zLfqPnfz{KokhQwOV8 z^n8u<*^u#f58ef)R4*=&ThZACf1aK{S+7vt)4;m*B| zbYGKq^D8gKis1HYTUZ}W@y9Ni>OJ0g3Y<_XMl;eJB?aFS8@@luIn(RHJXnx|C3vhr zedf+<0q^F9hr#PG`{HwoNX^=GQL_ng=Q(xBXu5uvGdm=`e2jUaXD^LTxL>ppjRR8g z5QoXQF?QLces6_2v>ff}6q$Q3r-wQ7$^YJ#LZw8t34KS)MF}X2WfvwDf|R^_oU`#> z_MPq*n_Yj$KhKh?s0za89f56zUd=3}s$>Bu`Yyz~qF?IZ8kP42IOnY^S6_%`jGZA^Ql-F^ZwpYI2n z&ri6eZ>hXyIH}34vG|0oQ>5qF&4Z;D!NDp;ZHX_8`N|Q(C;hcNEHxq*-uB^jM)PVP zx=k1rI)A=qOsgx3`d}AIdX-!6C2y2~+_iKa2{7|mp$?wK6G5-4zB!b}+ws(jPhVbe zk%S99nm2pa5iNN#SbMX!RpRHd7)^tb>nLi;cR0kWhuNw24b0=6j>e|W?#LgCdHdPzr4wJR}5>e`RDS2!2jn@KUsLE z@m+~gE+hx~_YVgg42pr5FZ}m6fByshB`glD#oe1FKc561=Fgu@W_mwPmeXJ09{%SK zW)ljHtoe+Ff3J@Gb9f}&SM<$FnZI}Q>o?$@q+?RTwuAm$`uA_36;T?w8tHu}|Mv&5 zNaWSv5dVLBh-elJJ#I(&hy2RFPZCW7k2rg~Rm+$pKxz`456=T?7#yLY5Tt_ca40MB z?x6bhjF>%=aA|GH2jbVTZ;;eLBfJ}im{4-rWy>OEy$4t|bwTeoLSPZFgj)#E^E!;q zVZc-^Fpv56!d~<%3w4&?Z#At4n=)@BLcBB+R|uvUBXeU|wAr-hny?YSZGbl|kOo4M zK?e95zFDq%Zs*lk4UzD2%-NbQI(m$){FEbHE|d%U-TB{^z696To*Z_s+^>3_pQ^zxlD-Yzf9ozsQePf&s&aQE(OWC>6pE z@|JmkS5?4gsf$9_Afhv$o$v2T$h`#@$=_T5JhLI^jXsLF|1E-5w%Ba6rn36;bw05Xb6x5C{2Qgp%8TF1pyN} zuNSz^U{hS1&2WDYgr^b0z`Pvo5V9RE)ky$%y;JE#z@FZlgK`ZoO$(X%t~m z8iDY-YX%T3h!B|Y+x#P}ucZCVgEp&SBG>8Y9P^g?d~h#5J$DTN5f*(?JFoKhn1mBI8`8L%n?p$6NqUSZE9d9Rvd*_z8wJ|eAK zL7*nNfSkbmI!@N#t6=ik;8iardLNkyYDPSf0+K{-yb3U|h!;*3#itjZxHP-^fKKx=}UK?lgho`*ce-f?>`99IEV|yB+vX_mjyT8$`Da(4yfniihhM_KY=ZM!D2*qPe$B*#THQlRJ7o8 z)qI3KT_XmN)%aa4C7!D2v23q@*XjsKxF6>QBSjw6dE71-r~^JTU4&;Z4p##G>viM( zA9{dapIEsUFP#HO{#^x?m41AC{6+$V?y0me;OEa1_@sql0H$y7-UTT!*V`vJBq-yP1hG z%-BKi2$VHQ>_{=rjBBvM5|CNln=t$x1P|hJ1$R9uI60BoHi<~XCRwRZRTd}szqgK# zCxyVAO{F@>%ZmXHY9hpoz9fj9az*mIA^x+YYPXW><^65?3Pc;7O=ZDf;pg($okd%zJ=*Rbg?q%_WY#SbLfQyfo`)|k{ zMC!`(hIYv|4R~FNBE3m>rUUIK@_;Ot{b6@d48zpZJu3g${=aXq z*$Ylz4JBd>H*OLr2Q2}s6ow$%JDvNT#ST$VLA+Nb=N!sG4EieSCjjN$EvR%xVB6m( z6FwvCEg#-&RhA}f_;!}a#5q9_fW8B^;djbdxfPqjvPT{ugsc!{bn8cUIY?bZ-#6{m zk!17A-TNDQz(N4lMUqXPU^^35`XrF>I+r1qX1mPycyM#<7rKxa5IuE1z!)EZBymc9zv4AHqC!O_vSRW->s*Vh#{|^2LVg=oZ&Uz?| z5#0)V=1IlH98Nrc9K;2b?jADpJ@Q#LbUafOdFpKFKB2G({s1 zAB$P%nsrtZmsZm1HJhSY=~qP1mb(a6QdvRhEz-ZAHyuelgf=;A2Jnan9b&X~_<;dz znWsBABi?WP%~wry3LEKDB*>f;dsSLm_HWdjG=`{oXz|p6F`OP9ePJj$;hd?D^>C@_ zjeB_y+H<5^#nhnxUQ2NnCItWAze)_hO1{il8I6Y=dUIfR%kR0j$&a1Eeq8El$KZ5h zdrvhO+08)*@b=SzE00gda0fp78w}Cf$hM4$Gh|B+ZAMxHl+S3zU2oQzvs4lJEy1dC z@!HKFhSB8~$4;KPsVaHky3sc)S9%^>nI+~N$_ zl7gCH;;9XgL#+fsw63L~c>6-o$M4DW7{17o>2O5VqWr&u--`&=B z9yUoWYyS6+(B`$Nm))?9ZoYm@qkG$pNcWT)9!7~x8kHy}DSH&nr;rM-WU+FDiY$l9USj{u-XL`U4 z(>7&j-Q0&}R!?0HJ%s|l^{*NUVmTTI*>5kXbaP@;jwJs8>isA}Q|FmUlyF;Im3u0^ zUfb*-fDdzA`N3mZ7{YswrfO7HpYY$ifk9Ooxtv=aOZ5Vzrz9WL$4hwD#lw1scBh;` zHs2OD=x)w)6_!d%F&@b2!@!W+yyA>foa%)?dnCUqB6S{!LfV}%Lmor)Ye1?_K>9q0 zoEE%R%U(A+A6{K(CO6e7oPtU0d{nv95RQme``!P!2|eT{w)>Y}^^=HB`I0FE{izGO zO#Lie&pjaq6{#0@U9_;-=N! z;e3yuN9%~qaOqPsQ2vrtfKD+U%pS71#q#_A1|8x!SnV~L(Z!6o-$7Dhe~`H{@jMZl zpW3g){M1l>w_sfS+#5PCLI>W39k4s^R|f0R`(SxJ7cO&>AK?Z|b?Tb_y`H)}vSJmk zTTmn)*~r#wa|c9Su0%Lx%3vvq;p z9DB()`M;dO8&?)D`p5KzKEZwm5_xHpws3|y$TqtN z>~A9b5E)7O@N?Eb0h-ef;RiUgxSdtNBQ{;h%?K5M`Lr0*FL}3!hthv`JzXyu98r>j zbH{{L`N?EDQTgRT=yRG5!EU^({RL_pS9@Q?c3K}SSG$pL8hh;)#|AJR;=wPaUgv}4 z_qB%}pwBz87`Tl%E78n(6qUU`eZHSm{^7-z@|C*J2~|Vaxav7(cVKk=2$FRvCkpde zB-pwmkrLUgk6h$|nzJ4lTq;jHX-J5eY){lmg22-fMk}}G`gD;0PDFf2*&lz&g_9}wwn?Q3T#E5Nfj-thKFmeEH2LZt)f z$mf0>uKpOmtBXx71`(ZbNKAbFYCKR@kowL8mgtF%1t1cBA@70g>UGkCy~Ri0>(lN- z3BJKQvv2B8YZ+_&z@q{<7W)&_k|2Y*w}H4IDaWaxmVSWVzE*uRFH~kR0YbwIQe^&f zwqoqgmzqI=J5R-2JgyX>ennVCU)VvUwZ!?CSMW+tLeo0wQy~uAkLn5md*7Kdb%xfH z$ti?HB)E7vs(M=pp|309JePiX$QZ6GCJT>?a2IOrcO3jkWr|3IZ=La)0Lv+4Kns*| z&3>I%NT8-q7Ml)W>I$J2?0OYQHu1pgtw1NYnO+Xzi=bzqqD}Gcd zxeJ9LqumI1wsLa}a43cB>h7;16M+eX^8c);d_TDDTebb#cT)5ibo5MdTt0ZrYY&c8 zl&-*Vfa{U?6jVN;l9c4dCptQSbX49bQ35ta6$q__vZca*VhdN+dA*(O!j6Q2WJXI; zou;Z->$umCgYyP62xC2xx9k;ZpUl$1JvbhE_F}$xd=jR<0(24*G7-C^?@N)#Y%>43 zl{qAk7ME|-L*Z<|7GEo_Czr3H$bU}m+i4+(592(0&%wO@)M1V42sYse2qn7#qp(N( zUl7(%MZY3f?Q+?48odV`k|BxJZvRpsdbD}t`)4;gTG^)d@p`^4WQ2)lQMFou{^NEB z%PxHO{ho<6sI1Qc2_@^rXp}M-o9k1RMb;YnW`{9F@51`4@vsC7fbJT~G2ncCnKSzu zC|3fkub%RS&X!hU8>U!w?&caNXTz7I*q92`^fCFfAP$o+`W#+tgm~%H0s=(Kw1k;{ z4keGch*~r8m>q;0%kAZnvN7UfJ9d*UuJjDLloGPr ztFru3qoX)sns~A0kRaEISzQg*%{N=BORlx+nvz}^cS)u+zVEcN58qr6m~YOsUeete zr^?x#^>E9};mD{pA?l)Iy2ThIEzWS~cyTyXRYeT+xEL`<0#8WVHP2pN3x)tP!NGwH ze~t&{ucq>MEZ^BH95PQ4ZfDXBzZO7==ysWpDV=f21E6a)6tO6=!Om;bDuGbZUkJkO zp2tpbW!V`|Cv9^|qw}@j)*)mQ%$HKC%C0&}%L{v$6WAof@}5vs-zUC6 zW0QKCpvHOu{RN(aNe<*gEk}V};HXz`By)<+G_( z$ov#<%X2&MW1NArf=NuuD~D4rx~@f&j-?yJ4EptE&f36kovZlBnE0JBPBu*PbxH#$ z&;t*I*OW}=<8k2BEE%&bdVr}SJnpzz4n@(i?;e4gB>*qf)m@wIqodTrX}5d+Df9yA z*s_6wZi@Ew_%p;`q6|V?$+$!Bcu&02nW$jCA;qPo$l&^D|DP2QdY2R+-X3i&3q56& ztnPD&_m98W4mwM6S~lPt3da()*a~X&CS_?*GHyCeIDUGm;WwBMSQT|a*Tphq)-Pfn z-@|B4K}~fwRgErkq8HO)9>2Cvuz2=#^1Sc81f-{YbN;p?EnyzKNA|W25Dh2N&V5! z>Y~h-OqcQs;zYOBd3{X_G9%vm z9hx_u_QC1$Dsj*5_>~f;Eki!c`CKU*KZg;W^9?k9{I|_N0@YhgyF7n*}T61AjdXfe1}e)O29m2nKxlD?9ZdoQ4h zG)pmdOIYlpG%b8|#;Of-9DjuU2dE-xJ^Lb$qYUJg3E z3LvqEt^-BU?0Z_kOiC`ahTRM>Z70Vg@jEJl&Me-V9yV9x<4tS>pk`VkUeiyYb-Wf} zUzTWg16CJ~na6J*Wvb6c7Oo>TAZDF0O-yUntwC!{2zzg5b#*SVbX%*6z_fkQJY78| z*u+!nDLNuurq>|Vyv3lPH0z-d5Y;LPyj0h)xG%X5ldMarqcqwEr z^T7ov(gpeV+UfWdeeC)&OE6{D8N)g)+mX?zM_>D)9lI69@an#Bi_?(Z08XPxC^f@e z9+i)V&vfL{-Ybrr;w{Ei_G(9u2XuD6Td!~DM8Ub{Bw+ z{0}z9SFH|V(bk=Rgqx^BUV4k_a%Y)ibd+@Kd7JT?vWkF7RnT%iuQkE#vk6=usCM{x zts^A)VT_~y2JHl*#Pn0H-taW2Y`SUm{jw7pw>bPU@d86=KUw!X_q~RyM+fXj*fnd0 zcj+md3OOq8jiXFsl+|k--!s+tx!Y_mE?;EY0G{Nx)bkQwzVqgYj(Au<643kH*;C+b zUB924tf~F!RLdZl6MyJHm66jr>xqY4!19bEXnCJrv zL1+FYGVyFzdnYeO#e|tAY{z{r5x<}WGyOy%>qr+(Ci3ltz6=`!jdT?9{a3E}^T_I% zKDj45aG@kRO)rva#aLKfx`n07Dqy<#{snU0p;znckZC?IkxdWAw;+>?tp7Yc=hu4N z13XfaEJEacyWC`+5IyqQLVht-uD*_}1OVbLeZw8Q=7ov+COqPpDtR0U zADVeHih$xp+QW z-jbryJ2%LH<`7xU-ga$>!@%%XvjMHG-bOQL0G4vCM6|=aC610%haMjz?`SjG%Bklc z^w617pDj-fn~>QjR^2!@WGG0VOsniPhQJuj#j~@|NNOYIeGEIfN@GQ-Oy53QBC>sZ zTDRBW`=?PSRUNS&x!WDTdTMFPctU9(w*A!E%J}%)&k)!u%BG#dc8p=Xd&;K!+;O>e zK1?7PTb-NtslFY1%O5K|sSbK}rM#vGCJh{nQ>LD4v1ik#O#MtIo+^`^!25=7CLMp_ z7*%1~_<81;uJPTAx65K>sS;WlAL8^7#8R&|E~HlLJrXrGI8#AlaHt2=33Q?A;;Sn~ zW+|s;_ix_oGb`OT;Wv|-ty2`<>c}QslKUXx%-B`C67OSWu6gYJaV`piGh@WtBWLm} zf9)*_Sv0Q~g)4u-S1re{84m-bS@IN)#7P=N^vl&0c*t5GG6vDDSk=6+s~Xg?{}85| zoy#AD!+SG6nxntGl!sw7-s%do6#4m8o!8LU#kWWLx}^fG4kgN8CkR$TMmROkbUF>S z#9obOEONJC7V_k@2y&IRDhv~lJ3BHEyy@iEOl1^5ed@`)Mm(chG10~g^^3ooMT1JP z>$i;d%Oy8JH1wkT;}bv?65Tbo*`hy@;rgrgIt}6U+J7#p76uh_I>Ga9ap|G%f*v$(@f%=dU*~iLFt-h7?MF7&$vGoWOasz(%q?cHrCyl(1OD=Pb zrN>U%e0~)#qWO$1;bG92yqyKL_@#-w4D6AG+jEeZx$>rVrMA!l@vfB~eU6+%tChY) z0D-d7T#-XN+2(1&D|+;+IzaB9SGntb*N#jpjDR{jSyyEy;I&=|`K`|Aop8lrG`KAF!V)5qcWU20*VCj2Lj784b5~RD- zlF{V{lcxL;Frf={KZfLJeFEtdXyXmlBYr{Q%h1d;{m`j_^=lPoDZsb3HliBj#6!@z zQ_tV#QiCdkA*q5{5H2WwFQ(9Z5+&~)8E162y_7fHkZ`DqWQZX$0Nse^7)By-0L`6F z_p2ShdY);r2<57eAmi2`bF)v;Py$B7X ziU~S%1Y{)w2m3oBdIroRB)K!qGIu8DcNV zr6KjwteSoHR!i=~pP$qgdE7B2U0RAa9ga{UB%Y)dukjEVbQ-F?q(q*ts6Q@3BS64Y z87>B9GTnfXewow$0%gh^2oB#e7LK5`M)k14oYqVRTch^L>y|88YFRcH%I z!PhJ?;03^7r@jn{aKH}g4gd|R$)Lb*^mU2gH;W*-MuXZce~gI%p(GRhn;HmS3DLrc zd;Av{QlbNBSujkm%ap;pLH$4YkDF0FI2Hz&5c%4#+K8`nUj zCPc0N@kj;lA>{J
    hua1e@+R67qid-A||Z{muXtG;OwdP21RXff{;02*Hal=*!Q zXyZ-6Zp;IFJKy4h5xc{13zWoj}!fcMtzE?07b>Z?%zVY=IU{=wn)B_s z_(**uT~XF|g5!V#jv0)H`sXx;99MGyoJBZ!z+V9sVJB<={Ol zB*9O;7ax-yF4zmD46c2`Q(*e&!BT+;(o^k&i|~u;-|Ffpzx}vL?*OoZWXLcFm2r~U zb>cu^h0~C#v^v?}_IM0ExisOrRV z5K>l@Wor;4q1Ga+AlV98B|tr6vgvYXkTp5$95bXNIR?D&1cw8;E)EPNNURZtrt!X0 zkxGeBkodP7@_p(Xll&e^$#dAjwM}P#3HLY=aMC zv>gs` z-~n)a)*F{#K2N$!M)0(mQ>fi)jzyV@65!mc(8SpdLk@xIUIGB^NTb}I{V8HMnM3HC z>s*#^pt5!JfjT|C`2pLfH0@BEcBGdSkei#_U#{d=aWIw;zb@So` z*4TxrAwHqz2FP}F)W_;*Yr3yk((gUfhms6k{;=M?cMABsr_V!s`uQnG@3bPghv@%Q>#4=ntC95YC-$cX`weG`TJIh54KVCkm;C^5kR{Mtu(2aCWVGTVJ-BUI zF8@veTx)BUpoIzSb(kPuk)p&{s`{6~MAs_WQwHO6Mjjp6C* zb?;194FKRjPBimuEwvR2buM`g9Tw0|hu-pJO8>rDn|8VWvxPnl1b=Gr9FSgn(NX74 zm?s(cR{>9>WDZd|@{KZm*?u3O0~%DD_SfC0F$ZL&^Ai>U?Y32g2_WvJu|^1V2P23D0~a%l|2W1ynoEC<_^{1zJ zeT!_V03`*~Cb8B|!8Xb}-e!&&(Xwhv*rYDc>{O2_{3VnjIuCeyxqf-wd9-b1`r0c( zIsMMtQSxcRPtTKBrT?8CSr9jPBCB*sp#LWojUC>FwfTh{;SA5t_-71k^Tv8eQ6mV6)=`_9w<0{w?M8+GzUH>VI)HHr@D|F^8cwSr4vZ`f%E+&IM1QVH$6spE}3%v-SD%L<# zp%W=_jV$Bcx!2vg^H;z=Ti?nTRXg%SO@g&2j(7OympILYBl{4!37~-@x8}n+cffb`1LVO*&}qw9lgbSw zqK8vNAJ%&4%-q~2_lvUc9&?{kpV?WZPu|U^FWFh$p2PY7{<846Ml$ZfO0J)!cCkE1@R~;Lo_g)zl@xxu$sGqB z3XDf9Zi32}{!4d&DS8M}SwoI_B0P4mI$>InVwgeALoeK$|6 zN<(Q#-P6co6pCux8!)x>*Et{LryO}Le?7~;9{T2^BkDCO;!v#!JCe^pKpI!G5dx*$ z^O@gx)F;^-b&&8jCQ6ebUMV$M7P$CW_T>6%{s79X!xW@~*=_r1`PVL&$v#butVP$B zVm7LYsd>T{1}e2|Gba+XOi60jl)vgV0ghLegrbkDazL{TcO_mNSStqr;FdbefF>&Y z-@KXby`~g*$H(@AM?Y#5vLa$%)#1h!Mt`EHQW53+3??^Ne77;2*-fgDy=(Ex1$={# z5L-CC+aB4#rAZwcFSm-Zwxe^yfAIK#X{j!QbD}75Jvz4OAMBe6U8V3hSy;?KOn*QZ z8%Yp^ob_(?CmrKIDG`dFwyz!>T>?S7T?Ct7OSDa|dU5OQfFF7BY9V(7EV}i`niNzQN!5^UR@*(>#Ku~^%bgrfy7*8DIOCZJU1!L(W0h292 zLpz52LH!^28jmSaqN_*j7>O=UULQoqu7M%J#d00eMwqx%YT7LKn=v=wg{W#~j?sz- zpWyYWh|Z8YIMh3CE@xRbgW!7wA5lQ6$<2PVn%8~_esrs-LCr1+@5l>D>qZbs*bi>f zj7Qd)qZ2xiJw6uI>klB**=Q`Kg{*>ofvuGw?A@_s~?sE0D2|C??;qUHx}qK4MVf z^e<}UDm=48<_v_o33$Qo9}=VUF8C$5gId<;xB1dcNJtSy-<#m{RbOQGRyhSjZ6dgO zOimqauYpbBFD*}9{QA^n33PdH0X(&=hF#}p2hP2}y9*B@((3}^Q?q7oa7|?sVGEcA zfe6snWF|@#$gC3g2SA|k&mNi?(Lca8Zu431Cpfvr%5)oyUsac;!I01FfIKW2#0rdq zAa11s_=IF28XEa7>(@r^#=@26e2Z*)jS<{ZCGcbG4kv;n(OF?&bca}bhc;G7LG>hm z9k{?*M!sn1)qshWvkFL}>TS>KgoV-wAhP4>?@%85B&*U!`%Af^xH|sN;-7)6>QP6! zBupkwybYga4TV#J<%it8CuI1;ee4bW;8`U~!s;;yHa0mPO#xTFWVdGz*;v?20940WoUTX4>ncIl>K;h`s6h|%5`kF^QkiWfd+0?dPP{nQ*7rA6u*IFA zH5y(6_G&I@Fmn*I1k$?gMd{fW{)S9j@9Z!V9j*Ym^95untd_6~=PL>yN;&{YXN_&h zm}K=RDVEI(=xY^Q%Qy}4!!F)l;roozLSO2MLX)HAc!iNPDC|hjQsvZ|hUbxihu)fw zPb(K9UWC(FqqnYPw zxQo{VjIFBEv?c_E2kRyTG7vp8^dm@kzC(1pDX6Pe_+bIUQ$Q)S?)>@3k~=*F$0IjX zNqym?9$I-Y`06;0>%@Koqb2FupvkrH;2Fi>Di1T!dgqC8O+ zO<_cx0HJu~3#!{f4i%VBSFizMk)nNlR6(sqT56+ZH(1kbg1`L1@`97U++qUCUSGG6 zn{t|ytt%^`mrnFxTl;hszmnDpImsdA{Gn)H-U=k$Js{H9`iQlDl!RYfCZpj2D;Wzv zoht+ltRx9o2uXN`I;?YCnl1h~vk;68$#_hGp2ZZ{^&f0w90zN<1Vve`xo{p0Wqs>` zf(bwf$Fh!jVwE@}l#&4TlxsEZwEuKpjnn7{=dcwR3pDaOcN&ieYXqk`nogqM3?GAQ z(GzfWZh!(GI<$z_g#Vt}|3fHG1J>PQXXwOFspVpWr)VAkOPz;+^$V#3f*rDlS${y# z7{hCIwW@UGTZj#n7thD@l^6rTgMKP9q?x-1ZlyI7_Uc*C-Kk;|G&ET9C@WR>fyRLz zBtk05Z$yM>xv9HASp$`uI1>gyf>y`}5MckVK@_b~ zGpTZ%^oZHZ7B5HGyP*its{(3OQB+D32TH0r(uLB{@{?)}Q67}CylVQ`W(ZBH%O6=% zWSHh$Ap;`&ttrl55kB5(8oGSDb;}<^PQ$!0=Qv>Y$g-7YFhXCsc9EXO8kG>XSJEkG z?UZ}o@LcdFxen4tx{84vywNg0X;AU{H;yJ4{d9$>1h+uc_m@#pDqU;n3Ibo`D)6CC zUYB<~@Hgv#)GHqX57-v9mzR?^DZ;OtU!KZkEfn^it;cRPpIr0I`-!HC{)Lm6U^~xi z2MyPykZ#n9rifna$3?a3CAMB=Km2AyGKDhlqb54XJR5GKzwS@3@JnG06RHiFS|qhj z-mQKB#dZkY0?h};P9OIafGXx~glLik*r$nRN1QfGNOi;JCvfpKi(l%;)wMR`-zUvw zEl1VUpNxBzQ)TesJ}DVZIjY71$Y()v<`sjomhr9&(YmJ!^n5UK$v|u?f4L7b?)B@q zff~$o4mo`zp@bl#OGARb3=l%ae)mYIcQH`S`~GZQnA+=E9L$Rl)`?!cP)e}{dI(y& zKf9zrCX-Z-{?>-nz1^xx<~2=f4IvDgIxVH$Yh{{cjf4!?!*JN)cn9WSIm71{$GYM4OwJJVW^_`o?VMbZ^2Yu6`0o(wOwCpZvd zfmA?!30JQPH$hOcTSDlWSzP8Kh+{>5AIJ;`Lo%WeGR~AwkU2k}=d40w!+D7ciKY}! zIeY%IzWB$w4U7hr{Eq^uPb=BgebkN+2b0`)AGwz8sIFfDbq$hpct*o7Cci66P#&?D zC^v+hz%(HCcAfVXZbPN(!jY!uUR`JA(@VqB*FH5W^fL+MoG3fu-oF!L5MzVf;_X|8 z*Q%nxx%^}b@Vobmhv7!Mi_-djVX%qH#GU|tn>a=nuhBCr?95io4|V5~^=<-V*EC~1 zL8v*b$$vKYWY6*&9!_&W(h46#JbnD)!RP_B=0Sh_OY!ZiWx+IO{8=Zm4Sa1e2MyNk ze7)yw`m~#k5?%X{{(#&@djE3s6R+de=&MRh+EgWcemzgUosFixUoHik@S{=mkQ!T< zb0>GjJAiYc?U0pr?gRL5(n;@6fU!vxM~Tu%UFAe@2GGIo<&y2)1G-Y73|gJ=X!(av zl8>Le)?^Qt@%)Y;75-3u_oxxXpX><+5U}9e$Ho>&_~S}GM5TUB;y>iBqlkVP!NVwp|$0LY0P zmR(MCHj%Zr)$~hEf_R;qLxeOM%F<@z_ei1yn7R_)&%7xVwt44(C43+*EatT{U*uUCl&f<%dZS5p>WIO6!YKFwL ziO`mn+tT{FB@ksT;RFc3F~Gck`Ds2aaaR8^;rHfp^d*q0^8F`Y$`fmO%kos$ z1lSYFIou?ks4(nsA|k#C<;#oT5vj!}-Bn?yVVb{Q9C0A%*d$&=urlp^Kt^+$kg`QN zQNHS#WGQDSotwh`C2AXUygc?52v>|j7Vc%+&{-Ocaxo_aR}G+d$COm< zR15M(fCVoNIlr7Ay;tp!BRYg!O!suM#P%}xmRuB8+{zcb{@@t7lyAJwuASh?i&yi= znyXz?2*|a(5{{lmMP#L?-?8N$5Lt5Q&{|d$*@@w5_FxJU3+Wr$BCU&xXARxjUYmKP zoJ;*b3%ter?~TBPCTiJQvbWWqDjc_Kb+|{BGLGugQE7GP{nNp{kdmH%TG_4REf|?C z>6VZ-84s)Or@IQVY_@MjGakR(`D_|yH2Npw{-FC8pxPvHzfJOz2t>`~>Ay*`Czz0_ zQuk|AO}{jqXy2H)fr`sRYk1dk$nXhg-qmAY|G=jg>ki_Ne3g}2qPp3!sWcH~o_xB? zr&iZbPAl-hkI0}RdvPh<>8crM!33lxv~zVDYOTYq$@6SYmmC&7ea9g8*TwTF@gxaX z@uR~~HH1vHb1%Ll9kw`J#cN6 zDJ#LrtNfgUx&7O}H;SymXh+PvPT=*899cs-awho=9KYxM_9RafArvwACAbm9If>cF zN7AHOY7q^{S{LTVb351h& z6|~=h0Bm<`^fI2=J~d5NjN)gapIWKp9@HM&%oT8`6EQ?jYAB5=X>ZKsMVm7QIqKmg zg7%N14&2!o^7JW^D_jM$`DgQK(ipwy7K_byG}>;{34tO9M# zCrk1GW6!Vr(~!?uSnj>PVmdwqw!`uq54=S0LBZXWK>^D#sZy#!hf#_LEy$k6>nBf5 z`0#LAuzA#r3O>zppI<&8eZR;yCb_#JD#x&7H1#&&T6X?vGV>qkF4h=Iar7xIEzyFUcOBzVMW zKb)96v&2q+R|WEI-B-(a@n&-fQ1Yt=1&6q9XE26nzeZm%*uy zQ(9*vV6r+2L3)c-!%dLi$Rf~@H!&BV9k3|-7hIP^pCpI(>V8>kKmi{ zbfu>P5Lb^Iz-gh&m~yb~_@R*FIx#=GHM{LLp)Y8&A6#Rzy4lO5>-eD=y z7do5nxXY=R)|%gJUw6{S*5lbfIuwo;q1Ewa`O{6+r)8G9I)f36Kg1YsroZsWPtZ0# zhe15~x9B1T>aU0qTKL;UzfQL{EWO3sJ~HN5?LDdR{8aAGesB1#eqtd<#v}L|pO<@C z=eWI9N|*cZt#F*R2;OX24wjGOEwtO+RN zd8V1m38{#rVMKzH;2KH=1{`s~rQguJ9pwtn%wI?nY^&uUZgHN-3rUycWHd|Laq=T)XBiBwH~ z`RKaTG8`>e;e_FAI)i{+Cg#)#3nEjsP2TjWiz8bL`EDLnvs2a3e4HLs)_>cBnyKS= z(jumhOxYgy28pRW^;mmk|MNk&6D?bPo42}jN-hCLiz*{p8Esu#-_N2uC9EuS~2Sy=havlHQ{?Ob@`;u z653K8y)U=+WKR*4^|HmnwVJ43c;L)Ayc)Yhp>W<@e*ZH``YYpf1RHTZ7Nu!;ZE*{YrkKvX)scrj1(;P<@ zwny;;5mUcS-!b~>GtV+PxC)u&B$>?ncJ3%P%Jh4+OlvaJ2S3Z?TIRhf3s~71ojo46 z6fTQ-L=rIQcjLSWs$&oBuYx;pXZ~4?@2FXdQT}5*IpiEU$K;TLWFHgvyVed9Zf@Yw zZXYRAj;tJEn;evvtIItz(3n(UD;(4J)Mmz5$6$rb)KgSF<@YGp%A>_cJv3R>E|fT2 zDK1_1yN~=ultzdoE$T zcIuqUDH1=>+eyWCwARM1SEG<7G6--Ra-XVyA>p;<{|aAYY%i6r`1G7n`dn>m|GgkQ zxA&gE{~|U#0M}enr5;xK&PIK|3N@s{dLq|NwAoa$u;WkYQ-}VxDs`QFooWBC9??a2 z)!W&1-jwkDS4$;VYgG3DCPxpBUxxyjgE_U_M409}^4-TQ5&zYk7zKL=n$K6=!m(G{ zvhA+hZU)H1PiZqyqIT%~8C!3jP@-oXaQ^(c~Fwf*5XZ&;>{r0RvjY5xK8bnQ~ih z>Zv~;W$9H{ebxMX3SlIMQO>cf|D3nt?;xFp_GGqs4K*2`=$-*J4Vkqk0OT(B@GenW zE&;zh{8*Dz^qWpgHrdt2$D~1@y4QB=iYC3qiPs&d1XlWkDNpylEI(lS!GTaxfBD%n zxcF`1V0!R#M_`$;MTJF41?lwO$W7Q8HK$`(tHhx_!0Q@&5W)rJxVZvCA;u70`b_SN zI4N|s7Ka@{^rc>Tyhfg!zyeT*%=<>bFb^cA^n7X1K^eJC(PP+!cOt!D}6877F}qQs}A#TobkKLY4`@CZA0znMr2ZBdtpiYMzEN` zPh@Fbt=^r*gmP)^31~jd|8yO>b~{ImTsb24NAW*j4}Ro7_^*@vl4J3*m`T=~>5V3X zZJ~jvABMp;8-7VKA7BP^qJ~)r$R-eJbk&N=DaL_+eVySOk791VZZ`ThUCewC*{e>D zPuFpLwsg?Ztoa73SYSj}>sRn;tSwq~W?L#Bnx$(l8V-wr_xZiF`^ZDA4X{85HPf(K zi`U$&f4*$rfHmp$gGQdCR^`;16?(ZT9f?Wq!=&ML=UtrIjY3_GV!=$+4K2JQH-HN& zd@_`dG9B40zy+CTmO;p|j}YPuT2BW!9>}(mH4LVsO%A0L1M7I=Ykpu=G@B*^W_6H2 z8$tpC=NjaktvK>^;%E3?dw~C|B+xS8%Vip_yj5P5sd~C3^E`hBO@?6mWDdZ4?Cyu4 zpkT#ivkDPe2UG=;4EaVK09B2#Tt19Oi49Rx+0L%x7@_zlWU2=ty4nQltPLAA$-jDU zDuGk@2^W>GzboL;Yt}jlw;vVi;Pen6kaF}MFyi>A&;tN7*cCC3)2Bt1!&rB_3+)0p zU5lIqWC$cA3xBG=8DQ3^!F~5!3l}n%|4lq2gv&c2<``7Jp+Q^~_a44uQeCj{CwN#) zkPfqJ&E+eD?5n1&-UxOvMD*m=Gm?L=_&vq7e)-oq~{6aY1c**pebpqsf4A zTdfeW;Ot`)P^~cN8Z*t^Yq?w_tz8%gJ+ejQ8eJ)34(x>5OqaWgRJf0ew7C-MixfKF zloGcdJCss<+@M}>?)9d)FUdr``H8OrE(Syp>K*p!dx;Wz4SgQw5SpGxmL$QN~EUBAhYeW}XFF_Bwl<6ZDSSpdNId%-Gg zuyPCdJ_Uy$e@qg~+m}a* z_gI7XTh}Xjg-Uy-=zr41-@6Z!M1}+u7~9%QgWck|OWi@nK)S0TD21w-bGSF?4myuR z{&F#8pSl8mxV{H**@IzGK4p&WA}JBhabm+}#c&`Vu};Qko0!9Tjs z#rO(UA_J5>(gKq8DbfNpo{CL7PAwg_dVTg{MQC%t$b`ybIIZp|!R1?=mT(uw)X zdm)v$rh*^CHg#Uqb}eP{=00x?kDKaxE1NHxs~Pk7+f@4i-xWYec8$-Y1|Rvm={uqz z#ImgE?t<1!Yw8`tbXh2t5)_K0sPsJ2d3loLuW|u#15X{+;D7$X7BrENv-+s7HpiiR zmwrdE-P{b&hdLVu?T0Vkj?#N@7wM#eQ;L+=T!}^;X)zyK=oGU z9q_`xp+Kx{RXu*CtYnL@0|gAf?|o~vynd5@chfBl1abj&dL;t{HR^}g-$`vk(=%26 zQ*$Ok8|I7KBn~QOazrLoy$_dVz|pq`5MmyP+xVs9lwK0|{mZ+bbe_tpvCM+8fbw@a z)c|JmB1;?pY|@q-z_-t1i$s{vasPml&;emN_Ck6fR+|0%H@BeW1e;$CiH6lzO#`iD zP6G+{5EZ?Dj2SaB*r%2=-wxtkygp#v9*}H4!YkjTDyJZbDlgJPF4ftycAwT0i!A`4 zWKW%7MQatJim|NHW?yP6WMhR0c+GqtS*p$2X;adUFN(QT9|*y>M{nI!3OU-8bAgtj zv+kCjeOd_b|j2AvuDOd!_qVK`jT?V|>BioCQ&+J%z z{ew^EWVE1ghrhlk+(ckBc)Ak#oLhZ>#E5s1C%phKkSn=3QDYYlDOBVd4Tv#0wZ zhOox}se^01rE^DCKc`$afqj}%(@?rqJDMCpkk|1&-89}}^LMpx0NR3ZyNk6%B#!bk0| zz*-3hwo`&b%2`CCAjPPh3F(41{C~a!Q3A1|KENr|yc~NpnqT(lKO3;PHBkQbK2}^+ zUxR6~hVhLpBR^B`r5~0_IadDCj&}jHi$=Az?=8AjEci9E|HqFADHeDGi|$3LNnT>> zm-5bCASb81y0UP7&mcMTmJZQAj2ZU>!N2GMr8p3YUDVrG76U>yn2-=9#lo^16C`xi z9{w900noexUH-YYhTIXr(a(*bbD_y2|4(7_&qq#$Xa~J3hI*)g%BE53g@LFC`fR>2 znM>3{>whm3^mip~=+B9pcvsB-fB*2`pP=N0h=eAS>56#TqS^cGug2}~`BfC=oyxZ? zNpbnpd9`od-6WGG-7@#fe{1tU2Ip8}uwBzTKT}OTq85)NK+I8Vipyx}-O$w=G`Grzeez+TY85i3D>t8U@ zo-xCOvhNS7y5}F0%%U615+0x%WKsqTWCfDNgC-Py{%>r>XUHi-CL{Bj4`#arU;+V< zMxRf73)aE{R<7_vByYSdc#HO*V73spa-^}WaRx|Geqf4E(E<|x{uiWHs>?Nx@3tz` zAF2BZRDuwXu2<>@yg_^hgE%r1cQ)?x=7giU|r4_?0;N@w)rC z9bhxpz<=t34)&6kYANG4Z_@%Ha{;i)y7U9qSpRRHI;d4i0xn@*5bc0jtAqpvgL6-MdWf{;MG+z~e1z`CtzOzGGHBueZB$K%&yrzc&f6oobV2}iM zQmB}E0PsTQ&~W7dRP&W!(xZt)LeK)CU1}RS!blBdq_!ww=i0VJYY2?A6sCEai zBPYKb%Uj(=2e%22V-Rni2n9cshFbIx=QgISEoP=bO1!@j1K_SxUI)C5DtdV!y`U>Y zJSzU|b9;0ccUNmGL-PCO2LfPaImRkYWF5fgOuD;S7qgCJ);?JI5hD+{e{`9-vW7hi zNYxDJa7s|jGn)vQ-1*4#1SrV}DK^s@wDl!wfykg3XBCetxpWA))dI}P01&Ng4<@t) zK*x<8aquH@=icA@fIL%Ra&)94eukjeY*L}~8P57uM!!P3sEG0e6sV=HWkTT4j=~5x`fv_xi}68!PX9!gd87f0 zA>Rz+IjWpFD|#@OKx>^+K2ToNe;l0vGwN5#8goA;?P{m+=TB3B7Z|eJHas*xM*sW) zQZdY(fd-32zy;3&b!5l981P4>C}x4c2p=|N-tz6|NLo2xqsALvbZugDjy@NB!7Hn~ zf!qk<(&?4u*;FBQD0vWMu&3KnO_0&oeFKDoTlVJDlpQ_1@f?CK{?QeKNUzNx&%Wp ztpLTFs?zE)(Bs8j2gdjTkajVrj~zS=i2WFP*}7fux;dhv4z-rAu0hUB&@aRTxn`HA zz>6}dB?~c?4nT^?`0aQV)Yo%4GAQc-M*7RtIT+D|ip6M;>pm_}I*lu@0ns@Z6wl<2 zg_ddv=d+p_w^bZgXo7-HeF|?~BjCFT)bXJu2z$$6O4qx<*mFg;mzTmwwAwt#I z_iV`q&ko^WdFGqTYw%Y8fLW6VlBY}qez+rW{L(GutPUX4ljZh?(_Jhf^pLnyAyz&i z%qJAmj*Gm$#C*cR6aguBU#ETh<>^+N)L6ck6`Lurbm+SpTs_C5SH9ThdjcB250j72 zf`0CTZ`cJ@e*snbJB<<*;+^OSnrcSykMMiJOt_wxrU`y38ng&AAvQvnf_DYnr;txF z+!)>n0^$;Z&MOHl4ojofufD{4y+jklCC`{>;8Rlm3ctW^g)1Wk?;w&--g>h45?vzW z^Ct*b=9@d+xJh+C`Z3tVQPgRx1GJCxI?{7L%6&f^I`yuKV+B(Hz)EY$%273Y=-80? z^*Ag0`a#{w?`xWn&SW3#+Gb7L1&?#x{0yl$<}FAzE`*eT&{uxSJ-JI`eQnIiGA9~S zL%lWN-~nbTm*8QrDu&+{K3e*r`c-g?oQUq zep~YVn=+KdV>TfW>p>=+EPde#PKbu9#drx3*B?fi85mJmvq{&>^tz;LCW_ZmdLNdM zx_$VWnK2NVV4(-J@M9??7o`_-cG;+Clj$QLZ$G3BTVppOmPb~k@&`delf4*Bj#L#6 zIFr9$8;F1SE^*^P8&~i4AdfKtMkHqSFPR@vMzIJVAcBm91r^*kk#s7RUif zuh5!zbBe#X)aEuV@y`MWy^B@BTTR=L6$)0qio>juWb4A8GN(8`UeNHl?4wo`b8*+D zd~O4G#ikB_dOu|y1T0y$W3eE@z@4v~6E!ddgF{{rwK1dJw-^~O&u)JjWOmGj&^1z~ zzZ2wSQchgkj0B3hrthE31hYfAOY66atUd3g&eDbpv@eqGZg<)>797;Awe^})a8P~v zZ8x-J(t{Rpw4pTCCIXr*Ey9QcxAXT;8SmXe!99>TXz(qC3xOaLvV*45_f37OMmoES z=dZe@#n_m5;o2k2&zL*H0>X)WX|Uq#^oA`5c5Bk?05z3$&$vI!e0lcM&uqhd4AIkTf@h%$6$MoA>*B`o4Js`svpV zSYrMNP{*H#QXxxf0s(*JBbkU_-NJ;iT(OjE7fZo``t(`Zfuj*pCaJ_hWu zK9aLG0^k@ln-z`0h|DYUOV5|a4smSfN$^CHc;kn%?}D$j7V5M)N{<8cbIVO698+Cy ztBkmNbH$~GYenRf1<}!J<5m3*OTzt1!o*0r?<=}eytv8V-bJn%ynr#v9m_?2R|`(S zd-vh#$TKP8pzcTB%T)lbXb2irrnDAJ6yEv#-H(!2<`LE+lrXc})&iNGoV1gsWJH;C zkd?zew2dTDfYK(Rcp@Hh7x&VjvyPC@Q9^AB7!BqD-TUDY=|bPB_+Kw!ata zGuf8zB1m6efPMP<=;z$}mQQti-pL&Mn!nVZL+8dWDMdyev8 zh3sldOs@&KJL)KRwo3?Q>5eUB$Mo9>1~?vI3-Z6uw=%Kla7=@1low=u=F)uTp)^I8 z``p_yte6}(ZK(6u$A?x2F$QIb7N0IHdC57BLNHkH5Z1Ulkf9<&Q#qCMRwd$4AkpG|&Y`NP@Q)|T8=9s8#3&A6*I zXU0qIt?+A#W9JVV{;R-6Zab6)!a9DY+Dwg-C6FrXSS_lG5jjg0Brsj@vB(Uk21IY|zL^WMk;UvEd8|tz53fBM6cWBQUn`N_98=#UvHem#H7E~l<5)mL zj<>wMY2qBkoXj4M3>Wj|TkP$s5`Oa|fCn|22%r3cA^c2qc^z-!(Vi}=GG=*!zb<~0 z;dZ+8E&2L!Cq1@?aLF=y30;ib>Em4frFG-WvxFV@skL*)C6u)K_2 zyRZ6(>yIuGu+zhl6W~@n^nTyEYF5(9q)WKdod(6Uj+FaE*T{;w`>B0CREaxi>gSqt zF$ogpS=G*e`(8LwaT1JtjZ=bYaG3gVzgIGG-1zds=)>X??T|U%M5(GopkKlxLhcv7 zPs1@o-gXCHU0_!h9=-fF-=Gk?+`+HqaojY($`D$#t^zmJ&BqF3AQ z;i|L4`aGNcF~z9i+q9a67qsw)sCa1)p~@sJ?qOdf#BDz1O?&!Hfqa&ZuQsr^Y-8xd zi+TZS!RI9RKQ7eEQ{1MVrH{DVU3P)0PTmrRO*%D?rs()G`N*&7cqv2tZQ?BIMUGB~ zyLiPMGIl!DE27J~ckSJWdPmB)?X%gxw79*Wu`fLNFwME^jesDQNFe{VM`BoBz_nvN zc^HBthtIk;$+@*cS-;W|(%E31(MK#Y)Kj9)4Sq$ln`H zR6-7?3PdWVZ&MS9Uw zD_`EH-B;I}yfW!{sz%}|j*UbDS>;mvgTHyJbe{IUFn@Ssw`U+r=h|7yfl4s+RQP%M zhg{tKtptT6MQ66l?*uwqQ3AVoUkSvpw=rdoBhEsLDcyyfwqhAfc6H$)FZ@)wJ_mLI zi8pRxs^2#-O-iD$;m>FMn#NgN3PgKQh9Qa?HROY%s`uS9_VJbY>T2urFdVj{HW9RK zWtIdC!Ol>Ik&(~MD_@bt%&y- zKk0<={9(M<48OxhW9K_FrnY|huwadT+v9)3Jb-wd8>92~-VLlE-~N!?H5wJW-4UMp zu$kS89faZE%PNv%=7@8SykGfjob40QJq|=YNzX8%(NV38Dql+^tk*Z|n{+%DvWc2v zc0{?K&9yW$E{tON<`CL3yu;Bsj0DAnb(jN*BC%d=k6pWJj`z8!XQh&zy|S0SH;MgX z4@d94`BW@JQZG1A)K z{J%EX5H-t3282`kA)vEy5=9ufGfuxjwzUi!3utmP5b)|P;H6eE<4>(p5%8KJnVft} zyCvDl8EZE8x`RD?L2$8@c>2XYx%I9&9SJsK@ORiOSJpOu+cPXi<|rHwC!2neDSJnlOip=)fDUyXfJvw#e4hz6~&Bm(UY`!-m5FFu-^*IoZPxqi0r)J?KTsum#zWtUXs}I)~oNhZ zcpm%^H#=`cAhXn1FYCVdhD@t}3D8M@B!)=~CZ%>Lm8` zVR2)LQsfAikaUIOxj!XJ)oe%XqpWt!as6DGZp%G-gquD@xo${Ve|#@u+Vg^X9pP1q zdkPsjV?cM5;o$V+cH|Gm>s7qH0oOjU?Z(KzXU5(cTFcQtlVV`nW34=}8^94{a@5>d9DZgv9@oh-F5MljfThRPF1;CQKViLpy^J9y%wy!K|RxulRE|KPzE{4sa!xcrXj0ZJb} zr@1!{p??kuF&4UE1-jG;VMV$RYz9Z<23`FyG~(AA^t25R{m2l7r!6F(_PWr6-19dU zk5gsQmRdGm-_HK9LC>5m&#)r)9Dj9LeK2oIbw9gGVI9okB+!834yYD66qKl*j z66cahixA5R9|Hg`uJ!)&hf9~Al<^)`ae5}Y^k(3nQx6%Sb3J%wNU@(wc3|FhH}Yp`T-+{et<emx8A>9#Mdpf zm7#&N!H-D$LAvQE6GL$X!@c=uYqIZ;J$)Ni=ugFy=7HpE? zW3AzzQqR@4x6F9T#Y3+6(Z(Y+p*-V!mV3KCF5DhS=k|uyH|hPSm;;?e6#ct*2QX zskM>%zL?D03vp=VE@?K%zbN-I6T@2?-(T_gVHgK>P-TtDb#pi)j;(%i9dd}{UANQC zHQh2_BSs%TabeMY_=4)bIeM@fOUHTWfwd+w$9; z;_|$zQ+m6Xv7!vy*CkhZ+kQAvRL1aCy~Q+31_z7!!HZ^!Uz+UAN5U%~4{yHP7CH_X zPCGZ*iyi)Zi9?RMO}77t{ancpA{w5%u~G2f9K_gADnuZOD`iVqW1{1SssTf zyb6}0(mOth(4fc2=IUye`K~%gv7amBRQ>S{{qG3M4SOS&E%i+b>l%fkuW;DGy(NMI zx{*Xf%GBx-az0^P$Hc?;-%k-p>J(iLSW`Lr_#02{(q6?Mm9o4p9t$IayT1xQaos$? zE7RQj8DjEfDO9wkxnZ@WSR_3QGpvz6nywu@cI0ybiKVwyBo zl~3Jl{$X2_EX`qQwUfPWCoYkd%tlQC&vqYk`plshDIlM+9ys-uJfdcYFOy`_JCkz| zh~d9y{MIDw$m!6G2Ky+N)N6zVe-3uL ze^=niGe_HY!B9nh6HbDAxTBxJEzq5krhpg9bc-w-dXBU?jqj!Ge2#(2mpYS**eOm;G@@HoX&cLAQ$c})U&S{esXP{qWejsOT z{ZjB)xAbV{i7lHBa}W@TPdcU z2gGXqmwWG;vJDn~VyOKFzl8qE2tnz|*jf%L&f`SRLanaXa%$!^wpVVHE8*S5Kwzo6GheH`I;+8q2ZjGPA{xsU$)^w0uKB43n5QOUr^EH#=PW>~S*)^%9lvHYbI|A&C3QQ}18 zEvRu#|IJTz+Jg{&&UTq znuSK`h4=$C5^N8ApAr}#!vubReX|14?7jHB6L%rxh(jQKT=1K{o~u~F#AR2$2@J>sW>C8yp>!k>Uv`*L}cdeC792(KB9 zw;m>dqp@CdCaBP7EVq90gu(g=E(M@U&WXLE_vks(pX|!qC4Vr-!#WH<_rGP_WJk<) zb(j+~3Az3Sw`n224VZH^GhaEsy|cS>v)%ah{Prs4&cGddM;^S;hwZ!b{?RowCipRW zNtXC!-t9z&wbVG>^Q!RA^{>0Vg%{3BeZk9IJ*3mKw;g?tbkNw!y+uy2#j2u(k$DYU zw_TIF2{uO^E;3TBD*UIP{C`Znby!th)b%ZhaOm!mP6ectZb4G%kQSt+8$`OhrKG!& z?(XhJy1U_9=YF2|eZTAS5A%4=K6|e<*PLVg#tMqrDgYDJZHtWi!Y#Xwm;)3z{&Nu+ zmGp3G{%C7HBgH)~Z+yoZ^IA0=RDzcMLT1*!Mah2wq`6l3> z)BsYh1qSDCu_8L6~

    #nE32RufBV2PQ6P=%xuJYrona|7r}{91+v z2yPv?+S$;>yg9V;Bt{8t<9XIixdw_Ov|O^ejeWKfC}fdT?q}E;V4sj$cb=dEL-E>Q z6KO}x1V`=4^I`V$X4Q^$+e7mqPGSKWga9EH?B1dDy>Gp-^B!O@25X3qUk#0aB?doe9OG0$N{VLH%GBF`p$EOiqA#ebzSh zl?<32k9xuJ)yX6voF7z9#m4&=@FZI!9p%0c~98YbJ*MkJSJdYx;w(#l;v z%F*{=OJKxOWz`YVmvEW_j3smEH1|h9YdZFO0QNaRn9d<&Bv1vJ)Sn;nRKYhXN3qH< z2GtC(lPb?7V6bALOS*W!m{~|W<$8XYn+GTwqs*|RV)L+UwX^czIQPNG-*U0|(2b{& zNa^Anst$4hqwaYqJ)M1!Rdj4B*5j+(&x0tdm~X$@Yrg=0jE)m$X#>RVzIcu{`NA0w z+@lT)M#)3wji;vZW6MXBLN-B6{3bvqIAew1$FS^D=tg~RVH`Ar)WBxkIU-3j^9q9kKdG&9xm>&)=G zcVb7inS{$Du#f^X=JXf%B8*Bjue}QxlctZ~gFLPScKMI>z@J>zY%`@s0P9q&=EM26 zQ$xC3tt2%!_?51K`tnRv(RRyboaO!dS}mOZy0tj#BMt6Cu2bn5bP z_gU-W{k52MjAVrw`(-nHjB$nAPMdfk-}V>7j=D%yYPR~^jf)hwOJxf>J%h^+$1<(h zG~F-f5wfnQcVdG)SNLbAu2;H=lMm~R@4{G<4baZ(AT>`_*YmNJM<#)&vei#BEm!Jn z-^{S%j#^MsPaNEy@0m*FDHTw_K4MdkrUP=})mpE^Js&b&$Rl6Xb6?&%JlM9S;qseIaxnoZc6pq*LG?k zi+5Vh$aXj}&4x#j4Y7<%eID|>urjFFG%rsTrjI}D9`m)i*f=QUyqL_ham3dCda27i zxC+bb7Y0Gw$MnbCzaIjr3R4#I%}X{z$?bgvTW?%}G{*c-AgyU5ENLV%3<~4XK8)KN zn=#vTI*Nyw`-mA+k=Yfl1#<(>Z!urq?Z90$k^-=f+bakR^<3o0L@~WIi-kiEDXyc5 zK{zJR)l_!jvFkK+4I<3=VINUw#hrSbuDt9xIrGr~Yzwb7#D7~Ee=_^sytHFj9zPzb z$s6D23O-78E45kubZ|xqnmzz&Ed@DjzRf<5j~nFAtXmq;)t~3*PtQ{6?#WJU^wP2| zdDM|C)xRO^aysjQ24VR!x=~Vtmd$ej7;<-2S=BwQne1pd=!wHfo>@#e+VNl2@%8W@ z;={|jk55~fJS*=<0;$h`$;|4K7EIuB67-R%FnxjQE4>5q@KAZ1MB@HwLrSs~ zq&nPy3P`e8U^9kvnl)dO0CwQ5*%iQgun$eZiVA|HffQIXdXR+$$B3o7R=%Bfyih$Ru~pvKfv>6S z1-7eT2#3f|4B?8$Ax_nm&wv6@^DoK>ARHv$Ffze@m3qzXUBI+EWim-9z}U+p2$a+x zQ=&kAowxi8^eLc%;=c)!dwM5UTw6SyAC5Mmbr9e0?`GV7;WJ<=#p!6L7F><06qAa8g+2K# z&TbB3T9jjf>`!Hj0mQ+*=vk>qxe6rU_z?2PIDuoQ`Aj>9h16ui?p5By1!(Mg$Z(a+ z!$C}F(hDnF6NN6A^hBUCv#6zHMjBXTc$Xu~dst=h-n+mYALiXcxse7-|DZqx|1V=$ zSApTK9dU?hCBzO3hNoOrXDRoz9Xa_;xfq=Z%U9l%)}74cuP*t7h)fvmPb%7T2j?UA z0{bnH3yl9>e-YP3I3;C*gGxFo$wARGIf)#Izg{c>WLkL`kmt_@r-;O3h z>@Vf}DA^zm0vacR}QHf-~Y#;=}D zbLN$o)sb1a`^CoH{EV1hWE9{=>&kn6S)@6-X#)}+h&(}=q#FOXiZ{7O!=xy2y0D!R zKn$*1T&r(Q|3d80ym*fY;^1_FaBZ;z+>XLjtf!ss!Oz_$Ximj3);6PW{pEfOLLZT> z>_%bQgtzffB@Z+qVT5weM2v>?BUm!j*o#$Pc?TIXHl2JuoH8qNa76S9eoHhR5z0+= zrF|@>TLzI~51m8hP6*}JPJ<^aHrP%RRGZ&vq>NyK77|*ToRV}P{b8GLcc}VI4i(EY z#bFf=@8bOp*x2X5@27gH=0FcH^P;Wsn`HxU_Y0G|RNP;SQNsGZX1zZ!0+L^@aY^?* zbAS_d`a+=;jYT;s2U|y|WZ1la75R)U_r6VSz&Mn1`ujGg;u7U37py@jr^-p$ax1+7 zDKtOg5cs4D6bqh@fXRTzu$7wGAzkDw#31ZLHwYov@>1bk-2hs_SH=YL~81{$yu?dDgmNsq&vu3E@SXv%Kue%oM9I$f|= zUz2K7f<3~!MWWjHtcE>=koc<+T(zhma?b+|57PulZ@TFiztFZNhdg;b(6CnLNhnjn zdW-6U9Jg|SnGv*uGcyrVdrft$?mPSs95xfGorY|HcqJ1W|B-=05#;@*<#x#G02Gg` zvz&}1sz;dAY`x$uEq_&}DQLcKu2KxKb>Dp^!}^hnjM<4-Sy2ECKLgckk^FncT9m zur*21jXiZCr}TI00q)_YH$0_rIDs_j}6*N-tEa|T0s&_HXb`Jti|OfZWpm?6O`y5^ zuv(i?VOCXr-IT)gw%N&j-7F4!vhmpsqwTAj5w_f{Dz4toW;x)lsgG=3ML%?aneqf{ zyPnuHDBcY4t0ZJGV32M3A=^}9rSswY4P@Qc{<)Ug;g=F$L?rQ2nc9;G+8(A#J(Sr{ zprT|ni3M=_C)-j*BYa1#EWP;`-0`aK>6Uf(B>H)yd<51`DR6!S(!}j>d;VhW=5-H^ z=^=AvwzDxlw^xMX8$IY;K_wn~?zWGti1qXy3`OIvum$WQ{WUF)gY&eH_e=P|0}RE+ zzmk~b1(V(d5#%>G{kBhqi_uV4udnU$&bf9@k}2Z?kePayU!;t@nH`<@x&nU11yoOd z4n@1j`f_YtIBU-Buxo<@pOC$|`K%K5SwoE{f4wRF49SL~qrK@Oy1Eie*^6`8w)mz> z4ESp1&ED`Ao~B2O2hd4>YYo^hLp!Jc@(Z71shgaiBtVYCrNXTZL?T-Nn6g`K%Za7L z?c7WagYlAN0|0I|W59Md{oD=^W5$-{_*~SyZpL2!hpHQ=o*$6GD8ypT1~?_4J|ouS z{Q2slJP7dw={VpqXfR7Zt^j%UjlgM2*UrLhb#|Zy<*>(f-S=;40$YaT(Pz~`TZ3TiaO5yj$hi2-437g>%L_S z9%BI-?-~Gyld$hFi`$oPW1h+X+;q<%LS%1@iFZ#E;x-{WMBcg;x}z#XE|$K5coY6< zVqNFJr$rmz^M-ag|G*yQ>zK#0^*H~b>bPg^WBdpfH>_Vn$#RBNV{~+_KSd%y%-$bf z_niioeu#JC^H@)*>TLSj2iBZuH*jZU{AM8{*}DwWT=)C54lkY8!%Gn8$a6oX{M7cb zFdC@P8BgV8*2z*j<7pxG&?=f9_p%PF54xxhAC^A`62A{57R-K}Tq8|CA^h4X`gOi1 z88CDOKY%X+n;RaRyB>=hB?uDOXM;{>f#A|^*Mxf&ZmV7Vf{5I)dh0f5whEYk^j63U z^KbUx<-a>-r^H^^i8AJQZvd#R{CmNfC0_Tyk)cnEYI1yi5_!w0xmk^&mRkPC^vsJl z(u-#CH*S-yt)UT}+j)KlOG1QU_Ed*gVvJsa!~=oo@YG~}_$3pdixhLCmAjTCi1F2g zdnSg<*~zs1af|fd5459VkgQ#}S4|=pS-ZBK21F-dvOWZw_R0YSgZAn* z=QZtVA2U^{#n;kq>9jkC@aj%}hahFB4?xQahc~-*Mx);~PkwZwyt-J0#LFvsXs`y%B|%u%4qPa%>S0t$FHM@`BMmZzJKOB*Ki}|`D@2IAp(jx zf^1p4s}FP%I9R;BN4|&)nZjKH^9Re7{{S6j-2fv^9=m^RPGsjsVFef&2?csQe{)+X zjFq?bCV^^2Kk_7K&FeD;uMu#sI_8ahjFH|XILYWxlQtz(B2kqcG3Dc1FaZIzrvT~d zz)J5-p4}KquvexAZa(bmPG5ZjX>(gEWj^NoP9EuoI!6g|_Y2*@uKvaS$PfU-Z955G zU~GbJlw!|<&$lDlrgZ>%BnwU380T7!(HZdI-S7wZ=QIQ9ygJMAs8KO_=1P-SpsSF` z;VDs%S19vJU7pGv=f3sj5#|x$b!S$g+}`eD>B%ViB?=?_$2e`5-SG^8p5hA`Jp+Dz zhK}=)HavLpKdd($t?2EmJ;Dg2ILqfFO>t|@^xW;AUimfaz3Ad!rlGFMv5;Hbx27lc5=Fcw=pdXcM&LZZ3~{^@rE*wq37P4u>ONyw*8(FY->VFBSw zZWyi?CB|!?sV0i8lwU!6F(dfY6WA1X%ZT)`bV_D`aLRoQ@nub@?cmKjEoRB63~=@r z^bxP4E3D#w^cxlJuWeJAm#M|W5`BVnQIQ0go65v&h5sNFae`7q6)b@d12uz-&HfTV zhqViiSv+rp!NO#kJROXoeetT*}8nSzUoS(4UO0EZzoeNw0BYIMyH; z{yt4jpsDyMr@bt}?aA!k&T~B|h7PeV`cSH!St1 z0_ls#a8bQV9-~iYg^}K^mBejzkzG+wHtK?wLuohZB9r=_qkKqRxs_!t!doAD%2S?i zx5HhwIXabO@tb4}Eh~Gwzo)78U7MMzQev*Cwbz_uJp{guMH{114D)ScG#5V9^{}|n z9yYtqeHx)5IAx|hsEa#k+_=X$wjx(TTnqQX=(ZR%pn)6Z|C3QrG$S}*Z>n+}WI^eS zAsF^o3ru2$4?y(RPZ0x$!+ae zb9R_1JRkRUygzScuGD~Vai3w11*!Vh)6MhR_a6@y72ghv29=tdiS0Z%1<38xW`{Fb zNxP6Avwe!Fdg~Sj>POiG;)w#+daaROie?GYTB`)^&O4PIiQI?}_z@x3*eYxhgVsgAU#q<|9HqP6eTXdC41K6#_~^J zXEx3`7y{2b!g;M0n<~wh=0tWzYNFI1>Kf))8e5mks1$p7H`5?jJk3kr^CE^gTDVt5 zO2zPT`Ra7^IgAdRD5p-d7&%%ErbXk}E9|7i(ib43&MUxQk9v&YjZp0qhK^m#pT$Bz zw_9N?!gXJWn2NTiYgen~rY%bF`0bq|O<>Afo+_DnT5C~GiZ(dx6GAw>e}u>HGCS>N zcO|1B87b>)@7v7cQcs3&^@$5ocwKVTj^Dls`8knT(8cxDBmH=PmvEm>`8`cts+FcgdO*jH&8lRDk=K)95>%>w1 zb}fj76f!)D`ayqafE}TTWiOm+pW;uwo}CY5)O<)&ozY@GpEc6bCQBhy#iY5}ZAKikwbK46_YV(yK<-Xk*H7MT~>X*yZss3xZQ#^BGpd)kr!^Qd4LW zL>P9Nfr-6;Xg)VJy}e5wMwL?M^FgH z94g$;!soUz{|^h`;H)lV>%5?SP5mNdyDd|GT0nl#fSDhYKhEcnP2t};lfP>i24w=8 zai~4&KhA;65u23xoWDA%=WX8|UtV$ArnwLH!w12Sdyq200&O$5chFSTfI&Ax;aKr| z7OE6Jh*TixURcryAxyG&&aHqB2kXio72a^bbj4uw9>ektLy~~I*GKSn7f=_(LtJJ- zA#6V_=+o|8XeeW4^~1c35+Kp@(&OW{`%X|r;pUHNEHdPDR3<>On|>;yXWn(f_9d(` z^|RIlBs0=I6^DYEzMuZwh%BnT<~u=WS(1%BCJ)6XJUUg|?hn4LtXsI7DB-Sv_pnr8 zsj&9B)9pNkbW@U{@cHhd5m@xRlqAL+=cLNA4*MlHiN8Slm-|dZcH|S^x*K(S>Zm`? z9RYflnlZTq_m^dfs|qZnD@s#rB+6Yfi!p~MA0v>orijL7ju_TQHA+O?YLwWQ_s zF%l;0oAKYIm1%-FPG$NLRdFWa?*A@HYH^+isrm&`hGyyOy@QAlC4~?Kohv2E~7T zjma(u9yi?1kI%?8DdSOqLlSEZjgO!i-X^)6=%1O`jo%DUS$0NasF3133M7w`<(i|m zwo=%AKdu}tvba6zWA`uXkS6*yVfS7ju@pr#UgCUj<41T#a!jQ>r3TldF0sRZ~quWA+T)H=M{bWc9fO!sy;qkK{t`B~gzl0?g6W zkqqYThPS=n=<()PECtB$)qRu<9v(+BjHOak)##1uz3sL9s3P{ob>eM}^RLj(CYy}+ ziCe2%Qg222Z7xo!i2@z_J3Nl_e3|`8KKij8Gt*_6A7CDlb${(Fwnn1f zY@4Aw*TUR0RsT^kTAoYph@v@M>m_yY@VA0V+2mpN*6kS@!+R`X5owsaXkn31|U#g$k&wc0EO0MR9U1cnKiD2rC=#J+VM!D+9c5Q3Bo)Yxq zb!KZ5bqM2nJ6)dy53vFZn}Vim@Cz~=oofc4VlgWSa?Cp(!b)#(O1>{@5Nwv8rM?qZ zzZ{u6n5b$x@7WoDE1%4=n#$@Uh%CL2&qXCa`f+^X|KZhUW(!HUV=zzub=RlaaLeCR z!&CwHM9wr-2St^`NM!gGiERpV| z6z7^X-PO(LVX!;cgx+drwVd1g)Gyg=2 zt{t(ALC`t&Ri8VbVQ<7qPYu0~LYFI*<-i8#d9{rHy>j1?s?>_hrETd)gP@JNZWDtb z3-FO+WToRa_Lit(Lz0P-Tynw-ztLmz{MMGBaM$8po@sdc&Hmp< z9HiOwXnwqE78GG|f14}@$Mc`Kl?PLfCWVl+Bb^eyS;fC(ynLP(HbrI?q_QKLjGgws zBlo?f$RdkZqHUGr%xtkHli2(+pM(!nA$Ip`u-lD)3u68Cq@;ojJ&=Kw7KzEkn~~GA zpzQ7w8)omP*Eyea8;z4=0)%q0B7~<_V+~NT+Z_|!7x2oxItLZ}Mm04%kENR4DZbIo$5XWA(R@bF%~(K*k*``lFRUx%D(UwDB(QF4vv zyIxac5W;{_kx=TNgg{tNroMvki0}25V%3|p8Ouh;;G|ncy|YPLqtM8X0a_`&Q;K9L ziuM)1N%+yf)0oAJSf-*;`mNY}>yssj6VN+V(N-_}0u@Bi!lDbjFIxA-|8KI8o(ut6 zp|WWb*MqWa7dlulS)%Yn)ye}_RTd`}j`_tG9Jq%Xu)yQ2lVIidxeLw{&kIV#sY(zc zDW)^&7D2AAwNa*neMKQu({j7b?fMV%{D7^fP%m}7`gOk)DrTNMP&(pfK4ldhDmjC6 z??cdO;nvBkQcQl`B%Zzr2e}7*??)Y-9S^JfiL@u2X(!WLDWwd^#XGN^3`#XR$DMsG-w|LFvH#9y_e`ynT>)LO%Ug&sq@-zgk&gd2hr2 zYW40VDQzYT)81RGrt5R$_mXMH%<1aUW?IA_*gfK7I_^^dSDREM=^rT63>U-I zgmw(&>4W))nI`CO@5f9YS-x(C4&36(^cxI?;GiiFRJ>bpnwBQgw|P{3OlX1+KPD<( zzHFb5^ERW=1wHZeP0EUt>F7H6)5~%*r0w8uW+};^T&^BgJgu%C&b>D1&quz+AAH3& znI%jN^0ZJ7`Ch-GzF))tR3ka%|$tx zmjj;_p}VSoIoYl7{PBQYt?S?8e?T@TK*S~)ba7;I`ACpucI-X-WT3Z&DN=JWZsfQK z{$!_$G9pFDlETvtp3G_0>mUg?oqEkf^?pb4pchAoCV$ij%ulC>Ucy;)t5)9uCEsMK z$DK)X58ZSraLrQqTtNW}8rr;P6cJ?kBikW7(nheN9giZnb=$*k2}m_zKby&m4dxVN z4Z=)7>)GAEtnAWph10GUDY#V^<3FJlm^UvIgm7`W%lY1OhL#wh8y0hX`6;a-uM0>j zRnUlxlKMTy1|a0K0L__dsG`UNaD}GRWH^l|SO2ns}nI1n?lTlUbrr1!@@@`nx~RG9w&{E3RN+)TUTh`xExx;vNVn&{6xXd z3}4=0oqR~85^ALJMhxInawoZH8HFwkNU|Uz3hBD=g?O@9dKa7#=+`!Y!G`Nr{12`Xd}42_Iztv-^|t{YWl*Kw5Y8Npk+*Ni5cA-gYX;vKXofC~Rj@8l2IscXXH?r8Zk5>y8K(6(n!pEVpx#2)E`-u#!UXPiSg#F0 zVo?5f`Y0R9nyjRj$~6wGpu!C;Kb`kM^QRes#*Jm1Z5 zfeE;u1z--Tuqw&%9wa>gO8E=~qJe$Vj5L+Uwcq+mdnhVX6d%R#Y78DXJ|1c*jx z-NU1Dn?SH;4CR?bvN^8#;@7;w*BYfw`u@jmf~ncqave0PHV1Cq%%CzZ=>CH{P#)=f zaEg*iWm~v9Bm{4fsvpE~aNY9kzB>!btQ8PKp|KJB?;W@>o~vEgXJET1m!eDIU{wQ` zsiT~v_sa8Lq!>Dq%VR1Q{7eBooS^m|@|x@p575_}XpZ^P#yVgl2h7eo;uO@nzbLTYV?*VxZ$m z2L&{3a$Q1YJy?`(IZ!?f4A47ypZdVVNqhwbq69j&c3jVdCU^2~1(w>z-Dv#klcWlb zaoQG~lp@q=|Gyqw9g8r0JxVY0{tI+ZBikWF}53TV%A8STkR_V z6@^Y90>8g;_wF{CI#EAJ;i?nVEBA}^NY{5j=0tR1@68TYkSFYZrMtfXKt%@tD?iRI zXLRUJGNE6gZ9e@q4)gfBpMHzQ4=@9=j_$(KEK(W_O*8a*g(tbV`pIqBdUfyR=RP1M zNxDh@_}v@b!RJab?zo8fdmyolFuGdp-?u+9zM${Rx-s~suYm72=4Re51MSGSEr!xv z5I(viJ=OJM?7kV9sgy6zbmt$XLU`*>k59z6UfUnN$us{9AkzJgxC!`vBg|4?0=5R1 z#Gm0ebOQC{Jf(a(JFCzc=g$1GCDq7+Tm#EaOv5(yEhtj8zmST;bhI_mW{8HtxPvss zxU2JS(E$-clsJ06mZg_CA&?q9jGM;B&TZt_8h|WSU3)LzkAW@?ja<@Azt{MlZGk}Y zLuh9<&1ElNkNCcJW*i>Z4|CyJd9c4kKFh-ogZ3{F-{iD<*I+0c)$T zYU0Wr|Kav6rWYmlC<&wIS+4I7K|BJ|L00c70T&j)j7xlBD+qRc(gvXA3c+~0rD_36 zS}qVO%L{VrVhPLM)reTqgpBa6KLQ$?h(s^p+6?vk*Y^fIV&Q%mSUby9=fke~Uyo!F zDZcycQh;18IY@(nocUsR`?$}GcY-WP8QFZ-h?~Ld@ba?~y;_-H1h`GA#w3VKg7U(N zgR|XjZ^a%0`8wYQ`13VwGE5@QNat!oMUF#F!V&?7oVNoyPvhZqn{S&g#!tTHJIG8Q z?nvMKAR9FhWnu6?`w0se{eeQ$iGz-Z+JKU5zZ7lJ#%Q0d&@JzCB<3A zEeR^4#u9c4(g$O-*KR4JF#EEa=Qw$zT1^b`-h7Tqe(8PuiR!G(77EJ5; zt!*wR>2gdmtLYifr{|S5?Bz%;Ix%Jk{_sZX_;oAIy=14jOHlFEX2gJ51q(~GcIGY0 z+pOzMMJ9OUc#s0eS~|wx-sWU~oEXa`3XW}U0KhFg8GffymsAS1-o%FZv|Lr z>}C@h<+sm4*b8^sz?gIB9!#jz>`&_rkY`il0B~a{3#Vc^zWbqiVo}~2SiIMp-wG_k zH*0kyf3&%L5B=0sUpU46dA|4j>N9#I_TmJ!UL86(fn3qZcTS|~*ay%r%7u}I-q#CRlBGHA`%W~P6|z5>c| zE)aUGxiwicSEwQd3K|l`)x(p2u+4b?6uX5G=Vo>;BCA6O^$b107W;zm5rdD+1e!N{ z5X4kcR;&HRd53f+9>q?eh0a}VVA-G*(-J=u?~kLuMKzoOEX!y1@sV_*q)ilg!fW=# z>o|3z#E8(+xJdt0Fd}a#NAVf4+RKIan+2gZ6IPmcFMFoV$SL9o+CWP|AF`aj=PvIspwq^Q+!3qrR;aOR6@S5n=>(W61%rNI5}E%YpZ>1BA>1#JF$6v zF80ULle+NdNY`yIPOzG`OnxJ31`jZ^euPlHY%}vVMmF$covf0B#?%)r^XW-AjqsJx zXJ%iyHtHo%Nw_${BSNI7?oYC8aNvv;HZDYM@sU*`KmKCdX?iqqMHT+NuCK@Mv^F#! zTZ-S(l(00gnPD0(`#OGl$Cto5DbY&qBrDDRErk-bG`z?+u^R&3tPAQ>6Zgl4(VDWZ~2ArjC@_9x6V6XS@bFT0Kpi>jLW3(bp8k+$J@B zRewQnD~+YDFQwr*@5QL7XwG{u9~;9ScZnt(^>oBevxqV zjC$Qc{(RHp6SD${+}gc4ADc|p8c9C(JA7$ywegF<4Y&Jubc-U+Z^xGPBA0|=L~4X= zB4^~o_V2YP+YU-=7R6ETnDypiWl1KFAY-nJ(6wWKRUK4yCikHtg)oOs!3x40_T6mz zkYg#mFfB>zaHB~il;hie;OnIuC5YAW(eikNHRz}D!& zBp^yu1m$|TkNbmdxbpku(7I04k%-U;oU5Zpqy=Z14vH0VxEw=jF?k&|{?qD$c85o7 zFT)90=@tLZN?R~$$b*v*V)G2(qU<$D&-vuUDOvl|*xLuQ7)xZ&&AoCJQ@`yi9!|-X z=1=2b1<@#xlf8W{_1{c5m%{Qj%F-m51E|;i>vaz~F0-%o&P=2_HJAwmf}&k zI<-Cl4kLmU=6P=cBJ(L)TY*-OX;0ZBSm5Ys5Hp%m>zkBF1cTWd4;+bpXS9pq!%EfUqi!t0 z!(3``kd;8&7x2;uDcpjvK6JOe&O0$t7+%g$iLJ3@Whjg&&+&VFYQf?&-sqL1jFr?C z!u_;9en#a`hOnG-x%m?nBE?OONQ#PSE0}I^ju=)WMnc9N!hx6>o%2fL6J!xew9@44 z7m^Hec^ReEcp^6tZ)%e9CKLP0s|@09V;WMk-N@nE1g@O-@CHa=^dRNWo-;elUbUE@ z&zs?wm3@q3P!OgIraXf zO%6Mi@?|mGp@6>14C`?wYOYG`D~i>OpMKlf)AMq5d3)D75q-4CEM!kc&9fGgte7*Y z`l$5NHk53>@IP(p7^7jYwcy%3cD6{aT$6{VWLe`^L;6&Bc@EIl-CO_uz1S-{e?oZ~h_iSZzB{ zp&G=-F)_a*x!!yt;#Z26CzEprxk-MV!~(fYPDdwyqEM}Y?3lokZE4Z|@C)~hZVC)}XFM$sc|XUTyC671-K*IUZ$%`PA8prihg%7a-0 z+bhZt`fESJ=%X3+9zHSs1h%dx&!-8u3d|3)LwrLmEZyHSQ-8p^0^KVwoPL4b)|bA7 zy_EfBuQx+O>FvBzC0+Ok{iNRg6P;*PM5Lfj8+{2UDI!#hYla!Q8Dj5p(z>x}6!iqK z&cuzU)?ZV|g#OTSiM}xCn~^MJU~V1w&4~BM<3#wptKJlEMh2$hr^asI99=>kLFFo3g;PG) zVi^mvZd)?tr5NaX>(_D+veHm0=z<+TCgM-)U*o3&KpM8y2H!6lY?dmWa~DvD7z3?< z7gSVAQMm{p!kBm%wASeRTKu5mIeh>4!@hze-ST>^>|X;#uyAz;+U;Iev;6HasxP#iBfv~mM@&ReC-7a;TcJ7FPLW_ikxqG zx@0jJQ`yo!6Z2ygO}y4PI6o7(g2uuD z5c`z_G4mT{q7a1G?I}hwS_d$VSST!Qp07#*E%2sLp1ntp7VfXUXy&+Hrh1;yvY?x% zUO4>6g@}4p5aG^)qqk_YD^qw=T1iqYy%*XLgqT6eyhvTJ-lipWTZ}PsUyOc$xeL*5 zG5Bh6V11r3@BS{@47_5Qh9>h^A{PZC|UyvNRmL3n$WK zbSecB-lmACMTS zph>A&C2@^1GC-&lIy@{?ZKhgCNfuV-;DI(`+&7L4eIIZS`z6Ui`ELXn1iI!L(7__^ z9tUu1oC*Kqx) zj5L`95Lwa%RJd{YsrMUpqQR4XZ)sU=@c$MYO+c*l&aO#p2R~Q-c7SpXgmN);!*JDK z-iks=uML2id;plcf59`#UgV{mFY9y!O6Ro`Ek2->HwR2_MqC5uK+qB?yHK~%kp8Lc zAaK?4`7vC%*GR|1`I*H8AEvP!sM}o|3RIhlU!Qd0cbIQ^49*4g>8hhJfT8CkP6G_6I-t zn`k0K+-!C-xDx54@48DzXg0|T@g_?d?;sItU&dj;CywP)A_Ye`&8}3=JZ;d}O$J(P z*ZaM~OYpz(fBYHI8p*c?_HR! zkWtW86^J!NTRYx?_O7O4yeT0o42j>}2LjgaEhb6VJYdv!4&!-;56e$@)^o=MW%u$a z)h+7tK#g{vArT?|N_)VB5Ge~tW-U5c>bQ!#>%rwjf*%L&KY0EB{V^&6 zYvSD>|0^OuC=zORVmp&3@c?wxpfcQcu-p9KzxNvU6;%XR-oFgS*{(<)FbGAyrCKbd zKLhCzb0CW(=ua3h^ZlOy0vJq5?3=rNy|jNBMBO6CNWeXbt{kcH3@+y^j-t$G>e#IR ziO>KK2=p1DhA!X762wDpr;WABrIZSiu>2+pU?Skk1#~+2XTKyF{@+XQIKz&Ka&9H? zhl{)jl24~YEdY4<831OlSl1yn`47hOKX1E9ay8lNPJ_RjplQ~nx~>CqktKWy{1^+s zi#D$p|7q|4az_G$!oe7JiVDA;H)$>!ehTbOSa{_JVbXajRZ*?ak1NlFU?MM;jt{vx zJi2yesZz4Q@#;q0xgUF98AnFpU6B1=S=|-!;PRBR0IK94(D`)810c31dxvd+jy@_+Z#vP1^k>Oe|HVdtjUE|JjY#3ID~py% zbvYr}w>dQ8||eE#!*)|JOChI z8c#B5a(p$p&WHHl8&5T!tP@MOh3wq3_0H_wAK&8;*iQM4Rd)h9Y8A$W&{i!M zFy1ZzjD8xi_idTX>(0}vKtyD0J6Z;EL~6}7HrLa%W?dG6%eQDrO!lZ zd^`0T)`B)7*+1y>GavcnRINLwT}CZib%`*i*!)}Qz50eisK*`jzRN(5r3(PywUeQS zFfh&d3&I8F@YHl+lJ)!u55SzwyswU@I;wJ*X8`IMT%0skum&4ESNq9~ngMF^_dkK= zV3#ci9eB=b`l!Z*RAR%FJWi;E%rDcl&lg&#{~45y^kylKs+R6;{ zuyxLPLBH-AGNMU7D$7`Yej5mV%63FxffG&*a^4}T%#dZ(O=w*8677Af{-h8c#s8T? zSx9cCHQLs;!4uCyYnbEr?IS(C)@$Ry`!(qCbZdMM8g}=))*uYiFQUbRc;SKKxaM!& zhl1uQxIe<1dhP!+d*xFShiMBZdUw{#KXdU%3wDxy?xsrbiczmaOJ!|=@qZo=@=ch>J-{(@4jAbO+h-)s zA`ier&NA7j^@ftF8{9$)&!E-%9n{E$V}ZtK`1TkswZlC7vy{Qi`oNeL_7#+a-Qz@{ zfA;5b9lwPYB^PYbd?e3MOH~fAiI!PF%~?F)XC~RUiaRaOkEQ%4tuOUT2$nzWNf7id zfOYZ^41ExT`zxnO@8p3gEd$jlxMo&ASb^&aS_6nIfPH$7bp#P@&@rgJk>g?s6v2m} zIz&7a7918`C`7!(_=SECeFy)tlYX=#TFiWPBt2ra>yGXuDTL}`Da2Uu^ksWM$PoSD zb3HI@iZ&o_b|U*@$kl?-GT&=x=ue*UxD;JAxRv2J4(i?bwM{MiDP;@b zR!8035acuAUi=~2OMP}2NP?C%i)^J3|BLOpDBY?;NfPrx2|WSICZ zPD%iD<`p`5Hv@@9GmgNBm+O?Zx|3fwkl3;h90-$td)(~?665rNW@>!xSkw(k#*g2` z*eFr757g!jAhV#O20Z^9 zr*dx4`4P3{4;nX#FyMGeFLd>cQq^{*Ap{BDXMX_^>fn5n=a=QfRXR)0SPp2T@&ZzV zbyOd1GILDu^I~w!AgImyzQuoys~w93brxG6$aqir4GuvW0mK<=50gF4OF~jW4RV>j z4NfipS-HsCI#Ej%;<%i4^t)^<;b<*UXe?Q2cFvPm9%vJt)&rxVrsNnwPBs+uR;j=r z`DPAWqK?P5%U{Sd6_^>1;1k@`#9EIhx4xBmt|pHy z(yjX;xU=`{B^=QS^#qfffKyqh5O3;VkB0ll)Z*+=0;QuAq>;JrqlO<|cK>pX#tI3i zhdLhfj59;I{=eehI;yHKY8O_d8>EqLkP-x?Lt45Um6DR~@JI<#(%p@MbO|UZC?QBU z0@96ierux`zxTcWedAsS!@)T0v-e(e&9&ByXFfAH3AY=auxjgOco|qnf!0Kn%1}!d zV~`M$6JVyY0`LMt@q``9bdTO=4`v*KnkmOXkr#eDFbqNnLwwDZlt-ZUPc_wysmyP` zaQEH+jLE;MG%88FUmPv;C|ao-D8BY+D=B(Ig{VIj7YG;AA(6WpHQAH1#k=Oae@+jw zcE(aENcPyl-N%j*()5@$0O>6|1h1aYg7zrgjTJ^8$Q*k_bL3a^RzcyL+GxO5I}ai= z&VH`HCi=tuNIE@VM)GAM^o)aqq3tS%QwXc%_)^LAJpv*67^tUnrH=XzD0Xhfj6kKO zn2%>)SCqFWy+Tx{nF?pl>z#*TZ+djF+sq7}h**3rrH(*@E!$l^i(ZwqE1Vd?Q zvj@|!!}JYVwFi1Z7}cP*Rw`Jz^8tQ2L&&uJ=^nRNL4#$+uEU_?318L1Sahf79x}3H z`7e~}$kAg$k^R}W!Vv~a5|^6&B3>J%I@_+Pw|Eni{nZGUwfma;DGb( zsz=D?)XHso%uE+-WFU@mC?!qjo2i0EwD_13VLp4Q^+LSPGiRSD zmv&bAMfVHTNt&IH9RZy97_}@;yr3>(oVb^60JJNF8*}ZK!`)&KSBA@Kz2-2qT(NPe+67fSA6i;x(sQR2D1UUlaXVz%A zs54>#GDsG=R4$J0^gQq1SvP`rhl}tY)7i?8t9y}w1+B~*%l1j9!X>%i|fepA* z8ldP>T@3l0^p^2+h#u~}gG4-)FAO=rfIAYI|Ixa_&byaqq|U8vXi5pf=W{ECrs z7coQ7^-(I=-8U1QJpn2yr&f{(Bb$5Spkmq{EEGS-4LfobJ*i+fb!7nuBaYZY|5G=~ z_Yq;3ruN=XsRcmXYK*aQ&lZRep&pcIsVt)#GR)74v(2QaXF-uIV`0xmjM^sX zcv9!qj5#|#r!wihEagJrX*z8@?`92bya%KXtU7(Bk~5OlvV^oyn`Z~$lwgGdU?|aQ1?>$b59tT(WF$llwIf5 z6LWXu)#{bF&U;U+&aT(k!(68f9F4R>hPi66gF9bUs^ufM(JY6ymw+TUw&7&c}0#$R-qbvw<0sTmrD6iJyiW2k#{ z;8MS*9ar9cu!FOdeU~AGCv*Qw;>3-cWFQ6xD3&NWhP!pezQ*cU6nrB#&n0%XMAoIz zhEGkuqnZDfEw$@r#0$;9``$6Q3egl%b6n(fZ|@mLI?OWOyb{98959o!@~-G3T$8K1 z>qmZq1^kcpt}(;0dvDVnj=0F}%3DkNP|(W)Q~)7I+@~@vEXol``FQFg#RUF7scN($sj6qLwRR{|yF-|40@ldTt9zhm03~gEDFYs+2{r&$C0Hn|Sh-QC2_G!> zH0Nrdfq1*4%Zo>C5(OuA+x1(};&b_FtC2q-tk zUwtN>@}BM;?OHs&o9*qdBN71Ns6}FHuCck7W`*7TZDcP5#7DTrpznopC~E`9wjKU^ zMq#IfE+n|Wf700ifPlG-^zrwm^<5>9rTJ!g%9wLTX?QI~NEptT#>RU(o;7IEm=6I>mR*{TQ%VV9_=;lK{i+Cj-(>=VqxHUr<}|GL(i`BRw4mVFT?o_- z#fQuBHJET1@4%q+L2*E+r|_~yE;L+ydFbdVVk)$QUJQ7uu>gC@_9P>b;g>MFa6VaM z;!iS(fnJY)O~S?1tWZq`Q-6uB4RPg94gakbl(?U|#PJm&KF7a;T#zeH07OzZCtFDN zpWD@52e^KI`xK@B=$0;k*)GSj6gXG-y-Rx12mtK#I4}dsZ`Jiv8x1`wmCEoW4CxzE zvz?B9$UDp+blKoOH?STI4`DF@>>d)`RVdrC;U5dtXx{<36%w;tntSP*DBok}N;I$j zoE*S@L+<)+HI(;AXLD?7&_-s*S)MDD_(i71b;+}(SJba^Id}WLHds!F4nO+5Sr=Z1 z3L`)VH0lU+5dnPnWXWtbAVwVO=K*M;Bxrg*{opBD zhrASndhT#n@&s-2q*A?xG=wk29m^*(h5#u(>9aLnxzpFkgZmsS$Q>)1r!c)xf|3P> z!x;a0yHJTSK(f^Qi6HRp6uy(!%$;?2Rp%gqDg}uEl$~sw8m>W$UOieJK%#{JQnXLec*S-wm4O7y&UieU%iceZ~CUYC3-zTJoFB$FNT=PmFOA+G2QCh5EX3=urQ{2BNQKFxXGeJ=Uy%*)vS{W~K) z?Bx%Y6@aw=yrRYz{^yoINBCU~3goiA{(8Y9{d;wQUnr%3`0V{&tNuqCj)XvkC!7Qa zNdA2kz+rt#Z^x-p_<2_RUWz^XAQI;^w5w)`D1Esb{XMzbKo;aL}GD zc?P(9x=@+TqmTib>sR97?@P;|;)SD$4R14m2KFIHQImv<_rO+Tp~Gd}v$^u&&$D{$ zATymWECdg<+5J5?&@V7P=r2UOhM+kZ@EQKQ8B+J$g_ryyr`{#^`9;N-|4d>8Aa)f{ zbGt45U;V>|y@iJMKfiIo66iERr|$drK)~43B+uYk$vmWcz;fT5)4w1tTQZuO%VkA3K{rd@&A1qpi}|-9I%i- zWk$6F@>`sEumImOaE?z{Bzo1VYGnom4-bj`ako? z2nSSCq6W1$gX1zv@LLVRFA*RVtbJC-7-g0KLV5Y?OQ7X1_oIv?#@LlBeA+ViMAe`OG(g`5cXLtgN z#!=A3(Nu3%ClXMhHY3yES$sc-aR0z<8+G1N4hP9k8jeIzM$$sK^vxHm3dBwMi3C7P z$4pkhto>e9f~xsF1kv8C2JJvcq1L`cH|FdCEMbffka2?XXIKM#$Af*qDs81}Rq77~ zyr$zp@@h;0Q2YTNX6tsWlH)_rShGksR~^B;O+}7%pl}&cuRDb6i#d>!b-vm6m%1y) zr+{DE66`|&4}t${Ex&t;(z$Va^;^@S+vg|VKS*_5r~O}lZImc)<9u#dl9j4CR(~qc zwL=DEZlw!~bgW-W5EQ&%05(uD{awO%;I><)zO-#JM^?5?wrefWfnrM4AXkt+bq2JK znn=y+1_(zVuf3pf9)5t-VpRVac)AcWYEkS_02b_P@j3g=ZXbpv_y|NSofuyir>xvZo92A&ydksXzs{lBMMNg}F@{ZKb8?Vda79wS>^JUE@lCVR7 zQ6UWBMQ;bJ@Iw$hmxIJN6)K>@={*1>g#opOuB+_`8yIUJSpqA&Thb@x13IF4ux0N8 zM$!=wu@*TTp8f@4n*UZO95Vh#48pmQCWNoIt-73>?uxeK!4xe9Le13or(C#W<0z* zP$kLfe46I$z4JBDUmojJ@~EijRatNs2>b7W%GW?4+Ht6YlN3h_`9Xr~czl?s2gsTo zCLY;MLcj*|V-V4%IC@iR_#$=V@T3-0(=hMEl+v`5;q4X+Gj#X5SgCEuhPHQ!`x6HA z)Y?P=B2@+=jwNTD^)?I*MGje@R#L|x97)`GZ~T?(N9ijtPX7?FuHJu}c_A``hK>-z z8WIdLZk5Iyf}peK^gNkKv@&}45Qw{5fDnG*$DVhfg0u@ToL4F#ZV*#B1m3WJPUiOg z<+On7u=Mr?%b-2z^sIy&h|V8)_K4fR*4VSQCJ4prRa`cx&P_!Nz~H;S{QQAGRltMi z{e2#b1yI!v*Ssw!)&&VXM5gsX({pipdJKf<*T5!R0f?*w?=#QAw!pgmz`XOjJ3|j0 ze7wEr3^HO2aX&q03y3=tBIZA)ZoNWw3d+G+;vb6oyigEIIbcNJvA^ltu*0$_-lG`g z^+K!orjutc*c;8EoI!kZ z^8#b#7nx6_UTQ~1JMx^l*yu^hQe*?t-|<74GV|hZiJ!-a9#;+-@b5^?y0sfF4Ml2& zMASjeFT+3<)~NrX`7KHXIOHSGPOK~+sWdKDFkL1qyJSp#`I9$PmB%ftb;fn>AFR;Y zwXoS?pQ<$i6>8>b`g&xo=(CAN-hL1oc|gc>a)@!g3QFHDWd@OkeFmT#rO`cr{qeUK z+I|TE6NAD9XoY-Oh{04{Z%V*bv{GhXp}7fo>Ez>TqzWs#LX$sWg!h^?$z0zTi4i)M zmYQi5w}%Q~s6h3{h1;}VLOsvEnNK0&ke0R*U z-G+{Nhw#}}_?tCQ0VgwV#Im-}?l!zJ`KFhr0SBf)-xCC?bmmznfRm4yaK-EPxyPta z+fuzNQhheI*OiTa1|e6yQOR>Z`^hcLF^D7m15)Aj?ur{i^)B?Q`;1AwoIK!wz6XRA;V@=d>)2mCh0N81iFzeO^9z+zX?Xm~K+sw=F?9wqtm)BC1qsv_#iyJvR- za47V)e|tHwb*iWPfsD~Ry<9trpW0Uu#gJuMeu%c3l--`iX{LF0@Xg-!#0&ChyF(rJE#6@z^nO2Z%AQZh5{!S4 z)@Ufmi}Ws4XeK?3JT#YiETlc2_dBKl#PV$ {^9Agy4$~ztcxQytxxHf}1=f3hk zFyIfsZGP+I*NA}IGTe-G9=C4z(fUFt?JfWummnoN0^kW`A!viQnB#U17xMt&*2(az zG*Expl<1)?MnwK25sKd1l81HMD-3T*Q4shBRc!msOJq=dK4HJ5iKQ0btHO5QO!N~= z1fVNh;a4$a*^LcE&Af&kg7B7Ax$H|?@LY!>r}J~-P!#G-Q%-j!AOp&Rk(c)2xHL%P zex3D&^yBdYdIInzIxlmpRoQukw+dotUrf4g6XDB{zFw;PhK=!&Av&{c6(t1gJ*IkvG5SN8A=7?}fNaH32V8Ck|cN9?T{GkI< z<;NJNZFR2lX+^1DmS#!Z6eBt38v&MLFJK|WJcHMf{`B<0&k=~;ODj%>jFmS^5*BwmEW02l zONYFa*vO~ILN@j&puHdVp{v`t+uqAgvAo?$O;UP6-nK6@NA2Wus1t-nj3Be0t!e!#nhhq}1Sx?fQPOr3rHWjB)8y5qt_I9#2X-STXU zaFc?sez-wDZ&~_@15tg{wFbBtVtX+)wj8|tVgyMQ5nfnX;ddUE!QI+^12%Ji>n+iz zIQQs@8sn9TCXC>p4u6*w%csr~&o(=6p*X{yPlbGp2X>Fr;)!j(-*n#q=Uj&SEp&Tt za#;vN)50_Q+(xK?J9~+?1uKhP!1yuYQ1t5aKby!Gn#~cV5+tt(Mq@^LZgg=fRIh7q z8CJ8(71NMfU25kEnt0iSBO2@x_6!GSuhj9 z<^eAUfH%kP%) z-jC0=Ij2j!sh(6@PN}^~xQcHhUI7Y+X|2GtgRYotuFf5n)Zs^V@0YI3`wQ&JCcLGuOE z1V3x#r1KYsmq1y8-jT%0|Nu<)k3%5pY6V`1N)}$2FbFCwqW()jF{}o zxA(rf1LEZ?R#XRJvpPm6WAFDlcN1Tc2Tzo*z2B$WRamtlv;;kcbbZ(KJ9 z-1k*CKCK6CX$+~m7S*rhs81=E#fS;kPG@{wlQR*J-EDVoS@6^qO z(B>>`UR{Uz8MV;C)N6ZpJI|qKUuIv=yC*5NQBlIfe{g@$Q{gP0@5!!4{Og*98Q)2q zU9$U89z3iE-m%uYulwu<`YNlOgo~@|B%f(`HK{B1>*Q^`DfVX?R%kG@)PMBoQ4#4F z%dkn~N=fLpmrFkn*?_KdF;zP2wNG6WtU2MY)t4(;)ywi{6mA|QtC%-t=E_^=1o-4Mc-7b%(|$vZZK2g|vd8F&MZlNaU!r=;Qpr@V+fA{3^^3*^ zR}Jk4c5YzWASZm|_pMSzlQCs5tYwpI>@G0)y7AZ|C}>c5SGh>b@y)c@N^+6H$l%Ix z)Lzks(#ZPgQYWXAnTpg@Nxy1k;}0`_g{r!kGPbGRp+>aB_53y2*SrlUGNtz=CZgkt zYD$xyrJmWZb-Ub7IBXm=TREh6sJAM?&#YTMzfydgb&P6k5MSkqpyHf(=EfE%*q*w% zNFXsJ8&{h+giQJL3LG+}k_p_=%zFLE7iWtjgV~no$)UZ32eKf2*D<8?i*QA8kt~?KS%U2e<*^u@jbXSL zojktf8=?%VD{~COhccV!(M{Fr`&la3knn!}M_CJ3icly!wH!At*JHr7`!G**$Z^bw zs#>|qCz9^+D(L!rzbg$V>SBUzoCBmG>$^^5+VOR8c|QkmF^Z=mWeCV>DD^@K`$P1v z+|`hNKLmW88VY_&0u%jdQN1YXw`dYrfA0MGlStSy4y7}tQ6Cs)P zvh=VP7gkqQwK_3|{ka;t_XlbbH1(UEH{kysbfGeM%TUc22eH3j6BGlgm2x;>|07(< zt8nVK4b;D&{QdohIFw2sn>$5&sliCNO-T^`{1GMw?&8l_D&K`{{)#&^K~nV3Bhs9 zagxOUeu{__3U+?sin~3}r4B&Tft_^)#w=-t>Hm9ffV2nbMA8e3S3ZwKGjcdBtM$b{ z#{28Bs6cA?nIHbS%~KKZCU7y{hiSY|3}Ttzpu)cYZ*l|Is6Tp?b^G>hd=-W&(e#IK z7#J9>@sIIm7v5dImB=#uLo~I5sEp>zPcB|fg#6+GyA*frjZ>F5d46j$^w$V-n2nu! z%!0~HWLhbaM>&r>xl#YR%Kt(a^kn7a zL;ohx?8>*gcH(uUlL!H>DQ1^59|rMt0Rq_xNs==(4G``F6xV@p8OuE5T8Xu z?Jb+yp9M`RZnU?$hmUfwK5Rf7i~WhMhxK@QyCmh7iR__^&IfXj-DP}Fi=gA(7Iwl? zQlURKBR0ich9{~j}X)^m_JnLmSLB3M)?f^08;gIe7wr9 zVeMEZ&_rm#5;4@EnwN*{sQBWS?_;6zZ91?E0B{9%v%+qz^ySeL-&t}*>DNnNo@#2^ zUHsZI2|6h{>N8oqxV*Jm`8!R+gcK(Y*MQ}TC(V8qLb_jG2g6&iDHR;=u~v`w&VNbK zm56((`g;|qu-SgOYx<-CKlMWpLm~!Q;`a5-qe^Ne2be4&Xy@iq0Wt;U8pWIoBPg%} z)x1P8-pkGLAzsZKLHSBC8rB`#^bvKPEacT<;1P`po{vk@mqi5(?k1(R($=}zKh&+S`2-~jz=k}Q?IJ7U=lC$|FL z>N>!Yu*g5kRlYn*5n;MdZu3vWmzOm0xV(5i0$glBa)^#JkBuCy2-d^E@UXeOy!>2| z8;S=B=R?2OdbPaKKJu2S&g~<6g`Wxx4_Bm_B%f@~S~E$=%RNuH4e9b0SK)*4YBq$*^<8d* zwq>?RAPNmiDIdtGzpVfM2w6=B+9eiLmRD9-yc(_Bgh5-5wa*M`=r*U8x}V{GYq`iF z0s$L0Xg6C<6V9!Q7So>Uktlfnc7!U>ujJ4WUWFOFiELtbzH*k;NSO&l`a}ELqW`zq zkdJwFzq3x?tFrYKXZ=d?X7Zq&p)d%_kCzjgu5+TJ!aXL;9}&6jx@n~P#0OxRpNy7J z?Mhv0X(WT~{)DMbJ_)Jtq}A5U8*ly%tG%5*(>Oz$!3K}*<@=%7jAH4gfZ#pMvFi(} zTG2os`lSkCi=;ckkSMrGc{<1M(3+N;1s<;qbgcHubs2b%jEnE)Q1}W`Mo_{MG1b~? zJ=_+$)4HvFP>y_GIjVDOB(~nqx-?=K;CE&#bh7cX!|pV%g2FqjmA%F&gz8ppYOOJ&-6+eEi_86is$zM$mr~lmTqZgBkF6Z^ z?<4^Z%LIetOv>GH2h;tjC+je$rr5i1>MTE1qpPbc$XDR=ZSfL&8pkP6%n(#WiXUwF z?@?fk{7MUFaxl;4i_7r))xsjs99Ur7ykWQV?1wP4IiRbw1e$rq$04hgK+C8*r=p^Q zR>A%eZ+dd;r#geKf#^iS9<9em9_hb?^^O;AS$d|Et7U9&XX+E^LHw%nwr9Daa5)+q zM-ufYGnmowqH^aynW-?hz>cM{eqf*!0>wJ-JMfo_N7@mp-{IKK>A7|Hd6Z`aB{ zN3}$7TMmvg2tV{2sj)Lb(ZM(GuX)4;?POw7zXM?yBmL` zY}9-)H$ZhilXLefptxKKN@ednHWEmqxzJU*1^T>M(xysNL-}@kHV*;4)h%5guX9WX zcfVlhrl!3LFkN?%-#e(by|MwlYzp@Hxt*(H&tAqmye6qXi71_oHmqI&PvW z7w75n&&S&Ws*;~yz6*|3oOr06TtK~zP`3+>n_?0V>o(@)EuM7edxO_$sOA@Ysi##1I_^Py*Pd61Sq|^a}YPJ!oq%AV^rNn z52H1+Foxxe8#Or6)v41B4*OGTa{-Afz-g^A%m=(EJwL$W!f5mueqib@FM0(Q30{lX zbD2lKTQ!tIcoH=5kkBHXdRqIj#ufWCN#REkv$m>k?>|G6F=rryAqMvm2 zS|{e+4_5v8wqIU**1mQ{g582FpylzFu)0u-3dvywS17l?3R1Igg03^vnQ?{NrHu8| z%Qv%Yj}+!Lb<>6?XjEM{k-FH_KjVWc$2a>8J|{GuXEHwye>e=efJT3NBWObshBgK;NamXL&vB%Djf7Dl(2M z7eI|wDaaxo+bC^P)KGYh3v2zhrql6YC5^X0hqJvaIoJbVe<<>4Gst67B0wSbNMO2-7y zt(%_j`tbKqI3Y@TK)_y*Sv5b;tUf$0y-0eygD7H6s(mpileM(6Lg&jTnse`jTEMJ<0uN-N zyN`Qm;nc~+CnYZqmh?h+0(Ksk60AfCi0wzGl{B~ks0WqH*fxg;$a9UlKdT4ikG?RYA(Q)oaI_;o{YDrb z$}0mhp{eXx`ia7Yh2>@eLuCe>`c+&+1_VZVR8>@UFg5wiV2Q)xet;}@CA9g0)Z{op z?#@?<7X-CeKU27f3)Ym90|8#x%Ml7N1hi(|J;$+Boh0s|osWcv-Far+95*q@Ean+7 za>ufy&brIM>|`4NU-ccGCqSdD0iiloi~2k++Ijjo+tHy(1kfB;!N)~DNZ z9T0p5ONVDvqbx%9YG_QZ+5AtjFkx#|b*`>ecTIZst-h{qmg&&Z{PfzLPa<@oq6Krl zZb#o?)Z0TRMP-vg1RqEG8FYZ9@R{E%k3Oi8%&wOPo9I8=5&QWGocz zoYmOP_PPf56JI!6uwwikR#%J1D&~n?W?hF!6pn1%1rjn>WnRlIV~aU7y;7ioq%0Bw z1(vucmg71M1y2h+cEH|<=zF@XAzZ0&xZNgv_EKbyGycL1?~i@jYLzHFwL0O_@-4eI z4KG{hDBppcWYxKYwX153EqUjD6qcgy6YE^N)Qelg!5<9eM-gTyrU^mJq$Ba1kF3DP zQ@k)RWLAmTez%i&zJ3ts%_dLwey>eSVRp74cdf}&uZO)$t6*Emb zKU;ru_SgnqBKPBwK>7yfhEeCb#RPhP8u>AsvBpkt@1D6yG%6<~IT+j)NgSh?eH%bTM5&slDoB;h5--eg+ON4;Oi zs`y$a(bH)A2kcsx6HIr!@emI@ObG!5`OVg)E-xx)h(8YUf8;Z(oYcw%3Lw-5#U z!5p(0sU>rVskSD>a=XrB=VExr1mkML&mbi=oS~P=dhBys6^>r}s#=}2U{*o*_<^CM z6;>v6*~DX+_c9&D<#ZR5EV2a@j`?(1*!#rQ(%hV}vC0z{#zp{aOcJh8$R);OO;6te z+`~Y+22u+;XRGCzSU8?NfpiGDtX0UqQL$c6RkTPr(kmWqK87Y4_7>7$;9Meh=V0e= z8(infPZw!8rf&YBivRY?4H6hF+F}*oMDJ^vt4u?lUo(5VGoKi8?Tu242Kpxm$l>YyX(^)1x(1H! zUHHM${$S$?4m?ra+83b%wk^yNEr0CEgzWxS+gcw(nx80>u2Jur+Z>}ircA?xb(QuV zP3c}sd2elpR88hUo?~cJz=hmB*?w(R;)@~eAm`&Vkwl!7J&QCP)GNWy)hJ=>t}Upy z74*%H$WC3qP-S6F^W~9pfi`3Ii$kOUoMAcZk;)Ii9VkD-)mg9iDYnR5Wpo%R`Qgye zWN&h)P~)+`O2tG1b{v3u+a`+>ya?>6y#}kvf}NEi07&1l^y#xOJ|m?frTOpIUZ$}S zJw7x%XMU2yP1^!m*zayoAktMbO5`oNBhWeRiTsQMq=!H{_ppowdU7S4TAzC zz60wp3z)yBB5}yr`E`CVDY@$DIXNGpP%JDMIN>%ECOxmS$`HloqPd*pJlV~PY ztDVKKmTFH4k!fW_e9}S^VzXIPV0Ab3Nf4?0spU~665ccFN^g%bVmq0zykz)VE@^S9 z9EkqaDEg4^C^dGs5iu_3(tOf}!kZ=qto+a%*rRxE`6o1oh(*k!-VwrfGBqgX46z^J zsx#lbtZ797A{^4W)32=SGyK7vC zu+#bJmA0vPVD1rjz9%M(y$QWVksJ9?1TRG@j;nfiV`(I-=7kMR^^Nl4){n(9!??Se zw!O|Nu8rooK2=Y~tY?Mf!UBIqm&hOs)Ulod0rm_5K*CnBd_bgzr<_W>)3lTEAj@i4 z0&W?>X5<#v1nvTZKdHQXigl{BzT=pzA7yu6d7gHf&;#st-fr}JLfpHH)BCMAe#I0m z)WG4Wp4FJ_YzE)oYrBy0Ai{lbWKWUJaI&#e#rvuo)4?J^&7uHFKxPN(ov%Z+u}w$J z=cV=TXo9@`^)AC@S=iYGsfz**wOQ%kZGO+&6}TK!?0i}L?>}M@_Pg<<{I%l=9bb2T4YWrJ_5d)WvN%Ci&MB5Gs)J?B8uVdUB zX^yN?&j*gpFF~Myz&G7@u@BmzfK~DXrr=J`9jw~DQi7Wtpbtj#V7|+w_VMXrVlk## z-P--`?ZrEZQh8P@CsnBI?uHJ+CIQKvLo6Eix^KF52%Rw2n0Hcjm9t9uMof0^jh6HY zG5xYA!KMw~zi_8{p6@jT!u38sp3;2Dc_k$ERJnQxRUlC+?vzh>e`BdK*J}DxAbr=s%8g6g4;)KIR~rTK?+o$rnH&S+Hw zdUF6hRtdlcUjPT#_2a?nMw6BR475S1^94YlP@xH0MQGH-hT+gi4ujfvinpf#&^LT? z_ygQE)fw=7)gVT(-)B`W8fP_4s)6xA$adBE#5;DG^^Xd_fEMFT+D+c4Rsbh#Gj<99 z#$>A=wP;lQt8c`WG_6-UZVru2$`T5JW|BolQy_d6xCVf{XKTqnFeW{^5rt0PSOc$x z0R3U2L&uFL`L`{Jy9QP8ZSnABw+)Js*z#uD%Y zK;kM0QIr933L;Aaj+?R-T_bJm{0gjzK^F)-wyTuByavL7PWo&s<;%?mC0>UBl1vI; zYd719a15FycC3c+ufG9s_)im8_j4dP-rlPf#u)%!aSrF3uttqQ4UE&X51RU@^8-Ww zU~GNV@^$x0Ny)CkO_t>H0$of0h{f#Gk<60Hldihzl&*dx@;7PAZ>d*?hK0r!ysl(s z`D(GTBK-~oTNr_^6ZCnnyk6rbff}DkPUrf|Jm(j<0+91Qjln&Zh zm%7+EFmq336mj=J{_g+snh%5LCa+U!G z9!+cjd&v4As(3aaP@g5dIiwvpM}&U)30!7~CA-2FcijbX4e$yff{?2HfyBy}E9-94 z&Hgr8d3$y#{gJx>zMNQBwNE7{x(p(o6)WRLnZl<_ifbT#fD9KC2BHDHHRNUbnj>!Q zktxn)L*83xEBiZrgBt~TRy0Fp8k&pN-e0|n5P-{GEeLOd#$7uov9J1Wo7%;4ngvy5 zJ^A-S1>3^+>tN?)GG0WhM?83)o9QqLqItdqH+&2%P7mW!kAv_KaQzD*bm$kseZ7^j z>L+9M770h~PBO34XI|K#jDy-3-?0O{{Qm8Mo}O=0nDZ9XM|IDi;x>dT@~aF_0I2I& zbM)ZS4FD(R^5~q%b(v2vL(EWK!4Dpg%{9`H*5)*N%4+*shl1Q8#CrN!dljlDo6=}j zctH0({Icu2I=tSVRJyTmLqmwUUK>j~U+!ec{UN~gDI}2?ko^Hbn+Vmz5@{0c&c$AO zE+_Niv8K*`;SvBiP4&4>OXz~Va$CS{QQ19 z`0Bnno*6R8Z2IZ<1S8{}uhm2i43wN%OyuXzwD|L-e$v#aSL*5SHwK~h7pL>QO(@Ue z&qMaI783hm_o^~`vbu|AgY2Wt(|gNoeyExeCcM@^+q>^np}d!A7N*w`W9iNXf?jVv zE;pTz>6Pp4i5FB0{SNp6X`n#!x1=F_jv)8|Q9?5>4h^A>RiT@2DYn7ft$(67*DS{K ztF;rT4GGr=nFLW=6$Y%vMK{LXS9z!RUg*37v`3Tav#95u{cyXkgiLjG-4wdjGX?ui zo^)Jyn7$XxqU;J0u`b^iuinGBafhR>?xe%Q&7RD8cy}nNE6I$_Em1z+X#GLi-Xom+ zQC6)#d2DHDf|xnY^yY)1dH~T4*E)Sr-_I1C?|`yf1=T$uN~xj^LSEEzve-0)St`rZ zm%>2RXI?gK@Hs!*tDidkLa&%=1=-`iNMkNav$D38v*3<=PD5XZM#+xwG{j9GM$h z;HV-Ix+{vx!xSZsH+nNu4|iV0^JUn=zUXPLId6TOXa1TKXtHp5q=A)!NAqT#U5{zr z_b5b1U#+kp3&lSa#Ly8J)IXi)i=$2y=8>QDqos+3=y)pJnfCie%;101iM$llrXazj zs)q~aWE=LF@F3>;_gV@hs_)O znQ(V8Z+7X2Vy(fhu9QT^ZO}ylO>w2-|)TrKgZE&}V?3Dy2Cj?;- zjpB#0EkJlSw)`{TG_!J;??jm#a?l-(YnuB?(UevDMYkH0xyPxnMuF9;>9m>rjBi(z zM>4$*zHn=5Y`IP>`6A~l@*1_jmLP2Kp`wA{eu0{5iMhoCl@|(T+Kyu>I$r&7IXx;? z%IV4no3C{fj+HxJBA%%ewd0%PWqPuE7REh!zkRG?)NeNM+Pz{idG}@G8mmIrxKzpvG25lLUxG2BbN1qHV_`z~XC(7o`00GRIuSzAc{o3??z><^q7RArF5 zzuelOx1@Xl?YPCoVTNbrdRH{)*)Ite+>g$x(r(NpOO19E#C!^%g zSR4`tiQ0X#N9X9QcykPG6(_l?IKQt;--er*DQpJ5$_$^HOdBys&RPS`A-XkREHlEz zk#e&=UajBxV&HMfA^{z0{PYX1OpdO_eJ06=OKepzYP{s#br4%P!v z#9jE`lLx-9mIpMSuY|?y9{><1(BOw6Ag}({^OaKICS(O)gV5iPYYMS->_Xwt<0JG% z&IiM8+`V<2bP47Ho_jQ2{z16M$6`_>$txgP*wa*flBD~HFs1KezW{L;{UVlve5RPc z;bkC}LjR5n;M*ww*yND*_tF4e>aNVp{rhDiIDoO<%!WMh+cE`f4YmVbyVv1$KGWYL zq`eP3syAq2PXDY%K_H7AJr;}LKdKB_1Hi#seWvusD1p>I90*u&C^+#SwQmM$|Nr(? b4}7%BKOiZi7=&H{|70W-?iGs}J^TLvsqpK+ literal 0 HcmV?d00001 diff --git a/docs/sqlflow/doc/mysql-setup.html b/docs/sqlflow/doc/mysql-setup.html new file mode 100644 index 0000000..8a50e36 --- /dev/null +++ b/docs/sqlflow/doc/mysql-setup.html @@ -0,0 +1,433 @@ + + + + + + + + + Run MySQL Server and Client in Docker Containers - Sqlflow + + + + + + + + + + + + + + + + + +

    + + + diff --git a/docs/sqlflow/doc/mysql-setup.md b/docs/sqlflow/doc/mysql-setup.md new file mode 100644 index 0000000..49c3d09 --- /dev/null +++ b/docs/sqlflow/doc/mysql-setup.md @@ -0,0 +1,138 @@ +# Run MySQL Server and Client in Docker Containers + +The document explains how to setup MySQL in our development environment. + +## Run MySQL Server in a Docker Container + +```bash +docker run --rm \ + -v /tmp/test1:/var/lib/mysql \ + --name mysql01 \ + -e MYSQL_ROOT_PASSWORD=root \ + -e MYSQL_ROOT_HOST='%' \ + -p 3306:3306 \ + -d \ + mysql/mysql-server:8.0 +``` + +- the `-v` option ensures that the database is saved on the host. The default directory where MySQL saves the database is `/var/lib/mysql`. This directory can be configured in `/etc/mysql/my.cnf`, as explained in [this post](https://www.mkyong.com/mysql/where-does-mysql-stored-the-data-in-my-harddisk/). By overlaying the directory `/tmp/test1` on the host to `/var/lib/mysql`, we "cheat" MySQL to save databases on the host. So, we can kill the container and restart it, and the database is still there. + + Please be aware that the directory on the host must be empty the first time we run the above command; otherwise, MySQL would fail to initialize. I figured out this problem after several failures using `docker logs`. + +- the `-e` option sets the root password of MySQL to "root". Feel free to set it to any password you like. + +- the second `-e` options sets `MYSQL_ROOT_HOST` to a [wildcard](https://github.com/docker-library/mysql/issues/241#issuecomment-263011059) so to allow clients connecting to the server via TCP/IP as the user "root". This trick works with MySQL 5.7 and 8.0, but not the most recent under-development version. + +- the `--name` option names the container to `mysql01`, which can be used to refer to this container. + +- the `-p` option maps the port 3306, on which the MySQL server listens, to the same port on the host, so that clients could connect to the server via TCP/IP. + +## Run MySQL Client in the Server Container + +```bash +docker exec -it mysql01 mysql -uroot -p +``` + +This command executes the command `mysql`, which is the command line tool of MySQL, in the container named `mysql01`. + +- The command line flags of `mysql` include `-u`, which specifies the username of MySQL, and `-p`, which makes MySQL prompts for the password. For this example, we should type the password "root", which was set in the previous command. + +- Please wait for a few seconds after the starting of the MySQL server container before we execute the client; otherwise, the startup of the client might fail. + +- Once we get into the MySQL client, we can type SQL commands, e.g., + + ```sql + show databases; + create database yi; + ``` + +## Run Client in a Different Container on the Same Host + +```bash +docker run --rm -it \ + -v /tmp/test1:/var/lib/mysql \ + mysql/mysql-server:8.0 \ + mysql -uroot -p +``` + +- The `-v` option maps the database directory on the host to the client container. This mapping is necessary because, by default, the client talks to the server via Unix socket `/var/lib/mysql/mysql.sock`, which is `/tmp/test1/mysql.sock` on the host. + +## Run Client in a Container on a Remote Host + +```bash +docker run \ + --rm -it \ + mysql/mysql-server:8.0 \ + mysql -h 192.168.1.3 -P 3306 -uroot -p +``` + +- the `-h` option tells the client where the server is running on. In this example, the given IP is the one of the host where I ran the MySQL server container. + +Please be aware that the above command works only if the server allows remote connections. + +## Run Python Client in a Container + +To write a Python client, we need to install the Python package `mysql-connector-python`. + +```Dockerfile +FROM python:2.7 +RUN pip install mysql-connector-python +``` + +Please be aware that some [documents](https://www.w3schools.com/python/python_mysql_getstarted.asp) says that we need to install `mysql-connector`. I tried; but the `mysql.connector.connect` call failed with the error `mysql.connector.errors.NotSupportedError: Authentication plugin 'caching_sha2_password' is not supported`. + +Build the Docker image: + +```bash +docker build -t sqlflow . +``` + +Run the image: + +```bash +docker run --rm -it sqlflow bash +``` + +and we can start Python and run the following Python code snippet + +``` +>>> import mysql.connector +>>> db = mysql.connector.connect(user='root', passwd='root', host='192.168.1.3') +>>> print(db) + +``` + +## Run a Go Client + +In order to connect to a database, you need to import the database's driver first. + +``` +export GOPATH=$HOME/go +go get -u github.com/go-sql-driver/mysql +``` + +`go run` the following file + +```go +package main + +import ( + "database/sql" + "github.com/go-sql-driver/mysql" + "log" +) + +func main() { + testConfig := &mysql.Config{ + User: "root", + Passwd: "root", + Net: "tcp", + Addr: "localhost:3306", + } + db, e := sql.Open("mysql", testConfig.FormatDSN()) + if e != nil { + log.Fatal(e) + } + defer db.Close() +} +``` diff --git a/docs/sqlflow/doc/pipe.html b/docs/sqlflow/doc/pipe.html new file mode 100644 index 0000000..167b719 --- /dev/null +++ b/docs/sqlflow/doc/pipe.html @@ -0,0 +1,374 @@ + + + + + + + + + Piping Responses - Sqlflow + + + + + + + + + + + + + + + + + +
    +
    + Sqlflow + +
    + + +
    +
    +

    This site uses Just the Docs, a documentation theme for Jekyll.

    +
    +
    +
    +
    +
    + +
    +
    + + Search +
    +
    +
    + + + + +
    +
    +
    + + + +
    +

    Piping Responses

    + +

    Streaming Responses

    + +

    As described in the overall design, a SQLFlow job could be a standard or an extended SQL statemnt, where an extended SQL statement will be translated into a Python program. Therefore, each job might generate up to the following data streams:

    + +
      +
    1. standard output, where each element is a line of text,
      2. +
    2. standard error, where each element is a line of text,
      3. +
    3. data rows, where the first element consists of fields name/types, and each of the rest is a row of data,
      4. +
    4. status, where the element could be pending, failed, and succeeded.
      5. +
    + +

    To create good user experience, we need to pipe these responses from SQLFlow jobs to Jupyter Notebook in real-time.

    + +

    Stages in the Pipe

    + +

    The pipe that streams outputs from SQLFlow jobs to the Jupyter Notebook consists of the following stages:

    + +
    Web browser 
+ ↑
+ | HTTP
+ ↓
+Jupyter Notebook server
+ ↑
+ | ZeroMQ streams: Shell, IOPub, stdin, Controls, Heartbeat
+ ↓
+iPython kernel
+ ↑
+ | IPython magic command framework
+ ↓
+SQLFlow magic command for Jupyter
+ ↑
+ | gRPC
+ ↓
+SQLFlow server
+ ↑
+ | Go channels
+ ↓
+SQLFlow job manager (Go functions)
+ ↑
+ | IPC with Go's standard library
+ ↓ 
+SQLFlow jobs
+
    + +

    In the above figure, from the SQLFlow magic command to the bottom layer are our work.

    + +

    Streaming

    + +

    We have two alternative ideas: multiple streams and a multiplexing stream. +We decided to use a multiplexing stream because we had a unsuccessful trial with the multiple streams idea: we make the job writes to various Go channels and forward each Go channel to a streaming gRPC call, as the following:

    + +

    Multiple streams

    + +

    The above figure shows that there are multiple streams between the Jupyter Notebook server and Jupyter kernels. According to the document, there are five: Shell, IOPub, stdin, Control, and Heartbeat. These streams are ZeroMQ streams. We don’t use ZeroMQ, but we can take the idea of having multiple parallel streams in the pipe.

    + +
    service SQLFlow {
+    rpc File(string sql) returns (int id) {}
+
+    rpc ReadStdout(int id) returns (stream string) {}
+    rpc ReadStderr(int id) returns (stream string) {}
+    rpc ReadData(int id) returns (stream Row) {}
+    rpc ReadStatus(int id) returns (stream int) {}
+}
+
    + +

    However, we realized that if the user doesn’t call any one of the SQLFlow.Read... call, there would be no forwarding from the Go channel to Jupyter, thus the job would block forever at writing.

    + +

    A Multiplexing Stream

    + +

    Another idea is multiplexing all streams into one. For example, we can have only one ZeroMQ stream, where each element is a polymorphic type – could be a text string or a data row.

    + +
    service SQLFlow {
+    rpc Run(string sql) returns (stream Response) {}
+}
+
+// Only one of the following fields should be set.
+message Response {
+    oneof record {
+        repeated string head = 1;             // Column names.
+        repeated google.protobuf.Any row = 2; // Cells in a row.
+        string log = 3;                       // A line from stderr or stdout.
+    }
+}
+
    + + + + +
    +
    +
    +
    + + + diff --git a/docs/sqlflow/doc/pipe.md b/docs/sqlflow/doc/pipe.md new file mode 100644 index 0000000..b54e497 --- /dev/null +++ b/docs/sqlflow/doc/pipe.md @@ -0,0 +1,91 @@ +# Piping Responses + + +## Streaming Responses + +As described in the [overall design](doc/design.md), a SQLFlow job could be a standard or an extended SQL statemnt, where an extended SQL statement will be translated into a Python program. Therefore, each job might generate up to the following data streams: + +1. standard output, where each element is a line of text, +1. standard error, where each element is a line of text, +1. data rows, where the first element consists of fields name/types, and each of the rest is a row of data, +1. status, where the element could be *pending*, *failed*, and *succeeded*. + +To create good user experience, we need to pipe these responses from SQLFlow jobs to Jupyter Notebook in real-time. + + +## Stages in the Pipe + +The pipe that streams outputs from SQLFlow jobs to the Jupyter Notebook consists of the following stages: + +``` +Web browser + ↑ + | HTTP + ↓ +Jupyter Notebook server + ↑ + | ZeroMQ streams: Shell, IOPub, stdin, Controls, Heartbeat + ↓ +iPython kernel + ↑ + | IPython magic command framework + ↓ +SQLFlow magic command for Jupyter + ↑ + | gRPC + ↓ +SQLFlow server + ↑ + | Go channels + ↓ +SQLFlow job manager (Go functions) + ↑ + | IPC with Go's standard library + ↓ +SQLFlow jobs +``` + +In the above figure, from the SQLFlow magic command to the bottom layer are our work. + + +## Streaming + +We have two alternative ideas: multiple streams and a multiplexing stream. +We decided to use a multiplexing stream because we had a unsuccessful trial with the multiple streams idea: we make the job writes to various Go channels and forward each Go channel to a streaming gRPC call, as the following: + +### Multiple streams + +The above figure shows that there are multiple streams between the Jupyter Notebook server and Jupyter kernels. According to the [document](https://jupyter-client.readthedocs.io/en/stable/messaging.html), there are five: Shell, IOPub, stdin, Control, and Heartbeat. These streams are ZeroMQ streams. We don't use ZeroMQ, but we can take the idea of having multiple parallel streams in the pipe. + + +```protobuf +service SQLFlow { + rpc File(string sql) returns (int id) {} + + rpc ReadStdout(int id) returns (stream string) {} + rpc ReadStderr(int id) returns (stream string) {} + rpc ReadData(int id) returns (stream Row) {} + rpc ReadStatus(int id) returns (stream int) {} +} +``` + +However, we realized that if the user doesn't call any one of the `SQLFlow.Read...` call, there would be no forwarding from the Go channel to Jupyter, thus the job would block forever at writing. + +## A Multiplexing Stream + +Another idea is multiplexing all streams into one. For example, we can have only one ZeroMQ stream, where each element is a polymorphic type -- could be a text string or a data row. + +```protobuf +service SQLFlow { + rpc Run(string sql) returns (stream Response) {} +} + +// Only one of the following fields should be set. +message Response { + oneof record { + repeated string head = 1; // Column names. + repeated google.protobuf.Any row = 2; // Cells in a row. + string log = 3; // A line from stderr or stdout. + } +} +``` diff --git a/docs/sqlflow/doc/quickstart.html b/docs/sqlflow/doc/quickstart.html new file mode 100644 index 0000000..b57d8d8 --- /dev/null +++ b/docs/sqlflow/doc/quickstart.html @@ -0,0 +1,387 @@ + + + + + + + + + Quick start - Sqlflow + + + + + + + + + + + + + + + + + +
    +
    + Sqlflow + +
    + + +
    +
    +

    This site uses Just the Docs, a documentation theme for Jekyll.

    +
    +
    +
    +
    +
    + +
    +
    + + Search +
    +
    +
    + + + + +
    +
    +
    + + + +
    +

    Quick start

    + +

    SQLFlow is currently under active development. For those who are interested in trying +it out, we have provided several demos. Play around with it. Any bug report and +issue are welcomed. :)

    + +

    Setup

    + +
      +
    1. Install Docker.
      2. +
    2. Set up a MySQL server following example/datasets/README.md.
      3. +
    3. Pull the latest SQLFlow Docker image: docker pull sqlflow/sqlflow:latest.
      4. +
    + +

    Demo 1: Jupyter Notebook

    + +
      +
    1. +

      Start a Docker container that runs sqlflowserver and Jupyter Notebook. If you are +using Docker for Linux, please change host.docker.internal:3306 to localhost:3306.

      + + 
      docker run --rm -it -p 8888:8888 sqlflow/sqlflow:latest \
+bash -c "sqlflowserver --db_user root --db_password root --db_address host.docker.internal:3306 &
+SQLFLOW_SERVER=localhost:50051 jupyter notebook --ip=0.0.0.0 --port=8888 --allow-root"
+
      + +

      If you are using Docker for Mac, please be aware the option --db_address host.docker.internal:3306 where +host.docker.internal translates to the host ip address as recommended here.

      + +

      If you are running MySQL on remote, please be aware that MySQL only allows connections from localhost +by default. Fix can be found here.

      +
      2. +
    2. +

      Open a Web browser and direct to localhost:8888 and input the token. Then you +can create notebooks. In a cell, you should be able to type

      + + 
      %%sqlflow
+select 1
+
      +
      3. +
    3. +

      Explore more examples at example.ipynb

      +
      4. +
    + +

    Demo 2: Command Line Prompt

    + +

    Start a Docker container that runs SQLFlow command line prompt. If you are using +Docker for Linux, please change host.docker.internal:3306 to localhost:3306.

    + +
    docker run -it --rm --net=host sqlflow/sqlflow:latest demo \
+--db_user root --db_password root --db_address host.docker.internal:3306
+
    + +

    You should be able to see the following prompt.

    + +
    sqlflow>
+
    + +

    Example

    + +
      +
    • Select data + 
      sqlflow> select * from iris.train limit 2;
+-----------------------------
++--------------+-------------+--------------+-------------+-------+
+| SEPAL LENGTH | SEPAL WIDTH | PETAL LENGTH | PETAL WIDTH | CLASS |
++--------------+-------------+--------------+-------------+-------+
+|          6.4 |         2.8 |          5.6 |         2.2 |     2 |
+|            5 |         2.3 |          3.3 |           1 |     1 |
++--------------+-------------+--------------+-------------+-------+
+
      +
      • +
    • Train a Tensorflow DNNClassifier + 
      sqlflow> SELECT *
+FROM iris.train
+TRAIN DNNClassifier
+WITH n_classes = 3, hidden_units = [10, 20]
+COLUMN sepal_length, sepal_width, petal_length, petal_width
+LABEL class
+INTO sqlflow_models.my_dnn_model;
+-----------------------------
+...
+Training set accuracy: 0.96721
+Done training
+
      +
      • +
    • Prediction using a trained model + 
      sqlflow> SELECT *
+FROM iris.test
+predict iris.predict.class
+USING sqlflow_models.my_dnn_model;
+
      +
      • +
    • Checkout prediction result + 
      sqlflow> select * from iris.predict limit 10;
+
      +
      • +
    + + + + +
    +
    +
    +
    + + + diff --git a/docs/sqlflow/doc/quickstart.md b/docs/sqlflow/doc/quickstart.md new file mode 100644 index 0000000..75a00a6 --- /dev/null +++ b/docs/sqlflow/doc/quickstart.md @@ -0,0 +1,93 @@ +# Quick start + +SQLFlow is currently under active development. For those who are interested in trying +it out, we have provided several demos. Play around with it. Any bug report and +issue are welcomed. :) + +## Setup + +1. Install [Docker](https://docs.docker.com/install/). +1. Set up a MySQL server following [example/datasets/README.md](/example/datasets/README.md). +1. Pull the latest SQLFlow Docker image: `docker pull sqlflow/sqlflow:latest`. + +## Demo 1: Jupyter Notebook + +1. Start a Docker container that runs sqlflowserver and Jupyter Notebook. If you are + using Docker for Linux, please change `host.docker.internal:3306` to `localhost:3306`. + + ``` + docker run --rm -it -p 8888:8888 sqlflow/sqlflow:latest \ + bash -c "sqlflowserver --db_user root --db_password root --db_address host.docker.internal:3306 & + SQLFLOW_SERVER=localhost:50051 jupyter notebook --ip=0.0.0.0 --port=8888 --allow-root" + ``` + + If you are using Docker for Mac, please be aware the option `--db_address host.docker.internal:3306` where + `host.docker.internal` translates to the host ip address as recommended [here](https://docs.docker.com/docker-for-mac/networking/). + + If you are running MySQL on remote, please be aware that MySQL only allows connections from localhost + by default. Fix can be found [here](https://stackoverflow.com/questions/14779104/how-to-allow-remote-connection-to-mysql). + +1. Open a Web browser and direct to `localhost:8888` and input the token. Then you +can create notebooks. In a cell, you should be able to type + + ``` + %%sqlflow + select 1 + ``` + +1. Explore more examples at [example.ipynb](/example/jupyter/example.ipynb) + +## Demo 2: Command Line Prompt + +Start a Docker container that runs SQLFlow command line prompt. If you are using +Docker for Linux, please change `host.docker.internal:3306` to `localhost:3306`. + +``` +docker run -it --rm --net=host sqlflow/sqlflow:latest demo \ +--db_user root --db_password root --db_address host.docker.internal:3306 +``` + +You should be able to see the following prompt. + +``` +sqlflow> +``` + +### Example + +- Select data +```sql +sqlflow> select * from iris.train limit 2; +----------------------------- ++--------------+-------------+--------------+-------------+-------+ +| SEPAL LENGTH | SEPAL WIDTH | PETAL LENGTH | PETAL WIDTH | CLASS | ++--------------+-------------+--------------+-------------+-------+ +| 6.4 | 2.8 | 5.6 | 2.2 | 2 | +| 5 | 2.3 | 3.3 | 1 | 1 | ++--------------+-------------+--------------+-------------+-------+ +``` +- Train a Tensorflow [DNNClassifier](https://www.tensorflow.org/api_docs/python/tf/estimator/DNNClassifier) +```sql +sqlflow> SELECT * +FROM iris.train +TRAIN DNNClassifier +WITH n_classes = 3, hidden_units = [10, 20] +COLUMN sepal_length, sepal_width, petal_length, petal_width +LABEL class +INTO sqlflow_models.my_dnn_model; +----------------------------- +... +Training set accuracy: 0.96721 +Done training +``` +- Prediction using a trained model +```sql +sqlflow> SELECT * +FROM iris.test +predict iris.predict.class +USING sqlflow_models.my_dnn_model; +``` +- Checkout prediction result +```sql +sqlflow> select * from iris.predict limit 10; +``` \ No newline at end of file diff --git a/docs/sqlflow/doc/sql_parser.html b/docs/sqlflow/doc/sql_parser.html new file mode 100644 index 0000000..61bd481 --- /dev/null +++ b/docs/sqlflow/doc/sql_parser.html @@ -0,0 +1,355 @@ + + + + + + + + + Extended SQL Parser Design - Sqlflow + + + + + + + + + + + + + + + + + +
    +
    + Sqlflow + +
    + + +
    +
    +

    This site uses Just the Docs, a documentation theme for Jekyll.

    +
    +
    +
    +
    +
    + +
    +
    + + Search +
    +
    +
    + + + + +
    +
    +
    + + + +
    +

    Extended SQL Parser Design

    + +

    This documentation explains the technical decision made in building a SQL +parser in Go. It is used to parsed the extended SELECT syntax of SQL that +integrates TensorFlow Estimators.

    + + + +

    Lexer and Parser Generator

    + +

    In 2001, when I was in graduate school, I defined an extended SQL syntax for +querying distributed relational databases, as part of the course project of +Distributed RDBMS by Prof. Li-zhu Zhou. I wrote the parser using bison (a +modern version of yacc) and flex (a modern version of +lex). yacc and lex generate C code; +bison and flex generate C++ code. However, this time, I’d use Go.

    + +

    I surveyed goyacc, a +standard Go tool. The usage is very similar to that of yacc and bison. +However, the Go toolchain doesn’t provide a tool like lex/flex.

    + +

    Google revealed golex, which is out of +maintenance.

    + +

    The Medium +post +recommends Ragel, which is a C++ +program and could generate Go lexer; however, it lacks documents.

    + +

    Handwritten Lexer and Parser

    + +

    Some documents, including this +one +recommends handwriting lexers. However, it doesn’t explain how to write the +parser.

    + +

    GoAcademy always provides high-quality tech blog posts. This +one is from the +author of InfluxDB. However, I +stopped at where it explains wrapping a SQL statement as a string by an +io.Reader, because it is obvious that we should keep the string as a string so +that that token strings could refer to the same memory storage of the SQL +statement.

    + +

    Following a link in the above GoAcademy post, I found Rob Pike’s excellent talk +on how to write a lexer in Go in 2011. Many works after that change Rob’s +implementation somehow but always lead to longer and less comprehensible +codebases.

    + +

    The Choice

    + +

    Therefore, I wrote the lexer and parser both following Rob Pike’s idea. After +few days work, I realized that:

    + +
      +
    1. I should borrow the idea from Rob to represent SQL statements as strings, but +not io.Reader as other work do,
      2. +
    2. but no need to use channels and goroutines at all, and
      3. +
    3. it is technically intractable to write a SQL lexer/parser manually.
      4. +
    + +

    So, I switched to write a lexer manually, and to generate the parser using +goyacc. During my work, I referred to this +example and +the official yacc +manual for details +about operator association and precedence.

    + + + + +
    +
    +
    +
    + + + diff --git a/docs/sqlflow/doc/sql_parser.md b/docs/sqlflow/doc/sql_parser.md new file mode 100644 index 0000000..d45f7ee --- /dev/null +++ b/docs/sqlflow/doc/sql_parser.md @@ -0,0 +1,65 @@ +# Extended SQL Parser Design + +This documentation explains the technical decision made in building a SQL +parser in Go. It is used to parsed the extended SELECT syntax of SQL that +integrates TensorFlow Estimators. + +## Related Work + +### Lexer and Parser Generator + +In 2001, when I was in graduate school, I defined an extended SQL syntax for +querying distributed relational databases, as part of the course project of +Distributed RDBMS by Prof. Li-zhu Zhou. I wrote the parser using [`bison` (a +modern version of `yacc`) and `flex` (a modern version of +`lex`)](http://dinosaur.compilertools.net/). `yacc` and `lex` generate C code; +`bison` and `flex` generate C++ code. However, this time, I'd use Go. + +I surveyed [`goyacc`](https://godoc.org/golang.org/x/tools/cmd/goyacc), a +standard Go tool. The usage is very similar to that of `yacc` and `bison`. +However, the Go toolchain doesn't provide a tool like `lex`/`flex`. + +Google revealed [`golex`](https://github.com/cznic/golex), which is out of +maintenance. + +The [Medium +post](https://medium.com/@mhamrah/lexing-with-ragel-and-parsing-with-yacc-using-go-81e50475f88f) +recommends [Ragel](http://www.colm.net/open-source/ragel/), which is a C++ +program and could generate Go lexer; however, it lacks documents. + +### Handwritten Lexer and Parser + +Some documents, including [this +one](https://hackthology.com/writing-a-lexer-in-go-with-lexmachine.html) +recommends handwriting lexers. However, it doesn't explain how to write the +parser. + +GoAcademy always provides high-quality tech blog posts. [This +one](https://blog.gopheracademy.com/advent-2014/parsers-lexers/) is from the +author of [InfluxDB](https://github.com/influxdata/influxdb). However, I +stopped at where it explains wrapping a SQL statement as a string by an +`io.Reader`, because it is obvious that we should keep the string as a string so +that that token strings could refer to the same memory storage of the SQL +statement. + +Following a link in the above GoAcademy post, I found Rob Pike's excellent talk +on how to write a lexer in Go in 2011. Many works after that change Rob's +implementation somehow but always lead to longer and less comprehensible +codebases. + +## The Choice + +Therefore, I wrote the lexer and parser both following Rob Pike's idea. After +few days work, I realized that: + +1. I should borrow the idea from Rob to represent SQL statements as strings, but + not `io.Reader` as other work do, +1. but no need to use channels and goroutines at all, and +1. it is technically intractable to write a SQL lexer/parser manually. + +So, I switched to write a lexer manually, and to generate the parser using +goyacc. During my work, I referred to [this +example](https://github.com/golang-samples/yacc/blob/master/simple/calc.y) and +the official [`yacc` +manual](https://www.epaperpress.com/lexandyacc/download/yacc.pdf) for details +about operator association and precedence. diff --git a/docs/sqlflow/doc/submitter.html b/docs/sqlflow/doc/submitter.html new file mode 100644 index 0000000..94549d4 --- /dev/null +++ b/docs/sqlflow/doc/submitter.html @@ -0,0 +1,377 @@ + + + + + + + + + Submitter - Sqlflow + + + + + + + + + + + + + + + + + +
    +
    + Sqlflow + +
    + + +
    +
    +

    This site uses Just the Docs, a documentation theme for Jekyll.

    +
    +
    +
    +
    +
    + +
    +
    + + Search +
    +
    +
    + + + + +
    +
    +
    + + + +
    +

    Submitter

    + +

    A submitter is a pluggable module in SQLFlow that is used to submit an ML job to a third party computation service.

    + +

    Workflow

    + +

    When a user types in an extended SQL statement, SQLFlow first parses and semantically verifies the statement. Then SQLFlow either runs the ML job locally or submits the ML job to a third party computation service.

    + +

    + +

    In the latter case, SQLFlow produces a job description (TrainDescription or PredictDescription) and hands it over to the submitter. For a training SQL, SQLFlow produces TrainDescription; for prediction SQL, SQLFlow produces PredDescription. The concrete definition of the description looks like the following

    + +
    type ColumnType struct {
+    Name             string // e.g. sepal_length
+    DatabaseTypeName string // e.g. FLOAT
+}
+
+// SELECT *
+// FROM iris.train
+// TRAIN DNNClassifier
+// WITH
+//   n_classes = 3,
+//   hidden_units = [10, 20]
+// COLUMN sepal_length, sepal_width, petal_length, petal_width
+// LABEL class
+// INTO sqlflow_models.my_dnn_model;
+type TrainDescription struct {
+    StandardSelect string       // e.g. SELECT * FROM iris.train
+    Estimator      string       // e.g. DNNClassifier
+    Attrs          map[string]string // e.g. {{"n_classes", "3"}, {"hidden_units", "[10, 20]"}}
+    X              []ColumnType // e.g. {{"sepal_length", "FLOAT"}, ...}
+    Y              ColumnType   // e.g. {"class", "INT"}
+    ModelName      string       // e.g. my_dnn_model
+}
+
+// SELECT *
+// FROM iris.test
+// PREDICT iris.predict.class
+// USING sqlflow_models.my_dnn_model;
+type PredDescription struct {
+    StandardSelect string // e.g. SELECT * FROM iris.test
+    TableName      string // e.g. iris.predict
+    ModelName      string // e.g. my_dnn_model
+}
+
    + +

    Submitter Interface

    + +

    The submitter interface should provide two functions Train and Predict. The detailed definition can be the following

    + +
    type Submitter interface {
+    // Train executes a ML training job and streams job's response through writer.
+    // A typical Train function should include
+    // - Loading the training data
+    // - Initializing the model
+    // - model.train
+    // - Saving the trained model to a persistent storage
+    Train(desc TrainDescription, writer PipeWriter) error
+    // Predict executes a ML predicting job and streams job's response through writer
+    // A typical Predict function should include
+    // - Loading the model from a persistent storage
+    // - Loading the prediction data
+    // - model.predict
+    // - Writing the prediction result to a table
+    Predict(desc PredictDescription, writer PipeWriter) error
+}
+
    + +

    Register a submitter

    + +

    A new submitter can be added as

    + +
    import (
+    ".../my_submitter"
+    ".../sqlflow/sql"
+)
+
+func main() {
+    // ...
+    sql.Register(my_submitter.NewSubmitter())
+    // ...
+    for {
+    	sql := recv()
+    	sql.Run(sql)
+    }
+}
+
    + +

    where sql.Register will put my_submitter instance to package level registry. During sql.Run, it will check whether there is a submitter registered. If there is, sql.Run will run either submitter.Train or submitter.Predict.

    + + + + +
    +
    +
    +
    + + + diff --git a/docs/sqlflow/doc/submitter.md b/docs/sqlflow/doc/submitter.md new file mode 100644 index 0000000..e2486c9 --- /dev/null +++ b/docs/sqlflow/doc/submitter.md @@ -0,0 +1,95 @@ +# Submitter + +A submitter is a pluggable module in SQLFlow that is used to submit an ML job to a third party computation service. + +## Workflow + +When a user types in an extended SQL statement, SQLFlow first parses and semantically verifies the statement. Then SQLFlow either runs the ML job locally or submits the ML job to a third party computation service. + +![](doc/figures/sqlflow-arch2.png) + +In the latter case, SQLFlow produces a job description (`TrainDescription` or `PredictDescription`) and hands it over to the submitter. For a training SQL, SQLFlow produces `TrainDescription`; for prediction SQL, SQLFlow produces `PredDescription`. The concrete definition of the description looks like the following + + +{% raw %} +```go +type ColumnType struct { + Name string // e.g. sepal_length + DatabaseTypeName string // e.g. FLOAT +} + +// SELECT * +// FROM iris.train +// TRAIN DNNClassifier +// WITH +// n_classes = 3, +// hidden_units = [10, 20] +// COLUMN sepal_length, sepal_width, petal_length, petal_width +// LABEL class +// INTO sqlflow_models.my_dnn_model; +type TrainDescription struct { + StandardSelect string // e.g. SELECT * FROM iris.train + Estimator string // e.g. DNNClassifier + Attrs map[string]string // e.g. {{"n_classes", "3"}, {"hidden_units", "[10, 20]"}} + X []ColumnType // e.g. {{"sepal_length", "FLOAT"}, ...} + Y ColumnType // e.g. {"class", "INT"} + ModelName string // e.g. my_dnn_model +} + +// SELECT * +// FROM iris.test +// PREDICT iris.predict.class +// USING sqlflow_models.my_dnn_model; +type PredDescription struct { + StandardSelect string // e.g. SELECT * FROM iris.test + TableName string // e.g. iris.predict + ModelName string // e.g. my_dnn_model +} +``` +{% endraw %} + +## Submitter Interface + +The submitter interface should provide two functions `Train` and `Predict`. The detailed definition can be the following + +```go +type Submitter interface { + // Train executes a ML training job and streams job's response through writer. + // A typical Train function should include + // - Loading the training data + // - Initializing the model + // - model.train + // - Saving the trained model to a persistent storage + Train(desc TrainDescription, writer PipeWriter) error + // Predict executes a ML predicting job and streams job's response through writer + // A typical Predict function should include + // - Loading the model from a persistent storage + // - Loading the prediction data + // - model.predict + // - Writing the prediction result to a table + Predict(desc PredictDescription, writer PipeWriter) error +} +``` + +## Register a submitter + +A new submitter can be added as + +```go +import ( + ".../my_submitter" + ".../sqlflow/sql" +) + +func main() { + // ... + sql.Register(my_submitter.NewSubmitter()) + // ... + for { + sql := recv() + sql.Run(sql) + } +} +``` + +where `sql.Register` will put `my_submitter` instance to package level registry. During `sql.Run`, it will check whether there is a submitter registered. If there is, `sql.Run` will run either `submitter.Train` or `submitter.Predict`. \ No newline at end of file diff --git a/docs/sqlflow/doc/syntax.html b/docs/sqlflow/doc/syntax.html new file mode 100644 index 0000000..49a1e59 --- /dev/null +++ b/docs/sqlflow/doc/syntax.html @@ -0,0 +1,448 @@ + + + + + + + + + SQLFlow: Design Doc - Sqlflow + + + + + + + + + + + + + + + + + +
    +
    + Sqlflow + +
    + + +
    +
    +

    This site uses Just the Docs, a documentation theme for Jekyll.

    +
    +
    +
    +
    +
    + +
    +
    + + Search +
    +
    +
    + + + + +
    +
    +
    + + + +
    +

    SQLFlow: Design Doc

    + +

    What is SQLFlow

    + +

    SQLFlow is a bridge that connects a SQL engine, for example, MySQL, Hive, SparkSQL, Oracle, or SQL Server, and TensorFlow and other machine learning toolkits. SQLFlow extends the SQL syntax to enable model training and inference.

    + + + +

    We could write simple machine learning prediction (or scoring) algorithms in SQL using operators like DOT_PRODUCT. However, this requires copy-n-pasting model parameters from the training program into SQL statements.

    + +

    Some proprietary SQL engines provide extensions to support machine learning.

    + +

    Microsoft SQL Server

    + +

    Microsoft SQL Server has the machine learning service that runs machine learning programs in R or Python as an external script:

    + +
    CREATE PROCEDURE generate_linear_model
+AS
+BEGIN
+    EXEC sp_execute_external_script
+    @language = N'R'
+    , @script = N'lrmodel <- rxLinMod(formula = distance ~ speed, data = CarsData);
+        trained_model <- data.frame(payload = as.raw(serialize(lrmodel, connection=NULL)));'
+    , @input_data_1 = N'SELECT [speed], [distance] FROM CarSpeed'
+    , @input_data_1_name = N'CarsData'
+    , @output_data_1_name = N'trained_model'
+    WITH RESULT SETS ((model varbinary(max)));
+END;
+
    + +

    A challenge to the users is that they need to know not only SQL but also R or Python, and they must be capable of writing machine learning programs in R or Python.

    + +

    Teradata SQL for DL

    + +

    Teradata also provides a RESTful service, which is callable from the extended SQL SELECT syntax.

    + +
    SELECT * FROM deep_learning_scorer(
+  ON (SELECT * FROM cc_data LIMIT 100)
+  URL('http://localhost:8000/api/v1/request')
+  ModelName('cc')
+  ModelVersion('1')
+  RequestType('predict')
+  columns('v1', 'v2', ..., 'amount')
+)
+
    + +

    The above syntax couples the deployment of the service (the URL in the above SQL statement) with the algorithm.

    + +

    Google BigQuery

    + +

    Google BigQuery enables machine learning in SQL by introducing the CREATE MODEL statement.

    + +
    CREATE MODEL dataset.model_name
+  OPTIONS(model_type='linear_reg', input_label_cols=['input_label'])
+AS SELECT * FROM input_table;
+
    + +

    Currently, BigQuery only supports two simple models: linear regression and logistic regression.

    + +

    Design Goal

    + +

    None of the above meets our requirement.

    + +

    First of all, we want to build an open source software. Also, we want it to be extensible:

    + +
      +
    • +

      We want it extensible to many SQL engines, instead of targeting any one of them. Therefore, we don’t want to build our syntax extension on top of user-defined functions (UDFs); otherwise, we’d have to implement them for each SQL engine.

      +
      • +
    • +

      We want the system extensible to support sophisticated machine learning models and toolkits, including TensorFlow for deep learning and xgboost for trees.

      +
      • +
    + +

    Another challenge is that we want SQLFlow to be flexible enough to configure and run cutting-edge algorithms, including specifying feature crosses. At the same time, we want SQLFlow easy to learn – at least, no Python or R code embedded in the SQL statements, and integrate hyperparameter estimation.

    + +

    We understand that a key to address the above challenges is the syntax of the SQL extension. To craft a highly-effective and easy-to-learn syntax, we need user feedback and fast iteration. Therefore, we’d start from a prototype that supports only MySQL and TensorFlow. We plan to support more SQL engines and machine learning toolkits later.

    + +

    Design Decisions

    + +

    As the beginning of the iteration, we propose an extension to the SQL SELECT statement. We are not going a new statement way like that BigQuery provides CREATE MODEL, because we want to maintain a loose couple between our system and the underlying SQL engine, and we cannot create the new data type for the SQL engine, like CREATE MODEL requires.

    + +

    We highly appreciate the work of TensorFlow Estimator, a high-level API for deep learning. The basic idea behind Estimator is to implement each deep learning model, and related training/testing/evaluating algorithms as a Python class derived from tf.estimator.Estimator. As we want to keep our SQL syntax simple, we would make the system extensible by calling estimators contributed by machine learning experts and written in Python.

    + +

    The SQL syntax must allow users to set Estimator attributes (parameters of the Python class’ constructor, and those of train, evaluate, or predict). Users can choose to use default values. We have a plan to integrate our hyperparameter estimation research into the system to optimize the default values.

    + +

    Though estimators derived from tf.estimator.Estimator run algorithms as TensorFlow graphs; SQLFlow doesn’t restrict that the underlying machine learning toolkit has to be TensorFlow. Indeed, as long as an estimator provides methods of train, evaluate, and predict, SQLFlow doesn’t care if it calls TensorFlow or xgboost. Precisely, what SQLFlow expect is an interface like the following:

    + +
    class AnEstimatorClass:
+  __init__(self, **kwargs)
+  train(self, **kwargs)
+  evaluate(self, **kwargs)
+  predict(self, **kwargs)
+
    + +

    We also want to reuse the feature columns API from Estimator, which allows users to columns of tables in a SQL engine to features to the model.

    + +

    Extended SQL Syntax

    + +

    Again, just as the beginning of the iteration, we propose the syntax for training as

    + +
    SELECT * FROM kaggle_credit_fraud_training_data
+LIMIT 1000
+TRAIN DNNClassifier       /* a pre-defined TensorFlow estimator, tf.estimator.DNNClassifier */
+WITH layers=[100, 200],   /* a parameter of the Estimator class constructor */
+     train.batch_size = 8 /* a parameter of the Estimator.train method */
+COLUMN *,                 /* all columns as raw features */
+       cross(v1, v9, v28) /* plus a derived (crossed) column */
+LABEL class
+INTO sqlflow_models.my_model_table;      /* saves trained model parameters and features into a table */
+
    + +

    We see the redundancy of * in two clauses: SELECT and COLUMN. The following alternative can avoid the redundancy, but cannot specify the label.

    + +
    SELECT *                  /* raw features or the label? */
+       corss(v1, v9, v28) /* derived featuers */
+FROM kaggle_credit_fraud_training_data
+
    + +

    Please be aware that we save the trained models into tables, instead of a variable maintained by the underlying SQL engine. To invent a new variable type to hold trained models, we’d make our system tightly integrated with the SQL engine, and harms the extensibility to other engines.

    + +

    The result table should include the following information:

    + +
      +
    1. The estimator name, e.g., DNNClassifier in this case.
      2. +
    2. Estimator attributes, e.g., layer and train.batch_size.
      3. +
    3. The feature mapping, e.g., * and cross(v1, v9, v28).
      4. +
    + +

    Similarly, to infer the class (fraud or regular), we could

    + +
    SELECT * FROM kaggle_credit_fraud_development_data
+PREDICT kaggle_credit_fraud_development_data.class
+USING sqlflow_models.my_model_table;
+
    + +

    System Architecture

    + +

    A Conceptual Overview

    + +

    In the prototype, we use the following architecture:

    + +
    SQL statement -> our SQL parser --standard SQL-> MySQL
+                                \-extended SQL-> code generator -> execution engine
+
    + +

    In the prototype, the code generator generates a Python program that trains or predicts. In either case,

    + +
      +
    1. it retrieves the data from MySQL via MySQL Connector Python API,
      2. +
    2. optionally, retrieves the model from MySQL,
      3. +
    3. trains the model or predicts using the trained model by calling the user specified TensorFlow estimator,
      4. +
    4. and writes the trained model or prediction results into a table.
      5. +
    + +

    Working with Jupyter Notebook and Kubernetes

    + +

    The following figures shows the system components and their runtime environment. The left part shows how to run the system on a PC/laptop, the right part shows how to run it on a Kubernetes cluster.

    + +

    + + + + +
    +
    +
    +
    + + + diff --git a/docs/sqlflow/doc/syntax.md b/docs/sqlflow/doc/syntax.md new file mode 100644 index 0000000..26694f0 --- /dev/null +++ b/docs/sqlflow/doc/syntax.md @@ -0,0 +1,160 @@ +# SQLFlow: Design Doc + +## What is SQLFlow + +SQLFlow is a bridge that connects a SQL engine, for example, MySQL, Hive, SparkSQL, Oracle, or SQL Server, and TensorFlow and other machine learning toolkits. SQLFlow extends the SQL syntax to enable model training and inference. + +## Related Work + +We could write simple machine learning prediction (or scoring) algorithms in SQL using operators like [`DOT_PRODUCT`](https://thenewstack.io/sql-fans-can-now-develop-ml-applications/). However, this requires copy-n-pasting model parameters from the training program into SQL statements. + +Some proprietary SQL engines provide extensions to support machine learning. + +### Microsoft SQL Server + +Microsoft SQL Server has the [machine learning service](https://docs.microsoft.com/en-us/sql/advanced-analytics/tutorials/rtsql-create-a-predictive-model-r?view=sql-server-2017) that runs machine learning programs in R or Python as an external script: + +```sql +CREATE PROCEDURE generate_linear_model +AS +BEGIN + EXEC sp_execute_external_script + @language = N'R' + , @script = N'lrmodel <- rxLinMod(formula = distance ~ speed, data = CarsData); + trained_model <- data.frame(payload = as.raw(serialize(lrmodel, connection=NULL)));' + , @input_data_1 = N'SELECT [speed], [distance] FROM CarSpeed' + , @input_data_1_name = N'CarsData' + , @output_data_1_name = N'trained_model' + WITH RESULT SETS ((model varbinary(max))); +END; +``` + +A challenge to the users is that they need to know not only SQL but also R or Python, and they must be capable of writing machine learning programs in R or Python. + +### Teradata SQL for DL + +Teradata also provides a [RESTful service](https://www.linkedin.com/pulse/sql-deep-learning-sql-dl-omri-shiv), which is callable from the extended SQL SELECT syntax. + +```sql +SELECT * FROM deep_learning_scorer( + ON (SELECT * FROM cc_data LIMIT 100) + URL('http://localhost:8000/api/v1/request') + ModelName('cc') + ModelVersion('1') + RequestType('predict') + columns('v1', 'v2', ..., 'amount') +) +``` + +The above syntax couples the deployment of the service (the URL in the above SQL statement) with the algorithm. + +### Google BigQuery + +Google [BigQuery](https://cloud.google.com/bigquery/docs/bigqueryml-intro) enables machine learning in SQL by introducing the `CREATE MODEL` statement. + +```sql +CREATE MODEL dataset.model_name + OPTIONS(model_type='linear_reg', input_label_cols=['input_label']) +AS SELECT * FROM input_table; +``` + +Currently, BigQuery only supports two simple models: linear regression and logistic regression. + +## Design Goal + +None of the above meets our requirement. + +First of all, we want to build an open source software. Also, we want it to be extensible: + +- We want it extensible to many SQL engines, instead of targeting any one of them. Therefore, we don't want to build our syntax extension on top of user-defined functions (UDFs); otherwise, we'd have to implement them for each SQL engine. + +- We want the system extensible to support sophisticated machine learning models and toolkits, including TensorFlow for deep learning and [xgboost](https://github.com/dmlc/xgboost) for trees. + +Another challenge is that we want SQLFlow to be flexible enough to configure and run cutting-edge algorithms, including specifying [feature crosses](https://www.tensorflow.org/api_docs/python/tf/feature_column/crossed_column). At the same time, we want SQLFlow easy to learn -- at least, no Python or R code embedded in the SQL statements, and integrate hyperparameter estimation. + +We understand that a key to address the above challenges is the syntax of the SQL extension. To craft a highly-effective and easy-to-learn syntax, we need user feedback and fast iteration. Therefore, we'd start from a prototype that supports only MySQL and TensorFlow. We plan to support more SQL engines and machine learning toolkits later. + +## Design Decisions + +As the beginning of the iteration, we propose an extension to the SQL SELECT statement. We are not going a new statement way like that BigQuery provides `CREATE MODEL`, because we want to maintain a loose couple between our system and the underlying SQL engine, and we cannot create the new data type for the SQL engine, like `CREATE MODEL` requires. + +We highly appreciate the work of [TensorFlow Estimator](https://www.tensorflow.org/guide/estimators), a high-level API for deep learning. The basic idea behind Estimator is to implement each deep learning model, and related training/testing/evaluating algorithms as a Python class derived from `tf.estimator.Estimator`. As we want to keep our SQL syntax simple, we would make the system extensible by calling estimators contributed by machine learning experts and written in Python. + +The SQL syntax must allow users to set Estimator attributes (parameters of the Python class' constructor, and those of `train`, `evaluate`, or `predict`). Users can choose to use default values. We have a plan to integrate our hyperparameter estimation research into the system to optimize the default values. + +Though estimators derived from `tf.estimator.Estimator` run algorithms as TensorFlow graphs; SQLFlow doesn't restrict that the underlying machine learning toolkit has to be TensorFlow. Indeed, as long as an estimator provides methods of `train`, `evaluate`, and `predict`, SQLFlow doesn't care if it calls TensorFlow or xgboost. Precisely, what SQLFlow expect is an interface like the following: + +```python +class AnEstimatorClass: + __init__(self, **kwargs) + train(self, **kwargs) + evaluate(self, **kwargs) + predict(self, **kwargs) +``` + +We also want to reuse the [feature columns API](https://www.tensorflow.org/guide/feature_columns) from Estimator, which allows users to columns of tables in a SQL engine to features to the model. + + +## Extended SQL Syntax + +Again, just as the beginning of the iteration, we propose the syntax for training as + +```sql +SELECT * FROM kaggle_credit_fraud_training_data +LIMIT 1000 +TRAIN DNNClassifier /* a pre-defined TensorFlow estimator, tf.estimator.DNNClassifier */ +WITH layers=[100, 200], /* a parameter of the Estimator class constructor */ + train.batch_size = 8 /* a parameter of the Estimator.train method */ +COLUMN *, /* all columns as raw features */ + cross(v1, v9, v28) /* plus a derived (crossed) column */ +LABEL class +INTO sqlflow_models.my_model_table; /* saves trained model parameters and features into a table */ +``` + +We see the redundancy of `*` in two clauses: `SELECT` and `COLUMN`. The following alternative can avoid the redundancy, but cannot specify the label. + +```sql +SELECT * /* raw features or the label? */ + corss(v1, v9, v28) /* derived featuers */ +FROM kaggle_credit_fraud_training_data +``` + +Please be aware that we save the trained models into tables, instead of a variable maintained by the underlying SQL engine. To invent a new variable type to hold trained models, we'd make our system tightly integrated with the SQL engine, and harms the extensibility to other engines. + +The result table should include the following information: + +1. The estimator name, e.g., `DNNClassifier` in this case. +1. Estimator attributes, e.g., `layer` and `train.batch_size`. +1. The feature mapping, e.g., `*` and `cross(v1, v9, v28)`. + +Similarly, to infer the class (fraud or regular), we could + +```sql +SELECT * FROM kaggle_credit_fraud_development_data +PREDICT kaggle_credit_fraud_development_data.class +USING sqlflow_models.my_model_table; +``` + +## System Architecture + +### A Conceptual Overview + +In the prototype, we use the following architecture: + +``` +SQL statement -> our SQL parser --standard SQL-> MySQL + \-extended SQL-> code generator -> execution engine +``` + +In the prototype, the code generator generates a Python program that trains or predicts. In either case, + +1. it retrieves the data from MySQL via [MySQL Connector Python API](https://dev.mysql.com/downloads/connector/python/), +1. optionally, retrieves the model from MySQL, +1. trains the model or predicts using the trained model by calling the user specified TensorFlow estimator, +1. and writes the trained model or prediction results into a table. + +### Working with Jupyter Notebook and Kubernetes + +The following figures shows the system components and their runtime environment. The left part shows how to run the system on a PC/laptop, the right part shows how to run it on a Kubernetes cluster. + +![](doc/figures/sqlflow-arch.png) diff --git a/docs/sqlflow/doc/walkthrough.html b/docs/sqlflow/doc/walkthrough.html new file mode 100644 index 0000000..a6d5369 --- /dev/null +++ b/docs/sqlflow/doc/walkthrough.html @@ -0,0 +1,362 @@ + + + + + + + + + SQLFlow: Code Walkthrough - Sqlflow + + + + + + + + + + + + + + + + + +
    +
    + Sqlflow + +
    + + +
    +
    +

    This site uses Just the Docs, a documentation theme for Jekyll.

    +
    +
    +
    +
    +
    + +
    +
    + + Search +
    +
    +
    + + + + +
    +
    +
    + + + +
    +

    SQLFlow: Code Walkthrough

    + +

    User Experience

    + +

    SQLFlow allows users to write SQL programs with extended syntax in Jupyter Notebook or a command-line tool.

    + +

    The following SQL statements train a TensorFlow model named DNNClassifier, which is a Python class derived from tf.estimator.Estimator:

    + +
    
+SELECT * FROM a_table TRAIN DNNClassifier WITH learning_rate=0.01 INTO sqlflow_models.my_model;
+
    + +

    And the following statement uses the trained model for prediction.

    + +
    SELECT * FROM b_table PREDICT b_table.predicted_label USING sqlflow_models.my_model;
+
    + +

    Please be aware that the part in the above statements before the extended keyword TRAIN and PREDICT is a standard SQL statement. This feature simplifies the implementation of the SQLFlow system.

    + +

    System Implementation

    + +

    If a SQL statement is of the standard syntax, SQLFlow throws it to the SQL engine and redirects the output to the user; otherwise, SQLFlow translates the statement of extended syntax into a Python program. Currently, it generates a program that throws the standard-syntax part of SELECT to MySQL, reads the results in the train-loop of a TensorFlow program. We will talk about how to extend SQLFlow to connect more SQL engines like Oracle, Hive, and SparkSQL, and generates more types of machine learning programs that calls distributed TensorFlow, PyTorch, and xgboost later. Before that, let us explain the system components.

    + +

    SQLFlow as a gRPC Server

    + +

    SQLFlow is a gRPC server, which can connect with multiple clients. A typical client is pysqlflow, the SQLFlow plugin for Jupyter Notebook server. Another once is a text-based client /cmd/sqlflowserver/main.go.

    + +
    Jupyter Notebook          ---(SQL statements)-->       SQLFlow gRPC server
+(SQLFlow magic command)   <--(a stream of messages)--
+
    + +

    The protobuf definition of the gRPC service is at /server/proto/sqlflow.proto. The return of the method SQLFlow.Run is a stream of Reponses, where each represents either a table header, a row, or a log message. The header and rows are usually from a standard SQL statement, for example, SELECT or DESCRIBE, and the log messages are usually from the running of a generated Python program.

    + +

    SQLFlow in the gRPC Server

    + +

    Once the SQLFlow server receives a batch of SQL statements via a gRPC call, it runs the following steps for each statement:

    + +
      +
    1. the parser to generate parsing result,
      2. +
    2. the verifier to verify the semantics given the parsing result,
      3. +
    3. the code generator to generate a Python program, or the submitter, from the parsing result,
      4. +
    4. the executor that runs the submitter locally.
      5. +
    + +

    Step 3. and 4. are only for a SQL statement of extended syntax; otherwise, SQLFlow server proxies the standard-syntax statement to the SQL engine.

    + +

    The executor calls Go’s standard package that captures the stdout and stderr from the submitter process and passing the result back to the gRPC client. Therefore, it is the responsibility of the submitter to print log messages to its stderr and stdout.

    + +

    Minimal Viable Product

    + +

    In the minimal viable product (MVP) of SQLFlow, the code generator generates a Python program consists of two parts:

    + +
      +
    1. throw the standard SELECT part in the extended-syntax statement to MySQL via ODBC, and
      2. +
    2. a loop that reads outputs from the run of the SELECT statement and trains the model (or, using a trained model to predict).
      3. +
    + +

    The training part calls TensorFlow to update the parameters of the model specified in the TRAIN clause.

    + +

    Extensibility

    + +

    By writing more code generators, we could extend SQLFlow to support more SQL engines, e.g., Hive and Oracle, and use machine learning toolkits, e.g., PyTorch and xgboost, in addition to TensorFlow, on various computing platforms. You are welcome to add more code generators such as

    + +
      +
    • codegen_distributed_tf.go to generate a submitter program similar to the MVP but runs a distributed TensorFlow training job.
      • +
    • codegen_kubernetes_tf.go to launch a distributed TensorFlow job on a Kubernetes cluster, other than running locally, in the same container as where SQLFlow gRPC server resides.
      • +
    • codegen_gcloud_pytorch.go to launch a submitter that calls PyTorch instead of TensorFlow for training on the Google Cloud.
      • +
    + +

    Job Management

    + +

    The current design of the gRPC interface assumes that the connection between the client, e.g., the Jupyter Notebook, and the SQLFlow server keeps alive during the running of the training program. This assumption is reasonable because even if the user closes her/his Web browser and disconnect to the Jupyter Notebook server, the connection between Jupyter to SQLFlow server might keep alive. However, this might not be robust enough if the Jupyter Notebook server runs on a user’s laptop and gets killed. In such a case, the gRPC server cannot stream the messages back to the client and would cause the failure of the submitter.

    + +

    A solution is to change the gRPC interface of SQLFlow server to have a method that files a job and returns immediately, and another method to get a batch of recent messages given a job ID. We will make a design for that soon.

    + + + + +
    +
    +
    +
    + + + diff --git a/docs/sqlflow/doc/walkthrough.md b/docs/sqlflow/doc/walkthrough.md new file mode 100644 index 0000000..66bf11d --- /dev/null +++ b/docs/sqlflow/doc/walkthrough.md @@ -0,0 +1,71 @@ +# SQLFlow: Code Walkthrough + +## User Experience + +SQLFlow allows users to write SQL programs with extended syntax in Jupyter Notebook or a command-line tool. + +The following SQL statements train a TensorFlow model named `DNNClassifier`, which is a Python class derived from [`tf.estimator.Estimator`](https://www.tensorflow.org/api_docs/python/tf/estimator/Estimator): + +```sql + +SELECT * FROM a_table TRAIN DNNClassifier WITH learning_rate=0.01 INTO sqlflow_models.my_model; +``` + +And the following statement uses the trained model for prediction. + +```sql +SELECT * FROM b_table PREDICT b_table.predicted_label USING sqlflow_models.my_model; +``` + +Please be aware that the part in the above statements before the extended keyword TRAIN and PREDICT is a standard SQL statement. This feature simplifies the implementation of the SQLFlow system. + +## System Implementation + +If a SQL statement is of the standard syntax, SQLFlow throws it to the SQL engine and redirects the output to the user; otherwise, SQLFlow translates the statement of extended syntax into a Python program. Currently, it generates a program that throws the standard-syntax part of SELECT to MySQL, reads the results in the train-loop of a TensorFlow program. We will talk about how to extend SQLFlow to connect more SQL engines like Oracle, Hive, and SparkSQL, and generates more types of machine learning programs that calls distributed TensorFlow, PyTorch, and xgboost later. Before that, let us explain the system components. + +### SQLFlow as a gRPC Server + +SQLFlow is a gRPC server, which can connect with multiple clients. A typical client is [pysqlflow](https://github.com/sql-machine-learning/pysqlflow), the SQLFlow plugin for Jupyter Notebook server. Another once is a text-based client [/cmd/sqlflowserver/main.go](/cmd/sqlflowserver/main.go). + +``` +Jupyter Notebook ---(SQL statements)--> SQLFlow gRPC server +(SQLFlow magic command) <--(a stream of messages)-- +``` + +The protobuf definition of the gRPC service is at [/server/proto/sqlflow.proto](/server/proto/sqlflow.proto). The return of the method `SQLFlow.Run` is a stream of `Reponse`s, where each represents either a table header, a row, or a log message. The header and rows are usually from a standard SQL statement, for example, SELECT or DESCRIBE, and the log messages are usually from the running of a generated Python program. + +### SQLFlow in the gRPC Server + +Once the SQLFlow server receives a batch of SQL statements via a gRPC call, it runs the following steps for each statement: + +1. the [parser](/sql/sql.y) to generate parsing result, +2. the [verifier](/sql/verifier.go) to verify the semantics given the parsing result, +3. the [code generator](/sql/codegen.go) to generate a Python program, or the *submitter*, from the parsing result, +4. the [executor](/sql/executor.go) that runs the submitter locally. + +Step 3. and 4. are only for a SQL statement of extended syntax; otherwise, SQLFlow server proxies the standard-syntax statement to the SQL engine. + +The executor calls Go's [standard package](https://godoc.org/os/exec) that captures the stdout and stderr from the submitter process and passing the result back to the gRPC client. Therefore, it is the responsibility of the submitter to print log messages to its stderr and stdout. + +### Minimal Viable Product + +In the minimal viable product (MVP) of SQLFlow, the code generator generates a Python program consists of two parts: + +1. throw the standard SELECT part in the extended-syntax statement to MySQL via ODBC, and +1. a loop that reads outputs from the run of the SELECT statement and trains the model (or, using a trained model to predict). + +The training part calls TensorFlow to update the parameters of the model specified in the TRAIN clause. + +### Extensibility + +By writing more code generators, we could extend SQLFlow to support more SQL engines, e.g., Hive and Oracle, and use machine learning toolkits, e.g., PyTorch and xgboost, in addition to TensorFlow, on various computing platforms. You are welcome to add more code generators such as + +- `codegen_distributed_tf.go` to generate a submitter program similar to the MVP but runs a distributed TensorFlow training job. +- `codegen_kubernetes_tf.go` to launch a distributed TensorFlow job on a Kubernetes cluster, other than running locally, in the same container as where SQLFlow gRPC server resides. +- `codegen_gcloud_pytorch.go` to launch a submitter that calls PyTorch instead of TensorFlow for training on the Google Cloud. + +### Job Management + +The current design of the gRPC interface assumes that the connection between the client, e.g., the Jupyter Notebook, and the SQLFlow server keeps alive during the running of the training program. This assumption is reasonable because even if the user closes her/his Web browser and disconnect to the Jupyter Notebook server, the connection between Jupyter to SQLFlow server might keep alive. However, this might not be robust enough if the Jupyter Notebook server runs on a user's laptop and gets killed. In such a case, the gRPC server cannot stream the messages back to the client and would cause the failure of the submitter. + +A solution is to change the gRPC interface of SQLFlow server to have a method that files a job and returns immediately, and another method to get a batch of recent messages given a job ID. We will make a design for that soon. diff --git a/docs/sqlflow/example/creditcard/README.md b/docs/sqlflow/example/creditcard/README.md new file mode 100644 index 0000000..623c61c --- /dev/null +++ b/docs/sqlflow/example/creditcard/README.md @@ -0,0 +1,149 @@ +# The Credit Card Fraud Detection Example + +The `downsample/creditcard.csv` file in this directory comes from [Kaggle](https://www.kaggle.com/mlg-ulb/creditcardfraud). + +## Down-sample + +To make it an example that can be checked into a git repo, I randomly picked up 2000 out from the 284,808 lines from the original dataset using the following command: + +```bash +gshuf -n 5000 /tmp/creditcard.csv > ./downsample/creditcard.csv +``` + +where `/tmp/creditcard.csv` is the one downloaded from Kaggle. + +If you want to down-sample by yourself on macOS, you might need to install `gshuf`. + +```bash +brew install coreutils +``` + +## Import into MySQL + +To develop/test/debug the TensorFlow program automatically generated by SQLFlow, which is supposed to be able to read data from MySQL, we'd import the `downsample/creditcard.csv` into MySQL. + +### Run MySQL Server + +For more about running MySQL in Docker containers, please refer to [Run MySQL Server and Client in Docker Containers](../../doc/mysql-setup.md). + +The following command starts the MySQL server: + +```bash +docker run --rm \ + -v $HOME/work/creditcarddb:/var/lib/mysql \ + --name mysql01 \ + -e MYSQL_ROOT_PASSWORD=root \ + -e MYSQL_ROOT_HOST='%' \ + -p 3306:3306 \ + -d \ + mysql/mysql-server:8.0 \ + mysqld --secure-file-priv="" +``` + +Please be aware that the above command doesn't use the default entrypoint of the Docker image; instead, it explicitly start `mysqld` in the container so that it could provide the command line option `--secure-file-priv=""`, which is required by `mysqlimport`. + +Run the following command to start a MySQL client: + +```bash +docker run --rm -it \ + -v $HOME/work/creditcarddb:/var/lib/mysql \ + mysql/mysql-server:8.0 \ + mysql -uroot -p +``` + +Remember to type the password `root` when prompted. Then input the following SQL program in the client to create the table `creditcard.creditcard`: + +```sql +CREATE DATABASE creditcard; +USE creditcard; +CREATE TABLE IF NOT EXISTS creditcard ( + time INT, + v1 float, + v2 float, + v3 float, + v4 float, + v5 float, + v6 float, + v7 float, + v8 float, + v9 float, + v10 float, + v11 float, + v12 float, + v13 float, + v14 float, + v15 float, + v16 float, + v17 float, + v18 float, + v19 float, + v20 float, + v21 float, + v22 float, + v23 float, + v24 float, + v25 float, + v26 float, + v27 float, + v28 float, + amount float, + class varchar(255)); +``` + +Before we could import the CSV file, we must follow the security policy of MySQL, which is wired and hard to understand, and copy the CSV file into the database directory: + +```bash +cp downsample/creditcard.csv ~/work/creditcarddb/creditcard/ +``` + +The following command runs `mysqlimport` in the `mysql/mysql` Docker image to import the data: + +```bash +docker run --rm -it \ + -v $HOME/work/creditcarddb:/var/lib/mysql \ + -v $HOME/work/creditcard:/var/lib/mysql-files \ + mysql:8.0 \ + mysqlimport \ + --ignore-lines=1 \ + --fields-terminated-by=, \ + --socket /var/lib/mysql/mysql.sock \ + -u root -p \ + creditcard creditcard.csv +``` + +Please be aware that `mysqlimport` doesn't exist in the `mysql/mysql-server` image. + +A message likes the following indicates the success of importing: + +``` +creditcard.creditcard: Records: 1999 Deleted: 0 Skipped: 0 Warnings: 0 +``` + +We can verify the imported result by running a SELECT statement like the following + +```bash +docker run --rm -it \ + -v $HOME/work/creditcarddb:/var/lib/mysql \ + mysql/mysql-server:8.0 \ + mysql -uroot -p \ + -e 'use creditcard; select time, v1, v28, amount, class from creditcard limit 10;' +``` + +which should print something like the following + +``` ++--------+-----------+------------+--------+-------+ +| time | v1 | v28 | amount | class | ++--------+-----------+------------+--------+-------+ +| 123590 | -0.341154 | 0.130741 | 69 | "0" | +| 135797 | -0.118853 | 0.11599 | 11.79 | "0" | +| 161249 | -1.04129 | -0.362025 | 7.38 | "0" | +| 72471 | 0.603151 | 0.0551408 | 352.04 | "0" | +| 120182 | 2.10678 | -0.0566284 | 0.89 | "0" | +| 160476 | 2.37223 | -0.0724277 | 6 | "0" | +| 135023 | 1.87987 | 0.0214597 | 20.05 | "0" | +| 117329 | -0.375213 | -0.130152 | 159 | "0" | +| 148637 | 1.83178 | -0.0378951 | 38.37 | "0" | +| 48293 | 1.26094 | 0.0275234 | 12 | "0" | ++--------+-----------+------------+--------+-------+ +``` diff --git a/docs/sqlflow/example/creditcard/index.html b/docs/sqlflow/example/creditcard/index.html new file mode 100644 index 0000000..ea22e98 --- /dev/null +++ b/docs/sqlflow/example/creditcard/index.html @@ -0,0 +1,427 @@ + + + + + + + + + The Credit Card Fraud Detection Example - Sqlflow + + + + + + + + + + + + + + + + + +
    +
    + Sqlflow + +
    + + +
    +
    +

    This site uses Just the Docs, a documentation theme for Jekyll.

    +
    +
    +
    +
    +
    + +
    +
    + + Search +
    +
    +
    + + + + +
    +
    +
    + + + +
    +

    The Credit Card Fraud Detection Example

    + +

    The downsample/creditcard.csv file in this directory comes from Kaggle.

    + +

    Down-sample

    + +

    To make it an example that can be checked into a git repo, I randomly picked up 2000 out from the 284,808 lines from the original dataset using the following command:

    + +
    gshuf -n 5000 /tmp/creditcard.csv  > ./downsample/creditcard.csv
+
    + +

    where /tmp/creditcard.csv is the one downloaded from Kaggle.

    + +

    If you want to down-sample by yourself on macOS, you might need to install gshuf.

    + +
    brew install coreutils
+
    + +

    Import into MySQL

    + +

    To develop/test/debug the TensorFlow program automatically generated by SQLFlow, which is supposed to be able to read data from MySQL, we’d import the downsample/creditcard.csv into MySQL.

    + +

    Run MySQL Server

    + +

    For more about running MySQL in Docker containers, please refer to Run MySQL Server and Client in Docker Containers.

    + +

    The following command starts the MySQL server:

    + +
    docker run --rm \
+   -v $HOME/work/creditcarddb:/var/lib/mysql \
+   --name mysql01 \
+   -e MYSQL_ROOT_PASSWORD=root \
+   -e MYSQL_ROOT_HOST='%' \
+   -p 3306:3306 \
+   -d \
+   mysql/mysql-server:8.0 \
+   mysqld --secure-file-priv=""
+
    + +

    Please be aware that the above command doesn’t use the default entrypoint of the Docker image; instead, it explicitly start mysqld in the container so that it could provide the command line option --secure-file-priv="", which is required by mysqlimport.

    + +

    Run the following command to start a MySQL client:

    + +
    docker run --rm -it \
+   -v $HOME/work/creditcarddb:/var/lib/mysql \
+   mysql/mysql-server:8.0 \
+   mysql -uroot -p
+
    + +

    Remember to type the password root when prompted. Then input the following SQL program in the client to create the table creditcard.creditcard:

    + +
    CREATE DATABASE creditcard;
+USE creditcard;
+CREATE TABLE IF NOT EXISTS creditcard (
+    time INT,
+	v1 float,
+	v2 float,
+	v3 float,
+	v4 float,
+	v5 float,
+	v6 float,
+	v7 float,
+	v8 float,
+	v9 float,
+	v10 float,
+	v11 float,
+	v12 float,
+	v13 float,
+	v14 float,
+	v15 float,
+	v16 float,
+	v17 float,
+	v18 float,
+	v19 float,
+	v20 float,
+	v21 float,
+	v22 float,
+	v23 float,
+	v24 float,
+	v25 float,
+	v26 float,
+	v27 float,
+	v28 float,
+	amount float,
+	class varchar(255));
+
    + +

    Before we could import the CSV file, we must follow the security policy of MySQL, which is wired and hard to understand, and copy the CSV file into the database directory:

    + +
    cp downsample/creditcard.csv ~/work/creditcarddb/creditcard/
+
    + +

    The following command runs mysqlimport in the mysql/mysql Docker image to import the data:

    + +
    docker run --rm -it \
+    -v $HOME/work/creditcarddb:/var/lib/mysql \
+	-v $HOME/work/creditcard:/var/lib/mysql-files \
+    mysql:8.0 \
+    mysqlimport \
+	  --ignore-lines=1  \
+      --fields-terminated-by=,  \
+      --socket /var/lib/mysql/mysql.sock \
+	  -u root -p \
+	  creditcard creditcard.csv 
+
    + +

    Please be aware that mysqlimport doesn’t exist in the mysql/mysql-server image.

    + +

    A message likes the following indicates the success of importing:

    + +
    creditcard.creditcard: Records: 1999  Deleted: 0  Skipped: 0  Warnings: 0
+
    + +

    We can verify the imported result by running a SELECT statement like the following

    + +
    docker run --rm -it \
+   -v $HOME/work/creditcarddb:/var/lib/mysql \
+   mysql/mysql-server:8.0 \
+   mysql -uroot -p \
+   -e 'use creditcard; select time, v1, v28, amount, class from creditcard limit 10;'
+
    + +

    which should print something like the following

    + +
    +--------+-----------+------------+--------+-------+
+| time   | v1        | v28        | amount | class |
++--------+-----------+------------+--------+-------+
+| 123590 | -0.341154 |   0.130741 |     69 | "0"   |
+| 135797 | -0.118853 |    0.11599 |  11.79 | "0"   |
+| 161249 |  -1.04129 |  -0.362025 |   7.38 | "0"   |
+|  72471 |  0.603151 |  0.0551408 | 352.04 | "0"   |
+| 120182 |   2.10678 | -0.0566284 |   0.89 | "0"   |
+| 160476 |   2.37223 | -0.0724277 |      6 | "0"   |
+| 135023 |   1.87987 |  0.0214597 |  20.05 | "0"   |
+| 117329 | -0.375213 |  -0.130152 |    159 | "0"   |
+| 148637 |   1.83178 | -0.0378951 |  38.37 | "0"   |
+|  48293 |   1.26094 |  0.0275234 |     12 | "0"   |
++--------+-----------+------------+--------+-------+
+
    + + + + +
    +
    +
    +
    + + + diff --git a/docs/sqlflow/example/creditcard/initialize_dataset_and_database.bash b/docs/sqlflow/example/creditcard/initialize_dataset_and_database.bash new file mode 100755 index 0000000..560e618 --- /dev/null +++ b/docs/sqlflow/example/creditcard/initialize_dataset_and_database.bash @@ -0,0 +1,127 @@ +#!/bin/bash +################################################################################ + +if [[ $(which docker) == "" ]]; then + echo "Cannot find Docker. " + echo "Please install it referring https://docs.docker.com/install" + exit +fi + +docker version > /dev/null +if [[ $? != "0" ]]; then + echo "The command docker version doesn't work as expected." + echo "Please check the reason and retry." + exit +fi + +if [[ $(which kaggle) == "" ]]; then + echo "Cannot find the command line tool kaggle. Try to install it ..."; + pip install kaggle --upgrade + if [[ $? != "0" ]]; then + echo "Failed to install kaggle." + echo "Please check https://github.com/Kaggle/kaggle-api" + exit + fi +fi + +if [[ ! -f $HOME/.kaggle/kaggle.json ]]; then + echo "Please sign up and get your API credentials from kaggle.com" + echo "Please check https://github.com/Kaggle/kaggle-api" + exit +fi + +export SQLFLOW_DB_DIR=$HOME/.sqlflow/example/db/creditcard +export SQLFLOW_DATABASE=creditcard + +echo "Cleanup legacy resource ..." +docker stop mysql-creditcard +docker rm mysql-creditcard +rm -rf $SQLFLOW_DB_DIR +mkdir -p $SQLFLOW_DB_DIR + +echo "Starting the MySQL server in a Docker container ..." +docker run --rm \ + -v $SQLFLOW_DB_DIR:/var/lib/mysql \ + --name mysql-creditcard \ + -e MYSQL_ROOT_PASSWORD=root \ + -e MYSQL_ROOT_HOST='%' \ + -p 3306:3306 \ + -d \ + mysql/mysql-server:8.0 \ + mysqld --secure-file-priv="" + +echo "Downloading the dataset ..." +kaggle datasets download -p /tmp --unzip \ + -d mlg-ulb/creditcardfraud + +if [[ ! -S $SQLFLOW_DB_DIR/mysql.sock ]]; then + echo "Cannot find " $SQLFLOW_DB_DIR/mysql.sock + echo "Waiting for 10 more seconds ..." + sleep 10 + if [[ ! -S $SQLFLOW_DB_DIR/mysql.sock ]]; then + echo "Seems that something went wrong." + exit + fi +fi + +echo "Creating the database and table creditcard.creditcard ..." +docker run --rm -it \ + -v $SQLFLOW_DB_DIR:/var/lib/mysql \ + mysql/mysql-server:8.0 \ + mysql --user root --password=root \ + -e 'CREATE DATABASE creditcard; +USE creditcard; +CREATE TABLE IF NOT EXISTS creditcard ( + time INT, + v1 float, + v2 float, + v3 float, + v4 float, + v5 float, + v6 float, + v7 float, + v8 float, + v9 float, + v10 float, + v11 float, + v12 float, + v13 float, + v14 float, + v15 float, + v16 float, + v17 float, + v18 float, + v19 float, + v20 float, + v21 float, + v22 float, + v23 float, + v24 float, + v25 float, + v26 float, + v27 float, + v28 float, + amount float, + class varchar(255));' + +if [[ $? != "0" ]]; then + echo "Failed to create table" + exit +fi + +if [[ ! -f /tmp/creditcard.csv ]]; then + echo "Failed to download the dataset" + exit +fi + +echo "Importing the CSV file ..." +mv /tmp/creditcard.csv $SQLFLOW_DB_DIR/creditcard +docker run --rm -it \ + -v $SQLFLOW_DB_DIR:/var/lib/mysql \ + mysql:8.0 \ + mysqlimport \ + --ignore-lines=1 \ + --fields-terminated-by=, \ + --socket /var/lib/mysql/mysql.sock \ + -u root --password=root \ + creditcard creditcard.csv diff --git a/docs/sqlflow/example/datasets/Dockerfile b/docs/sqlflow/example/datasets/Dockerfile new file mode 100644 index 0000000..7c7136d --- /dev/null +++ b/docs/sqlflow/example/datasets/Dockerfile @@ -0,0 +1,4 @@ +FROM mysql/mysql-server:8.0 + +COPY popularize_churn.sql /popularize_churn.sql +COPY popularize_iris.sql /popularize_iris.sql diff --git a/docs/sqlflow/example/datasets/README.md b/docs/sqlflow/example/datasets/README.md new file mode 100644 index 0000000..182575a --- /dev/null +++ b/docs/sqlflow/example/datasets/README.md @@ -0,0 +1,75 @@ +# The MySQL Server Container for Testing + +This directory contains a Dockerfile that builds a Docker image derived the MySQL Server 8.0 image, and includes SQL programs that popularize the following datasets: + +1. [Churn from Kaggle](https://www.kaggle.com/blastchar/telco-customer-churn) +1. [Irises classfication from TensorFlow](https://www.tensorflow.org/guide/premade_estimators#classifying_irises_an_overview) + +We can run a Docker container of it for unit testing. + +## Build + +```bash +cd example/datasets +docker build -t sqlflow:data . +``` + +## Run + +```bash +docker run --rm -d --name sqlflowdata \ + -p 3306:3306 \ + -e MYSQL_ROOT_PASSWORD=root \ + -e MYSQL_ROOT_HOST=% \ + sqlflow:data +``` + +## Popularize Datasets + +We need to manually popularize the databases and tables: + +```bash +docker exec -it sqlflowdata bash +``` + +To popularize the Churn dataset into `churn`: + +```bash +cat /popularize_churn.sql | mysql -uroot -proot +``` + +To popularize the Irises dataset into `iris`: + +```bash +cat /popularize_iris.sql | mysql -uroot -proot +``` + +To prepare database for storing machine learning models: + +```bash +echo "CREATE DATABASE IF NOT EXISTS sqlflow_models;" | mysql -uroot -proot +``` + +## Query + +In the container, run + +```bash +echo "select count(*) from churn.test;" | mysql -uroot -proot +``` + +should print the number of rows as the following + +``` +count(*) +10 +``` + +## Trouble shooting: + +1. It usually takes about 15 seconds to bring up the MySQL Server. If you try to connect it +before that, you may see the following error + +``` +ERROR 1045 (28000): Access denied for user 'root'@'localhost' (using password: YES) +``` diff --git a/docs/sqlflow/example/datasets/data_to_sql.awk b/docs/sqlflow/example/datasets/data_to_sql.awk new file mode 100644 index 0000000..36a9af1 --- /dev/null +++ b/docs/sqlflow/example/datasets/data_to_sql.awk @@ -0,0 +1,20 @@ +# To get the INSERT statements in popularize_table.sql, please apply +# this script to the dataset downloaded from +# https://www.kaggle.com/blastchar/telco-customer-churn. Then, +# manually remove the few lines that contains " ". Before applying +# this AWK program, please convert the downloaded CSV file from DOS +# format into UNIX format, and remove the first line of metadata. +BEGIN { + FS = ","; +} + +{ + printf("INSERT INTO %s VALUES(", table); + for (i = 1; i <= NF; i++) { + printf("\"%s\"", $i); + if (i < NF) { + printf(","); + } + } + printf(");\n"); +} diff --git a/docs/sqlflow/example/datasets/index.html b/docs/sqlflow/example/datasets/index.html new file mode 100644 index 0000000..c89c653 --- /dev/null +++ b/docs/sqlflow/example/datasets/index.html @@ -0,0 +1,358 @@ + + + + + + + + + The MySQL Server Container for Testing - Sqlflow + + + + + + + + + + + + + + + + + +
    +
    + Sqlflow + +
    + + +
    +
    +

    This site uses Just the Docs, a documentation theme for Jekyll.

    +
    +
    +
    +
    +
    + +
    +
    + + Search +
    +
    +
    + + + + +
    +
    +
    + + + +
    +

    The MySQL Server Container for Testing

    + +

    This directory contains a Dockerfile that builds a Docker image derived the MySQL Server 8.0 image, and includes SQL programs that popularize the following datasets:

    + +
      +
    1. Churn from Kaggle
      2. +
    2. Irises classfication from TensorFlow
      3. +
    + +

    We can run a Docker container of it for unit testing.

    + +

    Build

    + +
    cd example/datasets
+docker build -t sqlflow:data .
+
    + +

    Run

    + +
    docker run --rm -d --name sqlflowdata \
+   -p 3306:3306 \
+   -e MYSQL_ROOT_PASSWORD=root \
+   -e MYSQL_ROOT_HOST=% \
+   sqlflow:data
+
    + +

    Popularize Datasets

    + +

    We need to manually popularize the databases and tables:

    + +
    docker exec -it sqlflowdata bash
+
    + +

    To popularize the Churn dataset into churn:

    + +
    cat /popularize_churn.sql | mysql -uroot -proot
+
    + +

    To popularize the Irises dataset into iris:

    + +
    cat /popularize_iris.sql | mysql -uroot -proot
+
    + +

    To prepare database for storing machine learning models:

    + +
    echo "CREATE DATABASE IF NOT EXISTS sqlflow_models;" | mysql -uroot -proot
+
    + +

    Query

    + +

    In the container, run

    + +
    echo "select count(*) from churn.test;" | mysql -uroot -proot
+
    + +

    should print the number of rows as the following

    + +
    count(*)
+10
+
    + +

    Trouble shooting:

    + +
      +
    1. It usually takes about 15 seconds to bring up the MySQL Server. If you try to connect it +before that, you may see the following error
      2. +
    + +
    ERROR 1045 (28000): Access denied for user 'root'@'localhost' (using password: YES)
+
    + + + + +
    +
    +
    +
    + + + diff --git a/docs/sqlflow/example/fraud_detection/README.md b/docs/sqlflow/example/fraud_detection/README.md new file mode 100644 index 0000000..8885fc9 --- /dev/null +++ b/docs/sqlflow/example/fraud_detection/README.md @@ -0,0 +1,101 @@ +# Credit Card Fraud Detection + +Data collected from [here](https://www.kaggle.com/mlg-ulb/creditcardfraud) + +To run: `python premade_estimator.py`, which gives +``` +Adding numeric_columns: Time +Adding numeric_columns: V1 +Adding numeric_columns: V2 +Adding numeric_columns: V3 +Adding numeric_columns: V4 +Adding numeric_columns: V5 +Adding numeric_columns: V6 +Adding numeric_columns: V7 +Adding numeric_columns: V8 +Adding numeric_columns: V9 +Adding numeric_columns: V10 +Adding numeric_columns: V11 +Adding numeric_columns: V12 +Adding numeric_columns: V13 +Adding numeric_columns: V14 +Adding numeric_columns: V15 +Adding numeric_columns: V16 +Adding numeric_columns: V17 +Adding numeric_columns: V18 +Adding numeric_columns: V19 +Adding numeric_columns: V20 +Adding numeric_columns: V21 +Adding numeric_columns: V22 +Adding numeric_columns: V23 +Adding numeric_columns: V24 +Adding numeric_columns: V25 +Adding numeric_columns: V26 +Adding numeric_columns: V27 +Adding numeric_columns: V28 +Adding numeric_columns: Amount +INFO:tensorflow:Using default config. +WARNING:tensorflow:Using temporary folder as model directory: /tmp/tmpzj6rzutf +INFO:tensorflow:Using config: {'_num_worker_replicas': 1, '_global_id_in_cluster': 0, '_save_checkpoints_steps': None, '_tf_random_seed': None, '_evaluation_master': '', '_protocol': None, '_session_config': allow_soft_placement: true +graph_options { + rewrite_options { + meta_optimizer_iterations: ONE + } +} +, '_model_dir': '/tmp/tmpzj6rzutf', '_device_fn': None, '_cluster_spec': , '_save_summary_steps': 100, '_eval_distribute': None, '_keep_checkpoint_max': 5, '_master': '', '_train_distribute': None, '_log_step_count_steps': 100, '_num_ps_replicas': 0, '_experimental_distribute': None, '_save_checkpoints_secs': 600, '_service': None, '_task_type': 'worker', '_task_id': 0, '_keep_checkpoint_every_n_hours': 10000, '_is_chief': True} +INFO:tensorflow:Calling model_fn. +INFO:tensorflow:Done calling model_fn. +INFO:tensorflow:Create CheckpointSaverHook. +INFO:tensorflow:Graph was finalized. +INFO:tensorflow:Running local_init_op. +INFO:tensorflow:Done running local_init_op. +INFO:tensorflow:Saving checkpoints for 0 into /tmp/tmpzj6rzutf/model.ckpt. +INFO:tensorflow:loss = 892.3532, step = 1 +INFO:tensorflow:global_step/sec: 295.764 +INFO:tensorflow:loss = 0.00025717804, step = 101 (0.338 sec) +INFO:tensorflow:global_step/sec: 446.454 +INFO:tensorflow:loss = 0.00024299514, step = 201 (0.224 sec) +INFO:tensorflow:global_step/sec: 444.479 +INFO:tensorflow:loss = 0.00020988214, step = 301 (0.225 sec) +INFO:tensorflow:global_step/sec: 421.979 +INFO:tensorflow:loss = 0.00023327809, step = 401 (0.237 sec) +INFO:tensorflow:global_step/sec: 446.172 +INFO:tensorflow:loss = 0.00020405005, step = 501 (0.224 sec) +INFO:tensorflow:global_step/sec: 431.785 +INFO:tensorflow:loss = 0.00017921966, step = 601 (0.232 sec) +INFO:tensorflow:global_step/sec: 442.912 +INFO:tensorflow:loss = 0.00018488085, step = 701 (0.226 sec) +INFO:tensorflow:global_step/sec: 444.628 +INFO:tensorflow:loss = 0.00017795907, step = 801 (0.225 sec) +INFO:tensorflow:global_step/sec: 428.385 +INFO:tensorflow:loss = 0.00016896412, step = 901 (0.233 sec) +INFO:tensorflow:Saving checkpoints for 1000 into /tmp/tmpzj6rzutf/model.ckpt. +INFO:tensorflow:Loss for final step: 0.00016340206. +INFO:tensorflow:Calling model_fn. +WARNING:tensorflow:Trapezoidal rule is known to produce incorrect PR-AUCs; please switch to "careful_interpolation" instead. +WARNING:tensorflow:Trapezoidal rule is known to produce incorrect PR-AUCs; please switch to "careful_interpolation" instead. +INFO:tensorflow:Done calling model_fn. +INFO:tensorflow:Starting evaluation at 2018-10-16-01:02:49 +INFO:tensorflow:Graph was finalized. +INFO:tensorflow:Restoring parameters from /tmp/tmpzj6rzutf/model.ckpt-1000 +INFO:tensorflow:Running local_init_op. +INFO:tensorflow:Done running local_init_op. +INFO:tensorflow:Finished evaluation at 2018-10-16-01:02:49 +INFO:tensorflow:Saving dict for global step 1000: accuracy = 1.0, accuracy_baseline = 1.0, auc = 0.9999999, auc_precision_recall = 0.0, average_loss = 0.0001784083, global_step = 1000, label/mean = 0.0, loss = 0.0016056747, precision = 0.0, prediction/mean = 0.00017826515, recall = 0.0 +INFO:tensorflow:Saving 'checkpoint_path' summary for global step 1000: /tmp/tmpzj6rzutf/model.ckpt-1000 + +Test set accuracy: 1.00000 + +INFO:tensorflow:Calling model_fn. +INFO:tensorflow:Done calling model_fn. +INFO:tensorflow:Graph was finalized. +INFO:tensorflow:Restoring parameters from /tmp/tmpzj6rzutf/model.ckpt-1000 +INFO:tensorflow:Running local_init_op. +INFO:tensorflow:Done running local_init_op. + +Prediction is "Not Fraud" (100.0%), expected "Not Fraud" + +Prediction is "Not Fraud" (100.0%), expected "Not Fraud" + +Prediction is "Not Fraud" (100.0%), expected "Not Fraud" +``` diff --git a/docs/sqlflow/example/fraud_detection/creditcard.csv.test b/docs/sqlflow/example/fraud_detection/creditcard.csv.test new file mode 100644 index 0000000..134140e --- /dev/null +++ b/docs/sqlflow/example/fraud_detection/creditcard.csv.test @@ -0,0 +1,10 @@ +"Time","V1","V2","V3","V4","V5","V6","V7","V8","V9","V10","V11","V12","V13","V14","V15","V16","V17","V18","V19","V20","V21","V22","V23","V24","V25","V26","V27","V28","Amount","Class" +0,1.19185711131486,0.26615071205963,0.16648011335321,0.448154078460911,0.0600176492822243,-0.0823608088155687,-0.0788029833323113,0.0851016549148104,-0.255425128109186,-0.166974414004614,1.61272666105479,1.06523531137287,0.48909501589608,-0.143772296441519,0.635558093258208,0.463917041022171,-0.114804663102346,-0.183361270123994,-0.145783041325259,-0.0690831352230203,-0.225775248033138,-0.638671952771851,0.101288021253234,-0.339846475529127,0.167170404418143,0.125894532368176,-0.00898309914322813,0.0147241691924927,2.69,"0" +1,-0.966271711572087,-0.185226008082898,1.79299333957872,-0.863291275036453,-0.0103088796030823,1.24720316752486,0.23760893977178,0.377435874652262,-1.38702406270197,-0.0549519224713749,-0.226487263835401,0.178228225877303,0.507756869957169,-0.28792374549456,-0.631418117709045,-1.0596472454325,-0.684092786345479,1.96577500349538,-1.2326219700892,-0.208037781160366,-0.108300452035545,0.00527359678253453,-0.190320518742841,-1.17557533186321,0.647376034602038,-0.221928844458407,0.0627228487293033,0.0614576285006353,123.5,"0" +7,-0.644269442348146,1.41796354547385,1.0743803763556,-0.492199018495015,0.948934094764157,0.428118462833089,1.12063135838353,-3.80786423873589,0.615374730667027,1.24937617815176,-0.619467796121913,0.291474353088705,1.75796421396042,-1.32386521970526,0.686132504394383,-0.0761269994382006,-1.2221273453247,-0.358221569869078,0.324504731321494,-0.156741852488285,1.94346533978412,-1.01545470979971,0.057503529867291,-0.649709005559993,-0.415266566234811,-0.0516342969262494,-1.20692108094258,-1.08533918832377,40.8,"0" +14,-5.40125766315825,-5.45014783420644,1.18630463143652,1.73623880012095,3.04910587764025,-1.76340557365201,-1.55973769907953,0.160841747266769,1.23308974041888,0.345172827050629,0.917229867699146,0.970116716069048,-0.266567764915222,-0.479129929276704,-0.526608502569153,0.47200411177674,-0.725480944982201,0.075081351540202,-0.406866573198217,-2.19684802485647,-0.503600328973703,0.984459785590244,2.45858857639219,0.0421188969891572,-0.481630823956716,-0.621272013713977,0.392053289557744,0.949594245504846,46.8,"0" +18,1.16661638244228,0.502120087854101,-0.0673003143663533,2.26156923949128,0.428804194630708,0.0894735167274599,0.241146579907281,0.138081705243702,-0.989162395259052,0.922174967079328,0.744785789437272,-0.531377250104474,-2.1053464503177,1.12687010488156,0.00307532251490036,0.424424506356606,-0.454475291663708,-0.0988706272108058,-0.816597306905994,-0.307168508741084,0.0187018719918659,-0.0619722672279421,-0.103854921794533,-0.370415176650599,0.603200338674873,0.108555872801682,-0.0405207055560295,-0.0114178152991164,2.28,"0" +23,1.17328461017079,0.353497876646456,0.28390506526532,1.1335633178795,-0.172577181842493,-0.916053706932115,0.36902484539083,-0.327260242196767,-0.24665102789456,-0.0461393022689538,-0.143418527309124,0.979350375717585,1.49228543545219,0.101417525935881,0.76147754528118,-0.014584082470356,-0.511640116553888,-0.32505635460906,-0.390933797720026,0.0278779087328671,0.0670033044874043,0.227811927638182,-0.150487224821759,0.43504510338369,0.724824579577241,-0.337082059875827,0.0163683792669046,0.0300411908366385,41.88,"0" +23,-0.414288810090829,0.905437322625407,1.72745294417921,1.47347126657189,0.00744274117322988,-0.200330677416199,0.740228319420026,-0.029247400012072,-0.593392019124765,-0.346188231273199,-0.0121421884961019,0.786796316060126,0.635953883297271,-0.0863244719500975,0.0768036871984215,-1.4059193336148,0.775591738395162,-0.94288892749919,0.543969461627011,0.0973075910928853,0.0772374339988386,0.457330598656119,-0.038499724582026,0.642521902774992,-0.183891335259639,-0.277464019495649,0.182687486490454,0.152664644968975,33,"0" +26,-0.53538776309446,0.865267807575272,1.35107628772237,0.14757547449791,0.433680212077009,0.0869829381161816,0.693039311115721,0.179742260742556,-0.285641861230583,-0.482474471178567,0.871799580440785,0.853447434991487,-0.571821891415037,0.102252103048142,-1.51999120265821,-0.285912497081499,-0.309633386824141,-0.403901992939884,-0.823742994489464,-0.283263782156003,0.049525686687599,0.206536543024476,-0.18710806982075,0.000753014257497798,0.0981166059546266,-0.553470968822753,-0.0783055021726955,0.0254273775211197,1.77,"0" +29,0.996369531566045,-0.12258878723806,0.546819472613086,0.706579541087689,0.134559557409452,1.15699511200606,-0.294561331432968,0.407429099419853,0.337862663634561,-0.408150492552647,0.682646123949549,1.25371968484033,0.391917044082342,-0.0375108583595551,0.795301519477008,-1.6682105913656,1.32299488209244,-2.83003522867318,-1.46343168833485,-0.203461209817417,-0.0763018365480559,0.108866913420892,0.162230940339532,-0.575624047818567,0.109794914589208,0.373813367278979,0.0505521972760614,0.00510555920011407,20.53,"0" diff --git a/docs/sqlflow/example/fraud_detection/creditcard.csv.train b/docs/sqlflow/example/fraud_detection/creditcard.csv.train new file mode 100644 index 0000000..b205dec --- /dev/null +++ b/docs/sqlflow/example/fraud_detection/creditcard.csv.train @@ -0,0 +1,10 @@ +"Time","V1","V2","V3","V4","V5","V6","V7","V8","V9","V10","V11","V12","V13","V14","V15","V16","V17","V18","V19","V20","V21","V22","V23","V24","V25","V26","V27","V28","Amount","Class" +0,-1.3598071336738,-0.0727811733098497,2.53634673796914,1.37815522427443,-0.338320769942518,0.462387777762292,0.239598554061257,0.0986979012610507,0.363786969611213,0.0907941719789316,-0.551599533260813,-0.617800855762348,-0.991389847235408,-0.311169353699879,1.46817697209427,-0.470400525259478,0.207971241929242,0.0257905801985591,0.403992960255733,0.251412098239705,-0.018306777944153,0.277837575558899,-0.110473910188767,0.0669280749146731,0.128539358273528,-0.189114843888824,0.133558376740387,-0.0210530534538215,149.62,"0" +1,-1.35835406159823,-1.34016307473609,1.77320934263119,0.379779593034328,-0.503198133318193,1.80049938079263,0.791460956450422,0.247675786588991,-1.51465432260583,0.207642865216696,0.624501459424895,0.066083685268831,0.717292731410831,-0.165945922763554,2.34586494901581,-2.89008319444231,1.10996937869599,-0.121359313195888,-2.26185709530414,0.524979725224404,0.247998153469754,0.771679401917229,0.909412262347719,-0.689280956490685,-0.327641833735251,-0.139096571514147,-0.0553527940384261,-0.0597518405929204,378.66,"0" +2,-1.15823309349523,0.877736754848451,1.548717846511,0.403033933955121,-0.407193377311653,0.0959214624684256,0.592940745385545,-0.270532677192282,0.817739308235294,0.753074431976354,-0.822842877946363,0.53819555014995,1.3458515932154,-1.11966983471731,0.175121130008994,-0.451449182813529,-0.237033239362776,-0.0381947870352842,0.803486924960175,0.408542360392758,-0.00943069713232919,0.79827849458971,-0.137458079619063,0.141266983824769,-0.206009587619756,0.502292224181569,0.219422229513348,0.215153147499206,69.99,"0" +2,-0.425965884412454,0.960523044882985,1.14110934232219,-0.168252079760302,0.42098688077219,-0.0297275516639742,0.476200948720027,0.260314333074874,-0.56867137571251,-0.371407196834471,1.34126198001957,0.359893837038039,-0.358090652573631,-0.137133700217612,0.517616806555742,0.401725895589603,-0.0581328233640131,0.0686531494425432,-0.0331937877876282,0.0849676720682049,-0.208253514656728,-0.559824796253248,-0.0263976679795373,-0.371426583174346,-0.232793816737034,0.105914779097957,0.253844224739337,0.0810802569229443,3.67,"0" +4,1.22965763450793,0.141003507049326,0.0453707735899449,1.20261273673594,0.191880988597645,0.272708122899098,-0.00515900288250983,0.0812129398830894,0.464959994783886,-0.0992543211289237,-1.41690724314928,-0.153825826253651,-0.75106271556262,0.16737196252175,0.0501435942254188,-0.443586797916727,0.00282051247234708,-0.61198733994012,-0.0455750446637976,-0.21963255278686,-0.167716265815783,-0.270709726172363,-0.154103786809305,-0.780055415004671,0.75013693580659,-0.257236845917139,0.0345074297438413,0.00516776890624916,4.99,"0" +7,-0.89428608220282,0.286157196276544,-0.113192212729871,-0.271526130088604,2.6695986595986,3.72181806112751,0.370145127676916,0.851084443200905,-0.392047586798604,-0.410430432848439,-0.705116586646536,-0.110452261733098,-0.286253632470583,0.0743553603016731,-0.328783050303565,-0.210077268148783,-0.499767968800267,0.118764861004217,0.57032816746536,0.0527356691149697,-0.0734251001059225,-0.268091632235551,-0.204232669947878,1.0115918018785,0.373204680146282,-0.384157307702294,0.0117473564581996,0.14240432992147,93.2,"0" +9,-0.33826175242575,1.11959337641566,1.04436655157316,-0.222187276738296,0.49936080649727,-0.24676110061991,0.651583206489972,0.0695385865186387,-0.736727316364109,-0.366845639206541,1.01761446783262,0.836389570307029,1.00684351373408,-0.443522816876142,0.150219101422635,0.739452777052119,-0.540979921943059,0.47667726004282,0.451772964394125,0.203711454727929,-0.246913936910008,-0.633752642406113,-0.12079408408185,-0.385049925313426,-0.0697330460416923,0.0941988339514961,0.246219304619926,0.0830756493473326,3.68,"0" +10,1.44904378114715,-1.17633882535966,0.913859832832795,-1.37566665499943,-1.97138316545323,-0.62915213889734,-1.4232356010359,0.0484558879088564,-1.72040839292037,1.62665905834133,1.1996439495421,-0.671439778462005,-0.513947152539479,-0.0950450453999549,0.230930409124119,0.0319674667862076,0.253414715863197,0.854343814324194,-0.221365413645481,-0.387226474431156,-0.00930189652490052,0.313894410791098,0.0277401580170247,0.500512287104917,0.25136735874921,-0.129477953726618,0.0428498709381461,0.0162532619375515,7.8,"0" +10,0.38497821518095,0.616109459176472,-0.874299702595052,-0.0940186259679115,2.92458437838817,3.31702716826156,0.470454671805879,0.53824722837695,-0.558894612428441,0.30975539423728,-0.259115563735702,-0.326143233995877,-0.0900467227020648,0.362832368569793,0.928903660629178,-0.129486811402759,-0.809978925963589,0.359985390219981,0.70766382644648,0.12599157561542,0.049923685888971,0.238421512225103,0.00912986861262866,0.996710209581086,-0.767314827174801,-0.492208295340017,0.042472441919027,-0.0543373883732122,9.99,"0" diff --git a/docs/sqlflow/example/fraud_detection/creditcard_data.py b/docs/sqlflow/example/fraud_detection/creditcard_data.py new file mode 100644 index 0000000..439fea1 --- /dev/null +++ b/docs/sqlflow/example/fraud_detection/creditcard_data.py @@ -0,0 +1,59 @@ +# Reference: https://github.com/tensorflow/models/blob/master/samples/core/get_started/iris_data.py +# which released under the Apache License 2.0 +import pandas as pd +import tensorflow as tf + +TRAIN_DATA_FILE = './creditcard.csv.train' +TEST_DATA_FILE = './creditcard.csv.test' + +CSV_COLUMN_NAMES = ["Time"] + ["V"+str(i) for i in range(1,29)] + ["Amount", "Class"] + + +def load_data(y_name='Class'): + """Returns the iris dataset as (train_x, train_y), (test_x, test_y).""" + import os + import errno + for filename in (TRAIN_DATA_FILE, TEST_DATA_FILE): + if not os.path.exists(filename): + raise FileNotFoundError(errno.ENOENT, os.strerror(errno.ENOENT), filename) + + train = pd.read_csv(TRAIN_DATA_FILE, names=CSV_COLUMN_NAMES, header=0) + train_x, train_y = train, train.pop(y_name) + + test = pd.read_csv(TEST_DATA_FILE, names=CSV_COLUMN_NAMES, header=0) + test_x, test_y = test, test.pop(y_name) + + return (train_x, train_y), (test_x, test_y) + + +def train_input_fn(features, labels, batch_size): + """An input function for training""" + + # Convert the inputs to a Dataset. + dataset = tf.data.Dataset.from_tensor_slices((dict(features), labels)) + + # Shuffle, repeat, and batch the examples. + dataset = dataset.shuffle(1000).repeat().batch(batch_size) + + # Return the dataset. + return dataset + + +def eval_input_fn(features, labels, batch_size): + """An input function for evaluation or prediction""" + features=dict(features) + if labels is None: + # No labels, use only features. + inputs = features + else: + inputs = (features, labels) + + # Convert the inputs to a Dataset. + dataset = tf.data.Dataset.from_tensor_slices(inputs) + + # Batch the examples + assert batch_size is not None, "batch_size must not be None" + dataset = dataset.batch(batch_size) + + # Return the dataset. + return dataset diff --git a/docs/sqlflow/example/fraud_detection/index.html b/docs/sqlflow/example/fraud_detection/index.html new file mode 100644 index 0000000..0091873 --- /dev/null +++ b/docs/sqlflow/example/fraud_detection/index.html @@ -0,0 +1,388 @@ + + + + + + + + + Credit Card Fraud Detection - Sqlflow + + + + + + + + + + + + + + + + + +
    +
    + Sqlflow + +
    + + +
    +
    +

    This site uses Just the Docs, a documentation theme for Jekyll.

    +
    +
    +
    +
    +
    + +
    +
    + + Search +
    +
    +
    + + + + +
    +
    +
    + + + +
    +

    Credit Card Fraud Detection

    + +

    Data collected from here

    + +

    To run: python premade_estimator.py, which gives

    +
    Adding numeric_columns: Time
+Adding numeric_columns: V1
+Adding numeric_columns: V2
+Adding numeric_columns: V3
+Adding numeric_columns: V4
+Adding numeric_columns: V5
+Adding numeric_columns: V6
+Adding numeric_columns: V7
+Adding numeric_columns: V8
+Adding numeric_columns: V9
+Adding numeric_columns: V10
+Adding numeric_columns: V11
+Adding numeric_columns: V12
+Adding numeric_columns: V13
+Adding numeric_columns: V14
+Adding numeric_columns: V15
+Adding numeric_columns: V16
+Adding numeric_columns: V17
+Adding numeric_columns: V18
+Adding numeric_columns: V19
+Adding numeric_columns: V20
+Adding numeric_columns: V21
+Adding numeric_columns: V22
+Adding numeric_columns: V23
+Adding numeric_columns: V24
+Adding numeric_columns: V25
+Adding numeric_columns: V26
+Adding numeric_columns: V27
+Adding numeric_columns: V28
+Adding numeric_columns: Amount
+INFO:tensorflow:Using default config.
+WARNING:tensorflow:Using temporary folder as model directory: /tmp/tmpzj6rzutf
+INFO:tensorflow:Using config: {'_num_worker_replicas': 1, '_global_id_in_cluster': 0, '_save_checkpoints_steps': None, '_tf_random_seed': None, '_evaluation_master': '', '_protocol': None, '_session_config': allow_soft_placement: true
+graph_options {
+  rewrite_options {
+    meta_optimizer_iterations: ONE
+  }
+}
+, '_model_dir': '/tmp/tmpzj6rzutf', '_device_fn': None, '_cluster_spec': <tensorflow.python.training.server_lib.ClusterSpec object at 0x7f40f9b40438>, '_save_summary_steps': 100, '_eval_distribute': None, '_keep_checkpoint_max': 5, '_master': '', '_train_distribute': None, '_log_step_count_steps': 100, '_num_ps_replicas': 0, '_experimental_distribute': None, '_save_checkpoints_secs': 600, '_service': None, '_task_type': 'worker', '_task_id': 0, '_keep_checkpoint_every_n_hours': 10000, '_is_chief': True}
+INFO:tensorflow:Calling model_fn.
+INFO:tensorflow:Done calling model_fn.
+INFO:tensorflow:Create CheckpointSaverHook.
+INFO:tensorflow:Graph was finalized.
+INFO:tensorflow:Running local_init_op.
+INFO:tensorflow:Done running local_init_op.
+INFO:tensorflow:Saving checkpoints for 0 into /tmp/tmpzj6rzutf/model.ckpt.
+INFO:tensorflow:loss = 892.3532, step = 1
+INFO:tensorflow:global_step/sec: 295.764
+INFO:tensorflow:loss = 0.00025717804, step = 101 (0.338 sec)
+INFO:tensorflow:global_step/sec: 446.454
+INFO:tensorflow:loss = 0.00024299514, step = 201 (0.224 sec)
+INFO:tensorflow:global_step/sec: 444.479
+INFO:tensorflow:loss = 0.00020988214, step = 301 (0.225 sec)
+INFO:tensorflow:global_step/sec: 421.979
+INFO:tensorflow:loss = 0.00023327809, step = 401 (0.237 sec)
+INFO:tensorflow:global_step/sec: 446.172
+INFO:tensorflow:loss = 0.00020405005, step = 501 (0.224 sec)
+INFO:tensorflow:global_step/sec: 431.785
+INFO:tensorflow:loss = 0.00017921966, step = 601 (0.232 sec)
+INFO:tensorflow:global_step/sec: 442.912
+INFO:tensorflow:loss = 0.00018488085, step = 701 (0.226 sec)
+INFO:tensorflow:global_step/sec: 444.628
+INFO:tensorflow:loss = 0.00017795907, step = 801 (0.225 sec)
+INFO:tensorflow:global_step/sec: 428.385
+INFO:tensorflow:loss = 0.00016896412, step = 901 (0.233 sec)
+INFO:tensorflow:Saving checkpoints for 1000 into /tmp/tmpzj6rzutf/model.ckpt.
+INFO:tensorflow:Loss for final step: 0.00016340206.
+INFO:tensorflow:Calling model_fn.
+WARNING:tensorflow:Trapezoidal rule is known to produce incorrect PR-AUCs; please switch to "careful_interpolation" instead.
+WARNING:tensorflow:Trapezoidal rule is known to produce incorrect PR-AUCs; please switch to "careful_interpolation" instead.
+INFO:tensorflow:Done calling model_fn.
+INFO:tensorflow:Starting evaluation at 2018-10-16-01:02:49
+INFO:tensorflow:Graph was finalized.
+INFO:tensorflow:Restoring parameters from /tmp/tmpzj6rzutf/model.ckpt-1000
+INFO:tensorflow:Running local_init_op.
+INFO:tensorflow:Done running local_init_op.
+INFO:tensorflow:Finished evaluation at 2018-10-16-01:02:49
+INFO:tensorflow:Saving dict for global step 1000: accuracy = 1.0, accuracy_baseline = 1.0, auc = 0.9999999, auc_precision_recall = 0.0, average_loss = 0.0001784083, global_step = 1000, label/mean = 0.0, loss = 0.0016056747, precision = 0.0, prediction/mean = 0.00017826515, recall = 0.0
+INFO:tensorflow:Saving 'checkpoint_path' summary for global step 1000: /tmp/tmpzj6rzutf/model.ckpt-1000
+
+Test set accuracy: 1.00000
+
+INFO:tensorflow:Calling model_fn.
+INFO:tensorflow:Done calling model_fn.
+INFO:tensorflow:Graph was finalized.
+INFO:tensorflow:Restoring parameters from /tmp/tmpzj6rzutf/model.ckpt-1000
+INFO:tensorflow:Running local_init_op.
+INFO:tensorflow:Done running local_init_op.
+
+Prediction is "Not Fraud" (100.0%), expected "Not Fraud"
+
+Prediction is "Not Fraud" (100.0%), expected "Not Fraud"
+
+Prediction is "Not Fraud" (100.0%), expected "Not Fraud"
+
    + + + + +
    +
    +
    +
    + + + diff --git a/docs/sqlflow/example/fraud_detection/premade_estimator.py b/docs/sqlflow/example/fraud_detection/premade_estimator.py new file mode 100644 index 0000000..1bdcb5a --- /dev/null +++ b/docs/sqlflow/example/fraud_detection/premade_estimator.py @@ -0,0 +1,115 @@ +# Copyright 2016 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +"""An Example of a DNNClassifier for the Iris dataset.""" +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import argparse +import tensorflow as tf +# import tensorflow.contrib.eager as tfe +# tf.enable_eager_execution() + +import creditcard_data + + +parser = argparse.ArgumentParser() +parser.add_argument('--batch_size', default=100, type=int, help='batch size') +parser.add_argument('--train_steps', default=1000, type=int, + help='number of training steps') + +def main(argv): + args = parser.parse_args(argv[1:]) + + # Fetch the data + (train_x, train_y), (test_x, test_y) = creditcard_data.load_data() + + # Feature columns describe how to use the input. + my_feature_columns = [] + for key in train_x.keys(): + print("Adding numeric_columns: {}".format(key)) + my_feature_columns.append(tf.feature_column.numeric_column(key=key)) + + # Build 2 hidden layer DNN with 100, 100 units respectively. + classifier = tf.estimator.DNNClassifier( + feature_columns=my_feature_columns, + # Two hidden layers of 10 nodes each. + hidden_units=[100, 100], + # The model must choose between 3 classes. + n_classes=2) + + # Train the Model. + classifier.train( + input_fn=lambda:creditcard_data.train_input_fn(train_x, train_y, + args.batch_size), + steps=args.train_steps) + + # Evaluate the model. + eval_result = classifier.evaluate( + input_fn=lambda:creditcard_data.eval_input_fn(test_x, test_y, + args.batch_size)) + + print('\nTest set accuracy: {accuracy:0.5f}\n'.format(**eval_result)) + + # # Generate predictions from the model + expected = [0, 0, 0] + predict_x = { + 'Time': [80.0, 80.0, 81.0], + 'V1': [-0.655264276212716, -3.00773905109524, 1.11104797028704], + 'V2': [0.40989854283685295, 0.929514494266081, 0.215351844382971], + 'V3': [1.28915643808596, -0.42372085549803296, 0.374072563622289], + 'V4': [-0.32504325923176003, -1.3929516201707801, 1.1500006534851999], + 'V5': [0.545669467098008, 0.0822467824081475, -0.329767568535767], + 'V6': [-0.349810664410475, -0.515781701182148, -0.898687821706977], + 'V7': [0.648240481592264, 0.104713792402636, 0.30990664398233], + 'V8': [0.0360628816033704, 0.8253105658956079, -0.25527729116611103], + 'V9': [0.0787014163023305, 0.38545311035542207, -0.110655138734984], + 'V10': [-0.829167267726513, 0.0640002245721828, -0.0575788321387636], + 'V11': [-1.78165873189534, -1.29310671179606, -0.21796149670612197], + 'V12': [-0.412806782520344, 0.0621999429338998, 0.600591209357443], + 'V13': [-0.6262202895658521, 0.27110701307160395, 0.686379564975428], + 'V14': [-0.113361054850318, 0.328058800506629, 0.262161286299107], + 'V15': [-0.714754632820629, 0.7240976288015201, 0.9226769790285508], + 'V16': [0.120701513866132, 0.6383894438325529, 0.135235271944387], + 'V17': [-0.47428167431379603, -0.43008221783924894, -0.490174913186185], + 'V18': [-0.240594669556031, -0.7235902240912719, -0.392752898979319], + 'V19': [0.0408234763497192, -0.562830692998269, -0.315464926788021], + 'V20': [-0.0347759899640066, -0.15171422957684697, 0.031671787802255], + 'V21': [-0.155727058998072, -0.40924608012503005, -0.132555303999709], + 'V22': [-0.47739031832143397, -0.588831627173022, -0.493866878067896], + 'V23': [-0.126524992708454, 0.0680450889350803, -0.0107821837368997], + 'V24': [-0.443627776856693, -0.890869739213917, 0.380388564169084], + 'V25': [-0.012118142098906, -0.047641092795032, 0.45621418060635105], + 'V26': [0.143172915176852, 0.717616984694838, -0.572346756048474], + 'V27': [0.049782585303356, -0.0540936594394297, 0.0109814630026346], + 'V28': [0.118280286710879, -0.42425971753603997, 0.039295013148069], + 'Amount': [32.51, 39.95, 68.74] + } + + predictions = classifier.predict( + input_fn=lambda:creditcard_data.eval_input_fn(features=predict_x, + labels=None, + batch_size=args.batch_size)) + + template = ('\nPrediction is "{}" ({:.1f}%), expected "{}"') + for pred_dict, expec in zip(predictions, expected): + class_id = pred_dict['class_ids'][0] + probability = pred_dict['probabilities'][class_id] + print(template.format(["Not Fraud", "Fraud"][class_id], + 100 * probability, ["Not Fraud", "Fraud"][expec])) + + +if __name__ == '__main__': + tf.logging.set_verbosity(tf.logging.INFO) + tf.app.run(main) diff --git a/docs/sqlflow/example/jupyter/example.ipynb b/docs/sqlflow/example/jupyter/example.ipynb new file mode 100644 index 0000000..7dac56f --- /dev/null +++ b/docs/sqlflow/example/jupyter/example.ipynb @@ -0,0 +1,106 @@ +{ + "cells": [ + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "%reload_ext sqlflow.magic" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "%%sqlflow\n", + "show databases;" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "%%sqlflow\n", + "select *\n", + "from iris.train\n", + "limit 1" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "%%sqlflow\n", + "SELECT *\n", + "FROM iris.train\n", + "TRAIN DNNClassifier\n", + "WITH\n", + " n_classes = 3,\n", + " hidden_units = [10, 20]\n", + "COLUMN sepal_length, sepal_width, petal_length, petal_width\n", + "LABEL class\n", + "INTO sqlflow_models.my_dnn_model;" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "%%sqlflow\n", + "SELECT *\n", + "FROM iris.test\n", + "predict iris.predict.class\n", + "USING sqlflow_models.my_dnn_model;" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "%%sqlflow\n", + "SELECT *\n", + "FROM iris.predict\n", + "LIMIT 5;" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [] + } + ], + "metadata": { + "kernelspec": { + "display_name": "Python 3", + "language": "python", + "name": "python3" + }, + "language_info": { + "codemirror_mode": { + "name": "ipython", + "version": 3 + }, + "file_extension": ".py", + "mimetype": "text/x-python", + "name": "python", + "nbconvert_exporter": "python", + "pygments_lexer": "ipython3", + "version": "3.7.2" + } + }, + "nbformat": 4, + "nbformat_minor": 2 +} diff --git a/docs/sqlflow/server/README.md b/docs/sqlflow/server/README.md new file mode 100644 index 0000000..ce2dfc0 --- /dev/null +++ b/docs/sqlflow/server/README.md @@ -0,0 +1,14 @@ +# sqlflowserver +The gRPC proxy server of SQL engines + +## Prerequisite + +- Golang +- Docker +- GRPC + +## Running + +```bash +go generate && cd server && go run main.go +``` diff --git a/docs/sqlflow/sql/README.md b/docs/sqlflow/sql/README.md new file mode 100644 index 0000000..9771697 --- /dev/null +++ b/docs/sqlflow/sql/README.md @@ -0,0 +1,13 @@ +# SQLFlow + +## Build, Run, Test + +Before running the unit tests, we need to build and run a Docker container that +hosts a MySQL database following [this guide](../example/datasets/README.md). + +To build the parser using `goyacc` and run all unit tests, use the following +command: + +```bash +go get -d ./... && goyacc -p sql -o parser.go sql.y && go test -v +``` diff --git a/docs/sqlflow/sql/python/README.md b/docs/sqlflow/sql/python/README.md new file mode 100644 index 0000000..a5e59e6 --- /dev/null +++ b/docs/sqlflow/sql/python/README.md @@ -0,0 +1,13 @@ +# Python Code Template + +To run a test, say, `test_sql_data.py`, we need to + +1. Start a container that runs MySQL server with populated data following [this guide](https://github.com/sql-machine-learning/sqlflow/blob/develop/example/datasets/README.md). + +1. Run the tests in a SQLFlow container that has TensorFlow and `mysql_connector` installed: + + ```bash + docker run --rm -it --network="host" -v $PWD:/work -w /work sqlflow/sqlflow python test_sql_data.py + ``` + + where `--network="host"` allows processes running in the container to access the host's network, where the MySQL server container exposes its port. diff --git a/docs/sqlflow/sql/python/index.html b/docs/sqlflow/sql/python/index.html new file mode 100644 index 0000000..a67a902 --- /dev/null +++ b/docs/sqlflow/sql/python/index.html @@ -0,0 +1,305 @@ + + + + + + + + + Python Code Template - Sqlflow + + + + + + + + + + + + + + + + + +
    +
    + Sqlflow + +
    + + +
    +
    +

    This site uses Just the Docs, a documentation theme for Jekyll.

    +
    +
    +
    +
    +
    + +
    +
    + + Search +
    +
    +
    + + + + +
    +
    +
    + + + +
    +

    Python Code Template

    + +

    To run a test, say, test_sql_data.py, we need to

    + +
      +
    1. +

      Start a container that runs MySQL server with populated data following this guide.

      +
      2. +
    2. +

      Run the tests in a SQLFlow container that has TensorFlow and mysql_connector installed:

      + + 
      docker run --rm -it --network="host" -v $PWD:/work -w /work sqlflow/sqlflow python test_sql_data.py
+
      + +

      where --network="host" allows processes running in the container to access the host’s network, where the MySQL server container exposes its port.

      +
      3. +
    + + + + +
    +
    +
    +
    + + + diff --git a/docs/sqlflow/sql/python/sql_data.py b/docs/sqlflow/sql/python/sql_data.py new file mode 100644 index 0000000..49040ed --- /dev/null +++ b/docs/sqlflow/sql/python/sql_data.py @@ -0,0 +1,57 @@ +import tensorflow as tf +import mysql.connector + + +def connect(user, passwd, host, port): + """connect is a convenient shortcut to mysql.connector.connect. Also, + it makes it reaonable to import mysql.connector in this file, so + to make it self-complete as a template. + + """ + return mysql.connector.connect(user=user, + passwd=passwd, + host=host, + port=port) + + +def load(db, slct, label, features): + """load returns the training features and the labels. The features is +a dict from field names to data columns, and the label is a dict from +the label field name to the label data column. + + Args: + db: returned from mysql.connector.connect() + slct (str): SQL SELECT statement + label (str): the label field name as a string. + features (list of str or None): feature field names. + + Returns: + (features, label): maps from feature/label field names to columns. + + """ + cursor = db.cursor() + cursor.execute(slct) + f = [i[0] for i in cursor.description] # get field names. + c = list(zip(*cursor.fetchall())) # transpose rows into columns. + d = dict(zip(f, c)) # dict from field names to columns. + l = d.pop(label) + if features != None: + d = dict((k, d[k]) for k in features) + return d, l + + +def feature_columns(features): + """feature_columns returns a list of tf.feature_column. + + Note: currently, we are using a quick-and-hacky implementation + that assumes all fields are numeric_columns, and not supporting + feature derivations like tf.feature_column.cross_column. + + Args: + features: returned from load + + Returns: + list of tf.feature_column + + """ + return [tf.feature_column.numeric_column(key=k) for k in features.keys()] diff --git a/docs/sqlflow/sql/python/test_sql_data.py b/docs/sqlflow/sql/python/test_sql_data.py new file mode 100644 index 0000000..00ac8a1 --- /dev/null +++ b/docs/sqlflow/sql/python/test_sql_data.py @@ -0,0 +1,32 @@ +import sql_data +import unittest + + +class TestSQLData(unittest.TestCase): + def __init__(self, *args, **kwargs): + super(TestSQLData, self).__init__(*args, **kwargs) + self.db = sql_data.connect('root', 'root', 'localhost', 3306) + self.assertIsNotNone(self.db) + + def test_load(self): + f, l = sql_data.load( + self.db, 'SELECT * FROM iris.train LIMIT 3', 'class', None) + self.assertEqual(4, len(f.keys())) # 4 features + self.assertEqual(3, len(l)) # label column length + + def test_load_with_filter(self): + fs = ['sepal_length', 'petal_width'] + f, l = sql_data.load( + self.db, 'SELECT * FROM iris.train LIMIT 3', 'class', fs) + self.assertEqual(len(fs), len(f)) + self.assertEqual(3, len(l)) # label column length + + def test_feature_columns(self): + f, l = sql_data.load( + self.db, 'SELECT * FROM iris.train LIMIT 3', 'class', None) + c = sql_data.feature_columns(f) + self.assertEqual(4, len(c)) # 4 features + + +if __name__ == '__main__': + unittest.main() diff --git a/docs/sqlflow/sql/test b/docs/sqlflow/sql/test new file mode 100644 index 0000000..e69de29 diff --git a/docs/sqlflow/sqlfs/README.md b/docs/sqlflow/sqlfs/README.md new file mode 100644 index 0000000..01b7e3a --- /dev/null +++ b/docs/sqlflow/sqlfs/README.md @@ -0,0 +1,56 @@ +# `sqlfs` + +The package `sqlfs` provides an `io.ReadCloser` and an `io.WriteCloser` that treats a SQL database a filesystem, where each table in the database is like a file. The schema of the table is very simple -- it has only one column of BLOB type. All the rows consist the storage. + +`sqlfs` provides the following features. + +## Create a file + +To create a table named "hello" in a database "mydb" for writing, we can call `Create`. + +```go +f, e := sqlfs.Create(db, "mydb.hello") +f.Write([]byte("hello world!\n")) +f.Close() +``` + +where `db` comes from a call to `database/sql.Open`. + +## Append to a file + +```go +f, e := sqlfs.Append(db, "mydb.hello") +f.Write([]byte("hello world!\n") +f.Close() +``` + +## Read from a file + +```go +f, e := sqlfs.Open(db, "mydb.hello") +buf := make([]byte, 1024) +f.Read(buf) +f.Close() +``` + +## Remove a file + +```go +DropTable(db, "mydb.hello") +``` + +## Check if a file exists + +```go +HasTable(db, "mydb.hello") +``` + +## Other I/O operations + +Feel free to use standard packages `io`, `ioutil`, etc with `sqlfs`. For example, we can call `io.Copy` to copy everything from the standard input to a table. + +```go +f, e := sqlfs.Create(db, "mydb.hello") +io.Copy(f, os.Stdin) +f.Close() +``` From 4838ee8d57f230080bf814e3b2ef4c91b3e2aa9b Mon Sep 17 00:00:00 2001 From: mxz96102 Date: Tue, 16 Apr 2019 10:27:34 +0800 Subject: [PATCH 4/9] Add new style to doc and index --- _config.yml | 8 +- _layouts/doc.md | 13 ++ _layouts/index.html | 174 +++++++++++++++- docs/assets/js/search-data.json | 165 ++++++++++++++- docs/favicon.ico | Bin 0 -> 1150 bytes docs/feed.xml | 2 +- docs/index.html | 188 +++++++++++++++++- docs/pages/about.html | 6 +- docs/sqlflow/cmd/demo/index.html | 16 +- docs/sqlflow/doc/build-tensorflow.html | 62 +++--- docs/sqlflow/doc/build.html | 62 +++--- .../doc/close_producer_from_consumer.html | 62 +++--- .../doc/database_abstraction_layer.html | 62 +++--- docs/sqlflow/doc/mysql-setup.html | 62 +++--- docs/sqlflow/doc/pipe.html | 62 +++--- docs/sqlflow/doc/quickstart.html | 62 +++--- docs/sqlflow/doc/sql_parser.html | 62 +++--- docs/sqlflow/doc/submitter.html | 62 +++--- docs/sqlflow/doc/submitter.md | 2 +- docs/sqlflow/doc/syntax.html | 62 +++--- docs/sqlflow/doc/walkthrough.html | 62 +++--- docs/sqlflow/example/creditcard/index.html | 64 +++--- docs/sqlflow/example/datasets/index.html | 62 +++--- .../example/fraud_detection/index.html | 62 +++--- docs/sqlflow/index.html | 16 +- docs/sqlflow/server/index.html | 16 +- docs/sqlflow/sql/index.html | 16 +- docs/sqlflow/sql/python/index.html | 62 +++--- docs/sqlflow/sqlfs/index.html | 16 +- favicon.ico | Bin 0 -> 1150 bytes 30 files changed, 1112 insertions(+), 458 deletions(-) create mode 100644 _layouts/doc.md create mode 100644 docs/favicon.ico create mode 100644 favicon.ico diff --git a/_config.yml b/_config.yml index a7ef609..96ea267 100644 --- a/_config.yml +++ b/_config.yml @@ -4,7 +4,7 @@ classes: wide search_enabled: true -title: 'Sqlflow' +title: 'SQLFLOW' destination: 'docs' @@ -18,6 +18,12 @@ exclude: ['*.sql', '*.go', '.gitignore', '*.sh', '*.proto', '*.y'] # baseurl: "/sql-machine-learning.github.io/" +defaults: + - scope: + path: "sqlflow" + values: + layout: "doc" + footer: links: - label: "GitHub" diff --git a/_layouts/doc.md b/_layouts/doc.md new file mode 100644 index 0000000..7ad6ea5 --- /dev/null +++ b/_layouts/doc.md @@ -0,0 +1,13 @@ +--- +layout: default +--- + + + +{{content}} + + diff --git a/_layouts/index.html b/_layouts/index.html index e2b096f..e6b2e93 100644 --- a/_layouts/index.html +++ b/_layouts/index.html @@ -1,8 +1,8 @@ - + {{page.title}} + +

    SQLFlow Demo

    Before you start, you need to set up test MySQL server as described here

    @@ -354,6 +358,10 @@

    Predict

    +
    + Sqlflow + +
    + + +
    +
    +

    This site uses Just the Docs, a documentation theme for Jekyll.

    +
    +
    +
    +
    +
    + +
    +
    + + Search +
    +
    +
    + + + + +
    +
    +
    + + + +
    +

    Run MySQL Server and Client in Docker Containers

    + +

    The document explains how to setup MySQL in our development environment.

    + +

    Run MySQL Server in a Docker Container

    + +
    docker run --rm \
+   -v /tmp/test1:/var/lib/mysql \
+   --name mysql01 \
+   -e MYSQL_ROOT_PASSWORD=root \
+   -e MYSQL_ROOT_HOST='%' \
+   -p 3306:3306 \
+   -d \
+   mysql/mysql-server:8.0
+
    + +
      +
    • +

      the -v option ensures that the database is saved on the host. The default directory where MySQL saves the database is /var/lib/mysql. This directory can be configured in /etc/mysql/my.cnf, as explained in this post. By overlaying the directory /tmp/test1 on the host to /var/lib/mysql, we “cheat” MySQL to save databases on the host. So, we can kill the container and restart it, and the database is still there.

      + +

      Please be aware that the directory on the host must be empty the first time we run the above command; otherwise, MySQL would fail to initialize. I figured out this problem after several failures using docker logs.

      +
      • +
    • +

      the -e option sets the root password of MySQL to “root”. Feel free to set it to any password you like.

      +
      • +
    • +

      the second -e options sets MYSQL_ROOT_HOST to a wildcard so to allow clients connecting to the server via TCP/IP as the user “root”. This trick works with MySQL 5.7 and 8.0, but not the most recent under-development version.

      +
      • +
    • +

      the --name option names the container to mysql01, which can be used to refer to this container.

      +
      • +
    • +

      the -p option maps the port 3306, on which the MySQL server listens, to the same port on the host, so that clients could connect to the server via TCP/IP.

      +
      • +
    + +

    Run MySQL Client in the Server Container

    + +
    docker exec -it mysql01 mysql -uroot -p
+
    + +

    This command executes the command mysql, which is the command line tool of MySQL, in the container named mysql01.

    + +
      +
    • +

      The command line flags of mysql include -u, which specifies the username of MySQL, and -p, which makes MySQL prompts for the password. For this example, we should type the password “root”, which was set in the previous command.

      +
      • +
    • +

      Please wait for a few seconds after the starting of the MySQL server container before we execute the client; otherwise, the startup of the client might fail.

      +
      • +
    • +

      Once we get into the MySQL client, we can type SQL commands, e.g.,

      + + 
      show databases;
+create database yi;
+
      +
      • +
    + +

    Run Client in a Different Container on the Same Host

    + +
    docker run --rm -it \
+   -v /tmp/test1:/var/lib/mysql \
+   mysql/mysql-server:8.0 \
+   mysql -uroot -p
+
    + +
      +
    • The -v option maps the database directory on the host to the client container. This mapping is necessary because, by default, the client talks to the server via Unix socket /var/lib/mysql/mysql.sock, which is /tmp/test1/mysql.sock on the host.
      • +
    + +

    Run Client in a Container on a Remote Host

    + +
    docker run \
+   --rm -it \
+   mysql/mysql-server:8.0 \
+   mysql -h 192.168.1.3 -P 3306 -uroot -p
+
    + +
      +
    • the -h option tells the client where the server is running on. In this example, the given IP is the one of the host where I ran the MySQL server container.
      • +
    + +

    Please be aware that the above command works only if the server allows remote connections.

    + +

    Run Python Client in a Container

    + +

    To write a Python client, we need to install the Python package mysql-connector-python.

    + +
    FROM python:2.7
+RUN pip install mysql-connector-python
+
    + +

    Please be aware that some documents says that we need to install mysql-connector. I tried; but the mysql.connector.connect call failed with the error mysql.connector.errors.NotSupportedError: Authentication plugin 'caching_sha2_password' is not supported.

    + +

    Build the Docker image:

    + +
    docker build -t sqlflow .
+
    + +

    Run the image:

    + +
    docker run --rm -it sqlflow bash
+
    + +

    and we can start Python and run the following Python code snippet

    + +
    >>> import mysql.connector
+>>> db = mysql.connector.connect(user='root', passwd='root', host='192.168.1.3')
+>>> print(db)
+<mysql.connector.connection_cext.CMySQLConnection object at 0x7fbab9f3fed0>
+
    + +

    Run a Go Client

    + +

    In order to connect to a database, you need to import the database’s driver first.

    + +
    export GOPATH=$HOME/go
+go get -u github.com/go-sql-driver/mysql
+
    + +

    go run the following file

    + +
    package main
+
+import (
+	"database/sql"
+	"github.com/go-sql-driver/mysql"
+	"log"
+)
+
+func main() {
+	testConfig := &mysql.Config{
+		User:   "root",
+		Passwd: "root",
+		Net:    "tcp",
+		Addr:   "localhost:3306",
+	}
+	db, e := sql.Open("mysql", testConfig.FormatDSN())
+	if e != nil {
+		log.Fatal(e)
+	}
+	defer db.Close()
+}
+
    + + + + +
    +
    +
    +

    -
    - Sqlflow - -
    - - -
    -
    -

    This site uses Just the Docs, a documentation theme for Jekyll.

    -
    -
    -
    -
    -
    - -
    -
    - - Search -
    -
    -
    - - - - -
    -
    -
    - - - -
    -

    Run MySQL Server and Client in Docker Containers

    - -

    The document explains how to setup MySQL in our development environment.

    - -

    Run MySQL Server in a Docker Container

    - -
    docker run --rm \
-   -v /tmp/test1:/var/lib/mysql \
-   --name mysql01 \
-   -e MYSQL_ROOT_PASSWORD=root \
-   -e MYSQL_ROOT_HOST='%' \
-   -p 3306:3306 \
-   -d \
-   mysql/mysql-server:8.0
-
    - -
      -
    • -

      the -v option ensures that the database is saved on the host. The default directory where MySQL saves the database is /var/lib/mysql. This directory can be configured in /etc/mysql/my.cnf, as explained in this post. By overlaying the directory /tmp/test1 on the host to /var/lib/mysql, we “cheat” MySQL to save databases on the host. So, we can kill the container and restart it, and the database is still there.

      - -

      Please be aware that the directory on the host must be empty the first time we run the above command; otherwise, MySQL would fail to initialize. I figured out this problem after several failures using docker logs.

      -
      • -
    • -

      the -e option sets the root password of MySQL to “root”. Feel free to set it to any password you like.

      -
      • -
    • -

      the second -e options sets MYSQL_ROOT_HOST to a wildcard so to allow clients connecting to the server via TCP/IP as the user “root”. This trick works with MySQL 5.7 and 8.0, but not the most recent under-development version.

      -
      • -
    • -

      the --name option names the container to mysql01, which can be used to refer to this container.

      -
      • -
    • -

      the -p option maps the port 3306, on which the MySQL server listens, to the same port on the host, so that clients could connect to the server via TCP/IP.

      -
      • -
    - -

    Run MySQL Client in the Server Container

    - -
    docker exec -it mysql01 mysql -uroot -p
-
    - -

    This command executes the command mysql, which is the command line tool of MySQL, in the container named mysql01.

    - -
      -
    • -

      The command line flags of mysql include -u, which specifies the username of MySQL, and -p, which makes MySQL prompts for the password. For this example, we should type the password “root”, which was set in the previous command.

      -
      • -
    • -

      Please wait for a few seconds after the starting of the MySQL server container before we execute the client; otherwise, the startup of the client might fail.

      -
      • -
    • -

      Once we get into the MySQL client, we can type SQL commands, e.g.,

      - - 
      show databases;
-create database yi;
-
      -
      • -
    - -

    Run Client in a Different Container on the Same Host

    - -
    docker run --rm -it \
-   -v /tmp/test1:/var/lib/mysql \
-   mysql/mysql-server:8.0 \
-   mysql -uroot -p
-
    - -
      -
    • The -v option maps the database directory on the host to the client container. This mapping is necessary because, by default, the client talks to the server via Unix socket /var/lib/mysql/mysql.sock, which is /tmp/test1/mysql.sock on the host.
      • -
    - -

    Run Client in a Container on a Remote Host

    - -
    docker run \
-   --rm -it \
-   mysql/mysql-server:8.0 \
-   mysql -h 192.168.1.3 -P 3306 -uroot -p
-
    - -
      -
    • the -h option tells the client where the server is running on. In this example, the given IP is the one of the host where I ran the MySQL server container.
      • -
    - -

    Please be aware that the above command works only if the server allows remote connections.

    - -

    Run Python Client in a Container

    - -

    To write a Python client, we need to install the Python package mysql-connector-python.

    - -
    FROM python:2.7
-RUN pip install mysql-connector-python
-
    - -

    Please be aware that some documents says that we need to install mysql-connector. I tried; but the mysql.connector.connect call failed with the error mysql.connector.errors.NotSupportedError: Authentication plugin 'caching_sha2_password' is not supported.

    - -

    Build the Docker image:

    - -
    docker build -t sqlflow .
-
    - -

    Run the image:

    - -
    docker run --rm -it sqlflow bash
-
    - -

    and we can start Python and run the following Python code snippet

    - -
    >>> import mysql.connector
->>> db = mysql.connector.connect(user='root', passwd='root', host='192.168.1.3')
->>> print(db)
-<mysql.connector.connection_cext.CMySQLConnection object at 0x7fbab9f3fed0>
-
    - -

    Run a Go Client

    - -

    In order to connect to a database, you need to import the database’s driver first.

    - -
    export GOPATH=$HOME/go
-go get -u github.com/go-sql-driver/mysql
-
    - -

    go run the following file

    - -
    package main
-
-import (
-	"database/sql"
-	"github.com/go-sql-driver/mysql"
-	"log"
-)
-
-func main() {
-	testConfig := &mysql.Config{
-		User:   "root",
-		Passwd: "root",
-		Net:    "tcp",
-		Addr:   "localhost:3306",
-	}
-	db, e := sql.Open("mysql", testConfig.FormatDSN())
-	if e != nil {
-		log.Fatal(e)
-	}
-	defer db.Close()
-}
-
    - - - - -
    -
    -
    -

    + + diff --git a/docs/sqlflow/doc/build-tensorflow.html b/docs/sqlflow/doc/build-tensorflow.html index 9632551..6923538 100644 --- a/docs/sqlflow/doc/build-tensorflow.html +++ b/docs/sqlflow/doc/build-tensorflow.html @@ -6,8 +6,8 @@ - Build TensorFlow from Source Code using Docker - Sqlflow - + Build TensorFlow from Source Code using Docker - SQLFLOW + + - + @@ -30,7 +30,7 @@
    - Sqlflow + SQLFLOW
    @@ -273,9 +327,22 @@
    + +
    - @@ -348,9 +415,7 @@

    The Choice

    about operator association and precedence.

    - + diff --git a/docs/sqlflow/doc/submitter.html b/docs/sqlflow/doc/submitter.html index e87280e..55e080a 100644 --- a/docs/sqlflow/doc/submitter.html +++ b/docs/sqlflow/doc/submitter.html @@ -38,182 +38,72 @@ - -
  • - - sqlfs - -
    • - - - -
  • - - Build TensorFlow from Source Code using Docker - -
    • - - - -
  • - - Canonical Development Environment - -
    • - - - -
  • - - Closing the producer goroutine from the consumer - -
    • - - -
  • - - Compatibility with Various SQL Engines - -
    • - -
  • - - Run MySQL Server and Client in Docker Containers - -
    • - -
  • - - Piping Responses - -
    • - -
  • - - Quick start - -
    • - -
  • - - Extended SQL Parser Design - -
    • - -
  • - - Submitter - -
    • - -
  • - - SQLFlow: Design Doc - -
    • - -
  • - - SQLFlow: Code Walkthrough - -
    • - -
  • - - SQLFlow - -
    • - -
  • - - SQLFlow Demo - -
    • - - -
  • - - The Credit Card Fraud Detection Example - -
    • - + - -
  • - - The MySQL Server Container for Testing - -
    • - + - -
  • - - Credit Card Fraud Detection - -
    • - +
  • - sqlflowserver + Quick start
    • @@ -223,7 +113,90 @@
  • - SQLFlow + Design + + +
    • @@ -231,16 +204,97 @@ -
  • +
  • - Python Code Template + Contribute + + +
    • - - @@ -273,9 +327,22 @@
    + +
    - @@ -370,9 +437,7 @@

    Register a submitter

    where sql.Register will put my_submitter instance to package level registry. During sql.Run, it will check whether there is a submitter registered. If there is, sql.Run will run either submitter.Train or submitter.Predict.

    - + diff --git a/docs/sqlflow/doc/syntax.html b/docs/sqlflow/doc/syntax.html index 2342dbc..2a0e8cf 100644 --- a/docs/sqlflow/doc/syntax.html +++ b/docs/sqlflow/doc/syntax.html @@ -38,182 +38,72 @@ - -
  • - - sqlfs - -
    • - - - -
  • - - Build TensorFlow from Source Code using Docker - -
    • - - - -
  • - - Canonical Development Environment - -
    • - - - -
  • - - Closing the producer goroutine from the consumer - -
    • - - -
  • - - Compatibility with Various SQL Engines - -
    • - -
  • - - Run MySQL Server and Client in Docker Containers - -
    • - -
  • - - Piping Responses - -
    • - -
  • - - Quick start - -
    • - -
  • - - Extended SQL Parser Design - -
    • - -
  • - - Submitter - -
    • - -
  • - - SQLFlow: Design Doc - -
    • - -
  • - - SQLFlow: Code Walkthrough - -
    • - -
  • - - SQLFlow - -
    • - -
  • - - SQLFlow Demo - -
    • - - -
  • - - The Credit Card Fraud Detection Example - -
    • - + - -
  • - - The MySQL Server Container for Testing - -
    • - + - -
  • - - Credit Card Fraud Detection - -
    • - +
  • - sqlflowserver + Quick start
    • @@ -221,9 +111,92 @@ -
  • +
  • + + Design - SQLFlow + +
    • @@ -233,14 +206,95 @@
  • - Python Code Template + Contribute + + +
    • - - @@ -273,9 +327,22 @@
    + +
    - @@ -441,9 +508,7 @@

    Working with Jupyter Noteb

    - + diff --git a/docs/sqlflow/doc/walkthrough.html b/docs/sqlflow/doc/walkthrough.html index 57faefe..8802463 100644 --- a/docs/sqlflow/doc/walkthrough.html +++ b/docs/sqlflow/doc/walkthrough.html @@ -38,182 +38,72 @@ - -
  • - - sqlfs - -
    • - - - -
  • - - Build TensorFlow from Source Code using Docker - -
    • - - - -
  • - - Canonical Development Environment - -
    • - - - -
  • - - Closing the producer goroutine from the consumer - -
    • - - -
  • - - Compatibility with Various SQL Engines - -
    • - -
  • - - Run MySQL Server and Client in Docker Containers - -
    • - -
  • - - Piping Responses - -
    • - -
  • - - Quick start - -
    • - -
  • - - Extended SQL Parser Design - -
    • - -
  • - - Submitter - -
    • - -
  • - - SQLFlow: Design Doc - -
    • - -
  • - - SQLFlow: Code Walkthrough - -
    • - -
  • - - SQLFlow - -
    • - -
  • - - SQLFlow Demo - -
    • - - -
  • - - The Credit Card Fraud Detection Example - -
    • - + - -
  • - - The MySQL Server Container for Testing - -
    • - + - -
  • - - Credit Card Fraud Detection - -
    • - +
  • - sqlflowserver + Quick start
    • @@ -221,9 +111,92 @@ -
  • +
  • + + Design - SQLFlow + +
    • @@ -233,14 +206,95 @@
  • - Python Code Template + Contribute + + +
    • - - @@ -273,9 +327,22 @@
    + +
    - @@ -355,9 +422,7 @@

    Job Management

    A solution is to change the gRPC interface of SQLFlow server to have a method that files a job and returns immediately, and another method to get a batch of recent messages given a job ID. We will make a design for that soon.

    - + diff --git a/docs/sqlflow/example/creditcard/index.html b/docs/sqlflow/example/creditcard/index.html index 4efeac5..b37e82d 100644 --- a/docs/sqlflow/example/creditcard/index.html +++ b/docs/sqlflow/example/creditcard/index.html @@ -38,182 +38,72 @@ - -
  • - - sqlfs - -
    • - - - -
  • - - Build TensorFlow from Source Code using Docker - -
    • - - - -
  • - - Canonical Development Environment - -
    • - - - -
  • - - Closing the producer goroutine from the consumer - -
    • - - -
  • - - Compatibility with Various SQL Engines - -
    • - -
  • - - Run MySQL Server and Client in Docker Containers - -
    • - -
  • - - Piping Responses - -
    • - -
  • - - Quick start - -
    • - -
  • - - Extended SQL Parser Design - -
    • - -
  • - - Submitter - -
    • - -
  • - - SQLFlow: Design Doc - -
    • - -
  • - - SQLFlow: Code Walkthrough - -
    • - -
  • - - SQLFlow - -
    • - -
  • - - SQLFlow Demo - -
    • - - -
  • - - The Credit Card Fraud Detection Example - -
    • - + - -
  • - - The MySQL Server Container for Testing - -
    • - + - -
  • - - Credit Card Fraud Detection - -
    • - +
  • - sqlflowserver + Quick start
    • @@ -223,7 +113,90 @@
  • - SQLFlow + Design + + +
    • @@ -233,14 +206,95 @@
  • - Python Code Template + Contribute + + +
    • - - @@ -275,7 +329,11 @@
    - @@ -420,9 +478,7 @@

    Run MySQL Server

    - + diff --git a/docs/sqlflow/example/datasets/index.html b/docs/sqlflow/example/datasets/index.html index 5682bd8..c0ff7d8 100644 --- a/docs/sqlflow/example/datasets/index.html +++ b/docs/sqlflow/example/datasets/index.html @@ -38,182 +38,72 @@ - -
  • - - sqlfs - -
    • - - - -
  • - - Build TensorFlow from Source Code using Docker - -
    • - - - -
  • - - Canonical Development Environment - -
    • - - - -
  • - - Closing the producer goroutine from the consumer - -
    • - - -
  • - - Compatibility with Various SQL Engines - -
    • - -
  • - - Run MySQL Server and Client in Docker Containers - -
    • - -
  • - - Piping Responses - -
    • - -
  • - - Quick start - -
    • - -
  • - - Extended SQL Parser Design - -
    • - -
  • - - Submitter - -
    • - -
  • - - SQLFlow: Design Doc - -
    • - -
  • - - SQLFlow: Code Walkthrough - -
    • - -
  • - - SQLFlow - -
    • - -
  • - - SQLFlow Demo - -
    • - - -
  • - - The Credit Card Fraud Detection Example - -
    • - + - -
  • - - The MySQL Server Container for Testing - -
    • - + - -
  • - - Credit Card Fraud Detection - -
    • - +
  • - sqlflowserver + Quick start
    • @@ -223,7 +113,90 @@
  • - SQLFlow + Design + + +
    • @@ -233,14 +206,95 @@
  • - Python Code Template + Contribute + + +
    • - - @@ -275,7 +329,11 @@
    - @@ -351,9 +409,7 @@

    Trouble shooting:

    - + diff --git a/docs/sqlflow/example/fraud_detection/index.html b/docs/sqlflow/example/fraud_detection/index.html index 8d1dfa1..7a69db9 100644 --- a/docs/sqlflow/example/fraud_detection/index.html +++ b/docs/sqlflow/example/fraud_detection/index.html @@ -38,182 +38,72 @@ - -
  • - - sqlfs - -
    • - - - -
  • - - Build TensorFlow from Source Code using Docker - -
    • - - - -
  • - - Canonical Development Environment - -
    • - - - -
  • - - Closing the producer goroutine from the consumer - -
    • - - -
  • - - Compatibility with Various SQL Engines - -
    • - -
  • - - Run MySQL Server and Client in Docker Containers - -
    • - -
  • - - Piping Responses - -
    • - -
  • - - Quick start - -
    • - -
  • - - Extended SQL Parser Design - -
    • - -
  • - - Submitter - -
    • - -
  • - - SQLFlow: Design Doc - -
    • - -
  • - - SQLFlow: Code Walkthrough - -
    • - -
  • - - SQLFlow - -
    • - -
  • - - SQLFlow Demo - -
    • - - -
  • - - The Credit Card Fraud Detection Example - -
    • - + - -
  • - - The MySQL Server Container for Testing - -
    • - + - -
  • - - Credit Card Fraud Detection - -
    • - +
  • - sqlflowserver + Quick start
    • @@ -223,7 +113,90 @@
  • - SQLFlow + Design + + +
    • @@ -233,14 +206,95 @@
  • - Python Code Template + Contribute + + +
    • - - @@ -275,7 +329,11 @@
    - @@ -381,9 +439,7 @@

    Credit Card Fraud Detection

    - + diff --git a/docs/sqlflow/index.html b/docs/sqlflow/index.html index 46d6b8a..3a65bb8 100644 --- a/docs/sqlflow/index.html +++ b/docs/sqlflow/index.html @@ -38,182 +38,72 @@ - -
  • - - sqlfs - -
    • - - - -
  • - - Build TensorFlow from Source Code using Docker - -
    • - - - -
  • - - Canonical Development Environment - -
    • - - - -
  • - - Closing the producer goroutine from the consumer - -
    • - - -
  • - - Compatibility with Various SQL Engines - -
    • - -
  • - - Run MySQL Server and Client in Docker Containers - -
    • - -
  • - - Piping Responses - -
    • - -
  • - - Quick start - -
    • - -
  • - - Extended SQL Parser Design - -
    • - -
  • - - Submitter - -
    • - -
  • - - SQLFlow: Design Doc - -
    • - -
  • - - SQLFlow: Code Walkthrough - -
    • - -
  • - - SQLFlow - -
    • - -
  • - - SQLFlow Demo - -
    • - - -
  • - - The Credit Card Fraud Detection Example - -
    • - + - -
  • - - The MySQL Server Container for Testing - -
    • - + - -
  • - - Credit Card Fraud Detection - -
    • - +
  • - sqlflowserver + Quick start
    • @@ -223,7 +113,90 @@
  • - SQLFlow + Design + + +
    • @@ -233,14 +206,95 @@
  • - Python Code Template + Contribute + + +
    • - - @@ -275,7 +329,11 @@
    - @@ -309,9 +367,7 @@

    License

    Apache License 2.0

    - + diff --git a/docs/sqlflow/server/index.html b/docs/sqlflow/server/index.html index 8e6bdb1..b56f2d3 100644 --- a/docs/sqlflow/server/index.html +++ b/docs/sqlflow/server/index.html @@ -38,162 +38,72 @@ - -
  • - - sqlfs - -
    • - + - -
  • - - Build TensorFlow from Source Code using Docker - -
    • - + -
  • - - Canonical Development Environment - -
    • - -
  • - - Closing the producer goroutine from the consumer - -
    • - -
  • - - Compatibility with Various SQL Engines - -
    • - -
  • - - Run MySQL Server and Client in Docker Containers - -
    • - -
  • - - Piping Responses - -
    • - -
  • - - Quick start - -
    • - -
  • - - Extended SQL Parser Design - -
    • - -
  • - - Submitter - -
    • - -
  • - - SQLFlow: Design Doc - -
    • - -
  • - - SQLFlow: Code Walkthrough - -
    • - - -
  • - - SQLFlow - -
    • - + - -
  • - - SQLFlow Demo - -
    • - + - -
  • - - The Credit Card Fraud Detection Example - -
    • - +
  • - The MySQL Server Container for Testing + Quick start
    • @@ -203,17 +113,90 @@
  • - Credit Card Fraud Detection - -
    • - - - - - -
  • + Design - sqlflowserver + +
    • @@ -223,24 +206,95 @@
  • - SQLFlow - -
    • - - - - - -
  • + Contribute - Python Code Template + +
    • - - @@ -275,7 +329,11 @@
    - @@ -296,9 +354,7 @@

    Running

    - + diff --git a/docs/sqlflow/sql/index.html b/docs/sqlflow/sql/index.html index 6aa1b80..11a9bbd 100644 --- a/docs/sqlflow/sql/index.html +++ b/docs/sqlflow/sql/index.html @@ -38,172 +38,72 @@ - -
  • - - sqlfs - -
    • - - - -
  • - - Build TensorFlow from Source Code using Docker - -
    • - - - -
  • - - Canonical Development Environment - -
    • - + -
  • - - Closing the producer goroutine from the consumer - -
    • - -
  • - - Compatibility with Various SQL Engines - -
    • - -
  • - - Run MySQL Server and Client in Docker Containers - -
    • - -
  • - - Piping Responses - -
    • - -
  • - - Quick start - -
    • - -
  • - - Extended SQL Parser Design - -
    • - -
  • - - Submitter - -
    • - -
  • - - SQLFlow: Design Doc - -
    • - -
  • - - SQLFlow: Code Walkthrough - -
    • - -
  • - - SQLFlow - -
    • - - -
  • - - SQLFlow Demo - -
    • - + - -
  • - - The Credit Card Fraud Detection Example - -
    • - + - -
  • - - The MySQL Server Container for Testing - -
    • - +
  • - Credit Card Fraud Detection + Quick start
    • @@ -213,17 +113,90 @@
  • - sqlflowserver - -
    • - - - - - -
  • + Design - SQLFlow + +
    • @@ -233,14 +206,95 @@
  • - Python Code Template + Contribute + + +
    • - - @@ -275,7 +329,11 @@
    - @@ -284,7 +342,7 @@

    SQLFlow

    Build, Run, Test

    Before running the unit tests, we need to build and run a Docker container that -hosts a MySQL database following this guide.

    +hosts a MySQL database following this guide.

    To build the parser using goyacc and run all unit tests, use the following command:

    @@ -293,9 +351,7 @@

    Build, Run, Test

    - + diff --git a/docs/sqlflow/sql/python/index.html b/docs/sqlflow/sql/python/index.html index 42f4213..3ae31d8 100644 --- a/docs/sqlflow/sql/python/index.html +++ b/docs/sqlflow/sql/python/index.html @@ -38,172 +38,72 @@ - -
  • - - sqlfs - -
    • - - - -
  • - - Build TensorFlow from Source Code using Docker - -
    • - - - -
  • - - Canonical Development Environment - -
    • - + -
  • - - Closing the producer goroutine from the consumer - -
    • - -
  • - - Compatibility with Various SQL Engines - -
    • - -
  • - - Run MySQL Server and Client in Docker Containers - -
    • - -
  • - - Piping Responses - -
    • - -
  • - - Quick start - -
    • - -
  • - - Extended SQL Parser Design - -
    • - -
  • - - Submitter - -
    • - -
  • - - SQLFlow: Design Doc - -
    • - -
  • - - SQLFlow: Code Walkthrough - -
    • - -
  • - - SQLFlow - -
    • - - -
  • - - SQLFlow Demo - -
    • - + - -
  • - - The Credit Card Fraud Detection Example - -
    • - + - -
  • - - The MySQL Server Container for Testing - -
    • - +
  • - Credit Card Fraud Detection + Quick start
    • @@ -213,7 +113,90 @@
  • - sqlflowserver + Design + + +
    • @@ -223,24 +206,95 @@
  • - SQLFlow - -
    • - - - - - -
  • + Contribute - Python Code Template + +
    • - - @@ -275,7 +329,11 @@
    - @@ -285,7 +343,7 @@

    Python Code Template

    1. -

      Start a container that runs MySQL server with populated data following this guide.

      +

      Start a container that runs MySQL server with populated data following this guide.

    2. Run the tests in a SQLFlow container that has TensorFlow and mysql_connector installed:

      @@ -298,9 +356,7 @@

      Python Code Template

    - + diff --git a/docs/sqlflow/sqlfs/index.html b/docs/sqlflow/sqlfs/index.html index 2603a48..5310fea 100644 --- a/docs/sqlflow/sqlfs/index.html +++ b/docs/sqlflow/sqlfs/index.html @@ -38,182 +38,72 @@ - -
  • - - sqlfs - -
    • - - - -
  • - - Build TensorFlow from Source Code using Docker - -
    • - - - -
  • - - Canonical Development Environment - -
    • - - - -
  • - - Closing the producer goroutine from the consumer - -
    • - - -
  • - - Compatibility with Various SQL Engines - -
    • - -
  • - - Run MySQL Server and Client in Docker Containers - -
    • - -
  • - - Piping Responses - -
    • - -
  • - - Quick start - -
    • - -
  • - - Extended SQL Parser Design - -
    • - -
  • - - Submitter - -
    • - -
  • - - SQLFlow: Design Doc - -
    • - -
  • - - SQLFlow: Code Walkthrough - -
    • - -
  • - - SQLFlow - -
    • - -
  • - - SQLFlow Demo - -
    • - - -
  • - - The Credit Card Fraud Detection Example - -
    • - + - -
  • - - The MySQL Server Container for Testing - -
    • - + - -
  • - - Credit Card Fraud Detection - -
    • - +
  • - sqlflowserver + Quick start
    • @@ -223,7 +113,90 @@
  • - SQLFlow + Design + + +
    • @@ -233,14 +206,95 @@
  • - Python Code Template + Contribute + + +
    • - - @@ -275,7 +329,11 @@
    - @@ -331,9 +389,7 @@

    Other I/O operations

    - + From 35127beb5e775144cb2826cd6b20af9f2dc2592e Mon Sep 17 00:00:00 2001 From: mxz96102 Date: Thu, 18 Apr 2019 16:40:45 +0800 Subject: [PATCH 6/9] fix doc path --- _layouts/doc.md | 2 +- docs/feed.xml | 2 +- docs/sitemap.xml | 16 ++++++++++++++++ docs/sqlflow/doc/submitter.html | 2 +- docs/sqlflow/doc/submitter.md | 3 +-- docs/sqlflow/doc/syntax.html | 2 +- 6 files changed, 21 insertions(+), 6 deletions(-) diff --git a/_layouts/doc.md b/_layouts/doc.md index da2a2a8..cf4587c 100644 --- a/_layouts/doc.md +++ b/_layouts/doc.md @@ -10,6 +10,6 @@ layout: default a {color: #1BA2FF} -{{content | replace: "/example/" , "/sqlflow/example/"}} +{{content | replace: "/example/" , "/sqlflow/example/"| replace: "doc/figures/" , "/sqlflow/doc/figures/"}} diff --git a/docs/feed.xml b/docs/feed.xml index 0e9bcb0..bb5a4c3 100644 --- a/docs/feed.xml +++ b/docs/feed.xml @@ -1 +1 @@ -Jekyll2019-04-18T15:09:15+08:00/feed.xmlSQLFLOW \ No newline at end of file +Jekyll2019-04-18T16:37:07+08:00/feed.xmlSQLFLOW \ No newline at end of file diff --git a/docs/sitemap.xml b/docs/sitemap.xml index 8cfd28e..38c865d 100644 --- a/docs/sitemap.xml +++ b/docs/sitemap.xml @@ -73,6 +73,18 @@ /sqlflow/sqlfs/ +/gohive/git.apache.org/thrift.git/lib/go/thrift/ +2019-04-18T15:54:36+08:00 + + +/gohive/git.apache.org/thrift/lib/go/thrift/ +2019-04-18T15:54:36+08:00 + + +/gohive/hiveserver2/ +2019-04-18T15:54:36+08:00 + + /gohive/ 2019-04-12T10:17:18+08:00 @@ -81,6 +93,10 @@ 2019-04-12T10:17:18+08:00 +/gohive/thrift/ +2019-04-18T15:54:36+08:00 + + /gohive/vendor/git.apache.org/thrift.git/lib/go/thrift/ 2019-04-12T10:17:18+08:00 diff --git a/docs/sqlflow/doc/submitter.html b/docs/sqlflow/doc/submitter.html index 55e080a..b5c78d3 100644 --- a/docs/sqlflow/doc/submitter.html +++ b/docs/sqlflow/doc/submitter.html @@ -354,7 +354,7 @@

    Workflow

    When a user types in an extended SQL statement, SQLFlow first parses and semantically verifies the statement. Then SQLFlow either runs the ML job locally or submits the ML job to a third party computation service.

    -

    +

    In the latter case, SQLFlow produces a job description (TrainDescription or PredictDescription) and hands it over to the submitter. For a training SQL, SQLFlow produces TrainDescription; for prediction SQL, SQLFlow produces PredDescription. The concrete definition of the description looks like the following

    diff --git a/docs/sqlflow/doc/submitter.md b/docs/sqlflow/doc/submitter.md index b030418..db67a28 100644 --- a/docs/sqlflow/doc/submitter.md +++ b/docs/sqlflow/doc/submitter.md @@ -11,6 +11,7 @@ When a user types in an extended SQL statement, SQLFlow first parses and semanti In the latter case, SQLFlow produces a job description (`TrainDescription` or `PredictDescription`) and hands it over to the submitter. For a training SQL, SQLFlow produces `TrainDescription`; for prediction SQL, SQLFlow produces `PredDescription`. The concrete definition of the description looks like the following {% raw %} + ```go type ColumnType struct { Name string // e.g. sepal_length @@ -45,9 +46,7 @@ type PredDescription struct { ModelName string // e.g. my_dnn_model } ``` - {% endraw %} - ## Submitter Interface The submitter interface should provide two functions `Train` and `Predict`. The detailed definition can be the following diff --git a/docs/sqlflow/doc/syntax.html b/docs/sqlflow/doc/syntax.html index 2a0e8cf..10b282b 100644 --- a/docs/sqlflow/doc/syntax.html +++ b/docs/sqlflow/doc/syntax.html @@ -505,7 +505,7 @@

    Working with Jupyter Noteb

    The following figures shows the system components and their runtime environment. The left part shows how to run the system on a PC/laptop, the right part shows how to run it on a Kubernetes cluster.

    -

    +

    From 74b0c7c9f0c39646050288410aaaea6459d96e16 Mon Sep 17 00:00:00 2001 From: mxz96102 Date: Thu, 18 Apr 2019 16:44:54 +0800 Subject: [PATCH 7/9] add go hive --- .../thrift.git/lib/go/thrift/index.html | 13 +++++++++++++ .../git.apache.org/thrift/lib/go/thrift/index.html | 13 +++++++++++++ docs/gohive/hiveserver2/index.html | 13 +++++++++++++ docs/gohive/thrift/index.html | 13 +++++++++++++ 4 files changed, 52 insertions(+) create mode 100644 docs/gohive/git.apache.org/thrift.git/lib/go/thrift/index.html create mode 100644 docs/gohive/git.apache.org/thrift/lib/go/thrift/index.html create mode 100644 docs/gohive/hiveserver2/index.html create mode 100644 docs/gohive/thrift/index.html diff --git a/docs/gohive/git.apache.org/thrift.git/lib/go/thrift/index.html b/docs/gohive/git.apache.org/thrift.git/lib/go/thrift/index.html new file mode 100644 index 0000000..9cd02cf --- /dev/null +++ b/docs/gohive/git.apache.org/thrift.git/lib/go/thrift/index.html @@ -0,0 +1,13 @@ + + + + + + + + +

    Redirecting to https://github.com/sql-machine-learning/gohive

    + + diff --git a/docs/gohive/git.apache.org/thrift/lib/go/thrift/index.html b/docs/gohive/git.apache.org/thrift/lib/go/thrift/index.html new file mode 100644 index 0000000..9cd02cf --- /dev/null +++ b/docs/gohive/git.apache.org/thrift/lib/go/thrift/index.html @@ -0,0 +1,13 @@ + + + + + + + + +

    Redirecting to https://github.com/sql-machine-learning/gohive

    + + diff --git a/docs/gohive/hiveserver2/index.html b/docs/gohive/hiveserver2/index.html new file mode 100644 index 0000000..9cd02cf --- /dev/null +++ b/docs/gohive/hiveserver2/index.html @@ -0,0 +1,13 @@ + + + + + + + + +

    Redirecting to https://github.com/sql-machine-learning/gohive

    + + diff --git a/docs/gohive/thrift/index.html b/docs/gohive/thrift/index.html new file mode 100644 index 0000000..9cd02cf --- /dev/null +++ b/docs/gohive/thrift/index.html @@ -0,0 +1,13 @@ + + + + + + + + +

    Redirecting to https://github.com/sql-machine-learning/gohive

    + + From 4a16f86bfd6bc124e7baa3dfcf0d3866b9e0500a Mon Sep 17 00:00:00 2001 From: mxz96102 Date: Sat, 20 Apr 2019 11:32:48 +0800 Subject: [PATCH 8/9] build after merge --- docs/assets/js/search-data.json | 340 ++++---- docs/doc/contribute.html | 358 -------- docs/doc/contribute/build-tensorflow.html | 438 ---------- docs/doc/contribute/build.html | 415 --------- .../close_producer_from_consumer.html | 456 ---------- docs/doc/contribute/mysql-setup.html | 442 ---------- docs/doc/contribute/submitter.html | 387 --------- docs/doc/design.html | 358 -------- .../design/database_abstraction_layer.html | 392 --------- docs/doc/design/figures/sqlflow-arch.graffle | Bin 3985 -> 0 bytes docs/doc/design/figures/sqlflow-arch.png | Bin 99876 -> 0 bytes docs/doc/design/figures/sqlflow-arch2.dot | 50 -- docs/doc/design/figures/sqlflow-arch2.png | Bin 96987 -> 0 bytes docs/doc/design/pipe.html | 383 --------- docs/doc/design/sql_parser.html | 364 -------- docs/doc/design/syntax.html | 457 ---------- docs/doc/design/walkthrough.html | 371 -------- docs/feed.xml | 2 +- docs/index.html | 8 +- docs/pages/about.html | 649 +++++++------- docs/robots.txt | 2 +- docs/sitemap.xml | 22 +- .../doc/contribute/build-tensorflow.md | 140 --- docs/sqlflow/doc/contribute/build.md | 133 --- .../close_producer_from_consumer.md | 166 ---- .../doc/contribute/figures/sqlflow-arch2.png | Bin 96987 -> 0 bytes docs/sqlflow/doc/contribute/mysql-setup.md | 143 ---- docs/sqlflow/doc/contribute/nav.md | 7 - docs/sqlflow/doc/contribute/submitter.md | 98 --- .../doc/design/database_abstraction_layer.md | 97 --- .../doc/design/figures/sqlflow-arch.png | Bin 99876 -> 0 bytes docs/sqlflow/doc/design/nav.md | 7 - docs/sqlflow/doc/design/pipe.md | 96 --- docs/sqlflow/doc/design/sql_parser.md | 70 -- docs/sqlflow/doc/design/syntax.md | 165 ---- docs/sqlflow/doc/design/walkthrough.md | 76 -- docs/sqlflow/doc/quickstart.html | 800 +++++++++--------- docs/sqlflow/doc/quickstart.md | 93 -- 38 files changed, 900 insertions(+), 7085 deletions(-) delete mode 100644 docs/doc/contribute.html delete mode 100644 docs/doc/contribute/build-tensorflow.html delete mode 100644 docs/doc/contribute/build.html delete mode 100644 docs/doc/contribute/close_producer_from_consumer.html delete mode 100644 docs/doc/contribute/mysql-setup.html delete mode 100644 docs/doc/contribute/submitter.html delete mode 100644 docs/doc/design.html delete mode 100644 docs/doc/design/database_abstraction_layer.html delete mode 100644 docs/doc/design/figures/sqlflow-arch.graffle delete mode 100644 docs/doc/design/figures/sqlflow-arch.png delete mode 100644 docs/doc/design/figures/sqlflow-arch2.dot delete mode 100644 docs/doc/design/figures/sqlflow-arch2.png delete mode 100644 docs/doc/design/pipe.html delete mode 100644 docs/doc/design/sql_parser.html delete mode 100644 docs/doc/design/syntax.html delete mode 100644 docs/doc/design/walkthrough.html delete mode 100644 docs/sqlflow/doc/contribute/build-tensorflow.md delete mode 100644 docs/sqlflow/doc/contribute/build.md delete mode 100644 docs/sqlflow/doc/contribute/close_producer_from_consumer.md delete mode 100644 docs/sqlflow/doc/contribute/figures/sqlflow-arch2.png delete mode 100644 docs/sqlflow/doc/contribute/mysql-setup.md delete mode 100644 docs/sqlflow/doc/contribute/nav.md delete mode 100644 docs/sqlflow/doc/contribute/submitter.md delete mode 100644 docs/sqlflow/doc/design/database_abstraction_layer.md delete mode 100644 docs/sqlflow/doc/design/figures/sqlflow-arch.png delete mode 100644 docs/sqlflow/doc/design/nav.md delete mode 100644 docs/sqlflow/doc/design/pipe.md delete mode 100644 docs/sqlflow/doc/design/sql_parser.md delete mode 100644 docs/sqlflow/doc/design/syntax.md delete mode 100644 docs/sqlflow/doc/design/walkthrough.md delete mode 100644 docs/sqlflow/doc/quickstart.md diff --git a/docs/assets/js/search-data.json b/docs/assets/js/search-data.json index 1db16fa..1a732b7 100644 --- a/docs/assets/js/search-data.json +++ b/docs/assets/js/search-data.json @@ -1,171 +1,171 @@ { - "0": { - "id": "0", - "title": "About Us", - "content": "About Add info about the product. Add info about the team. Add info about the contact.", - "url": "http://localhost:4000/pages/about.html", - "relUrl": "/pages/about.html" - }, - "1": { - "id": "1", - "title": "Contribute", - "content": "Contribute Dev, build and run!", - "url": "/doc_index/contribute.html", - "relUrl": "/doc_index/contribute.html" - }, - "2": { - "id": "2", - "title": "Design", - "content": "Design How we design SQLFLow", - "url": "/doc_index/design.html", - "relUrl": "/doc_index/design.html" - }, - "3": { - "id": "3", - "title": "SQLFlow: Bridging Data and AI", - "content": "", - "url": "/", - "relUrl": "/" - }, - "4": { - "id": "4", - "title": "Build TensorFlow from Source Code using Docker", - "content": "Build TensorFlow from Source Code using Docker To contribute to TensorFlow, we need to build TensorFlow from source code. The official guide is great. However, it interleaves the native building process and that using Docker and makes it confusing because packages needed by the former are not by the latter. Also, we found some useful tricks to start Docker containers in practices that are not in the official guide. Hence this document. Build the Pip Package in TensorFlow Development Container On either Mac or Linux, or any other OS, we don’t have to install and configure the building tools; instead, we can use a Docker image where all tools have been installed and properly configured. Get the Docker image containing all the building tools: docker pull tensorflow/tensorflow:latest-devel Then, let’s get the source code. On any OS, please install git using the native package manager. For example, on Ubuntu, please sudo apt-get install git or, on Mac, brew install git Then, use the git just installed, let’s clone tensorflow source code: git clone --recursive https://github.com/tensorflow/tensorflow cd tensorflow By default, we will be on the master branch. Feel free to do you change in your feature branches, or switch to a release branch, for example: git checkout v1.11.0 git checkout -b v1.11.0 Then, let us start a Docker container running the tensorflow/tensorflow:latest-devel image: docker run --rm -it -w /tensorflow -v $PWD:/tensorflow -v $HOME/.cache:/root/.cache -e "HOST_PERMS=$(id -u):$(id -g)" tensorflow/tensorflow:latest-devel /bin/bash -w /tensorflow brings us to the /tensorflow directory in the container once after we start it. -v $PWD:/tensorflow maps the current directory, which is the just cloned TensorFlow source directory on the host, to /tensorflow in the container. -v $HOME/.cache:/root/.cache maps the Bazel temporary directory on the host into the container, so the intermediate files generated by Bazel running in the container are actually saved on the host. This allows us to interrupt the container during the build and restart it later to resume the building. e "HOST_PERMS=$(id -u):$(id -g)" passes the user identification on the host into the container as an environment variable. We can reset the mode of files generated in the container to this user identity. From now on, we will be working in the container. Let us first configure the building: ./configure Usually, I would simply choose all the default options by hitting enter all the way down. Build the pip package builder: bazel build --config=opt //tensorflow/tools/pip_package:build_pip_package Build the pip package and save it into /tensorflow: ./bazel-bin/tensorflow/tools/pip_package/build_pip_package /tensorflow Change the generated wheel file into the usual mode as on the host: chown $HOST_PERMS /mnt/tensorflow-*.whl Install the Pip Pacakge in TensorFlow Development Container Let’s now try the new pip package. First, we need to uninstall the current tensorflow pip package, then we install the newly built one: pip uninstall tensorflow pip install /tensorflow/tensorflow-*.whl Now, we can verify if the new package works. First, we need to switch to a directory out from /tensorflow, so we don’t import from the source directory: cd /tmp # other directories also work then, we can try import the newly installed tensorflow package and verify it: python >>> import tensorflow >>> print(tensorflow.__version__) Now, let us quit from Python and from the Docker container. We should see the tensorflow-*.whl file on the host in the current directory. Install the Pip Package in a Clean Python Package After we quit from the development container, we should see the wheel file in the TensorFlow source directory on the host. Now, we can start a Python container and install the Pip package in it. Start the Python container docker run --rm -it -v $PWD:/tensorflow python:2.7 bash Install the pip pacakge in the container pip install /tensorflow/tensorflow*.whl Try TensorFlow by starting Python in the container python >>> import tensorflow as tf >>> print(tf.__version__)", - "url": "/sqlflow/doc/build-tensorflow.html", - "relUrl": "/sqlflow/doc/build-tensorflow.html" - }, - "5": { - "id": "5", - "title": "Canonical Development Environment", - "content": "Canonical Development Environment Referring to this example, we create a canonical development environment for Go and Python programmers using Docker. Editing on Host When we use this Docker image for daily development work, the source code relies on the host computer instead of the container. The source code includes this repo and all its dependencies, for example, the Go package google.golang.org/grpc. Code-on-the-host allows us to run our favorite editors (Emacs, VIM, Eclipse, and more) on the host. Please free to rely on editors add-ons to analyze the source code for auto-completion. Building in Container We build a Docker image that contains development tools: The Python interpreter The Go compiler The protobuf compiler The protobuf to Go compiler extension The protobuf to Python compiler extension Because this repo contains Go code, please make sure that you have the directory structure required by Go. On my laptop computer, I have export GOPATH=$HOME/go You could have your $GOPATH pointing to any directory you like. Given $GOPATH$ set, we could git clone the source code of our project by running: go get github.com/sql-machine-learning/sqlflow Change the directory to our project root, and we can use go get to retrieve and update Go dependencies. cd $GOPATH/src/github.com/sql-machine-learning/sqlflow go get -u -t ./... Note -t instructs get to also download the packages required to build the tests for the specified packages. As all Git users would do, we run git pull from time to time to sync up with others’ work. If somebody added new dependencies, we might need to run go -u ./... after git pull to update dependencies. To build this project, we need the protobuf compiler, Go compiler, Python interpreter, gRPC extension to the protobuf compiler. To ease the installation and configuration of these tools, we provided a Dockerfile to install them into a Docker image. To build the Docker image: docker build -t sqlflow:dev -f Dockerfile.dev . Development Build and Test We build and test the project inside the docker container. To run the container, we need to map the $GOPATH directory on the host into the /go directory in the container, because the Dockerfile configures /go as the $GOPATH in the container: docker run --rm -it -v $GOPATH:/go -w /go/src/github.com/sql-machine-learning/sqlflow sqlflow:dev bash Inside the Docker container, start a MySQL server in the background service mysql start& run all the tests as go generate ./... go install ./... go test -v ./... where go generate invokes the protoc command to translate server/sqlflow.proto into server/sqlflow.pb.go and go test -v builds and run unit tests. Release The above build process currently generates two binary files in $GOPATH/bin on the host. To package them into a Docker image, please run docker build -t sqlflow -f ./Dockerfile $GOPATH/bin To publish the released Docker image to our official DockerHub docker tag sqlflow sqlflow/sqlflow:latest docker push sqlflow/sqlflow:latest Demo: Command line Prompt The demo requires a MySQL server instance with populated data. If we don’t, we could follow example/datasets/README.md to start one on the host. After setting up MySQL, run the following inside the Docker container go run cmd/demo/demo.go --db_user root --db_password root --db_address host.docker.internal:3306 You should be able to see the following prompt sqlflow>", - "url": "/sqlflow/doc/build.html", - "relUrl": "/sqlflow/doc/build.html" - }, - "6": { - "id": "6", - "title": "Closing the producer goroutine from the consumer", - "content": "Closing the producer goroutine from the consumer The producer-and-consumer pattern is well used in Go concurrent programming. When the consumer stops, we want to gracefully stop the producer as well. Problem When a gRPC server receives a streaming request, it usually calls a function that returns a channel, reads the result from that channel and send the result to the client one by one. Take the following code for instance: upon receiving a request, the main goroutine Service calls launchJob. launchJob starts a separate goroutine as an anonymous function call and returns a channel. In the anonymous function, items will be sent to channel. And Service on the otherside of the channel will reads from it. func Service(req *Request, stream *StreamResponse) error { result := launchJob(req.Content) for r := range result { if e := stream.Send(result); e != nil { // should we signal the running goroutine so it will stop sending? return e } } } func launchJob(content string) chan Item { c := make(chan Item) go func() { defer close(c) acquireScarceResources() defer releaseScarceResources() ... // if stream.Send(result) returns an error and the Service returns, this will be blocked c <- Item{} ... }() return c } There is a major problem in this implementation. As pointed out by the comment, if the Send in Service returns an error, the Service function will return, leaving the anonymous function being blocked on c <- Item{} forever. This problem is important because the leaking goroutine usually owns scarce system resources such as network connection and memory. Solution: pipeline explicit cancellation Inspired by this blog post section Explicit cancellation, we can signal the cancellation via closing on a separate channel. And we can follow the terminology as io.Pipe. package sql import ( "errors" ) var ErrClosedPipe = errors.New("pipe: write on closed pipe") // pipe follows the design at https://blog.golang.org/pipelines // - wrCh: chan for piping data // - done: chan for signaling Close from Reader to Writer type pipe struct { wrCh chan interface{} done chan struct{} } // PipeReader reads real data type PipeReader struct { p *pipe } // PipeWriter writes real data type PipeWriter struct { p *pipe } // Pipe creates a synchronous in-memory pipe. // // It is safe to call Read and Write in parallel with each other or with Close. // Parallel calls to Read and parallel calls to Write are also safe: // the individual calls will be gated sequentially. func Pipe() (*PipeReader, *PipeWriter) { p := &pipe{ wrCh: make(chan interface{}), done: make(chan struct{})} return &PipeReader{p}, &PipeWriter{p} } // Close closes the reader; subsequent writes to the func (r *PipeReader) Close() { close(r.p.done) } // ReadAll returns the data chan. The caller should // use it as `for r := range pr.ReadAll()` func (r *PipeReader) ReadAll() chan interface{} { return r.p.wrCh } // Close closes the writer; subsequent ReadAll from the // read half of the pipe will return a closed channel. func (w *PipeWriter) Close() { close(w.p.wrCh) } // Write writes the item to the underlying data stream. // It returns ErrClosedPipe when the data stream is closed. func (w *PipeWriter) Write(item interface{}) error { select { case w.p.wrCh <- item: return nil case <-w.p.done: return ErrClosedPipe } } And the consumer and producer be can implemented as func Service(req *Request, stream *StreamResponse) error { pr := launchJob(req.Content) defer pr.Close() for r := range pr.ReadAll() { if e := stream.Send(r); e != nil { return e } } } func launchJob(content string) PipeReader { pr, pw := Pipe() go func() { defer pw.Close() if err := pw.Write(Item{}); err != nil { return } } return pr } Further Reading Google Form: Channel send timeout Go by Example: Timeouts Google I/O 2013 - Advanced Go Concurrency Patterns Go Concurrency Patterns Talk Go Concurrency Patterns: Pipelines and cancellation", - "url": "/sqlflow/doc/close_producer_from_consumer.html", - "relUrl": "/sqlflow/doc/close_producer_from_consumer.html" - }, - "7": { - "id": "7", - "title": "Compatibility with Various SQL Engines", - "content": "Compatibility with Various SQL Engines SQLFlow interacts with SQL engines like MySQL and Hive, while different SQL engines use variants of SQL syntax, it is important for SQLFlow to have an abstraction layer that hides such differences. SQLFlow calls Go’s standard database API. The submitter programs generated by SQLFlow call Python’s database API. Both APIs abstract the interface to various SQL engines; however, they are insufficient for SQLFlow to work. In this document, we examine all interactions between SQLFlow and the SQL engine so to identify what SQLFlow authors have to abstract in addition to calling Go’s and Python’s database APIs. Data Operations in Go Data Retrieval The basic idea of SQLFlow is to extend the SELECT statement of SQL to have the TRAIN and PREDICT clauses. For more discussion, please refer to the syntax design. SQLFlow translates such “extended SQL statements” into submitter programs, which forward the part from SELECT to TRAIN or PREDICT, which we call the “standard part”, to the SQL engine. SQLFlow also accepts the SELECT statement without TRAIN or PREDICt clauses and would forward such “standard statements” to the engine. It is noticeable that the “standard part” or “standard statements” are not standardized. For example, various engines use different syntax for joining. MySQL: SELECT pet.name, comment FROM pet, event WHERE pet.name =event.name; with keyword WHERE . Hive: SELECT pet.name, comment FROM pet JOIN event ON (pet.name =event.name) with keyword JOIN and ON. ODPS and SQLite use either INNER JOIN or OUTER JOIN. Fortunately, as SQLFlow forwards the above parts to the engine, it doesn’t have to care much about the differences above. Metadata Retrieval To verify the semantics of users’ inputs, SQLFlow needs to retrieve the schema of tables. For example, the input might be SELECT name, age, income FROM employee TRAIN DNNRegressor WITH hidden_layers=[10,50,10] COLUMN name, agee LABEL income; In the above example, the user misspelled the field name age in the COLUMN clause as “agee”. SQLFlow must be able to find that out. To do that, SQLFlow needs to query the field names from the SQL engine. However, different engines use various syntax. For example: MySQL: DESCRIBE/DESC employee; Hive: DESCRIBE FORMATTED employee; ODPS: DESC employee; SQLite: PRAGMA table_info([employee]); The returned data format varies too. Our solution to avoid such differences is not-to-use-them; instead, SQLFlow retrieves the table schema by running a query like SELECT * FROM employee LIMIT 1; and inferring field types using the mechanism called DatabaseTypeName provided by SQL engines drivers beneath the Go’s standard database API. Prepare Prediction Table A SQLFlow prediction job writes its prediction results into a table. It prepares the prediction table by Dropping previous prediction table DROP TABLE IF EXISTS my_table; Creating table with schema CREATE TABLE my_table (name1, type1, name2 type2); Most SQL engines, including MySQL, Hive, ODPS, SQLite, support both statements. Translate Database Column Type to TensorFlow Feature Column Type After retrieving database column type name through DatabaseTypeName, we can derive TensorFlow’s feature column type via a mapping such as {"FLOAT", "DOUBLE"} -> tf.numeric_column. Save Model SQLFlow saves trained ML model by dumping the serialized the model directory into a table. It first creates a table by CREATE TABLE IF NOT EXISTS %s (id INT AUTO_INCREMENT, block BLOB, PRIMARY KEY (id)) and insert blobs by INSERT INTO %s (block) VALUES(?). Note that Hive and ODPS doesn’t have BLOB type, we need to use BINARY (docs at ODPS, Hive) instead. Also, note that Hive and ODPS doesn’t support AUTO_INCREMENT, we need to implemented auto increment logic in sqlfs. Load Model SQLFlow loads trained ML model by reading rows in a table and deserializing the blob to a model directory. It reads rows by running SELECT block FROM %s ORDER BY id, which is supported by most databases. Data Operations in Python Connect to SQL Engines Thanks to the Python database API, connecting to different databases follows a similar API. conn = mysql.connector.connect(user='scott', password='password', host='127.0.0.1', database='employees') conn = sqlite3.connect('path/to/your/sqlite/file') conn = pyhive.connect('localhost') cursor = conn.cursor() cursor.execute('select * from my_table;') Insert Prediction Result into Prediction Table Python database API provides execute_many(sql, value) to insert multiple values at once. So one can prepare the following insertion statement. Please be aware that MySQL and SQLite use INSERT INTO to insert rows while Hive and ODPS use INSERT INTO TABLE. -- MySQL, SQLite INSERT INTO table_name VALUES (value1, value2, value3, ...); -- Hive, ODPS INSERT INTO TABLE table_name VALUES (value1, value2, value3, ...);", - "url": "/sqlflow/doc/database_abstraction_layer.html", - "relUrl": "/sqlflow/doc/database_abstraction_layer.html" - }, - "8": { - "id": "8", - "title": "Run MySQL Server and Client in Docker Containers", - "content": "Run MySQL Server and Client in Docker Containers The document explains how to setup MySQL in our development environment. Run MySQL Server in a Docker Container docker run --rm -v /tmp/test1:/var/lib/mysql --name mysql01 -e MYSQL_ROOT_PASSWORD=root -e MYSQL_ROOT_HOST='%' -p 3306:3306 -d mysql/mysql-server:8.0 the -v option ensures that the database is saved on the host. The default directory where MySQL saves the database is /var/lib/mysql. This directory can be configured in /etc/mysql/my.cnf, as explained in this post. By overlaying the directory /tmp/test1 on the host to /var/lib/mysql, we “cheat” MySQL to save databases on the host. So, we can kill the container and restart it, and the database is still there. Please be aware that the directory on the host must be empty the first time we run the above command; otherwise, MySQL would fail to initialize. I figured out this problem after several failures using docker logs. the -e option sets the root password of MySQL to “root”. Feel free to set it to any password you like. the second -e options sets MYSQL_ROOT_HOST to a wildcard so to allow clients connecting to the server via TCP/IP as the user “root”. This trick works with MySQL 5.7 and 8.0, but not the most recent under-development version. the --name option names the container to mysql01, which can be used to refer to this container. the -p option maps the port 3306, on which the MySQL server listens, to the same port on the host, so that clients could connect to the server via TCP/IP. Run MySQL Client in the Server Container docker exec -it mysql01 mysql -uroot -p This command executes the command mysql, which is the command line tool of MySQL, in the container named mysql01. The command line flags of mysql include -u, which specifies the username of MySQL, and -p, which makes MySQL prompts for the password. For this example, we should type the password “root”, which was set in the previous command. Please wait for a few seconds after the starting of the MySQL server container before we execute the client; otherwise, the startup of the client might fail. Once we get into the MySQL client, we can type SQL commands, e.g., show databases; create database yi; Run Client in a Different Container on the Same Host docker run --rm -it -v /tmp/test1:/var/lib/mysql mysql/mysql-server:8.0 mysql -uroot -p The -v option maps the database directory on the host to the client container. This mapping is necessary because, by default, the client talks to the server via Unix socket /var/lib/mysql/mysql.sock, which is /tmp/test1/mysql.sock on the host. Run Client in a Container on a Remote Host docker run --rm -it mysql/mysql-server:8.0 mysql -h 192.168.1.3 -P 3306 -uroot -p the -h option tells the client where the server is running on. In this example, the given IP is the one of the host where I ran the MySQL server container. Please be aware that the above command works only if the server allows remote connections. Run Python Client in a Container To write a Python client, we need to install the Python package mysql-connector-python. FROM python:2.7 RUN pip install mysql-connector-python Please be aware that some documents says that we need to install mysql-connector. I tried; but the mysql.connector.connect call failed with the error mysql.connector.errors.NotSupportedError: Authentication plugin 'caching_sha2_password' is not supported. Build the Docker image: docker build -t sqlflow . Run the image: docker run --rm -it sqlflow bash and we can start Python and run the following Python code snippet >>> import mysql.connector >>> db = mysql.connector.connect(user='root', passwd='root', host='192.168.1.3') >>> print(db) <mysql.connector.connection_cext.CMySQLConnection object at 0x7fbab9f3fed0> Run a Go Client In order to connect to a database, you need to import the database’s driver first. export GOPATH=$HOME/go go get -u github.com/go-sql-driver/mysql go run the following file package main import ( "database/sql" "github.com/go-sql-driver/mysql" "log" ) func main() { testConfig := &mysql.Config{ User: "root", Passwd: "root", Net: "tcp", Addr: "localhost:3306", } db, e := sql.Open("mysql", testConfig.FormatDSN()) if e != nil { log.Fatal(e) } defer db.Close() }", - "url": "/sqlflow/doc/mysql-setup.html", - "relUrl": "/sqlflow/doc/mysql-setup.html" - }, - "9": { - "id": "9", - "title": "Piping Responses", - "content": "Piping Responses Streaming Responses As described in the overall design, a SQLFlow job could be a standard or an extended SQL statemnt, where an extended SQL statement will be translated into a Python program. Therefore, each job might generate up to the following data streams: standard output, where each element is a line of text, standard error, where each element is a line of text, data rows, where the first element consists of fields name/types, and each of the rest is a row of data, status, where the element could be pending, failed, and succeeded. To create good user experience, we need to pipe these responses from SQLFlow jobs to Jupyter Notebook in real-time. Stages in the Pipe The pipe that streams outputs from SQLFlow jobs to the Jupyter Notebook consists of the following stages: Web browser ↑ | HTTP ↓ Jupyter Notebook server ↑ | ZeroMQ streams: Shell, IOPub, stdin, Controls, Heartbeat ↓ iPython kernel ↑ | IPython magic command framework ↓ SQLFlow magic command for Jupyter ↑ | gRPC ↓ SQLFlow server ↑ | Go channels ↓ SQLFlow job manager (Go functions) ↑ | IPC with Go's standard library ↓ SQLFlow jobs In the above figure, from the SQLFlow magic command to the bottom layer are our work. Streaming We have two alternative ideas: multiple streams and a multiplexing stream. We decided to use a multiplexing stream because we had a unsuccessful trial with the multiple streams idea: we make the job writes to various Go channels and forward each Go channel to a streaming gRPC call, as the following: Multiple streams The above figure shows that there are multiple streams between the Jupyter Notebook server and Jupyter kernels. According to the document, there are five: Shell, IOPub, stdin, Control, and Heartbeat. These streams are ZeroMQ streams. We don’t use ZeroMQ, but we can take the idea of having multiple parallel streams in the pipe. service SQLFlow { rpc File(string sql) returns (int id) {} rpc ReadStdout(int id) returns (stream string) {} rpc ReadStderr(int id) returns (stream string) {} rpc ReadData(int id) returns (stream Row) {} rpc ReadStatus(int id) returns (stream int) {} } However, we realized that if the user doesn’t call any one of the SQLFlow.Read... call, there would be no forwarding from the Go channel to Jupyter, thus the job would block forever at writing. A Multiplexing Stream Another idea is multiplexing all streams into one. For example, we can have only one ZeroMQ stream, where each element is a polymorphic type – could be a text string or a data row. service SQLFlow { rpc Run(string sql) returns (stream Response) {} } // Only one of the following fields should be set. message Response { oneof record { repeated string head = 1; // Column names. repeated google.protobuf.Any row = 2; // Cells in a row. string log = 3; // A line from stderr or stdout. } }", - "url": "/sqlflow/doc/pipe.html", - "relUrl": "/sqlflow/doc/pipe.html" - }, - "10": { - "id": "10", - "title": "Quick start", - "content": "Quick start SQLFlow is currently under active development. For those who are interested in trying it out, we have provided several demos. Play around with it. Any bug report and issue are welcomed. :) Setup Install Docker. Set up a MySQL server following example/datasets/README.md. Pull the latest SQLFlow Docker image: docker pull sqlflow/sqlflow:latest. Demo 1: Jupyter Notebook Start a Docker container that runs sqlflowserver and Jupyter Notebook. If you are using Docker for Linux, please change host.docker.internal:3306 to localhost:3306. docker run --rm -it -p 8888:8888 sqlflow/sqlflow:latest bash -c "sqlflowserver --db_user root --db_password root --db_address host.docker.internal:3306 & SQLFLOW_SERVER=localhost:50051 jupyter notebook --ip=0.0.0.0 --port=8888 --allow-root" If you are using Docker for Mac, please be aware the option --db_address host.docker.internal:3306 where host.docker.internal translates to the host ip address as recommended here. If you are running MySQL on remote, please be aware that MySQL only allows connections from localhost by default. Fix can be found here. Open a Web browser and direct to localhost:8888 and input the token. Then you can create notebooks. In a cell, you should be able to type %%sqlflow select 1 Explore more examples at example.ipynb Demo 2: Command Line Prompt Start a Docker container that runs SQLFlow command line prompt. If you are using Docker for Linux, please change host.docker.internal:3306 to localhost:3306. docker run -it --rm --net=host sqlflow/sqlflow:latest demo --db_user root --db_password root --db_address host.docker.internal:3306 You should be able to see the following prompt. sqlflow> Example Select data sqlflow> select * from iris.train limit 2; -- +--+-+--+-+-+ | SEPAL LENGTH | SEPAL WIDTH | PETAL LENGTH | PETAL WIDTH | CLASS | +--+-+--+-+-+ | 6.4 | 2.8 | 5.6 | 2.2 | 2 | | 5 | 2.3 | 3.3 | 1 | 1 | +--+-+--+-+-+ Train a Tensorflow DNNClassifier sqlflow> SELECT * FROM iris.train TRAIN DNNClassifier WITH n_classes = 3, hidden_units = [10, 20] COLUMN sepal_length, sepal_width, petal_length, petal_width LABEL class INTO sqlflow_models.my_dnn_model; -- ... Training set accuracy: 0.96721 Done training Prediction using a trained model sqlflow> SELECT * FROM iris.test predict iris.predict.class USING sqlflow_models.my_dnn_model; Checkout prediction result sqlflow> select * from iris.predict limit 10;", - "url": "/sqlflow/doc/quickstart.html", - "relUrl": "/sqlflow/doc/quickstart.html" - }, - "11": { - "id": "11", - "title": "Extended SQL Parser Design", - "content": "Extended SQL Parser Design This documentation explains the technical decision made in building a SQL parser in Go. It is used to parsed the extended SELECT syntax of SQL that integrates TensorFlow Estimators. Related Work Lexer and Parser Generator In 2001, when I was in graduate school, I defined an extended SQL syntax for querying distributed relational databases, as part of the course project of Distributed RDBMS by Prof. Li-zhu Zhou. I wrote the parser using bison (a modern version of yacc) and flex (a modern version of lex). yacc and lex generate C code; bison and flex generate C++ code. However, this time, I’d use Go. I surveyed goyacc, a standard Go tool. The usage is very similar to that of yacc and bison. However, the Go toolchain doesn’t provide a tool like lex/flex. Google revealed golex, which is out of maintenance. The Medium post recommends Ragel, which is a C++ program and could generate Go lexer; however, it lacks documents. Handwritten Lexer and Parser Some documents, including this one recommends handwriting lexers. However, it doesn’t explain how to write the parser. GoAcademy always provides high-quality tech blog posts. This one is from the author of InfluxDB. However, I stopped at where it explains wrapping a SQL statement as a string by an io.Reader, because it is obvious that we should keep the string as a string so that that token strings could refer to the same memory storage of the SQL statement. Following a link in the above GoAcademy post, I found Rob Pike’s excellent talk on how to write a lexer in Go in 2011. Many works after that change Rob’s implementation somehow but always lead to longer and less comprehensible codebases. The Choice Therefore, I wrote the lexer and parser both following Rob Pike’s idea. After few days work, I realized that: I should borrow the idea from Rob to represent SQL statements as strings, but not io.Reader as other work do, but no need to use channels and goroutines at all, and it is technically intractable to write a SQL lexer/parser manually. So, I switched to write a lexer manually, and to generate the parser using goyacc. During my work, I referred to this example and the official yacc manual for details about operator association and precedence.", - "url": "/sqlflow/doc/sql_parser.html", - "relUrl": "/sqlflow/doc/sql_parser.html" - }, - "12": { - "id": "12", - "title": "Submitter", - "content": "Submitter A submitter is a pluggable module in SQLFlow that is used to submit an ML job to a third party computation service. Workflow When a user types in an extended SQL statement, SQLFlow first parses and semantically verifies the statement. Then SQLFlow either runs the ML job locally or submits the ML job to a third party computation service. In the latter case, SQLFlow produces a job description (TrainDescription or PredictDescription) and hands it over to the submitter. For a training SQL, SQLFlow produces TrainDescription; for prediction SQL, SQLFlow produces PredDescription. The concrete definition of the description looks like the following {% raw %} type ColumnType struct { Name string // e.g. sepal_length DatabaseTypeName string // e.g. FLOAT } // SELECT * // FROM iris.train // TRAIN DNNClassifier // WITH // n_classes = 3, // hidden_units = [10, 20] // COLUMN sepal_length, sepal_width, petal_length, petal_width // LABEL class // INTO sqlflow_models.my_dnn_model; type TrainDescription struct { StandardSelect string // e.g. SELECT * FROM iris.train Estimator string // e.g. DNNClassifier Attrs map[string]string // e.g. {{"n_classes", "3"}, {"hidden_units", "[10, 20]"}} X []ColumnType // e.g. {{"sepal_length", "FLOAT"}, ...} Y ColumnType // e.g. {"class", "INT"} ModelName string // e.g. my_dnn_model } // SELECT * // FROM iris.test // PREDICT iris.predict.class // USING sqlflow_models.my_dnn_model; type PredDescription struct { StandardSelect string // e.g. SELECT * FROM iris.test TableName string // e.g. iris.predict ModelName string // e.g. my_dnn_model } {% endraw %} Submitter Interface The submitter interface should provide two functions Train and Predict. The detailed definition can be the following type Submitter interface { // Train executes a ML training job and streams job's response through writer. // A typical Train function should include // - Loading the training data // - Initializing the model // - model.train // - Saving the trained model to a persistent storage Train(desc TrainDescription, writer PipeWriter) error // Predict executes a ML predicting job and streams job's response through writer // A typical Predict function should include // - Loading the model from a persistent storage // - Loading the prediction data // - model.predict // - Writing the prediction result to a table Predict(desc PredictDescription, writer PipeWriter) error } Register a submitter A new submitter can be added as import ( ".../my_submitter" ".../sqlflow/sql" ) func main() { // ... sql.Register(my_submitter.NewSubmitter()) // ... for { sql := recv() sql.Run(sql) } } where sql.Register will put my_submitter instance to package level registry. During sql.Run, it will check whether there is a submitter registered. If there is, sql.Run will run either submitter.Train or submitter.Predict.", - "url": "/sqlflow/doc/submitter.html", - "relUrl": "/sqlflow/doc/submitter.html" - }, - "13": { - "id": "13", - "title": "SQLFlow: Design Doc", - "content": "SQLFlow: Design Doc What is SQLFlow SQLFlow is a bridge that connects a SQL engine, for example, MySQL, Hive, SparkSQL, Oracle, or SQL Server, and TensorFlow and other machine learning toolkits. SQLFlow extends the SQL syntax to enable model training and inference. Related Work We could write simple machine learning prediction (or scoring) algorithms in SQL using operators like DOT_PRODUCT. However, this requires copy-n-pasting model parameters from the training program into SQL statements. Some proprietary SQL engines provide extensions to support machine learning. Microsoft SQL Server Microsoft SQL Server has the machine learning service that runs machine learning programs in R or Python as an external script: CREATE PROCEDURE generate_linear_model AS BEGIN EXEC sp_execute_external_script @language = N'R' , @script = N'lrmodel <- rxLinMod(formula = distance ~ speed, data = CarsData); trained_model <- data.frame(payload = as.raw(serialize(lrmodel, connection=NULL)));' , @input_data_1 = N'SELECT [speed], [distance] FROM CarSpeed' , @input_data_1_name = N'CarsData' , @output_data_1_name = N'trained_model' WITH RESULT SETS ((model varbinary(max))); END; A challenge to the users is that they need to know not only SQL but also R or Python, and they must be capable of writing machine learning programs in R or Python. Teradata SQL for DL Teradata also provides a RESTful service, which is callable from the extended SQL SELECT syntax. SELECT * FROM deep_learning_scorer( ON (SELECT * FROM cc_data LIMIT 100) URL('http://localhost:8000/api/v1/request') ModelName('cc') ModelVersion('1') RequestType('predict') columns('v1', 'v2', ..., 'amount') ) The above syntax couples the deployment of the service (the URL in the above SQL statement) with the algorithm. Google BigQuery Google BigQuery enables machine learning in SQL by introducing the CREATE MODEL statement. CREATE MODEL dataset.model_name OPTIONS(model_type='linear_reg', input_label_cols=['input_label']) AS SELECT * FROM input_table; Currently, BigQuery only supports two simple models: linear regression and logistic regression. Design Goal None of the above meets our requirement. First of all, we want to build an open source software. Also, we want it to be extensible: We want it extensible to many SQL engines, instead of targeting any one of them. Therefore, we don’t want to build our syntax extension on top of user-defined functions (UDFs); otherwise, we’d have to implement them for each SQL engine. We want the system extensible to support sophisticated machine learning models and toolkits, including TensorFlow for deep learning and xgboost for trees. Another challenge is that we want SQLFlow to be flexible enough to configure and run cutting-edge algorithms, including specifying feature crosses. At the same time, we want SQLFlow easy to learn – at least, no Python or R code embedded in the SQL statements, and integrate hyperparameter estimation. We understand that a key to address the above challenges is the syntax of the SQL extension. To craft a highly-effective and easy-to-learn syntax, we need user feedback and fast iteration. Therefore, we’d start from a prototype that supports only MySQL and TensorFlow. We plan to support more SQL engines and machine learning toolkits later. Design Decisions As the beginning of the iteration, we propose an extension to the SQL SELECT statement. We are not going a new statement way like that BigQuery provides CREATE MODEL, because we want to maintain a loose couple between our system and the underlying SQL engine, and we cannot create the new data type for the SQL engine, like CREATE MODEL requires. We highly appreciate the work of TensorFlow Estimator, a high-level API for deep learning. The basic idea behind Estimator is to implement each deep learning model, and related training/testing/evaluating algorithms as a Python class derived from tf.estimator.Estimator. As we want to keep our SQL syntax simple, we would make the system extensible by calling estimators contributed by machine learning experts and written in Python. The SQL syntax must allow users to set Estimator attributes (parameters of the Python class’ constructor, and those of train, evaluate, or predict). Users can choose to use default values. We have a plan to integrate our hyperparameter estimation research into the system to optimize the default values. Though estimators derived from tf.estimator.Estimator run algorithms as TensorFlow graphs; SQLFlow doesn’t restrict that the underlying machine learning toolkit has to be TensorFlow. Indeed, as long as an estimator provides methods of train, evaluate, and predict, SQLFlow doesn’t care if it calls TensorFlow or xgboost. Precisely, what SQLFlow expect is an interface like the following: class AnEstimatorClass: __init__(self, **kwargs) train(self, **kwargs) evaluate(self, **kwargs) predict(self, **kwargs) We also want to reuse the feature columns API from Estimator, which allows users to columns of tables in a SQL engine to features to the model. Extended SQL Syntax Again, just as the beginning of the iteration, we propose the syntax for training as SELECT * FROM kaggle_credit_fraud_training_data LIMIT 1000 TRAIN DNNClassifier /* a pre-defined TensorFlow estimator, tf.estimator.DNNClassifier */ WITH layers=[100, 200], /* a parameter of the Estimator class constructor */ train.batch_size = 8 /* a parameter of the Estimator.train method */ COLUMN *, /* all columns as raw features */ cross(v1, v9, v28) /* plus a derived (crossed) column */ LABEL class INTO sqlflow_models.my_model_table; /* saves trained model parameters and features into a table */ We see the redundancy of * in two clauses: SELECT and COLUMN. The following alternative can avoid the redundancy, but cannot specify the label. SELECT * /* raw features or the label? */ corss(v1, v9, v28) /* derived featuers */ FROM kaggle_credit_fraud_training_data Please be aware that we save the trained models into tables, instead of a variable maintained by the underlying SQL engine. To invent a new variable type to hold trained models, we’d make our system tightly integrated with the SQL engine, and harms the extensibility to other engines. The result table should include the following information: The estimator name, e.g., DNNClassifier in this case. Estimator attributes, e.g., layer and train.batch_size. The feature mapping, e.g., * and cross(v1, v9, v28). Similarly, to infer the class (fraud or regular), we could SELECT * FROM kaggle_credit_fraud_development_data PREDICT kaggle_credit_fraud_development_data.class USING sqlflow_models.my_model_table; System Architecture A Conceptual Overview In the prototype, we use the following architecture: SQL statement -> our SQL parser --standard SQL-> MySQL -extended SQL-> code generator -> execution engine In the prototype, the code generator generates a Python program that trains or predicts. In either case, it retrieves the data from MySQL via MySQL Connector Python API, optionally, retrieves the model from MySQL, trains the model or predicts using the trained model by calling the user specified TensorFlow estimator, and writes the trained model or prediction results into a table. Working with Jupyter Notebook and Kubernetes The following figures shows the system components and their runtime environment. The left part shows how to run the system on a PC/laptop, the right part shows how to run it on a Kubernetes cluster.", - "url": "/sqlflow/doc/syntax.html", - "relUrl": "/sqlflow/doc/syntax.html" - }, - "14": { - "id": "14", - "title": "SQLFlow: Code Walkthrough", - "content": "SQLFlow: Code Walkthrough User Experience SQLFlow allows users to write SQL programs with extended syntax in Jupyter Notebook or a command-line tool. The following SQL statements train a TensorFlow model named DNNClassifier, which is a Python class derived from tf.estimator.Estimator: SELECT * FROM a_table TRAIN DNNClassifier WITH learning_rate=0.01 INTO sqlflow_models.my_model; And the following statement uses the trained model for prediction. SELECT * FROM b_table PREDICT b_table.predicted_label USING sqlflow_models.my_model; Please be aware that the part in the above statements before the extended keyword TRAIN and PREDICT is a standard SQL statement. This feature simplifies the implementation of the SQLFlow system. System Implementation If a SQL statement is of the standard syntax, SQLFlow throws it to the SQL engine and redirects the output to the user; otherwise, SQLFlow translates the statement of extended syntax into a Python program. Currently, it generates a program that throws the standard-syntax part of SELECT to MySQL, reads the results in the train-loop of a TensorFlow program. We will talk about how to extend SQLFlow to connect more SQL engines like Oracle, Hive, and SparkSQL, and generates more types of machine learning programs that calls distributed TensorFlow, PyTorch, and xgboost later. Before that, let us explain the system components. SQLFlow as a gRPC Server SQLFlow is a gRPC server, which can connect with multiple clients. A typical client is pysqlflow, the SQLFlow plugin for Jupyter Notebook server. Another once is a text-based client /cmd/sqlflowserver/main.go. Jupyter Notebook (SQL statements)--> SQLFlow gRPC server (SQLFlow magic command) <--(a stream of messages)-- The protobuf definition of the gRPC service is at /server/proto/sqlflow.proto. The return of the method SQLFlow.Run is a stream of Reponses, where each represents either a table header, a row, or a log message. The header and rows are usually from a standard SQL statement, for example, SELECT or DESCRIBE, and the log messages are usually from the running of a generated Python program. SQLFlow in the gRPC Server Once the SQLFlow server receives a batch of SQL statements via a gRPC call, it runs the following steps for each statement: the parser to generate parsing result, the verifier to verify the semantics given the parsing result, the code generator to generate a Python program, or the submitter, from the parsing result, the executor that runs the submitter locally. Step 3. and 4. are only for a SQL statement of extended syntax; otherwise, SQLFlow server proxies the standard-syntax statement to the SQL engine. The executor calls Go’s standard package that captures the stdout and stderr from the submitter process and passing the result back to the gRPC client. Therefore, it is the responsibility of the submitter to print log messages to its stderr and stdout. Minimal Viable Product In the minimal viable product (MVP) of SQLFlow, the code generator generates a Python program consists of two parts: throw the standard SELECT part in the extended-syntax statement to MySQL via ODBC, and a loop that reads outputs from the run of the SELECT statement and trains the model (or, using a trained model to predict). The training part calls TensorFlow to update the parameters of the model specified in the TRAIN clause. Extensibility By writing more code generators, we could extend SQLFlow to support more SQL engines, e.g., Hive and Oracle, and use machine learning toolkits, e.g., PyTorch and xgboost, in addition to TensorFlow, on various computing platforms. You are welcome to add more code generators such as codegen_distributed_tf.go to generate a submitter program similar to the MVP but runs a distributed TensorFlow training job. codegen_kubernetes_tf.go to launch a distributed TensorFlow job on a Kubernetes cluster, other than running locally, in the same container as where SQLFlow gRPC server resides. codegen_gcloud_pytorch.go to launch a submitter that calls PyTorch instead of TensorFlow for training on the Google Cloud. Job Management The current design of the gRPC interface assumes that the connection between the client, e.g., the Jupyter Notebook, and the SQLFlow server keeps alive during the running of the training program. This assumption is reasonable because even if the user closes her/his Web browser and disconnect to the Jupyter Notebook server, the connection between Jupyter to SQLFlow server might keep alive. However, this might not be robust enough if the Jupyter Notebook server runs on a user’s laptop and gets killed. In such a case, the gRPC server cannot stream the messages back to the client and would cause the failure of the submitter. A solution is to change the gRPC interface of SQLFlow server to have a method that files a job and returns immediately, and another method to get a batch of recent messages given a job ID. We will make a design for that soon.", - "url": "/sqlflow/doc/walkthrough.html", - "relUrl": "/sqlflow/doc/walkthrough.html" - }, - "15": { - "id": "15", - "title": "SQLFlow", - "content": "SQLFlow SQLFlow enriches SQL systems, e.g., MySQL, Hive, SparkSQL, with the capability of machine learning, using TensorFlow. We slightly extend the syntax of the SELECT statement of SQL to support model training and prediction. How to Use Quick start Extended SQL syntax How to Contribute Build from source code. The walkthrough of the source code The choice of parser generator Questions and Feedback Your feedback is our motivation to move on. Please let us know your questions, concerns, and issues by filing Github Issues. License Apache License 2.0", - "url": "/sqlflow/", - "relUrl": "/sqlflow/" - }, - "16": { - "id": "16", - "title": "SQLFlow Demo", - "content": "SQLFlow Demo Before you start, you need to set up test MySQL server as described here Start SQLFlow as go get -d ./... && go run main.go -logdir="/path/to/logs/" -loglevel="info" Regular SQL Statements sqlflow> SHOW DATABASES; +--+ | Database | +--+ | churn | | information_schema | | iris | | mysql | | performance_schema | | sqlflow_models | | sqlfs | | sys | +--+ sqlflow> SELECT * FROM iris.train LIMIT 1; +--+-+--+-+-+ | sepal_length | sepal_width | petal_length | petal_width | class | +--+-+--+-+-+ | 6.4 | 2.8 | 5.6 | 2.2 | 2 | +--+-+--+-+-+ Train sqlflow> SELECT * FROM iris.train TRAIN DNNClassifier WITH n_classes = 3, hidden_units = [10, 20] COLUMN sepal_length, sepal_width, petal_length, petal_width LABEL class INTO sqlflow_models.my_dnn_model; -- 2018/12/16 15:03:54 tensorflowCmd: run in Docker container Job success Predict sqlflow> SELECT * FROM iris.test PREDICT iris.predict.class USING sqlflow_models.my_dnn_model; -- 2018/12/16 15:05:58 tensorflowCmd: run in Docker container Job success The prediction is stored at iris.predict, you can look them up at sqlflow> SELECT * FROM iris.predict LIMIT 10; -- +--+-+--+-+-+ | sepal_length | sepal_width | petal_length | petal_width | class | +--+-+--+-+-+ | 6.4 | 2.8 | 5.6 | 2.2 | 2 | | 5 | 2.3 | 3.3 | 1 | 1 | | 4.9 | 2.5 | 4.5 | 1.7 | 2 | | 4.9 | 3.1 | 1.5 | 0.1 | 0 | | 5.7 | 3.8 | 1.7 | 0.3 | 0 | | 4.4 | 3.2 | 1.3 | 0.2 | 0 | | 5.4 | 3.4 | 1.5 | 0.4 | 0 | | 6.9 | 3.1 | 5.1 | 2.3 | 2 | | 6.7 | 3.1 | 4.4 | 1.4 | 1 | | 5.1 | 3.7 | 1.5 | 0.4 | 0 | +--+-+--+-+-+", - "url": "/sqlflow/cmd/demo/", - "relUrl": "/sqlflow/cmd/demo/" - }, - "17": { - "id": "17", - "title": "The Credit Card Fraud Detection Example", - "content": "The Credit Card Fraud Detection Example The downsample/creditcard.csv file in this directory comes from Kaggle. Down-sample To make it an example that can be checked into a git repo, I randomly picked up 2000 out from the 284,808 lines from the original dataset using the following command: gshuf -n 5000 /tmp/creditcard.csv > ./downsample/creditcard.csv where /tmp/creditcard.csv is the one downloaded from Kaggle. If you want to down-sample by yourself on macOS, you might need to install gshuf. brew install coreutils Import into MySQL To develop/test/debug the TensorFlow program automatically generated by SQLFlow, which is supposed to be able to read data from MySQL, we’d import the downsample/creditcard.csv into MySQL. Run MySQL Server For more about running MySQL in Docker containers, please refer to Run MySQL Server and Client in Docker Containers. The following command starts the MySQL server: docker run --rm -v $HOME/work/creditcarddb:/var/lib/mysql --name mysql01 -e MYSQL_ROOT_PASSWORD=root -e MYSQL_ROOT_HOST='%' -p 3306:3306 -d mysql/mysql-server:8.0 mysqld --secure-file-priv="" Please be aware that the above command doesn’t use the default entrypoint of the Docker image; instead, it explicitly start mysqld in the container so that it could provide the command line option --secure-file-priv="", which is required by mysqlimport. Run the following command to start a MySQL client: docker run --rm -it -v $HOME/work/creditcarddb:/var/lib/mysql mysql/mysql-server:8.0 mysql -uroot -p Remember to type the password root when prompted. Then input the following SQL program in the client to create the table creditcard.creditcard: CREATE DATABASE creditcard; USE creditcard; CREATE TABLE IF NOT EXISTS creditcard ( time INT, v1 float, v2 float, v3 float, v4 float, v5 float, v6 float, v7 float, v8 float, v9 float, v10 float, v11 float, v12 float, v13 float, v14 float, v15 float, v16 float, v17 float, v18 float, v19 float, v20 float, v21 float, v22 float, v23 float, v24 float, v25 float, v26 float, v27 float, v28 float, amount float, class varchar(255)); Before we could import the CSV file, we must follow the security policy of MySQL, which is wired and hard to understand, and copy the CSV file into the database directory: cp downsample/creditcard.csv ~/work/creditcarddb/creditcard/ The following command runs mysqlimport in the mysql/mysql Docker image to import the data: docker run --rm -it -v $HOME/work/creditcarddb:/var/lib/mysql -v $HOME/work/creditcard:/var/lib/mysql-files mysql:8.0 mysqlimport --ignore-lines=1 --fields-terminated-by=, --socket /var/lib/mysql/mysql.sock -u root -p creditcard creditcard.csv Please be aware that mysqlimport doesn’t exist in the mysql/mysql-server image. A message likes the following indicates the success of importing: creditcard.creditcard: Records: 1999 Deleted: 0 Skipped: 0 Warnings: 0 We can verify the imported result by running a SELECT statement like the following docker run --rm -it -v $HOME/work/creditcarddb:/var/lib/mysql mysql/mysql-server:8.0 mysql -uroot -p -e 'use creditcard; select time, v1, v28, amount, class from creditcard limit 10;' which should print something like the following +--+--++--+-+ | time | v1 | v28 | amount | class | +--+--++--+-+ | 123590 | -0.341154 | 0.130741 | 69 | "0" | | 135797 | -0.118853 | 0.11599 | 11.79 | "0" | | 161249 | -1.04129 | -0.362025 | 7.38 | "0" | | 72471 | 0.603151 | 0.0551408 | 352.04 | "0" | | 120182 | 2.10678 | -0.0566284 | 0.89 | "0" | | 160476 | 2.37223 | -0.0724277 | 6 | "0" | | 135023 | 1.87987 | 0.0214597 | 20.05 | "0" | | 117329 | -0.375213 | -0.130152 | 159 | "0" | | 148637 | 1.83178 | -0.0378951 | 38.37 | "0" | | 48293 | 1.26094 | 0.0275234 | 12 | "0" | +--+--++--+-+", - "url": "/sqlflow/example/creditcard/", - "relUrl": "/sqlflow/example/creditcard/" - }, - "18": { - "id": "18", - "title": "The MySQL Server Container for Testing", - "content": "The MySQL Server Container for Testing This directory contains a Dockerfile that builds a Docker image derived the MySQL Server 8.0 image, and includes SQL programs that popularize the following datasets: Churn from Kaggle Irises classfication from TensorFlow We can run a Docker container of it for unit testing. Build cd example/datasets docker build -t sqlflow:data . Run docker run --rm -d --name sqlflowdata -p 3306:3306 -e MYSQL_ROOT_PASSWORD=root -e MYSQL_ROOT_HOST=% sqlflow:data Popularize Datasets We need to manually popularize the databases and tables: docker exec -it sqlflowdata bash To popularize the Churn dataset into churn: cat /popularize_churn.sql | mysql -uroot -proot To popularize the Irises dataset into iris: cat /popularize_iris.sql | mysql -uroot -proot To prepare database for storing machine learning models: echo "CREATE DATABASE IF NOT EXISTS sqlflow_models;" | mysql -uroot -proot Query In the container, run echo "select count(*) from churn.test;" | mysql -uroot -proot should print the number of rows as the following count(*) 10 Trouble shooting: It usually takes about 15 seconds to bring up the MySQL Server. If you try to connect it before that, you may see the following error ERROR 1045 (28000): Access denied for user 'root'@'localhost' (using password: YES)", - "url": "/sqlflow/example/datasets/", - "relUrl": "/sqlflow/example/datasets/" - }, - "19": { - "id": "19", - "title": "Credit Card Fraud Detection", - "content": "Credit Card Fraud Detection Data collected from here To run: python premade_estimator.py, which gives Adding numeric_columns: Time Adding numeric_columns: V1 Adding numeric_columns: V2 Adding numeric_columns: V3 Adding numeric_columns: V4 Adding numeric_columns: V5 Adding numeric_columns: V6 Adding numeric_columns: V7 Adding numeric_columns: V8 Adding numeric_columns: V9 Adding numeric_columns: V10 Adding numeric_columns: V11 Adding numeric_columns: V12 Adding numeric_columns: V13 Adding numeric_columns: V14 Adding numeric_columns: V15 Adding numeric_columns: V16 Adding numeric_columns: V17 Adding numeric_columns: V18 Adding numeric_columns: V19 Adding numeric_columns: V20 Adding numeric_columns: V21 Adding numeric_columns: V22 Adding numeric_columns: V23 Adding numeric_columns: V24 Adding numeric_columns: V25 Adding numeric_columns: V26 Adding numeric_columns: V27 Adding numeric_columns: V28 Adding numeric_columns: Amount INFO:tensorflow:Using default config. WARNING:tensorflow:Using temporary folder as model directory: /tmp/tmpzj6rzutf INFO:tensorflow:Using config: {'_num_worker_replicas': 1, '_global_id_in_cluster': 0, '_save_checkpoints_steps': None, '_tf_random_seed': None, '_evaluation_master': '', '_protocol': None, '_session_config': allow_soft_placement: true graph_options { rewrite_options { meta_optimizer_iterations: ONE } } , '_model_dir': '/tmp/tmpzj6rzutf', '_device_fn': None, '_cluster_spec': <tensorflow.python.training.server_lib.ClusterSpec object at 0x7f40f9b40438>, '_save_summary_steps': 100, '_eval_distribute': None, '_keep_checkpoint_max': 5, '_master': '', '_train_distribute': None, '_log_step_count_steps': 100, '_num_ps_replicas': 0, '_experimental_distribute': None, '_save_checkpoints_secs': 600, '_service': None, '_task_type': 'worker', '_task_id': 0, '_keep_checkpoint_every_n_hours': 10000, '_is_chief': True} INFO:tensorflow:Calling model_fn. INFO:tensorflow:Done calling model_fn. INFO:tensorflow:Create CheckpointSaverHook. INFO:tensorflow:Graph was finalized. INFO:tensorflow:Running local_init_op. INFO:tensorflow:Done running local_init_op. INFO:tensorflow:Saving checkpoints for 0 into /tmp/tmpzj6rzutf/model.ckpt. INFO:tensorflow:loss = 892.3532, step = 1 INFO:tensorflow:global_step/sec: 295.764 INFO:tensorflow:loss = 0.00025717804, step = 101 (0.338 sec) INFO:tensorflow:global_step/sec: 446.454 INFO:tensorflow:loss = 0.00024299514, step = 201 (0.224 sec) INFO:tensorflow:global_step/sec: 444.479 INFO:tensorflow:loss = 0.00020988214, step = 301 (0.225 sec) INFO:tensorflow:global_step/sec: 421.979 INFO:tensorflow:loss = 0.00023327809, step = 401 (0.237 sec) INFO:tensorflow:global_step/sec: 446.172 INFO:tensorflow:loss = 0.00020405005, step = 501 (0.224 sec) INFO:tensorflow:global_step/sec: 431.785 INFO:tensorflow:loss = 0.00017921966, step = 601 (0.232 sec) INFO:tensorflow:global_step/sec: 442.912 INFO:tensorflow:loss = 0.00018488085, step = 701 (0.226 sec) INFO:tensorflow:global_step/sec: 444.628 INFO:tensorflow:loss = 0.00017795907, step = 801 (0.225 sec) INFO:tensorflow:global_step/sec: 428.385 INFO:tensorflow:loss = 0.00016896412, step = 901 (0.233 sec) INFO:tensorflow:Saving checkpoints for 1000 into /tmp/tmpzj6rzutf/model.ckpt. INFO:tensorflow:Loss for final step: 0.00016340206. INFO:tensorflow:Calling model_fn. WARNING:tensorflow:Trapezoidal rule is known to produce incorrect PR-AUCs; please switch to "careful_interpolation" instead. WARNING:tensorflow:Trapezoidal rule is known to produce incorrect PR-AUCs; please switch to "careful_interpolation" instead. INFO:tensorflow:Done calling model_fn. INFO:tensorflow:Starting evaluation at 2018-10-16-01:02:49 INFO:tensorflow:Graph was finalized. INFO:tensorflow:Restoring parameters from /tmp/tmpzj6rzutf/model.ckpt-1000 INFO:tensorflow:Running local_init_op. INFO:tensorflow:Done running local_init_op. INFO:tensorflow:Finished evaluation at 2018-10-16-01:02:49 INFO:tensorflow:Saving dict for global step 1000: accuracy = 1.0, accuracy_baseline = 1.0, auc = 0.9999999, auc_precision_recall = 0.0, average_loss = 0.0001784083, global_step = 1000, label/mean = 0.0, loss = 0.0016056747, precision = 0.0, prediction/mean = 0.00017826515, recall = 0.0 INFO:tensorflow:Saving 'checkpoint_path' summary for global step 1000: /tmp/tmpzj6rzutf/model.ckpt-1000 Test set accuracy: 1.00000 INFO:tensorflow:Calling model_fn. INFO:tensorflow:Done calling model_fn. INFO:tensorflow:Graph was finalized. INFO:tensorflow:Restoring parameters from /tmp/tmpzj6rzutf/model.ckpt-1000 INFO:tensorflow:Running local_init_op. INFO:tensorflow:Done running local_init_op. Prediction is "Not Fraud" (100.0%), expected "Not Fraud" Prediction is "Not Fraud" (100.0%), expected "Not Fraud" Prediction is "Not Fraud" (100.0%), expected "Not Fraud"", - "url": "/sqlflow/example/fraud_detection/", - "relUrl": "/sqlflow/example/fraud_detection/" - }, - "20": { - "id": "20", - "title": "sqlflowserver", - "content": "sqlflowserver The gRPC proxy server of SQL engines Prerequisite Golang Docker GRPC Running go generate && cd server && go run main.go", - "url": "/sqlflow/server/", - "relUrl": "/sqlflow/server/" - }, - "21": { - "id": "21", - "title": "SQLFlow", - "content": "SQLFlow Build, Run, Test Before running the unit tests, we need to build and run a Docker container that hosts a MySQL database following this guide. To build the parser using goyacc and run all unit tests, use the following command: go get -d ./... && goyacc -p sql -o parser.go sql.y && go test -v", - "url": "/sqlflow/sql/", - "relUrl": "/sqlflow/sql/" - }, - "22": { - "id": "22", - "title": "Python Code Template", - "content": "Python Code Template To run a test, say, test_sql_data.py, we need to Start a container that runs MySQL server with populated data following this guide. Run the tests in a SQLFlow container that has TensorFlow and mysql_connector installed: docker run --rm -it --network="host" -v $PWD:/work -w /work sqlflow/sqlflow python test_sql_data.py where --network="host" allows processes running in the container to access the host’s network, where the MySQL server container exposes its port.", - "url": "/sqlflow/sql/python/", - "relUrl": "/sqlflow/sql/python/" - }, - "23": { - "id": "23", - "title": "sqlfs", - "content": "sqlfs The package sqlfs provides an io.ReadCloser and an io.WriteCloser that treats a SQL database a filesystem, where each table in the database is like a file. The schema of the table is very simple – it has only one column of BLOB type. All the rows consist the storage. sqlfs provides the following features. Create a file To create a table named “hello” in a database “mydb” for writing, we can call Create. f, e := sqlfs.Create(db, "mydb.hello") f.Write([]byte("hello world! n")) f.Close() where db comes from a call to database/sql.Open. Append to a file f, e := sqlfs.Append(db, "mydb.hello") f.Write([]byte("hello world! n") f.Close() Read from a file f, e := sqlfs.Open(db, "mydb.hello") buf := make([]byte, 1024) f.Read(buf) f.Close() Remove a file DropTable(db, "mydb.hello") Check if a file exists HasTable(db, "mydb.hello") Other I/O operations Feel free to use standard packages io, ioutil, etc with sqlfs. For example, we can call io.Copy to copy everything from the standard input to a table. f, e := sqlfs.Create(db, "mydb.hello") io.Copy(f, os.Stdin) f.Close()", - "url": "/sqlflow/sqlfs/", - "relUrl": "/sqlflow/sqlfs/" - } - -} \ No newline at end of file + "0": { + "id": "0", + "title": "About Us", + "content": "About Add info about the product. Add info about the team. Add info about the contact.", + "url": "/pages/about.html", + "relUrl": "/pages/about.html" + }, + "1": { + "id": "1", + "title": "Contribute", + "content": "Contribute Dev, build and run!", + "url": "/doc_index/contribute.html", + "relUrl": "/doc_index/contribute.html" + }, + "2": { + "id": "2", + "title": "Design", + "content": "Design How we design SQLFLow", + "url": "/doc_index/design.html", + "relUrl": "/doc_index/design.html" + }, + "3": { + "id": "3", + "title": "SQLFlow: Bridging Data and AI", + "content": "", + "url": "/", + "relUrl": "/" + }, + "4": { + "id": "4", + "title": "Quick start", + "content": "Quick start SQLFlow is currently under active development. For those who are interested in trying it out, we have provided several demos. Play around with it. Any bug report and issue are welcomed. :) Setup Install Docker. Set up a MySQL server following example/datasets/README.md. Pull the latest SQLFlow Docker image: docker pull sqlflow/sqlflow:latest. Demo 1: Jupyter Notebook Start a Docker container that runs sqlflowserver and Jupyter Notebook. If you are using Docker for Linux, please change host.docker.internal:3306 to localhost:3306. docker run --rm -it -p 8888:8888 sqlflow/sqlflow:latest bash -c "sqlflowserver --db_user root --db_password root --db_address host.docker.internal:3306 & SQLFLOW_SERVER=localhost:50051 jupyter notebook --ip=0.0.0.0 --port=8888 --allow-root" If you are using Docker for Mac, please be aware the option --db_address host.docker.internal:3306 where host.docker.internal translates to the host ip address as recommended here. If you are running MySQL on remote, please be aware that MySQL only allows connections from localhost by default. Fix can be found here. Open a Web browser and direct to localhost:8888 and input the token. Then you can create notebooks. In a cell, you should be able to type %%sqlflow select 1 Explore more examples at example.ipynb Demo 2: Command Line Prompt Start a Docker container that runs SQLFlow command line prompt. If you are using Docker for Linux, please change host.docker.internal:3306 to localhost:3306. docker run -it --rm --net=host sqlflow/sqlflow:latest demo --db_user root --db_password root --db_address host.docker.internal:3306 You should be able to see the following prompt. sqlflow> Example Select data sqlflow> select * from iris.train limit 2; -- +--+-+--+-+-+ | SEPAL LENGTH | SEPAL WIDTH | PETAL LENGTH | PETAL WIDTH | CLASS | +--+-+--+-+-+ | 6.4 | 2.8 | 5.6 | 2.2 | 2 | | 5 | 2.3 | 3.3 | 1 | 1 | +--+-+--+-+-+ Train a Tensorflow DNNClassifier sqlflow> SELECT * FROM iris.train TRAIN DNNClassifier WITH n_classes = 3, hidden_units = [10, 20] COLUMN sepal_length, sepal_width, petal_length, petal_width LABEL class INTO sqlflow_models.my_dnn_model; -- ... Training set accuracy: 0.96721 Done training Prediction using a trained model sqlflow> SELECT * FROM iris.test predict iris.predict.class USING sqlflow_models.my_dnn_model; Checkout prediction result sqlflow> select * from iris.predict limit 10;", + "url": "/sqlflow/doc/quickstart.html", + "relUrl": "/sqlflow/doc/quickstart.html" + }, + "5": { + "id": "5", + "title": "Build TensorFlow from Source Code using Docker", + "content": "Build TensorFlow from Source Code using Docker To contribute to TensorFlow, we need to build TensorFlow from source code. The official guide is great. However, it interleaves the native building process and that using Docker and makes it confusing because packages needed by the former are not by the latter. Also, we found some useful tricks to start Docker containers in practices that are not in the official guide. Hence this document. Build the Pip Package in TensorFlow Development Container On either Mac or Linux, or any other OS, we don’t have to install and configure the building tools; instead, we can use a Docker image where all tools have been installed and properly configured. Get the Docker image containing all the building tools: docker pull tensorflow/tensorflow:latest-devel Then, let’s get the source code. On any OS, please install git using the native package manager. For example, on Ubuntu, please sudo apt-get install git or, on Mac, brew install git Then, use the git just installed, let’s clone tensorflow source code: git clone --recursive https://github.com/tensorflow/tensorflow cd tensorflow By default, we will be on the master branch. Feel free to do you change in your feature branches, or switch to a release branch, for example: git checkout v1.11.0 git checkout -b v1.11.0 Then, let us start a Docker container running the tensorflow/tensorflow:latest-devel image: docker run --rm -it -w /tensorflow -v $PWD:/tensorflow -v $HOME/.cache:/root/.cache -e "HOST_PERMS=$(id -u):$(id -g)" tensorflow/tensorflow:latest-devel /bin/bash -w /tensorflow brings us to the /tensorflow directory in the container once after we start it. -v $PWD:/tensorflow maps the current directory, which is the just cloned TensorFlow source directory on the host, to /tensorflow in the container. -v $HOME/.cache:/root/.cache maps the Bazel temporary directory on the host into the container, so the intermediate files generated by Bazel running in the container are actually saved on the host. This allows us to interrupt the container during the build and restart it later to resume the building. e "HOST_PERMS=$(id -u):$(id -g)" passes the user identification on the host into the container as an environment variable. We can reset the mode of files generated in the container to this user identity. From now on, we will be working in the container. Let us first configure the building: ./configure Usually, I would simply choose all the default options by hitting enter all the way down. Build the pip package builder: bazel build --config=opt //tensorflow/tools/pip_package:build_pip_package Build the pip package and save it into /tensorflow: ./bazel-bin/tensorflow/tools/pip_package/build_pip_package /tensorflow Change the generated wheel file into the usual mode as on the host: chown $HOST_PERMS /mnt/tensorflow-*.whl Install the Pip Pacakge in TensorFlow Development Container Let’s now try the new pip package. First, we need to uninstall the current tensorflow pip package, then we install the newly built one: pip uninstall tensorflow pip install /tensorflow/tensorflow-*.whl Now, we can verify if the new package works. First, we need to switch to a directory out from /tensorflow, so we don’t import from the source directory: cd /tmp # other directories also work then, we can try import the newly installed tensorflow package and verify it: python >>> import tensorflow >>> print(tensorflow.__version__) Now, let us quit from Python and from the Docker container. We should see the tensorflow-*.whl file on the host in the current directory. Install the Pip Package in a Clean Python Package After we quit from the development container, we should see the wheel file in the TensorFlow source directory on the host. Now, we can start a Python container and install the Pip package in it. Start the Python container docker run --rm -it -v $PWD:/tensorflow python:2.7 bash Install the pip pacakge in the container pip install /tensorflow/tensorflow*.whl Try TensorFlow by starting Python in the container python >>> import tensorflow as tf >>> print(tf.__version__)", + "url": "/sqlflow/doc/build-tensorflow.html", + "relUrl": "/sqlflow/doc/build-tensorflow.html" + }, + "6": { + "id": "6", + "title": "Canonical Development Environment", + "content": "Canonical Development Environment Referring to this example, we create a canonical development environment for Go and Python programmers using Docker. Editing on Host When we use this Docker image for daily development work, the source code relies on the host computer instead of the container. The source code includes this repo and all its dependencies, for example, the Go package google.golang.org/grpc. Code-on-the-host allows us to run our favorite editors (Emacs, VIM, Eclipse, and more) on the host. Please free to rely on editors add-ons to analyze the source code for auto-completion. Building in Container We build a Docker image that contains development tools: The Python interpreter The Go compiler The protobuf compiler The protobuf to Go compiler extension The protobuf to Python compiler extension Because this repo contains Go code, please make sure that you have the directory structure required by Go. On my laptop computer, I have export GOPATH=$HOME/go You could have your $GOPATH pointing to any directory you like. Given $GOPATH$ set, we could git clone the source code of our project by running: go get github.com/sql-machine-learning/sqlflow Change the directory to our project root, and we can use go get to retrieve and update Go dependencies. cd $GOPATH/src/github.com/sql-machine-learning/sqlflow go get -u -t ./... Note -t instructs get to also download the packages required to build the tests for the specified packages. As all Git users would do, we run git pull from time to time to sync up with others’ work. If somebody added new dependencies, we might need to run go -u ./... after git pull to update dependencies. To build this project, we need the protobuf compiler, Go compiler, Python interpreter, gRPC extension to the protobuf compiler. To ease the installation and configuration of these tools, we provided a Dockerfile to install them into a Docker image. To build the Docker image: docker build -t sqlflow:dev -f Dockerfile.dev . Development Build and Test We build and test the project inside the docker container. To run the container, we need to map the $GOPATH directory on the host into the /go directory in the container, because the Dockerfile configures /go as the $GOPATH in the container: docker run --rm -it -v $GOPATH:/go -w /go/src/github.com/sql-machine-learning/sqlflow sqlflow:dev bash Inside the Docker container, start a MySQL server in the background service mysql start& run all the tests as go generate ./... go install ./... go test -v ./... where go generate invokes the protoc command to translate server/sqlflow.proto into server/sqlflow.pb.go and go test -v builds and run unit tests. Release The above build process currently generates two binary files in $GOPATH/bin on the host. To package them into a Docker image, please run docker build -t sqlflow -f ./Dockerfile $GOPATH/bin To publish the released Docker image to our official DockerHub docker tag sqlflow sqlflow/sqlflow:latest docker push sqlflow/sqlflow:latest Demo: Command line Prompt The demo requires a MySQL server instance with populated data. If we don’t, we could follow example/datasets/README.md to start one on the host. After setting up MySQL, run the following inside the Docker container go run cmd/demo/demo.go --db_user root --db_password root --db_address host.docker.internal:3306 You should be able to see the following prompt sqlflow>", + "url": "/sqlflow/doc/build.html", + "relUrl": "/sqlflow/doc/build.html" + }, + "7": { + "id": "7", + "title": "Closing the producer goroutine from the consumer", + "content": "Closing the producer goroutine from the consumer The producer-and-consumer pattern is well used in Go concurrent programming. When the consumer stops, we want to gracefully stop the producer as well. Problem When a gRPC server receives a streaming request, it usually calls a function that returns a channel, reads the result from that channel and send the result to the client one by one. Take the following code for instance: upon receiving a request, the main goroutine Service calls launchJob. launchJob starts a separate goroutine as an anonymous function call and returns a channel. In the anonymous function, items will be sent to channel. And Service on the otherside of the channel will reads from it. func Service(req *Request, stream *StreamResponse) error { result := launchJob(req.Content) for r := range result { if e := stream.Send(result); e != nil { // should we signal the running goroutine so it will stop sending? return e } } } func launchJob(content string) chan Item { c := make(chan Item) go func() { defer close(c) acquireScarceResources() defer releaseScarceResources() ... // if stream.Send(result) returns an error and the Service returns, this will be blocked c <- Item{} ... }() return c } There is a major problem in this implementation. As pointed out by the comment, if the Send in Service returns an error, the Service function will return, leaving the anonymous function being blocked on c <- Item{} forever. This problem is important because the leaking goroutine usually owns scarce system resources such as network connection and memory. Solution: pipeline explicit cancellation Inspired by this blog post section Explicit cancellation, we can signal the cancellation via closing on a separate channel. And we can follow the terminology as io.Pipe. package sql import ( "errors" ) var ErrClosedPipe = errors.New("pipe: write on closed pipe") // pipe follows the design at https://blog.golang.org/pipelines // - wrCh: chan for piping data // - done: chan for signaling Close from Reader to Writer type pipe struct { wrCh chan interface{} done chan struct{} } // PipeReader reads real data type PipeReader struct { p *pipe } // PipeWriter writes real data type PipeWriter struct { p *pipe } // Pipe creates a synchronous in-memory pipe. // // It is safe to call Read and Write in parallel with each other or with Close. // Parallel calls to Read and parallel calls to Write are also safe: // the individual calls will be gated sequentially. func Pipe() (*PipeReader, *PipeWriter) { p := &pipe{ wrCh: make(chan interface{}), done: make(chan struct{})} return &PipeReader{p}, &PipeWriter{p} } // Close closes the reader; subsequent writes to the func (r *PipeReader) Close() { close(r.p.done) } // ReadAll returns the data chan. The caller should // use it as `for r := range pr.ReadAll()` func (r *PipeReader) ReadAll() chan interface{} { return r.p.wrCh } // Close closes the writer; subsequent ReadAll from the // read half of the pipe will return a closed channel. func (w *PipeWriter) Close() { close(w.p.wrCh) } // Write writes the item to the underlying data stream. // It returns ErrClosedPipe when the data stream is closed. func (w *PipeWriter) Write(item interface{}) error { select { case w.p.wrCh <- item: return nil case <-w.p.done: return ErrClosedPipe } } And the consumer and producer be can implemented as func Service(req *Request, stream *StreamResponse) error { pr := launchJob(req.Content) defer pr.Close() for r := range pr.ReadAll() { if e := stream.Send(r); e != nil { return e } } } func launchJob(content string) PipeReader { pr, pw := Pipe() go func() { defer pw.Close() if err := pw.Write(Item{}); err != nil { return } } return pr } Further Reading Google Form: Channel send timeout Go by Example: Timeouts Google I/O 2013 - Advanced Go Concurrency Patterns Go Concurrency Patterns Talk Go Concurrency Patterns: Pipelines and cancellation", + "url": "/sqlflow/doc/close_producer_from_consumer.html", + "relUrl": "/sqlflow/doc/close_producer_from_consumer.html" + }, + "8": { + "id": "8", + "title": "Compatibility with Various SQL Engines", + "content": "Compatibility with Various SQL Engines SQLFlow interacts with SQL engines like MySQL and Hive, while different SQL engines use variants of SQL syntax, it is important for SQLFlow to have an abstraction layer that hides such differences. SQLFlow calls Go’s standard database API. The submitter programs generated by SQLFlow call Python’s database API. Both APIs abstract the interface to various SQL engines; however, they are insufficient for SQLFlow to work. In this document, we examine all interactions between SQLFlow and the SQL engine so to identify what SQLFlow authors have to abstract in addition to calling Go’s and Python’s database APIs. Data Operations in Go Data Retrieval The basic idea of SQLFlow is to extend the SELECT statement of SQL to have the TRAIN and PREDICT clauses. For more discussion, please refer to the syntax design. SQLFlow translates such “extended SQL statements” into submitter programs, which forward the part from SELECT to TRAIN or PREDICT, which we call the “standard part”, to the SQL engine. SQLFlow also accepts the SELECT statement without TRAIN or PREDICt clauses and would forward such “standard statements” to the engine. It is noticeable that the “standard part” or “standard statements” are not standardized. For example, various engines use different syntax for joining. MySQL: SELECT pet.name, comment FROM pet, event WHERE pet.name =event.name; with keyword WHERE . Hive: SELECT pet.name, comment FROM pet JOIN event ON (pet.name =event.name) with keyword JOIN and ON. ODPS and SQLite use either INNER JOIN or OUTER JOIN. Fortunately, as SQLFlow forwards the above parts to the engine, it doesn’t have to care much about the differences above. Metadata Retrieval To verify the semantics of users’ inputs, SQLFlow needs to retrieve the schema of tables. For example, the input might be SELECT name, age, income FROM employee TRAIN DNNRegressor WITH hidden_layers=[10,50,10] COLUMN name, agee LABEL income; In the above example, the user misspelled the field name age in the COLUMN clause as “agee”. SQLFlow must be able to find that out. To do that, SQLFlow needs to query the field names from the SQL engine. However, different engines use various syntax. For example: MySQL: DESCRIBE/DESC employee; Hive: DESCRIBE FORMATTED employee; ODPS: DESC employee; SQLite: PRAGMA table_info([employee]); The returned data format varies too. Our solution to avoid such differences is not-to-use-them; instead, SQLFlow retrieves the table schema by running a query like SELECT * FROM employee LIMIT 1; and inferring field types using the mechanism called DatabaseTypeName provided by SQL engines drivers beneath the Go’s standard database API. Prepare Prediction Table A SQLFlow prediction job writes its prediction results into a table. It prepares the prediction table by Dropping previous prediction table DROP TABLE IF EXISTS my_table; Creating table with schema CREATE TABLE my_table (name1, type1, name2 type2); Most SQL engines, including MySQL, Hive, ODPS, SQLite, support both statements. Translate Database Column Type to TensorFlow Feature Column Type After retrieving database column type name through DatabaseTypeName, we can derive TensorFlow’s feature column type via a mapping such as {"FLOAT", "DOUBLE"} -> tf.numeric_column. Save Model SQLFlow saves trained ML model by dumping the serialized the model directory into a table. It first creates a table by CREATE TABLE IF NOT EXISTS %s (id INT AUTO_INCREMENT, block BLOB, PRIMARY KEY (id)) and insert blobs by INSERT INTO %s (block) VALUES(?). Note that Hive and ODPS doesn’t have BLOB type, we need to use BINARY (docs at ODPS, Hive) instead. Also, note that Hive and ODPS doesn’t support AUTO_INCREMENT, we need to implemented auto increment logic in sqlfs. Load Model SQLFlow loads trained ML model by reading rows in a table and deserializing the blob to a model directory. It reads rows by running SELECT block FROM %s ORDER BY id, which is supported by most databases. Data Operations in Python Connect to SQL Engines Thanks to the Python database API, connecting to different databases follows a similar API. conn = mysql.connector.connect(user='scott', password='password', host='127.0.0.1', database='employees') conn = sqlite3.connect('path/to/your/sqlite/file') conn = pyhive.connect('localhost') cursor = conn.cursor() cursor.execute('select * from my_table;') Insert Prediction Result into Prediction Table Python database API provides execute_many(sql, value) to insert multiple values at once. So one can prepare the following insertion statement. Please be aware that MySQL and SQLite use INSERT INTO to insert rows while Hive and ODPS use INSERT INTO TABLE. -- MySQL, SQLite INSERT INTO table_name VALUES (value1, value2, value3, ...); -- Hive, ODPS INSERT INTO TABLE table_name VALUES (value1, value2, value3, ...);", + "url": "/sqlflow/doc/database_abstraction_layer.html", + "relUrl": "/sqlflow/doc/database_abstraction_layer.html" + }, + "9": { + "id": "9", + "title": "Run MySQL Server and Client in Docker Containers", + "content": "Run MySQL Server and Client in Docker Containers The document explains how to setup MySQL in our development environment. Run MySQL Server in a Docker Container docker run --rm -v /tmp/test1:/var/lib/mysql --name mysql01 -e MYSQL_ROOT_PASSWORD=root -e MYSQL_ROOT_HOST='%' -p 3306:3306 -d mysql/mysql-server:8.0 the -v option ensures that the database is saved on the host. The default directory where MySQL saves the database is /var/lib/mysql. This directory can be configured in /etc/mysql/my.cnf, as explained in this post. By overlaying the directory /tmp/test1 on the host to /var/lib/mysql, we “cheat” MySQL to save databases on the host. So, we can kill the container and restart it, and the database is still there. Please be aware that the directory on the host must be empty the first time we run the above command; otherwise, MySQL would fail to initialize. I figured out this problem after several failures using docker logs. the -e option sets the root password of MySQL to “root”. Feel free to set it to any password you like. the second -e options sets MYSQL_ROOT_HOST to a wildcard so to allow clients connecting to the server via TCP/IP as the user “root”. This trick works with MySQL 5.7 and 8.0, but not the most recent under-development version. the --name option names the container to mysql01, which can be used to refer to this container. the -p option maps the port 3306, on which the MySQL server listens, to the same port on the host, so that clients could connect to the server via TCP/IP. Run MySQL Client in the Server Container docker exec -it mysql01 mysql -uroot -p This command executes the command mysql, which is the command line tool of MySQL, in the container named mysql01. The command line flags of mysql include -u, which specifies the username of MySQL, and -p, which makes MySQL prompts for the password. For this example, we should type the password “root”, which was set in the previous command. Please wait for a few seconds after the starting of the MySQL server container before we execute the client; otherwise, the startup of the client might fail. Once we get into the MySQL client, we can type SQL commands, e.g., show databases; create database yi; Run Client in a Different Container on the Same Host docker run --rm -it -v /tmp/test1:/var/lib/mysql mysql/mysql-server:8.0 mysql -uroot -p The -v option maps the database directory on the host to the client container. This mapping is necessary because, by default, the client talks to the server via Unix socket /var/lib/mysql/mysql.sock, which is /tmp/test1/mysql.sock on the host. Run Client in a Container on a Remote Host docker run --rm -it mysql/mysql-server:8.0 mysql -h 192.168.1.3 -P 3306 -uroot -p the -h option tells the client where the server is running on. In this example, the given IP is the one of the host where I ran the MySQL server container. Please be aware that the above command works only if the server allows remote connections. Run Python Client in a Container To write a Python client, we need to install the Python package mysql-connector-python. FROM python:2.7 RUN pip install mysql-connector-python Please be aware that some documents says that we need to install mysql-connector. I tried; but the mysql.connector.connect call failed with the error mysql.connector.errors.NotSupportedError: Authentication plugin 'caching_sha2_password' is not supported. Build the Docker image: docker build -t sqlflow . Run the image: docker run --rm -it sqlflow bash and we can start Python and run the following Python code snippet >>> import mysql.connector >>> db = mysql.connector.connect(user='root', passwd='root', host='192.168.1.3') >>> print(db) <mysql.connector.connection_cext.CMySQLConnection object at 0x7fbab9f3fed0> Run a Go Client In order to connect to a database, you need to import the database’s driver first. export GOPATH=$HOME/go go get -u github.com/go-sql-driver/mysql go run the following file package main import ( "database/sql" "github.com/go-sql-driver/mysql" "log" ) func main() { testConfig := &mysql.Config{ User: "root", Passwd: "root", Net: "tcp", Addr: "localhost:3306", } db, e := sql.Open("mysql", testConfig.FormatDSN()) if e != nil { log.Fatal(e) } defer db.Close() }", + "url": "/sqlflow/doc/mysql-setup.html", + "relUrl": "/sqlflow/doc/mysql-setup.html" + }, + "10": { + "id": "10", + "title": "Piping Responses", + "content": "Piping Responses Streaming Responses As described in the overall design, a SQLFlow job could be a standard or an extended SQL statemnt, where an extended SQL statement will be translated into a Python program. Therefore, each job might generate up to the following data streams: standard output, where each element is a line of text, standard error, where each element is a line of text, data rows, where the first element consists of fields name/types, and each of the rest is a row of data, status, where the element could be pending, failed, and succeeded. To create good user experience, we need to pipe these responses from SQLFlow jobs to Jupyter Notebook in real-time. Stages in the Pipe The pipe that streams outputs from SQLFlow jobs to the Jupyter Notebook consists of the following stages: Web browser ↑ | HTTP ↓ Jupyter Notebook server ↑ | ZeroMQ streams: Shell, IOPub, stdin, Controls, Heartbeat ↓ iPython kernel ↑ | IPython magic command framework ↓ SQLFlow magic command for Jupyter ↑ | gRPC ↓ SQLFlow server ↑ | Go channels ↓ SQLFlow job manager (Go functions) ↑ | IPC with Go's standard library ↓ SQLFlow jobs In the above figure, from the SQLFlow magic command to the bottom layer are our work. Streaming We have two alternative ideas: multiple streams and a multiplexing stream. We decided to use a multiplexing stream because we had a unsuccessful trial with the multiple streams idea: we make the job writes to various Go channels and forward each Go channel to a streaming gRPC call, as the following: Multiple streams The above figure shows that there are multiple streams between the Jupyter Notebook server and Jupyter kernels. According to the document, there are five: Shell, IOPub, stdin, Control, and Heartbeat. These streams are ZeroMQ streams. We don’t use ZeroMQ, but we can take the idea of having multiple parallel streams in the pipe. service SQLFlow { rpc File(string sql) returns (int id) {} rpc ReadStdout(int id) returns (stream string) {} rpc ReadStderr(int id) returns (stream string) {} rpc ReadData(int id) returns (stream Row) {} rpc ReadStatus(int id) returns (stream int) {} } However, we realized that if the user doesn’t call any one of the SQLFlow.Read... call, there would be no forwarding from the Go channel to Jupyter, thus the job would block forever at writing. A Multiplexing Stream Another idea is multiplexing all streams into one. For example, we can have only one ZeroMQ stream, where each element is a polymorphic type – could be a text string or a data row. service SQLFlow { rpc Run(string sql) returns (stream Response) {} } // Only one of the following fields should be set. message Response { oneof record { repeated string head = 1; // Column names. repeated google.protobuf.Any row = 2; // Cells in a row. string log = 3; // A line from stderr or stdout. } }", + "url": "/sqlflow/doc/pipe.html", + "relUrl": "/sqlflow/doc/pipe.html" + }, + "11": { + "id": "11", + "title": "Extended SQL Parser Design", + "content": "Extended SQL Parser Design This documentation explains the technical decision made in building a SQL parser in Go. It is used to parsed the extended SELECT syntax of SQL that integrates TensorFlow Estimators. Related Work Lexer and Parser Generator In 2001, when I was in graduate school, I defined an extended SQL syntax for querying distributed relational databases, as part of the course project of Distributed RDBMS by Prof. Li-zhu Zhou. I wrote the parser using bison (a modern version of yacc) and flex (a modern version of lex). yacc and lex generate C code; bison and flex generate C++ code. However, this time, I’d use Go. I surveyed goyacc, a standard Go tool. The usage is very similar to that of yacc and bison. However, the Go toolchain doesn’t provide a tool like lex/flex. Google revealed golex, which is out of maintenance. The Medium post recommends Ragel, which is a C++ program and could generate Go lexer; however, it lacks documents. Handwritten Lexer and Parser Some documents, including this one recommends handwriting lexers. However, it doesn’t explain how to write the parser. GoAcademy always provides high-quality tech blog posts. This one is from the author of InfluxDB. However, I stopped at where it explains wrapping a SQL statement as a string by an io.Reader, because it is obvious that we should keep the string as a string so that that token strings could refer to the same memory storage of the SQL statement. Following a link in the above GoAcademy post, I found Rob Pike’s excellent talk on how to write a lexer in Go in 2011. Many works after that change Rob’s implementation somehow but always lead to longer and less comprehensible codebases. The Choice Therefore, I wrote the lexer and parser both following Rob Pike’s idea. After few days work, I realized that: I should borrow the idea from Rob to represent SQL statements as strings, but not io.Reader as other work do, but no need to use channels and goroutines at all, and it is technically intractable to write a SQL lexer/parser manually. So, I switched to write a lexer manually, and to generate the parser using goyacc. During my work, I referred to this example and the official yacc manual for details about operator association and precedence.", + "url": "/sqlflow/doc/sql_parser.html", + "relUrl": "/sqlflow/doc/sql_parser.html" + }, + "12": { + "id": "12", + "title": "Submitter", + "content": "Submitter A submitter is a pluggable module in SQLFlow that is used to submit an ML job to a third party computation service. Workflow When a user types in an extended SQL statement, SQLFlow first parses and semantically verifies the statement. Then SQLFlow either runs the ML job locally or submits the ML job to a third party computation service. In the latter case, SQLFlow produces a job description (TrainDescription or PredictDescription) and hands it over to the submitter. For a training SQL, SQLFlow produces TrainDescription; for prediction SQL, SQLFlow produces PredDescription. The concrete definition of the description looks like the following {% raw %} type ColumnType struct { Name string // e.g. sepal_length DatabaseTypeName string // e.g. FLOAT } // SELECT * // FROM iris.train // TRAIN DNNClassifier // WITH // n_classes = 3, // hidden_units = [10, 20] // COLUMN sepal_length, sepal_width, petal_length, petal_width // LABEL class // INTO sqlflow_models.my_dnn_model; type TrainDescription struct { StandardSelect string // e.g. SELECT * FROM iris.train Estimator string // e.g. DNNClassifier Attrs map[string]string // e.g. {{"n_classes", "3"}, {"hidden_units", "[10, 20]"}} X []ColumnType // e.g. {{"sepal_length", "FLOAT"}, ...} Y ColumnType // e.g. {"class", "INT"} ModelName string // e.g. my_dnn_model } // SELECT * // FROM iris.test // PREDICT iris.predict.class // USING sqlflow_models.my_dnn_model; type PredDescription struct { StandardSelect string // e.g. SELECT * FROM iris.test TableName string // e.g. iris.predict ModelName string // e.g. my_dnn_model } {% endraw %} Submitter Interface The submitter interface should provide two functions Train and Predict. The detailed definition can be the following type Submitter interface { // Train executes a ML training job and streams job's response through writer. // A typical Train function should include // - Loading the training data // - Initializing the model // - model.train // - Saving the trained model to a persistent storage Train(desc TrainDescription, writer PipeWriter) error // Predict executes a ML predicting job and streams job's response through writer // A typical Predict function should include // - Loading the model from a persistent storage // - Loading the prediction data // - model.predict // - Writing the prediction result to a table Predict(desc PredictDescription, writer PipeWriter) error } Register a submitter A new submitter can be added as import ( ".../my_submitter" ".../sqlflow/sql" ) func main() { // ... sql.Register(my_submitter.NewSubmitter()) // ... for { sql := recv() sql.Run(sql) } } where sql.Register will put my_submitter instance to package level registry. During sql.Run, it will check whether there is a submitter registered. If there is, sql.Run will run either submitter.Train or submitter.Predict.", + "url": "/sqlflow/doc/submitter.html", + "relUrl": "/sqlflow/doc/submitter.html" + }, + "13": { + "id": "13", + "title": "SQLFlow: Design Doc", + "content": "SQLFlow: Design Doc What is SQLFlow SQLFlow is a bridge that connects a SQL engine, for example, MySQL, Hive, SparkSQL, Oracle, or SQL Server, and TensorFlow and other machine learning toolkits. SQLFlow extends the SQL syntax to enable model training and inference. Related Work We could write simple machine learning prediction (or scoring) algorithms in SQL using operators like DOT_PRODUCT. However, this requires copy-n-pasting model parameters from the training program into SQL statements. Some proprietary SQL engines provide extensions to support machine learning. Microsoft SQL Server Microsoft SQL Server has the machine learning service that runs machine learning programs in R or Python as an external script: CREATE PROCEDURE generate_linear_model AS BEGIN EXEC sp_execute_external_script @language = N'R' , @script = N'lrmodel <- rxLinMod(formula = distance ~ speed, data = CarsData); trained_model <- data.frame(payload = as.raw(serialize(lrmodel, connection=NULL)));' , @input_data_1 = N'SELECT [speed], [distance] FROM CarSpeed' , @input_data_1_name = N'CarsData' , @output_data_1_name = N'trained_model' WITH RESULT SETS ((model varbinary(max))); END; A challenge to the users is that they need to know not only SQL but also R or Python, and they must be capable of writing machine learning programs in R or Python. Teradata SQL for DL Teradata also provides a RESTful service, which is callable from the extended SQL SELECT syntax. SELECT * FROM deep_learning_scorer( ON (SELECT * FROM cc_data LIMIT 100) URL('http://localhost:8000/api/v1/request') ModelName('cc') ModelVersion('1') RequestType('predict') columns('v1', 'v2', ..., 'amount') ) The above syntax couples the deployment of the service (the URL in the above SQL statement) with the algorithm. Google BigQuery Google BigQuery enables machine learning in SQL by introducing the CREATE MODEL statement. CREATE MODEL dataset.model_name OPTIONS(model_type='linear_reg', input_label_cols=['input_label']) AS SELECT * FROM input_table; Currently, BigQuery only supports two simple models: linear regression and logistic regression. Design Goal None of the above meets our requirement. First of all, we want to build an open source software. Also, we want it to be extensible: We want it extensible to many SQL engines, instead of targeting any one of them. Therefore, we don’t want to build our syntax extension on top of user-defined functions (UDFs); otherwise, we’d have to implement them for each SQL engine. We want the system extensible to support sophisticated machine learning models and toolkits, including TensorFlow for deep learning and xgboost for trees. Another challenge is that we want SQLFlow to be flexible enough to configure and run cutting-edge algorithms, including specifying feature crosses. At the same time, we want SQLFlow easy to learn – at least, no Python or R code embedded in the SQL statements, and integrate hyperparameter estimation. We understand that a key to address the above challenges is the syntax of the SQL extension. To craft a highly-effective and easy-to-learn syntax, we need user feedback and fast iteration. Therefore, we’d start from a prototype that supports only MySQL and TensorFlow. We plan to support more SQL engines and machine learning toolkits later. Design Decisions As the beginning of the iteration, we propose an extension to the SQL SELECT statement. We are not going a new statement way like that BigQuery provides CREATE MODEL, because we want to maintain a loose couple between our system and the underlying SQL engine, and we cannot create the new data type for the SQL engine, like CREATE MODEL requires. We highly appreciate the work of TensorFlow Estimator, a high-level API for deep learning. The basic idea behind Estimator is to implement each deep learning model, and related training/testing/evaluating algorithms as a Python class derived from tf.estimator.Estimator. As we want to keep our SQL syntax simple, we would make the system extensible by calling estimators contributed by machine learning experts and written in Python. The SQL syntax must allow users to set Estimator attributes (parameters of the Python class’ constructor, and those of train, evaluate, or predict). Users can choose to use default values. We have a plan to integrate our hyperparameter estimation research into the system to optimize the default values. Though estimators derived from tf.estimator.Estimator run algorithms as TensorFlow graphs; SQLFlow doesn’t restrict that the underlying machine learning toolkit has to be TensorFlow. Indeed, as long as an estimator provides methods of train, evaluate, and predict, SQLFlow doesn’t care if it calls TensorFlow or xgboost. Precisely, what SQLFlow expect is an interface like the following: class AnEstimatorClass: __init__(self, **kwargs) train(self, **kwargs) evaluate(self, **kwargs) predict(self, **kwargs) We also want to reuse the feature columns API from Estimator, which allows users to columns of tables in a SQL engine to features to the model. Extended SQL Syntax Again, just as the beginning of the iteration, we propose the syntax for training as SELECT * FROM kaggle_credit_fraud_training_data LIMIT 1000 TRAIN DNNClassifier /* a pre-defined TensorFlow estimator, tf.estimator.DNNClassifier */ WITH layers=[100, 200], /* a parameter of the Estimator class constructor */ train.batch_size = 8 /* a parameter of the Estimator.train method */ COLUMN *, /* all columns as raw features */ cross(v1, v9, v28) /* plus a derived (crossed) column */ LABEL class INTO sqlflow_models.my_model_table; /* saves trained model parameters and features into a table */ We see the redundancy of * in two clauses: SELECT and COLUMN. The following alternative can avoid the redundancy, but cannot specify the label. SELECT * /* raw features or the label? */ corss(v1, v9, v28) /* derived featuers */ FROM kaggle_credit_fraud_training_data Please be aware that we save the trained models into tables, instead of a variable maintained by the underlying SQL engine. To invent a new variable type to hold trained models, we’d make our system tightly integrated with the SQL engine, and harms the extensibility to other engines. The result table should include the following information: The estimator name, e.g., DNNClassifier in this case. Estimator attributes, e.g., layer and train.batch_size. The feature mapping, e.g., * and cross(v1, v9, v28). Similarly, to infer the class (fraud or regular), we could SELECT * FROM kaggle_credit_fraud_development_data PREDICT kaggle_credit_fraud_development_data.class USING sqlflow_models.my_model_table; System Architecture A Conceptual Overview In the prototype, we use the following architecture: SQL statement -> our SQL parser --standard SQL-> MySQL -extended SQL-> code generator -> execution engine In the prototype, the code generator generates a Python program that trains or predicts. In either case, it retrieves the data from MySQL via MySQL Connector Python API, optionally, retrieves the model from MySQL, trains the model or predicts using the trained model by calling the user specified TensorFlow estimator, and writes the trained model or prediction results into a table. Working with Jupyter Notebook and Kubernetes The following figures shows the system components and their runtime environment. The left part shows how to run the system on a PC/laptop, the right part shows how to run it on a Kubernetes cluster.", + "url": "/sqlflow/doc/syntax.html", + "relUrl": "/sqlflow/doc/syntax.html" + }, + "14": { + "id": "14", + "title": "SQLFlow: Code Walkthrough", + "content": "SQLFlow: Code Walkthrough User Experience SQLFlow allows users to write SQL programs with extended syntax in Jupyter Notebook or a command-line tool. The following SQL statements train a TensorFlow model named DNNClassifier, which is a Python class derived from tf.estimator.Estimator: SELECT * FROM a_table TRAIN DNNClassifier WITH learning_rate=0.01 INTO sqlflow_models.my_model; And the following statement uses the trained model for prediction. SELECT * FROM b_table PREDICT b_table.predicted_label USING sqlflow_models.my_model; Please be aware that the part in the above statements before the extended keyword TRAIN and PREDICT is a standard SQL statement. This feature simplifies the implementation of the SQLFlow system. System Implementation If a SQL statement is of the standard syntax, SQLFlow throws it to the SQL engine and redirects the output to the user; otherwise, SQLFlow translates the statement of extended syntax into a Python program. Currently, it generates a program that throws the standard-syntax part of SELECT to MySQL, reads the results in the train-loop of a TensorFlow program. We will talk about how to extend SQLFlow to connect more SQL engines like Oracle, Hive, and SparkSQL, and generates more types of machine learning programs that calls distributed TensorFlow, PyTorch, and xgboost later. Before that, let us explain the system components. SQLFlow as a gRPC Server SQLFlow is a gRPC server, which can connect with multiple clients. A typical client is pysqlflow, the SQLFlow plugin for Jupyter Notebook server. Another once is a text-based client /cmd/sqlflowserver/main.go. Jupyter Notebook (SQL statements)--> SQLFlow gRPC server (SQLFlow magic command) <--(a stream of messages)-- The protobuf definition of the gRPC service is at /server/proto/sqlflow.proto. The return of the method SQLFlow.Run is a stream of Reponses, where each represents either a table header, a row, or a log message. The header and rows are usually from a standard SQL statement, for example, SELECT or DESCRIBE, and the log messages are usually from the running of a generated Python program. SQLFlow in the gRPC Server Once the SQLFlow server receives a batch of SQL statements via a gRPC call, it runs the following steps for each statement: the parser to generate parsing result, the verifier to verify the semantics given the parsing result, the code generator to generate a Python program, or the submitter, from the parsing result, the executor that runs the submitter locally. Step 3. and 4. are only for a SQL statement of extended syntax; otherwise, SQLFlow server proxies the standard-syntax statement to the SQL engine. The executor calls Go’s standard package that captures the stdout and stderr from the submitter process and passing the result back to the gRPC client. Therefore, it is the responsibility of the submitter to print log messages to its stderr and stdout. Minimal Viable Product In the minimal viable product (MVP) of SQLFlow, the code generator generates a Python program consists of two parts: throw the standard SELECT part in the extended-syntax statement to MySQL via ODBC, and a loop that reads outputs from the run of the SELECT statement and trains the model (or, using a trained model to predict). The training part calls TensorFlow to update the parameters of the model specified in the TRAIN clause. Extensibility By writing more code generators, we could extend SQLFlow to support more SQL engines, e.g., Hive and Oracle, and use machine learning toolkits, e.g., PyTorch and xgboost, in addition to TensorFlow, on various computing platforms. You are welcome to add more code generators such as codegen_distributed_tf.go to generate a submitter program similar to the MVP but runs a distributed TensorFlow training job. codegen_kubernetes_tf.go to launch a distributed TensorFlow job on a Kubernetes cluster, other than running locally, in the same container as where SQLFlow gRPC server resides. codegen_gcloud_pytorch.go to launch a submitter that calls PyTorch instead of TensorFlow for training on the Google Cloud. Job Management The current design of the gRPC interface assumes that the connection between the client, e.g., the Jupyter Notebook, and the SQLFlow server keeps alive during the running of the training program. This assumption is reasonable because even if the user closes her/his Web browser and disconnect to the Jupyter Notebook server, the connection between Jupyter to SQLFlow server might keep alive. However, this might not be robust enough if the Jupyter Notebook server runs on a user’s laptop and gets killed. In such a case, the gRPC server cannot stream the messages back to the client and would cause the failure of the submitter. A solution is to change the gRPC interface of SQLFlow server to have a method that files a job and returns immediately, and another method to get a batch of recent messages given a job ID. We will make a design for that soon.", + "url": "/sqlflow/doc/walkthrough.html", + "relUrl": "/sqlflow/doc/walkthrough.html" + }, + "15": { + "id": "15", + "title": "SQLFlow", + "content": "SQLFlow SQLFlow enriches SQL systems, e.g., MySQL, Hive, SparkSQL, with the capability of machine learning, using TensorFlow. We slightly extend the syntax of the SELECT statement of SQL to support model training and prediction. How to Use Quick start Extended SQL syntax How to Contribute Build from source code. The walkthrough of the source code The choice of parser generator Questions and Feedback Your feedback is our motivation to move on. Please let us know your questions, concerns, and issues by filing Github Issues. License Apache License 2.0", + "url": "/sqlflow/", + "relUrl": "/sqlflow/" + }, + "16": { + "id": "16", + "title": "SQLFlow Demo", + "content": "SQLFlow Demo Before you start, you need to set up test MySQL server as described here Start SQLFlow as go get -d ./... && go run main.go -logdir="/path/to/logs/" -loglevel="info" Regular SQL Statements sqlflow> SHOW DATABASES; +--+ | Database | +--+ | churn | | information_schema | | iris | | mysql | | performance_schema | | sqlflow_models | | sqlfs | | sys | +--+ sqlflow> SELECT * FROM iris.train LIMIT 1; +--+-+--+-+-+ | sepal_length | sepal_width | petal_length | petal_width | class | +--+-+--+-+-+ | 6.4 | 2.8 | 5.6 | 2.2 | 2 | +--+-+--+-+-+ Train sqlflow> SELECT * FROM iris.train TRAIN DNNClassifier WITH n_classes = 3, hidden_units = [10, 20] COLUMN sepal_length, sepal_width, petal_length, petal_width LABEL class INTO sqlflow_models.my_dnn_model; -- 2018/12/16 15:03:54 tensorflowCmd: run in Docker container Job success Predict sqlflow> SELECT * FROM iris.test PREDICT iris.predict.class USING sqlflow_models.my_dnn_model; -- 2018/12/16 15:05:58 tensorflowCmd: run in Docker container Job success The prediction is stored at iris.predict, you can look them up at sqlflow> SELECT * FROM iris.predict LIMIT 10; -- +--+-+--+-+-+ | sepal_length | sepal_width | petal_length | petal_width | class | +--+-+--+-+-+ | 6.4 | 2.8 | 5.6 | 2.2 | 2 | | 5 | 2.3 | 3.3 | 1 | 1 | | 4.9 | 2.5 | 4.5 | 1.7 | 2 | | 4.9 | 3.1 | 1.5 | 0.1 | 0 | | 5.7 | 3.8 | 1.7 | 0.3 | 0 | | 4.4 | 3.2 | 1.3 | 0.2 | 0 | | 5.4 | 3.4 | 1.5 | 0.4 | 0 | | 6.9 | 3.1 | 5.1 | 2.3 | 2 | | 6.7 | 3.1 | 4.4 | 1.4 | 1 | | 5.1 | 3.7 | 1.5 | 0.4 | 0 | +--+-+--+-+-+", + "url": "/sqlflow/cmd/demo/", + "relUrl": "/sqlflow/cmd/demo/" + }, + "17": { + "id": "17", + "title": "The Credit Card Fraud Detection Example", + "content": "The Credit Card Fraud Detection Example The downsample/creditcard.csv file in this directory comes from Kaggle. Down-sample To make it an example that can be checked into a git repo, I randomly picked up 2000 out from the 284,808 lines from the original dataset using the following command: gshuf -n 5000 /tmp/creditcard.csv > ./downsample/creditcard.csv where /tmp/creditcard.csv is the one downloaded from Kaggle. If you want to down-sample by yourself on macOS, you might need to install gshuf. brew install coreutils Import into MySQL To develop/test/debug the TensorFlow program automatically generated by SQLFlow, which is supposed to be able to read data from MySQL, we’d import the downsample/creditcard.csv into MySQL. Run MySQL Server For more about running MySQL in Docker containers, please refer to Run MySQL Server and Client in Docker Containers. The following command starts the MySQL server: docker run --rm -v $HOME/work/creditcarddb:/var/lib/mysql --name mysql01 -e MYSQL_ROOT_PASSWORD=root -e MYSQL_ROOT_HOST='%' -p 3306:3306 -d mysql/mysql-server:8.0 mysqld --secure-file-priv="" Please be aware that the above command doesn’t use the default entrypoint of the Docker image; instead, it explicitly start mysqld in the container so that it could provide the command line option --secure-file-priv="", which is required by mysqlimport. Run the following command to start a MySQL client: docker run --rm -it -v $HOME/work/creditcarddb:/var/lib/mysql mysql/mysql-server:8.0 mysql -uroot -p Remember to type the password root when prompted. Then input the following SQL program in the client to create the table creditcard.creditcard: CREATE DATABASE creditcard; USE creditcard; CREATE TABLE IF NOT EXISTS creditcard ( time INT, v1 float, v2 float, v3 float, v4 float, v5 float, v6 float, v7 float, v8 float, v9 float, v10 float, v11 float, v12 float, v13 float, v14 float, v15 float, v16 float, v17 float, v18 float, v19 float, v20 float, v21 float, v22 float, v23 float, v24 float, v25 float, v26 float, v27 float, v28 float, amount float, class varchar(255)); Before we could import the CSV file, we must follow the security policy of MySQL, which is wired and hard to understand, and copy the CSV file into the database directory: cp downsample/creditcard.csv ~/work/creditcarddb/creditcard/ The following command runs mysqlimport in the mysql/mysql Docker image to import the data: docker run --rm -it -v $HOME/work/creditcarddb:/var/lib/mysql -v $HOME/work/creditcard:/var/lib/mysql-files mysql:8.0 mysqlimport --ignore-lines=1 --fields-terminated-by=, --socket /var/lib/mysql/mysql.sock -u root -p creditcard creditcard.csv Please be aware that mysqlimport doesn’t exist in the mysql/mysql-server image. A message likes the following indicates the success of importing: creditcard.creditcard: Records: 1999 Deleted: 0 Skipped: 0 Warnings: 0 We can verify the imported result by running a SELECT statement like the following docker run --rm -it -v $HOME/work/creditcarddb:/var/lib/mysql mysql/mysql-server:8.0 mysql -uroot -p -e 'use creditcard; select time, v1, v28, amount, class from creditcard limit 10;' which should print something like the following +--+--++--+-+ | time | v1 | v28 | amount | class | +--+--++--+-+ | 123590 | -0.341154 | 0.130741 | 69 | "0" | | 135797 | -0.118853 | 0.11599 | 11.79 | "0" | | 161249 | -1.04129 | -0.362025 | 7.38 | "0" | | 72471 | 0.603151 | 0.0551408 | 352.04 | "0" | | 120182 | 2.10678 | -0.0566284 | 0.89 | "0" | | 160476 | 2.37223 | -0.0724277 | 6 | "0" | | 135023 | 1.87987 | 0.0214597 | 20.05 | "0" | | 117329 | -0.375213 | -0.130152 | 159 | "0" | | 148637 | 1.83178 | -0.0378951 | 38.37 | "0" | | 48293 | 1.26094 | 0.0275234 | 12 | "0" | +--+--++--+-+", + "url": "/sqlflow/example/creditcard/", + "relUrl": "/sqlflow/example/creditcard/" + }, + "18": { + "id": "18", + "title": "The MySQL Server Container for Testing", + "content": "The MySQL Server Container for Testing This directory contains a Dockerfile that builds a Docker image derived the MySQL Server 8.0 image, and includes SQL programs that popularize the following datasets: Churn from Kaggle Irises classfication from TensorFlow We can run a Docker container of it for unit testing. Build cd example/datasets docker build -t sqlflow:data . Run docker run --rm -d --name sqlflowdata -p 3306:3306 -e MYSQL_ROOT_PASSWORD=root -e MYSQL_ROOT_HOST=% sqlflow:data Popularize Datasets We need to manually popularize the databases and tables: docker exec -it sqlflowdata bash To popularize the Churn dataset into churn: cat /popularize_churn.sql | mysql -uroot -proot To popularize the Irises dataset into iris: cat /popularize_iris.sql | mysql -uroot -proot To prepare database for storing machine learning models: echo "CREATE DATABASE IF NOT EXISTS sqlflow_models;" | mysql -uroot -proot Query In the container, run echo "select count(*) from churn.test;" | mysql -uroot -proot should print the number of rows as the following count(*) 10 Trouble shooting: It usually takes about 15 seconds to bring up the MySQL Server. If you try to connect it before that, you may see the following error ERROR 1045 (28000): Access denied for user 'root'@'localhost' (using password: YES)", + "url": "/sqlflow/example/datasets/", + "relUrl": "/sqlflow/example/datasets/" + }, + "19": { + "id": "19", + "title": "Credit Card Fraud Detection", + "content": "Credit Card Fraud Detection Data collected from here To run: python premade_estimator.py, which gives Adding numeric_columns: Time Adding numeric_columns: V1 Adding numeric_columns: V2 Adding numeric_columns: V3 Adding numeric_columns: V4 Adding numeric_columns: V5 Adding numeric_columns: V6 Adding numeric_columns: V7 Adding numeric_columns: V8 Adding numeric_columns: V9 Adding numeric_columns: V10 Adding numeric_columns: V11 Adding numeric_columns: V12 Adding numeric_columns: V13 Adding numeric_columns: V14 Adding numeric_columns: V15 Adding numeric_columns: V16 Adding numeric_columns: V17 Adding numeric_columns: V18 Adding numeric_columns: V19 Adding numeric_columns: V20 Adding numeric_columns: V21 Adding numeric_columns: V22 Adding numeric_columns: V23 Adding numeric_columns: V24 Adding numeric_columns: V25 Adding numeric_columns: V26 Adding numeric_columns: V27 Adding numeric_columns: V28 Adding numeric_columns: Amount INFO:tensorflow:Using default config. WARNING:tensorflow:Using temporary folder as model directory: /tmp/tmpzj6rzutf INFO:tensorflow:Using config: {'_num_worker_replicas': 1, '_global_id_in_cluster': 0, '_save_checkpoints_steps': None, '_tf_random_seed': None, '_evaluation_master': '', '_protocol': None, '_session_config': allow_soft_placement: true graph_options { rewrite_options { meta_optimizer_iterations: ONE } } , '_model_dir': '/tmp/tmpzj6rzutf', '_device_fn': None, '_cluster_spec': <tensorflow.python.training.server_lib.ClusterSpec object at 0x7f40f9b40438>, '_save_summary_steps': 100, '_eval_distribute': None, '_keep_checkpoint_max': 5, '_master': '', '_train_distribute': None, '_log_step_count_steps': 100, '_num_ps_replicas': 0, '_experimental_distribute': None, '_save_checkpoints_secs': 600, '_service': None, '_task_type': 'worker', '_task_id': 0, '_keep_checkpoint_every_n_hours': 10000, '_is_chief': True} INFO:tensorflow:Calling model_fn. INFO:tensorflow:Done calling model_fn. INFO:tensorflow:Create CheckpointSaverHook. INFO:tensorflow:Graph was finalized. INFO:tensorflow:Running local_init_op. INFO:tensorflow:Done running local_init_op. INFO:tensorflow:Saving checkpoints for 0 into /tmp/tmpzj6rzutf/model.ckpt. INFO:tensorflow:loss = 892.3532, step = 1 INFO:tensorflow:global_step/sec: 295.764 INFO:tensorflow:loss = 0.00025717804, step = 101 (0.338 sec) INFO:tensorflow:global_step/sec: 446.454 INFO:tensorflow:loss = 0.00024299514, step = 201 (0.224 sec) INFO:tensorflow:global_step/sec: 444.479 INFO:tensorflow:loss = 0.00020988214, step = 301 (0.225 sec) INFO:tensorflow:global_step/sec: 421.979 INFO:tensorflow:loss = 0.00023327809, step = 401 (0.237 sec) INFO:tensorflow:global_step/sec: 446.172 INFO:tensorflow:loss = 0.00020405005, step = 501 (0.224 sec) INFO:tensorflow:global_step/sec: 431.785 INFO:tensorflow:loss = 0.00017921966, step = 601 (0.232 sec) INFO:tensorflow:global_step/sec: 442.912 INFO:tensorflow:loss = 0.00018488085, step = 701 (0.226 sec) INFO:tensorflow:global_step/sec: 444.628 INFO:tensorflow:loss = 0.00017795907, step = 801 (0.225 sec) INFO:tensorflow:global_step/sec: 428.385 INFO:tensorflow:loss = 0.00016896412, step = 901 (0.233 sec) INFO:tensorflow:Saving checkpoints for 1000 into /tmp/tmpzj6rzutf/model.ckpt. INFO:tensorflow:Loss for final step: 0.00016340206. INFO:tensorflow:Calling model_fn. WARNING:tensorflow:Trapezoidal rule is known to produce incorrect PR-AUCs; please switch to "careful_interpolation" instead. WARNING:tensorflow:Trapezoidal rule is known to produce incorrect PR-AUCs; please switch to "careful_interpolation" instead. INFO:tensorflow:Done calling model_fn. INFO:tensorflow:Starting evaluation at 2018-10-16-01:02:49 INFO:tensorflow:Graph was finalized. INFO:tensorflow:Restoring parameters from /tmp/tmpzj6rzutf/model.ckpt-1000 INFO:tensorflow:Running local_init_op. INFO:tensorflow:Done running local_init_op. INFO:tensorflow:Finished evaluation at 2018-10-16-01:02:49 INFO:tensorflow:Saving dict for global step 1000: accuracy = 1.0, accuracy_baseline = 1.0, auc = 0.9999999, auc_precision_recall = 0.0, average_loss = 0.0001784083, global_step = 1000, label/mean = 0.0, loss = 0.0016056747, precision = 0.0, prediction/mean = 0.00017826515, recall = 0.0 INFO:tensorflow:Saving 'checkpoint_path' summary for global step 1000: /tmp/tmpzj6rzutf/model.ckpt-1000 Test set accuracy: 1.00000 INFO:tensorflow:Calling model_fn. INFO:tensorflow:Done calling model_fn. INFO:tensorflow:Graph was finalized. INFO:tensorflow:Restoring parameters from /tmp/tmpzj6rzutf/model.ckpt-1000 INFO:tensorflow:Running local_init_op. INFO:tensorflow:Done running local_init_op. Prediction is "Not Fraud" (100.0%), expected "Not Fraud" Prediction is "Not Fraud" (100.0%), expected "Not Fraud" Prediction is "Not Fraud" (100.0%), expected "Not Fraud"", + "url": "/sqlflow/example/fraud_detection/", + "relUrl": "/sqlflow/example/fraud_detection/" + }, + "20": { + "id": "20", + "title": "sqlflowserver", + "content": "sqlflowserver The gRPC proxy server of SQL engines Prerequisite Golang Docker GRPC Running go generate && cd server && go run main.go", + "url": "/sqlflow/server/", + "relUrl": "/sqlflow/server/" + }, + "21": { + "id": "21", + "title": "SQLFlow", + "content": "SQLFlow Build, Run, Test Before running the unit tests, we need to build and run a Docker container that hosts a MySQL database following this guide. To build the parser using goyacc and run all unit tests, use the following command: go get -d ./... && goyacc -p sql -o parser.go sql.y && go test -v", + "url": "/sqlflow/sql/", + "relUrl": "/sqlflow/sql/" + }, + "22": { + "id": "22", + "title": "Python Code Template", + "content": "Python Code Template To run a test, say, test_sql_data.py, we need to Start a container that runs MySQL server with populated data following this guide. Run the tests in a SQLFlow container that has TensorFlow and mysql_connector installed: docker run --rm -it --network="host" -v $PWD:/work -w /work sqlflow/sqlflow python test_sql_data.py where --network="host" allows processes running in the container to access the host’s network, where the MySQL server container exposes its port.", + "url": "/sqlflow/sql/python/", + "relUrl": "/sqlflow/sql/python/" + }, + "23": { + "id": "23", + "title": "sqlfs", + "content": "sqlfs The package sqlfs provides an io.ReadCloser and an io.WriteCloser that treats a SQL database a filesystem, where each table in the database is like a file. The schema of the table is very simple – it has only one column of BLOB type. All the rows consist the storage. sqlfs provides the following features. Create a file To create a table named “hello” in a database “mydb” for writing, we can call Create. f, e := sqlfs.Create(db, "mydb.hello") f.Write([]byte("hello world! n")) f.Close() where db comes from a call to database/sql.Open. Append to a file f, e := sqlfs.Append(db, "mydb.hello") f.Write([]byte("hello world! n") f.Close() Read from a file f, e := sqlfs.Open(db, "mydb.hello") buf := make([]byte, 1024) f.Read(buf) f.Close() Remove a file DropTable(db, "mydb.hello") Check if a file exists HasTable(db, "mydb.hello") Other I/O operations Feel free to use standard packages io, ioutil, etc with sqlfs. For example, we can call io.Copy to copy everything from the standard input to a table. f, e := sqlfs.Create(db, "mydb.hello") io.Copy(f, os.Stdin) f.Close()", + "url": "/sqlflow/sqlfs/", + "relUrl": "/sqlflow/sqlfs/" + } + +} diff --git a/docs/doc/contribute.html b/docs/doc/contribute.html deleted file mode 100644 index 6dc7851..0000000 --- a/docs/doc/contribute.html +++ /dev/null @@ -1,358 +0,0 @@ - - - - - - - - - Contribute - Sqlflow - - - - - - - - - - - - - - - - - -
    -
    - Sqlflow - -
    - - -
    -
    -

    This site uses Just the Docs, a documentation theme for Jekyll.

    -
    -
    -
    -
    -
    - -
    -
    - - Search -
    -
    -
    - - - - -
    -
    -
    - - - -
    -

    Dev, build and run!

    - - - - -
    -

    Table of contents

    - - - -
    -
    -
    -
    - - - diff --git a/docs/doc/contribute/build-tensorflow.html b/docs/doc/contribute/build-tensorflow.html deleted file mode 100644 index c570145..0000000 --- a/docs/doc/contribute/build-tensorflow.html +++ /dev/null @@ -1,438 +0,0 @@ - - - - - - - - - Build TensorFlow from Source Code using Docker - Sqlflow - - - - - - - - - - - - - - - - - -
    -
    - Sqlflow - -
    - - -
    -
    -

    This site uses Just the Docs, a documentation theme for Jekyll.

    -
    -
    -
    -
    -
    - -
    -
    - - Search -
    -
    -
    - - - - -
    -
    -
    - - - - - -
    -

    Build TensorFlow from Source Code using Docker

    - -

    To contribute to TensorFlow, we need to build TensorFlow from source code. The official guide is great. However, it interleaves the native building process and that using Docker and makes it confusing because packages needed by the former are not by the latter. Also, we found some useful tricks to start Docker containers in practices that are not in the official guide. Hence this document.

    - -
    - -

    Build the Pip Package in TensorFlow Development Container

    - -

    On either Mac or Linux, or any other OS, we don’t have to install and configure the building tools; instead, we can use a Docker image where all tools have been installed and properly configured.

    - -
      -
    1. -

      Get the Docker image containing all the building tools:

      - - 
      docker pull tensorflow/tensorflow:latest-devel
-
      -
      2. -
    2. -

      Then, let’s get the source code. On any OS, please install git using the native package manager. For example, on Ubuntu, please

      - - 
      sudo apt-get install git
-
      - -

      or, on Mac,

      - - 
      brew install git
-
      - -

      Then, use the git just installed, let’s clone tensorflow source code:

      - - 
      git clone --recursive https://github.com/tensorflow/tensorflow
-cd tensorflow
-
      - -

      By default, we will be on the master branch. Feel free to do you change in your feature branches, or switch to a release branch, for example:

      - - 
      git checkout v1.11.0
-git checkout -b v1.11.0
-
      -
      3. -
    3. -

      Then, let us start a Docker container running the tensorflow/tensorflow:latest-devel image:

      - - 
      docker run --rm -it \
-   -w /tensorflow \
-   -v $PWD:/tensorflow \
-   -v $HOME/.cache:/root/.cache \
-   -e "HOST_PERMS=$(id -u):$(id -g)" \
-   tensorflow/tensorflow:latest-devel \
-   /bin/bash
-
      - -
        -
      • -w /tensorflow brings us to the /tensorflow directory in the container once after we start it.
        • -
      • -v $PWD:/tensorflow maps the current directory, which is the just cloned TensorFlow source directory on the host, to /tensorflow in the container.
        • -
      • -v $HOME/.cache:/root/.cache maps the Bazel temporary directory on the host into the container, so the intermediate files generated by Bazel running in the container are actually saved on the host. This allows us to interrupt the container during the build and restart it later to resume the building.
        • -
      • e "HOST_PERMS=$(id -u):$(id -g)" passes the user identification on the host into the container as an environment variable. We can reset the mode of files generated in the container to this user identity.
        • -
      -
      4. -
    4. -

      From now on, we will be working in the container. Let us first configure the building:

      - - 
      ./configure
-
      - -

      Usually, I would simply choose all the default options by hitting enter all the way down.

      -
      5. -
    5. -

      Build the pip package builder:

      - - 
      bazel build --config=opt //tensorflow/tools/pip_package:build_pip_package
-
      -
      6. -
    6. -

      Build the pip package and save it into /tensorflow:

      - - 
      ./bazel-bin/tensorflow/tools/pip_package/build_pip_package /tensorflow
-
      -
      7. -
    7. -

      Change the generated wheel file into the usual mode as on the host:

      - - 
      chown $HOST_PERMS /mnt/tensorflow-*.whl
-
      -
      8. -
    - -

    Install the Pip Pacakge in TensorFlow Development Container

    - -
      -
    1. -

      Let’s now try the new pip package. First, we need to uninstall the current tensorflow pip package, then we install the newly built one:

      - - 
      pip uninstall tensorflow
-pip install /tensorflow/tensorflow-*.whl
-
      -
      2. -
    2. -

      Now, we can verify if the new package works. First, we need to switch to a directory out from /tensorflow, so we don’t import from the source directory:

      - - 
      cd /tmp  # other directories also work
-
      - -

      then, we can try import the newly installed tensorflow package and verify it:

      - - 
      python
->>> import tensorflow
->>> print(tensorflow.__version__)
-
      -
      3. -
    3. -

      Now, let us quit from Python and from the Docker container. We should see the tensorflow-*.whl file on the host in the current directory.

      -
      4. -
    - -

    Install the Pip Package in a Clean Python Package

    - -

    After we quit from the development container, we should see the wheel file in the TensorFlow source directory on the host. Now, we can start a Python container and install the Pip package in it.

    - -
      -
    1. -

      Start the Python container

      - - 
      docker run --rm -it -v $PWD:/tensorflow python:2.7 bash
-
      -
      2. -
    2. -

      Install the pip pacakge in the container

      - - 
      pip install /tensorflow/tensorflow*.whl
-
      -
      3. -
    3. -

      Try TensorFlow by starting Python in the container

      - - 
      python
->>> import tensorflow as tf
->>> print(tf.__version__)
-
      -
      4. -
    - - - - -
    -
    -
    -
    - - - diff --git a/docs/doc/contribute/build.html b/docs/doc/contribute/build.html deleted file mode 100644 index 00d5886..0000000 --- a/docs/doc/contribute/build.html +++ /dev/null @@ -1,415 +0,0 @@ - - - - - - - - - Canonical Development Environment - Sqlflow - - - - - - - - - - - - - - - - - -
    -
    - Sqlflow - -
    - - -
    -
    -

    This site uses Just the Docs, a documentation theme for Jekyll.

    -
    -
    -
    -
    -
    - -
    -
    - - Search -
    -
    -
    - - - - -
    -
    -
    - - - - - -
    -

    Canonical Development Environment

    - -

    Referring to this example, -we create a canonical development environment for Go and Python programmers using Docker.

    - -

    Editing on Host

    - -

    When we use this Docker image for daily development work, the source code relies -on the host computer instead of the container. The source code includes this repo -and all its dependencies, for example, the Go package google.golang.org/grpc. -Code-on-the-host allows us to run our favorite editors (Emacs, VIM, Eclipse, and more) -on the host. Please free to rely on editors add-ons to analyze the source code -for auto-completion.

    - -

    Building in Container

    - -

    We build a Docker image that contains development tools:

    - -
      -
    1. The Python interpreter
      2. -
    2. The Go compiler
      3. -
    3. The protobuf compiler
      4. -
    4. The protobuf to Go compiler extension
      5. -
    5. The protobuf to Python compiler extension
      6. -
    - -

    Because this repo contains Go code, please make sure that you have the directory -structure required by Go. On my laptop computer, I have

    - -
    export GOPATH=$HOME/go
-
    - -

    You could have your $GOPATH pointing to any directory you like.

    - -

    Given $GOPATH$ set, we could git clone the source code of our project by running:

    - -
    go get github.com/sql-machine-learning/sqlflow
-
    - -

    Change the directory to our project root, and we can use go get to retrieve -and update Go dependencies.

    - -
    cd $GOPATH/src/github.com/sql-machine-learning/sqlflow
-go get -u -t ./...
-
    - -

    Note -t instructs get to also download the packages required to build -the tests for the specified packages.

    - -

    As all Git users would do, we run git pull from time to time to sync up with -others’ work. If somebody added new dependencies, we might need to run go -u ./... -after git pull to update dependencies.

    - -

    To build this project, we need the protobuf compiler, Go compiler, Python interpreter, -gRPC extension to the protobuf compiler. To ease the installation and configuration -of these tools, we provided a Dockerfile to install them into a Docker image. -To build the Docker image:

    - -
    docker build -t sqlflow:dev -f Dockerfile.dev .
-
    - -

    Development

    - -

    Build and Test

    - -

    We build and test the project inside the docker container.

    - -

    To run the container, we need to map the $GOPATH directory on the host into the -/go directory in the container, because the Dockerfile configures /go as -the $GOPATH in the container:

    - -
    docker run --rm -it -v $GOPATH:/go \
-    -w /go/src/github.com/sql-machine-learning/sqlflow \
-    sqlflow:dev bash
-
    - -

    Inside the Docker container, start a MySQL server in the background

    - -
    service mysql start&
-
    - -

    run all the tests as

    - -
    go generate ./...
-go install ./...
-go test -v ./...
-
    - -

    where go generate invokes the protoc command to translate server/sqlflow.proto -into server/sqlflow.pb.go and go test -v builds and run unit tests.

    - -

    Release

    - -

    The above build process currently generates two binary files in -$GOPATH/bin on the host. To package them into a Docker image, -please run

    - -
    docker build -t sqlflow -f ./Dockerfile $GOPATH/bin
-
    - -

    To publish the released Docker image to our official DockerHub

    -
    docker tag sqlflow sqlflow/sqlflow:latest
-docker push sqlflow/sqlflow:latest
-
    - -

    Demo: Command line Prompt

    - -

    The demo requires a MySQL server instance with populated data. If we don’t, we could -follow example/datasets/README.md to start one on the host. -After setting up MySQL, run the following inside the Docker container

    - -
    go run cmd/demo/demo.go --db_user root --db_password root --db_address host.docker.internal:3306
-
    - -

    You should be able to see the following prompt

    - -
    sqlflow>
-
    - - - - -
    -
    -
    -
    - - - diff --git a/docs/doc/contribute/close_producer_from_consumer.html b/docs/doc/contribute/close_producer_from_consumer.html deleted file mode 100644 index 5085ee8..0000000 --- a/docs/doc/contribute/close_producer_from_consumer.html +++ /dev/null @@ -1,456 +0,0 @@ - - - - - - - - - Closing the producer goroutine from the consumer - Sqlflow - - - - - - - - - - - - - - - - - -
    -
    - Sqlflow - -
    - - -
    -
    -

    This site uses Just the Docs, a documentation theme for Jekyll.

    -
    -
    -
    -
    -
    - -
    -
    - - Search -
    -
    -
    - - - - -
    -
    -
    - - - - - -
    -

    Closing the producer goroutine from the consumer

    - -

    The producer-and-consumer pattern is well used in Go concurrent programming. When -the consumer stops, we want to gracefully stop the producer as well.

    - -

    Problem

    - -

    When a gRPC server receives a streaming request, it usually calls a -function that returns a channel, -reads the result from that channel and send the result to the client one by one.

    - -

    Take the following code for instance: upon receiving a request, the main goroutine -Service calls launchJob. launchJob starts a separate goroutine as an anonymous -function call and returns a channel. In the anonymous function, items will be sent to -channel. And Service on the otherside of the channel will reads from it.

    - -
    func Service(req *Request, stream *StreamResponse) error {
-  result := launchJob(req.Content)
-  for r := range result {
-    if e := stream.Send(result); e != nil {
-      // should we signal the running goroutine so it will stop sending?
-      return e
-    }
-  }
-}
-
-func launchJob(content string) chan Item {
-  c := make(chan Item)
-  
-  go func() {
-    defer close(c)
-
-    acquireScarceResources()
-    defer releaseScarceResources()
-
-    ...
-    // if stream.Send(result) returns an error and the Service returns, this will be blocked
-    c <- Item{}
-    ...
-  }()
-  
-  return c
-}
-
    - -

    There is a major problem in this implementation. As pointed out by the comment, -if the Send in Service returns an error, the Service function will return, -leaving the anonymous function being blocked on c <- Item{} forever.

    - -

    This problem is important because the leaking goroutine usually owns scarce system -resources such as network connection and memory.

    - -

    Solution: pipeline explicit cancellation

    - -

    Inspired by this blog post section -Explicit cancellation, we can signal the cancellation via closing on a separate -channel. And we can follow the terminology as io.Pipe.

    - -
    package sql
-
-import (
-	"errors"
-)
-
-var ErrClosedPipe = errors.New("pipe: write on closed pipe")
-
-// pipe follows the design at https://blog.golang.org/pipelines
-// - wrCh: chan for piping data
-// - done: chan for signaling Close from Reader to Writer
-type pipe struct {
-	wrCh chan interface{}
-	done chan struct{}
-}
-
-// PipeReader reads real data
-type PipeReader struct {
-	p *pipe
-}
-
-// PipeWriter writes real data
-type PipeWriter struct {
-	p *pipe
-}
-
-// Pipe creates a synchronous in-memory pipe.
-//
-// It is safe to call Read and Write in parallel with each other or with Close.
-// Parallel calls to Read and parallel calls to Write are also safe:
-// the individual calls will be gated sequentially.
-func Pipe() (*PipeReader, *PipeWriter) {
-	p := &pipe{
-		wrCh: make(chan interface{}),
-		done: make(chan struct{})}
-	return &PipeReader{p}, &PipeWriter{p}
-}
-
-// Close closes the reader; subsequent writes to the
-func (r *PipeReader) Close() {
-	close(r.p.done)
-}
-
-// ReadAll returns the data chan. The caller should
-// use it as `for r := range pr.ReadAll()`
-func (r *PipeReader) ReadAll() chan interface{} {
-	return r.p.wrCh
-}
-
-// Close closes the writer; subsequent ReadAll from the
-// read half of the pipe will return a closed channel.
-func (w *PipeWriter) Close() {
-	close(w.p.wrCh)
-}
-
-// Write writes the item to the underlying data stream.
-// It returns ErrClosedPipe when the data stream is closed.
-func (w *PipeWriter) Write(item interface{}) error {
-	select {
-	case w.p.wrCh <- item:
-		return nil
-	case <-w.p.done:
-		return ErrClosedPipe
-	}
-}
-
    - -

    And the consumer and producer be can implemented as

    - -
    func Service(req *Request, stream *StreamResponse) error {
-  pr := launchJob(req.Content)
-  defer pr.Close()
-  for r := range pr.ReadAll() {
-    if e := stream.Send(r); e != nil {
-      return e
-    }
-  }
-}
-
-func launchJob(content string) PipeReader {
-	pr, pw := Pipe()
-	go func() {
-		defer pw.Close()
-		
-		if err := pw.Write(Item{}); err != nil {
-			return
-		}
-	}
-	return pr
-}
-
    - -

    Further Reading

    - -
      -
    1. Google Form: Channel send timeout
      2. -
    2. Go by Example: Timeouts
      3. -
    3. Google I/O 2013 - Advanced Go Concurrency Patterns
      4. -
    4. Go Concurrency Patterns Talk
      5. -
    5. Go Concurrency Patterns: Pipelines and cancellation
      6. -
    - - - - -
    -
    -
    -
    - - - diff --git a/docs/doc/contribute/mysql-setup.html b/docs/doc/contribute/mysql-setup.html deleted file mode 100644 index 2d5fe1a..0000000 --- a/docs/doc/contribute/mysql-setup.html +++ /dev/null @@ -1,442 +0,0 @@ - - - - - - - - - Run MySQL Server and Client in Docker Containers - Sqlflow - - - - - - - - - - - - - - - - - -
    -
    - Sqlflow - -
    - - -
    -
    -

    This site uses Just the Docs, a documentation theme for Jekyll.

    -
    -
    -
    -
    -
    - -
    -
    - - Search -
    -
    -
    - - - - -
    -
    -
    - - - - - -
    -

    Run MySQL Server and Client in Docker Containers

    - -

    The document explains how to setup MySQL in our development environment.

    - -

    Run MySQL Server in a Docker Container

    - -
    docker run --rm \
-   -v /tmp/test1:/var/lib/mysql \
-   --name mysql01 \
-   -e MYSQL_ROOT_PASSWORD=root \
-   -e MYSQL_ROOT_HOST='%' \
-   -p 3306:3306 \
-   -d \
-   mysql/mysql-server:8.0
-
    - -
      -
    • -

      the -v option ensures that the database is saved on the host. The default directory where MySQL saves the database is /var/lib/mysql. This directory can be configured in /etc/mysql/my.cnf, as explained in this post. By overlaying the directory /tmp/test1 on the host to /var/lib/mysql, we “cheat” MySQL to save databases on the host. So, we can kill the container and restart it, and the database is still there.

      - -

      Please be aware that the directory on the host must be empty the first time we run the above command; otherwise, MySQL would fail to initialize. I figured out this problem after several failures using docker logs.

      -
      • -
    • -

      the -e option sets the root password of MySQL to “root”. Feel free to set it to any password you like.

      -
      • -
    • -

      the second -e options sets MYSQL_ROOT_HOST to a wildcard so to allow clients connecting to the server via TCP/IP as the user “root”. This trick works with MySQL 5.7 and 8.0, but not the most recent under-development version.

      -
      • -
    • -

      the --name option names the container to mysql01, which can be used to refer to this container.

      -
      • -
    • -

      the -p option maps the port 3306, on which the MySQL server listens, to the same port on the host, so that clients could connect to the server via TCP/IP.

      -
      • -
    - -

    Run MySQL Client in the Server Container

    - -
    docker exec -it mysql01 mysql -uroot -p
-
    - -

    This command executes the command mysql, which is the command line tool of MySQL, in the container named mysql01.

    - -
      -
    • -

      The command line flags of mysql include -u, which specifies the username of MySQL, and -p, which makes MySQL prompts for the password. For this example, we should type the password “root”, which was set in the previous command.

      -
      • -
    • -

      Please wait for a few seconds after the starting of the MySQL server container before we execute the client; otherwise, the startup of the client might fail.

      -
      • -
    • -

      Once we get into the MySQL client, we can type SQL commands, e.g.,

      - - 
      show databases;
-create database yi;
-
      -
      • -
    - -

    Run Client in a Different Container on the Same Host

    - -
    docker run --rm -it \
-   -v /tmp/test1:/var/lib/mysql \
-   mysql/mysql-server:8.0 \
-   mysql -uroot -p
-
    - -
      -
    • The -v option maps the database directory on the host to the client container. This mapping is necessary because, by default, the client talks to the server via Unix socket /var/lib/mysql/mysql.sock, which is /tmp/test1/mysql.sock on the host.
      • -
    - -

    Run Client in a Container on a Remote Host

    - -
    docker run \
-   --rm -it \
-   mysql/mysql-server:8.0 \
-   mysql -h 192.168.1.3 -P 3306 -uroot -p
-
    - -
      -
    • the -h option tells the client where the server is running on. In this example, the given IP is the one of the host where I ran the MySQL server container.
      • -
    - -

    Please be aware that the above command works only if the server allows remote connections.

    - -

    Run Python Client in a Container

    - -

    To write a Python client, we need to install the Python package mysql-connector-python.

    - -
    FROM python:2.7
-RUN pip install mysql-connector-python
-
    - -

    Please be aware that some documents says that we need to install mysql-connector. I tried; but the mysql.connector.connect call failed with the error mysql.connector.errors.NotSupportedError: Authentication plugin 'caching_sha2_password' is not supported.

    - -

    Build the Docker image:

    - -
    docker build -t sqlflow .
-
    - -

    Run the image:

    - -
    docker run --rm -it sqlflow bash
-
    - -

    and we can start Python and run the following Python code snippet

    - -
    >>> import mysql.connector
->>> db = mysql.connector.connect(user='root', passwd='root', host='192.168.1.3')
->>> print(db)
-<mysql.connector.connection_cext.CMySQLConnection object at 0x7fbab9f3fed0>
-
    - -

    Run a Go Client

    - -

    In order to connect to a database, you need to import the database’s driver first.

    - -
    export GOPATH=$HOME/go
-go get -u github.com/go-sql-driver/mysql
-
    - -

    go run the following file

    - -
    package main
-
-import (
-	"database/sql"
-	"github.com/go-sql-driver/mysql"
-	"log"
-)
-
-func main() {
-	testConfig := &mysql.Config{
-		User:   "root",
-		Passwd: "root",
-		Net:    "tcp",
-		Addr:   "localhost:3306",
-	}
-	db, e := sql.Open("mysql", testConfig.FormatDSN())
-	if e != nil {
-		log.Fatal(e)
-	}
-	defer db.Close()
-}
-
    - - - - -
    -
    -
    -
    - - - diff --git a/docs/doc/contribute/submitter.html b/docs/doc/contribute/submitter.html deleted file mode 100644 index 4eb5847..0000000 --- a/docs/doc/contribute/submitter.html +++ /dev/null @@ -1,387 +0,0 @@ - - - - - - - - - Submitter - Sqlflow - - - - - - - - - - - - - - - - - -
    -
    - Sqlflow - -
    - - -
    -
    -

    This site uses Just the Docs, a documentation theme for Jekyll.

    -
    -
    -
    -
    -
    - -
    -
    - - Search -
    -
    -
    - - - - -
    -
    -
    - - - - - -
    -

    Submitter

    - -

    A submitter is a pluggable module in SQLFlow that is used to submit an ML job to a third party computation service.

    - -

    Workflow

    - -

    When a user types in an extended SQL statement, SQLFlow first parses and semantically verifies the statement. Then SQLFlow either runs the ML job locally or submits the ML job to a third party computation service.

    - -

    - -

    In the latter case, SQLFlow produces a job description (TrainDescription or PredictDescription) and hands it over to the submitter. For a training SQL, SQLFlow produces TrainDescription; for prediction SQL, SQLFlow produces PredDescription. The concrete definition of the description looks like the following

    - -
    type ColumnType struct {
-    Name             string // e.g. sepal_length
-    DatabaseTypeName string // e.g. FLOAT
-}
-
-// SELECT *
-// FROM iris.train
-// TRAIN DNNClassifier
-// WITH
-//   n_classes = 3,
-//   hidden_units = [10, 20]
-// COLUMN sepal_length, sepal_width, petal_length, petal_width
-// LABEL class
-// INTO sqlflow_models.my_dnn_model;
-
-type TrainDescription struct {
-    StandardSelect string       // e.g. SELECT * FROM iris.train
-    Estimator      string       // e.g. DNNClassifier
-    Attrs          map[string]string // e.g. {{"n_classes", "3"}, {"hidden_units", "[10, 20]"}}
-    X              []ColumnType // e.g. {{"sepal_length", "FLOAT"}, ...}
-    Y              ColumnType   // e.g. {"class", "INT"}
-    ModelName      string       // e.g. my_dnn_model
-}
-
-// SELECT *
-// FROM iris.test
-// PREDICT iris.predict.class
-// USING sqlflow_models.my_dnn_model;
-type PredDescription struct {
-    StandardSelect string // e.g. SELECT * FROM iris.test
-    TableName      string // e.g. iris.predict
-    ModelName      string // e.g. my_dnn_model
-}
-
    - -

    Submitter Interface

    - -

    The submitter interface should provide two functions Train and Predict. The detailed definition can be the following

    - -
    type Submitter interface {
-    // Train executes a ML training job and streams job's response through writer.
-    // A typical Train function should include
-    // - Loading the training data
-    // - Initializing the model
-    // - model.train
-    // - Saving the trained model to a persistent storage
-    Train(desc TrainDescription, writer PipeWriter) error
-    // Predict executes a ML predicting job and streams job's response through writer
-    // A typical Predict function should include
-    // - Loading the model from a persistent storage
-    // - Loading the prediction data
-    // - model.predict
-    // - Writing the prediction result to a table
-    Predict(desc PredictDescription, writer PipeWriter) error
-}
-
    - -

    Register a submitter

    - -

    A new submitter can be added as

    - -
    import (
-    ".../my_submitter"
-    ".../sqlflow/sql"
-)
-
-func main() {
-    // ...
-    sql.Register(my_submitter.NewSubmitter())
-    // ...
-    for {
-    	sql := recv()
-    	sql.Run(sql)
-    }
-}
-
    - -

    where sql.Register will put my_submitter instance to package level registry. During sql.Run, it will check whether there is a submitter registered. If there is, sql.Run will run either submitter.Train or submitter.Predict.

    - - - - -
    -
    -
    -
    - - - diff --git a/docs/doc/design.html b/docs/doc/design.html deleted file mode 100644 index 5061dc7..0000000 --- a/docs/doc/design.html +++ /dev/null @@ -1,358 +0,0 @@ - - - - - - - - - Design - Sqlflow - - - - - - - - - - - - - - - - - -
    -
    - Sqlflow - -
    - - -
    -
    -

    This site uses Just the Docs, a documentation theme for Jekyll.

    -
    -
    -
    -
    -
    - -
    -
    - - Search -
    -
    -
    - - - - -
    -
    -
    - - - -
    -

    How we design SQLFLow

    - - - - -
    -

    Table of contents

    - - - -
    -
    -
    -
    - - - diff --git a/docs/doc/design/database_abstraction_layer.html b/docs/doc/design/database_abstraction_layer.html deleted file mode 100644 index f9d97e6..0000000 --- a/docs/doc/design/database_abstraction_layer.html +++ /dev/null @@ -1,392 +0,0 @@ - - - - - - - - - Compatibility with Various SQL Engines - Sqlflow - - - - - - - - - - - - - - - - - -
    -
    - Sqlflow - -
    - - -
    -
    -

    This site uses Just the Docs, a documentation theme for Jekyll.

    -
    -
    -
    -
    -
    - -
    -
    - - Search -
    -
    -
    - - - - -
    -
    -
    - - - - - -
    -

    Compatibility with Various SQL Engines

    - -

    SQLFlow interacts with SQL engines like MySQL and Hive, while different SQL engines use variants of SQL syntax, it is important for SQLFlow to have an abstraction layer that hides such differences.

    - -

    SQLFlow calls Go’s standard database API. The submitter programs generated by SQLFlow call Python’s database API. Both APIs abstract the interface to various SQL engines; however, they are insufficient for SQLFlow to work. In this document, we examine all interactions between SQLFlow and the SQL engine so to identify what SQLFlow authors have to abstract in addition to calling Go’s and Python’s database APIs.

    - -

    Data Operations in Go

    - -

    Data Retrieval

    - -

    The basic idea of SQLFlow is to extend the SELECT statement of SQL to have the TRAIN and PREDICT clauses. For more discussion, please refer to the syntax design. SQLFlow translates such “extended SQL statements” into submitter programs, which forward the part from SELECT to TRAIN or PREDICT, which we call the “standard part”, to the SQL engine. SQLFlow also accepts the SELECT statement without TRAIN or PREDICt clauses and would forward such “standard statements” to the engine. It is noticeable that the “standard part” or “standard statements” are not standardized. For example, various engines use different syntax for joining.

    - -
      -
    • MySQL: SELECT pet.name, comment FROM pet, event WHERE pet.name =event.name; with keyword WHERE .
      • -
    • Hive: SELECT pet.name, comment FROM pet JOIN event ON (pet.name =event.name) with keyword JOIN and ON.
      • -
    • ODPS and SQLite use either INNER JOIN or OUTER JOIN.
      • -
    - -

    Fortunately, as SQLFlow forwards the above parts to the engine, it doesn’t have to care much about the differences above.

    - -

    Metadata Retrieval

    - -

    To verify the semantics of users’ inputs, SQLFlow needs to retrieve the schema of tables. For example, the input might be

    - -
    SELECT name, age, income FROM employee TRAIN DNNRegressor WITH hidden_layers=[10,50,10] COLUMN name, agee LABEL income;
-
    - -

    In the above example, the user misspelled the field name age in the COLUMN clause as “agee”. SQLFlow must be able to find that out.

    - -

    To do that, SQLFlow needs to query the field names from the SQL engine. However, different engines use various syntax. For example:

    - -
      -
    • MySQL: DESCRIBE/DESC employee;
      • -
    • Hive: DESCRIBE FORMATTED employee;
      • -
    • ODPS: DESC employee;
      • -
    • SQLite: PRAGMA table_info([employee]);
      • -
    - -

    The returned data format varies too. Our solution to avoid such differences is not-to-use-them; instead, SQLFlow retrieves the table schema by running a query like SELECT * FROM employee LIMIT 1; and inferring field types using the mechanism called DatabaseTypeName provided by SQL engines drivers beneath the Go’s standard database API.

    - -

    Prepare Prediction Table

    - -

    A SQLFlow prediction job writes its prediction results into a table. It prepares the prediction table by

    - -
      -
    1. Dropping previous prediction table DROP TABLE IF EXISTS my_table;
      2. -
    2. Creating table with schema CREATE TABLE my_table (name1, type1, name2 type2);
      3. -
    - -

    Most SQL engines, including MySQL, Hive, ODPS, SQLite, support both statements.

    - -

    Translate Database Column Type to TensorFlow Feature Column Type

    - -

    After retrieving database column type name through DatabaseTypeName, we can derive TensorFlow’s feature column type via a mapping such as {"FLOAT", "DOUBLE"} -> tf.numeric_column.

    - -

    Save Model

    - -

    SQLFlow saves trained ML model by dumping the serialized the model directory into a table. It first creates a table by CREATE TABLE IF NOT EXISTS %s (id INT AUTO_INCREMENT, block BLOB, PRIMARY KEY (id)) and insert blobs by INSERT INTO %s (block) VALUES(?).

    - -

    Note that Hive and ODPS doesn’t have BLOB type, we need to use BINARY (docs at ODPS, Hive) instead.

    - -

    Also, note that Hive and ODPS doesn’t support AUTO_INCREMENT, we need to implemented auto increment logic in sqlfs.

    - -

    Load Model

    - -

    SQLFlow loads trained ML model by reading rows in a table and deserializing the blob to a model directory.

    - -

    It reads rows by running SELECT block FROM %s ORDER BY id, which is supported by most databases.

    - -

    Data Operations in Python

    - -

    Connect to SQL Engines

    - -

    Thanks to the Python database API, connecting to different databases follows a similar API.

    - -
    conn = mysql.connector.connect(user='scott', password='password',
-                               host='127.0.0.1',
-                               database='employees')
-conn = sqlite3.connect('path/to/your/sqlite/file')
-conn = pyhive.connect('localhost')
-
-cursor = conn.cursor()
-cursor.execute('select * from my_table;')
-
    - -

    Insert Prediction Result into Prediction Table

    - -

    Python database API provides execute_many(sql, value) to insert multiple values at once. So one can prepare the following insertion statement. Please be aware that MySQL and SQLite use INSERT INTO to insert rows while Hive and ODPS use INSERT INTO TABLE.

    - -
    -- MySQL, SQLite
-INSERT INTO table_name VALUES (value1, value2, value3, ...);
--- Hive, ODPS
-INSERT INTO TABLE table_name VALUES (value1, value2, value3, ...);
-
    - - - - - -
    -
    -
    -
    - - - diff --git a/docs/doc/design/figures/sqlflow-arch.graffle b/docs/doc/design/figures/sqlflow-arch.graffle deleted file mode 100644 index 1a9b5418a59a55332c7765e59ae259df44f9b059..0000000000000000000000000000000000000000 GIT binary patch literal 0 HcmV?d00001 literal 3985 zcmV;C4{q=uiwFP!000030PS6CciOlX{XF>3YKiib|>wuyxOFbHff;kZ5M0Z zT-S!*e`>#ZVXpu6dTsMxyGJ|iug81Jz;nY$IsUYLc(9|azi2kM1_O^Zo4f5@ z<@oTR)mB6v&F0>Rb!GiLiUu#6&CAQnhApic4)06Duz4Kt0Slt5Ly_4Fk)Y9yy6d8t zQSO=Qh;O^D6TM#hWAlPty_Utfjva~XpGRh!P5HgJYzKiYKm4&7MuDjRYf)T-_kFh) z@Zq3wBtE?j?6WhE(amOJlCnnBYG6&&1sjx7oHE&Lj+4=CZ-3CYdu)df zMVUddHigO@y8gwi+>4cYZ})r_-Nrr?)o7hFktpju&54OXz**vFH}x}msp6L!F7M}2 zMr^^3{=g%)_Ha`|!ur1u~B3!^${3AN>&-lbN0bx?l);5nK-hXcgq6TA| z6a*((pLyJl(r}wx(NPA4%_)3GCG7LS{l$IT6ZLlexF_ZnBsfbtZb?DgThXuX{yXpM z7o@WPZhPMj_W73nar@&zYwvvP{OFTP(CL?Vp3~bA|Lz`rLf_6$_dk1E;`+`WX}9sa z@87}|Ea3!67Ty~~}q^Q(Wz&Z4_C4>ITHe{2f|ZMnbLyzImHdP(lTDMAfP zybOX>U8ja&Y$%3>uWu++y2S~T1kHr^35I!e{c_^=&X?LTZ9Z5=_M6T0FnGf3?vd|Z z-HhZinu~KD13DR3oyh&c4(%%z9Qa-Kb5i*!O=rf8aWXLiZFT<{hEXa(6bxB&YO`%S z7jg*pyYblMIw{jhB@$C<6;7{D{q?uHp{Umz%5Ow;YePXuHLf$xn&hzK*L?& zYhXJfqf{C(f_bksA>rZ=6Z2g+9C-Fs%dx$Z$|1$|l~;;i%_c3RMIJGi+!AT zvq?chxv)|QM!FWCi5$l zfk&#}G%JUYlH9XtCrJq@B#o7yWT4UnxYYibi!ZZ1k?OaUcN$b}kPU?p^SaP@V;c)m zX=7mZx&Xp-uBALs(rn?m-Q}0rrb`>m;)2;^kZs>td{b7+4i9`5oY-A=m@jPPB~rVl z+$7~2TS1V!i%A8=%*113gKSNc{$xGdxymIzv%N4GAtr`x+x6xYq?Tv@#qCDtWw(>{ z`J7u8d{h=*4FOo2hk(py;E>W5)NyqV(HJN-+~%6MkqI4b>!; zMkysk(=8(w(O6v43kv!MUTU3-IcIksGr3Wl{fx@eW}gX7Q+V0V8k?W6!qoK!D!}(b z1(JHJbs|Ce49DaWx|uoh)W6h*ZV}D0utl+{85mi4kBg^zd^E{UKLE@Yt=>$akBn%l zmr{buyhj%5l((2-s1rnIsAKz~8~zt{bGv5{O z&bb|gEK-%pSo?tu*`L?K*5Muxgz=x9fORn;onFA0FF%}mLnhv1tABUc&w(xG;oXV7 z_`Wl+gKj7KNvJCRB81~NYRWf^_4v(H<(sC(CYlwS=z45I)YwF%mTwfrCR7YmalUTg z+tDzvz0P+>ei6y%B+NgvornkO+FE>fp@x-?bEYZ-cfgdcd{`VP!O-ulwcjX%fcFBs ze?4dZnh>sFkTTHcyB5^b0_sVtuAb0p)zfTcRIo8k4C)Eg(>=@|Dyt`xkY}Wx)SA_k zZh?9N^#tk()YEF!lUi9l6&Y)}>S^|~>4FWYCs0rKvw~AuJy9Js41_Fd8k%O9hFloL zw4ov)s|doDX=$n^wUl8#F)5{+bxSFuij<;ALk2cf%hHiabx=~Eq#lEk+HbdyJ8M3V z*eT~1Af8q!p70&SldNGA4|fXNsXEYJr%GA*n7T%UJSfC20hIJ-~sTS^`J+HxuIyQAM-Q=pa;-< zfYAHGPL%}c+U4kr*`?H5iUzzKxaE(DXx z3&BOkT1oVaxhD`38#sXgE5QpRKwrvAuo#Y|Cr)6&Tj1~tqlxpPj_>XuI!cFuu% zf}L|9m>%xVIStn&m|*7|=%xG9OK3@Yi3&foP{Y)SCH^BSHY^-3ZBpy0sU=jmT2d>h zC8BDEhABprS_ZYS1#$`G(qoWI-&nxkf9$L|QQ$RQ4ti;&dZ~JNRZXo$FO`U)f{kWT z5KJJL?qSAohvm6A*oscnhGiN;E2)Mq=vC833v^HI2@|$z)=b!}@`Oo@SSBTgNmYXo zSf2yQ^r$wldfc%+Mf~px^~5~!Q8WefBal!ll~C1VtYWhk2~{GG7KvgH5(*?#g-OI+ z*XStH8yc}pim6GksuA)e*XXcOw}PTo6qJnMhD_CV#RLvvz#)w1d(t-sUL7ZWu@1lk;63Y; zzA^R)#;eC3#B>7eJw)t@S*?63rF7Qb+dbFW>35e0JaI3p2Oh+90^mJF;2k7#Q;d8* z^Ogy^)ng6;<^Xfg@R)rJIAoctM;t^?1H?T<#Jw91t|BqFTjHD!8Ad9$g+ZQXPzzM| zo;hSz4r-yPBzaL-@`WW|2x68Y@0?fHm{RpcaV@1hs(m--QvdoDFlx zva&%fsA)XS&2p&j&2os|A)tkxs3+Jh2hs^5T7X)be;B;7SVFWWv2>$Wf`AsFl@|Wd z#id9ktm>(F79*APq9ZC^5cdn@5(KdT`E<(_iR$tR z*CL)F3FH-5$JJx(B&KrG#UwoqLxnN~gFE8yp{;iSbgj$RStu8!yxCg2Wm z_pF0g#L%lFcm@1SJwWdn$F7KRS4Zp$__cb#-7^hckzj7+kb@9u0J-NFxgy3}9g!;_ zE*b#tnFg*Px++(Da1J~M5QhM}30?HL%y8#i=pG%!U5lAB)d8OZx@Ja=%5of>IR{fK6=%-@8Zn$yqkGOASFjI8LN@mY1U+Nw1XH9MdwgoK(H*ar|t$ys; z4(qe*CX$)mRsDxO5q`|W1zlMn&sFg%RwplVdtSXUbx<)i@G9^s@ap2cYRQ>1r6V?# z=OZb4pcHCoPmMxBngWFah2|-=&HRuDayj~;bEhFBDK=v2Gz5)Ws`XGGR=~a!poS+v z4Ttt1;sZ$s{lPGLU~Hfh>(NLl9cH628wI`uDfLju(m7=@(js9yM%EwfJ}V8+A&XSBvLwgii`#??YI&~1 z!o{of*=3gtWj~kv@3R>&G=ppg^jMmqz!IYVd>GB=i7nsl+Yu{4d?DbCv~b+O;qP64 z*A3IRpTr2sjAm2&789&$=XN-bg{N9bnmlVB1+Ewlaw%sipsC{>`ApL79PT{fmpLgg z-Gy3`t1xd3Y)3R{DnVRpEge)o8Ce*M+HREL?xa6AWh=?Q!rJ%P&g#MJjkb?nR?VTjZo!{k`X;@A`IT~MvK_F@!xAUY0W(=Tlg1zMN%l}5rpK7<%P%(-`8KkK$ z^`lYLl)tVSD;%s+P5#5k#qgaBBpE`Ps;$ZOlcAS-M3^@a&3S6ybU7ObaOx>)$=$d4 zpyk=8`%FFtnx+|fuQv7GgjYlN)MF>Cmxf$X<{_fXGv-JwYlMunNvJPZjx_mfd&5Fs r6vwW4QoNXZ*;g6jpPev*K7X=@wBJ|Ru2IHP{6b- diff --git a/docs/doc/design/figures/sqlflow-arch.png b/docs/doc/design/figures/sqlflow-arch.png deleted file mode 100644 index 0a5a20236049dee1546fa2f1357d6f3320edaf8b..0000000000000000000000000000000000000000 GIT binary patch literal 0 HcmV?d00001 literal 99876 zcmeFZWmuG3-!}|6bcaYYlpxX|-AFfzbax{SQUgdxry!k*l!SDr!T<_L2na|?iXdI@ z8r^m6=epl7_wgRb`{CIK8{uST&huQe*8l&jHBlOB^0-(OSV%}nxQYt*wUCe?QbLGFdXIyTgNt4q3krpbx>;HaYu%Uo)f|;w{Wv_^{{hsh9X|q%-qG(LyVpt zaiYKe`+c1rcGiEN$=Usn+X6SpiTDdAHwPEzU#|_GDvCHNtYziy;^>KZd2MGq4{;vR zUl0EG;lI!J`&lwBPOffN?(X1e;`jb|%0G|&`|;{-c2?m25T_C6{^RlgJofL$YuLJY zxPY;Cv$Ifi_ONmTC;Q{se_r7K{D?nZOOz8aT>lup-|x+@qu|XE#}eiI>x~x23LmRv zL_(58QoJwq&H<-+Iw!S(TM9Ov-7kM<4QKKH<@b`V=e+Hn}B1my?h% zi%4O_YJy8XT_LE)q%0bX9W7ovXAO5Qb=|j>EA-#g{l{ca2!yJ(5Um%%n*M{66$VzTgQ+~e`zrz8+m63f-$2Kk zvu8cDR?>-i%8X0|{rfAPc(LP{fU?o{W5MOo_RInN+0DB}ftP1={pMUmP?!hwjr}K9 zyU%qS`GxZFcs$2Zi5nU=lZg<8;J?4L5|L7&`ET`a)m~V2Z+N}hu!uWG$uGG%SwShvD3buj^B(dvCeLy79Gdj^I<&6AWF$;* z^p7veB%(YbMunbqE3PSzt;tvy&rsI_0nCKXYm81?9cB(WN&G#)z=;Ncu3TDA0x$B)NKTb zt&7C5i>$njFDq9|-ien4WCy@^2tP`g z(JH?_zk{7D4Y@Fm&1aEX8_oUBXFu_dTUjs5|8!r-8Z-7_(!ix@WB$qawzwI^Zr|Oe z^{mft9Yj#ENL)vR)~Rg1_C`^xNnD@CyC2QiILv+Ri{3izeu_PR5;>J8bt6`%)>?JF z{e0Iwh8^Wl!fhooJZS$tb2g2bpWy1yb8n>-mLxi6hK~aU-BdEpN|8eZm>t)S-!)BrV`edhZ@43A=t>ZbrUq!tvh8yy}h^)%JthsW2V; zqw14G_DtgQbf&KFx`YZ1;-p`U;`0?3X;A{Jx=fW4-a1SPqhgazNX#^R@hjtXU6foe zN>+1=sX$r>r=Q}>N8+9lv}Vn7Km4R4ay;iAKf(cXOx`FGR$J*)j0)aLPkA;f-IYkoAI(uN=l%nh;_kRxdVGXhl*AZL28u) zFMh}iz~`FW@4<%!7Pgy^x|l;YvNP>hZ7m+=NzwU)A`7@p@xJLYK>jl1p4xq)Y; z6jnP69UbXb{FPIo_~NhDu}Rrn=VJj|P+D5+3xd&Amo*>GIS+tzXsS$>+U(UqP+w z<DnN{@ASFoC4ZKuytu}*lYZWQgNlapSf6$?!+1WS3|^Z!L(fmz zc3QJ56(!t9cWM3ir1nV@m>7_W6qHb`@ATiBZ}vRc#0e|1+^y1>Z}H|d|NK=yD6n+5 zUp$S9{T=wWv$(3|(-iO9G+&@8z$#TB^--OoA1zY!`r=@$)<#{g&F}5@LWfCzEah{i zqH6r4zG1c~%?7PvC95Byp`N2u$MF(Z1hqJFF}MT*Uq-S;=G~vvJUMn2z}-qyTX2{x zNA$kC*{`!wJY^~K*nf2h0zBkeoSz5F<2}@Jkc}5vT!x78qkchFU?|JJ>w2%m5IRn zl=o)iW8-5B0sW~;qfhul4@=atS#$mOqsSOuYBhV<;ofMW)5tTut(eL0w9ovD5J!VQrYvf?lc5~9kWVx`oqhK0acp&z}(pN6LoUEhe{ zbpQPJQKn%;)^B4j#N_S9y4Ed&WTDzb%b!M_eP@w0)WXtf2d+ssw zgoh0+J#{V%U%L34>&vH0yvK?Zoj-5*ppd4lKo>vNpPc*gnt&b)BO_3Vqa;%{8zAMj z{Tjyneo~bPrGz;z*H<$sj4Rqq__@w``mL(WSD&J)dO{0(=jnV{B&-Jgx<_$T zLIujMrkM~leF2v*F+o@6Dct-Dp98Ne4pnE*m>=}DeI|YNobZcDa%pWfn=w7X5BC{m zf~0&7n)a#ccqSQiUMo{t<)rXES8ma{0ns^op(_1v*Y)l>I-iS!sm(zv&dUE{eLHgT z9b4bCo5WO0SR}VUz1GI2s4(w)Ce_3B7Xe%u{ZDOYS$&_4`}$0aY+c)!TfMrVK43Q z=Zm&+H4#+`en*z^5>;%qZU(wm`w04*nFD1`U5CY`)0BE^B=n0L>uy+YLw!z+2gK7D zsS^X(6e+rU`u1FD4_wPuTkW$jlib5qVE@OpA}$vdTrQCrLxv7ye@Km;Ed#>rhqjFt30K+s;-|aLka2V7(COT9$$?4ZS-Z?-1?maHkT&7bs z=_7IZGkI}z-^hEO34pIO@qn{;l}0Tq;grq^pif&XP}-dgf+BTyYpUuw0F6Tc3)6vd zPenn+`sSN2{dA?O<0=W8vsI%+C8NM<_e)y~JX=I(sL^ixw(92j_YR$gvdU(q1R4_V zls$}KZW22UxzoXm(}Pc!(VXoGZBLGyjU(`=$J!yX3cGVnM#tM~2}j_>H3uakzeflU zj8M)EpBM@v4QFkajZhjV~McEV|{@$uO#%fmS?>;8!53_XU* z*>b(w2`5lKHzzBNL@verPh16JG(!aY+aoMexZ_THwx{c2yL)j%lXZ2g?xA2Zx}Cdno`K^;h|T)y`ETS&M-V;4 zU(V-?e@)9V-fG-7_4E_1Ibk}WDr}^eg955Y+@r_U)1ol3BvzY{`yUbzJ-P7WgC50E zg=owsJcCz$4~^Dkv6`fx<(5bI(_xe$IRrT{k~5#*3m8`wXp$&;z3O zHbbrD=ep6W6h*|eg5X^{vklnR_V<4^SSb67)m7M5weNv-scv;a>-E*Sb4|L~{&wx? zrWSN6k(rLTlZdsDWjxT_#u0%g0lN_C;vwDZb7Q#s&7FWPGsGVNcwv#3fO%^b%HcMx z=RD#0i61soYZH&?#aj8&s76VieK&P1btl8PeiKC49wt~T;=KJB@kr;oKCNWJ{$R~4 z@mk<~Cjb>Lq<-#=S!n<3M-}?v)6Cf(NuPfPXX6$cq`cNHv)&tsA6J(!%6s@G$bTuk z_|BuNdX4gW!E#DRc}voZ>ONX4lWHsy196iMjnV3dHpjd3V0~&yD>` zAH05e$Md+svJZayV`IN=DdiM=Tcbh8*jIM`4ORaIo!8~gd%%?Vva033lFdT>c3Y^> zl)c_SdN>B8S^@dc5N{-xd8HQz%AaO!w;P#7X18^w+d|3lZP|=iE$LYd8#%i;uex zGR7@v4?Jk2hl!^dO0Bca7ZacL)lj^(|1`GpQav|^j&l2z+pJSVRP{!lNSSE_sRb=q zn&cE*>CYNYg>I5b_w7@fWbiw^pb)Omt1Y3>f0}zryp4AHb6mZz{MtQeme^gU?xj{o zU;qsHAzZwR2b)Sa&SE?0-Md7Fc~{Z*Lda|pT@UX&P~GH}c=H4w((e{KrbNnN$O5Y4R#Yeh0*S!EH`zinggsw#Z0upmzx@bLL&>yLiQ=IG%e&W3!C*s{8G3`!gcSw1wYV z&=TRIn5d{R4z)O#wR;)daz2phf$rtL4`1kjk_2(WY@wbmAjOIppn0OQ^Q05EUnTz} zxD)>wI*ZJn98E~W;rI6QN`$6Zmhxs&XJyWB{b7u_f<;y*jO>a0uGRgMCP;O`e zLZ|H~3eW(G6=n)GXtPzVmq)YYW4kl8FL7u@b^VTa+0&cMy^23deBqkJDtwXPceoif zbMH1QuP&cdU?b#1xCXy!j;P)&Di%Km&Tcn;tqh+n^ZsqWi3)Nq$D*-a{mmY@Teik!AKtcbpR;5wwB%ufUO*kFHIQVZ{-UMA= z;q10c!1sx42I8pO)r5p2R0%9nw8y!Ep4|&7bp4Ted$7T4N1s_Fs3Y;pi!UfWo86!u zf|ZmMSTcHBwP}4cH~y`HS|YtHRjs>5xSJTjE}J$u>%xF{x?jGzS(9<+YiED-&1o3- zGw!#F7EU2#GHr1#9elh>2t}yWh=P*y0X>_mz3vDMH$h4lMUPTSP# zE0g=YJm+NRx_sx!9d*+KqO$|$2N3Jmb~2p{m-PA0?Mew@q>Qq>0%8JoO3se`ugp=b zD7Z`|!`&*3+w@bRdzoBjAzL=MoApk!k5Xxt2jX=?U-v(XwWxRY!Y!oFnZfPnt_`B9 zH5}xw4drW#7e9NadiNvq|Kf@shOFRpY%~ceLkFv?XR!xYS{M?B3jihAjG&GvH${t_ zA)30QSrB$czWiDZEYBt_Y%0?w38e&qRGxt;@$fu40YC z%;>(9ROHy$#HgrwrGrB@7)DJ{t7@;6dRH6k#yR`@$&$pH*~Z6(pHH}yt!q@x57vhD z7c93qjqR})&?MVp_O3RupRO|!h@mJR8^+by8v}`>`NyYzY3u!|hl$cPzq1cOgZJAq4N> z2G>apDD}N1g7M?M0Ji%m8{}*{evdNW{)hJbrw*+k05E2eL3C~VzcgLMuRzIv1yLLi zw|V>D7{@$oj?AihUpgLSkV zB2>W(q|L3hc;(fZfv+JM89re^NT;z=c*IZ*0XS|m!^xXiIjIT3Xe2KXoj+)8BaPu) zxB0El#nzBGY0w&++Y=L9MYGdD(tqJSKuVGVQaP6?Hwz|&VGpQ#%Gtu6MW$ULZhvwB zxtV&WOg`HWlHzCI7<|7*F&6>wz~lGR_E5eRs7dysjmO^?Oz>#M%L5zc+@cLVKFYdn zPjd@YARz5{gV`d?6~Lk?E=qpK1RxnG{rK8fED1x=YVmT)5PV#4HWOdfex{^fXGj0U z$E!>vE{&>;#cVo*k>fN~zPb3zgyT`a?vS{?aK8kg+nHIX(r~lh=ZCo`?;ycTjeKdf zD&w}F-j`>G)P)zWXGc3*q8EETc>B^bjwe<_X+i+1KJZ?#Bf0$o0oVBgy{xoHSps_t z*b8)9Ba@}u;8MUQh{T0ORt;~{@6+T^XLRnC__p~~VH^Iy@Sho=54J9N6aRE1Hh|6J zQ9GU3++QWc+DG+Ytt5b1OGTKJ7T1G@#gh zZW8!-f93shdIk+13HhTuxMU*`?1T|a=D@5cyi|aMS;IDc=Jtq1yjY1d_AQPHuJ0Y! z?X@lo;$q+Dnrc-?X6o#c(=2);2{$KSX#%0BSZO3T=xQUJoX0Y<-=N*UX`!YWxzL9w zx<1nU2fZ5BArHdNf$~G>+xcsw-#zj1vP5agTsz!s<8bX`*t4RIMtTv|Elw&da8&(4 z?=<|UD*~{jmzLlnFKmUa(?EkWjp|VGY2dwrK6|F&MZ-0>0)ne_tGl(?>YV34<65K* zy}Aa{{d)6~B^dq9z~g!EG7;-kK9?_sqv&{4-_n_Q>d4SK%&n2(}<$R#@n!4-SERDNFiGKbZ= z=^2!!S9E$JFfv1v9x#!_+o~^KsS?(v$eQB%UtOGX#C z2p_VDv|sEi+s!prvr6ZE;8HeQ0gO~bHfp#Czy=%Msu_( z4TW$gTfAL>F0Pu^1BX<dECZ17k(` zda6W`H()_)%0J-55S830KQS&;!i~VaE_Cv-exfv>%m{Z}9%Z)l_Ui6<^>j+!`KlLM zGP(Tpma)e1^$_2Q}OMeb*1C zOuWUl&-Hy}Ie)xuzyK&a+VxpQp?|>} zp(fGG{#CI#&a}L^-(^St*!2b|ILg~WOEUw~>-}Jb-d#wb0s6Nu3h$>ny2U5idviA+ zuC>wii(LO0 zc}T{mPd}VkD@JdLyR9gAKG`gNX(*7zreF6?XgP*=0^iPDAOjFc;dCjKF193_9N&Hf zJPC)5^-beEPUA3oeKnJ+Emn9!_`fSNfEVQ~Ci?Z(P&BHUXNZr-@gf%z1d}paH@$GO zu>uYaip5Lo^Vx(080rz8bV-)apx*j0@Q>aniq8s$c-KgvHoDrzXBS{*++R zt`*bgM+uQrICq<(Y|9%JPC?|Vs~>pA2P@c1pX&GaXhT=F?z3qM6Ct3A{`~(0RbpCt zifmeHC;uB7@j{@HmfJQ@{~H<+L*Rg%oe};2f&(N!69P1{^CIi!f5AYopNxQ^_|=Oz z{08~|8M3jH2LPhZ!R=A{e~~fAk^x@XnAt7A_kRE&V8~M=vr3IszSlGRFCw#oIs$h6 ze|V!nsB937lOe#o9n3FjfLP-SP%0GWEZ6SGe;Gt|vj4C4&<`c$8X7uc!HuHil_`@T zTp`gL{o$P;TT!f%!Bef6RSIyQPnskU0FOYoKUwP&YX!*a39(7ix932@I8$Lz1d=e3 zeGD-kGUsu%t6-6ZE*O#OPvfm4_Z%iufL~i8^t4xp7FZ_;As|6g5VF5MdQJR&X^9|O zRkuzGhJt!UnK?n6$$GA9emXlhGr)!!IPfg^`MGX+x-XLvt5oY)Qn~y)2uH_-TZVXm z-@7GpE`G?Dz;majug~t9CksAc0_chjX4K-PnhgXy@zFh7F_6HX6ShJ@C!W$sDqs3@7NNq1Q?B@nqy}b9(6h|IqiIBxS?XIIfaH0^Se&bff zhsFp2`-yMyK$LvQW9O$y*C$R|F+yY}ylp`kVo{fXStLb^ogpiXwxW{Y;dsbb#e3xu zPsS|`0u@ZU*a}gc$XpWV+qG|TTVEkXajLG#q1Y*!=AIq|7}Q!1!c^G-iZF;Dq9$A0Gxd<6jQP+yoT z`D6G4yKW350?6;#r{v-OwiW4M0zkSFjy&GO*pMZ|J>C=c+@xaFulqKCJ4%2-<+gSK zFlzFOfv1clHe91pE1qUuF|;X{jFi@+A*!?(EAYh?5z<-jr*!eK9)7Xv_@#A5D3Gd! z^L7=P2nAl;LxMMu2X9a<-k`ZxMK>Wo#x>&m-mMZ1sbT)m#}+%|l=Aw?gH@Uww&RY| zgLM(PsFhaWCC`w)Kc>O!H+6bJKhHOPXKFypGYcdqF2uvPZcedRnkd;qd}xQRkY9|5 z8`jda9zK4Gir+_YM;VDy{br#3HU2TM2vzA;gJlmj%(8% za~CRit4B)&CoQo)LAdA?Jz!ZehDZh_Z#-BNw2v!;|7aItDOK;pBRHW_x~#9eXX~K$ z=B*ix`=!#}O%sPPJ<{$Z0t5HmfAU2*8MrGU zm_!QV1>XZK`ais&BDCTC#lqS*%Q)&aTgN!sVJSs1l%sf_fz#3{ulin_^yI!Z!S_@y znald_AM86|l5{Xzg2iEOL0ML*5MJ|0K?Ftpb;C*t7+u^Hpa2zMuJEGzj8av=t7oWU z;niJKDHHUdJJ`R~SsTl-r%FpNED#dslNWKn_$?fyV5~gT5>^TSM{|I_1w;Qx-re6G z?*jRJnrihoFTzJQbZOjmy)?UxQ7O+Vi37pHypL!fAB*1C{o{Rmi%G@T=VB381=k#= z;u_Q)`}NjM6LorqZ{bc+S9aVThHWJx1RX0BLp)+CmghYXn_hy)jGe1ozXlu}q-$J> zdelo&?EOB8sY=Y_o5owY0zM0a|U{vhj;`ovro!^^wrYp}96d1QOpHeeQi82OM1|Fhzck6V=?`{E6odiA=RBv2&>S}a8ErjX%S;jx@R}%QWI&E1PO8?fN!tJg30RxA5W02BodPzzTmD zTqNsWo3upKNn6|GzOGbl-uDxK7Q}|w4DAUyWW$+ZK^Ti?U7iqBjLN%)CJ%!zdLnER z2|M@apJLx4<+6CeeH(RJpwLnV;2+1%4;jB^IMfN_XE-i_3PUlNLnq8g=Yk*_10&%u z)HMz`)ZUn?%DoQu04*0o326$ylEF~DvC=g_Qvpf=84*2DVXN+b*!Jb4!f2VD2IkjK zX;y8&C%!bjsTZuN#Gt|1--3vU9Zksw<(9Eee(U*eOA?4VAku6rFmxIP8>(^<%=`C| zrf@514dokiT#dj>zkrxdQ;W>W@~T!9%Q8hy>8Bhb(u&|3qC0_`0M?7;WJ=|B?D5&WJ0G$Po%(u(x~Rz*6GTQ zxG@rXVPnjn<#W{6-+)FK{hs+7pgDR%iEMrahnqp3W@;0%5BE5&gU)7d&L9RoNlhfPn=vg-O4}`mbs>*#Keo;f{l#)CP!@fo!bg3jr67=SE%`z$%qL2->P> zSh>FXeqAK}6svg|sYGAKZ29g?@g+#w*->;{ot61t{21_SxBW;Tdj`B;B_M1S0htF8 z2|}zo@ILjck`CtoA=r6FY3zJ{>-oBYGMM>q-jQ%0K{pEeO*MZHQc`^9+Hj^ea!4^3 zi1MWXnb3-51;A`8u-IyT3H}R#=M~Vnn`~uGTa_&krI3Fa>7SuIwjZDyvkg_IVw~N; zCO4>T*%kI(juqg8gJdJ8`}#=gFQZ(XgcZ03;}`kR42huj^lax9pj8#V(v1@!Tsk+&1tW;ofd4+0jhz?Sv_E*aPd97o!T{)vAo4U;C*tu9ioiM`$@%W` zc%kaCMw+_Bvj#%*;l zj{EED$9rb=k08t$W_s7wv1t$z3 zJ$&`*L+l^(8wdG-*~CV)b*~#6$W92u5Hqd+K7G0#j2!w1gj?wl(k_Qy7fw7#tAAsA zhVY54M{_FBy&Sq6@!|9^Gi?m7VqXxQWdb1;{yUfbczi(4^_=HuKQYgCLW$5A_hPsX z04*>VV=#)Gu3jab{6V$3Y_?qpIb-?^Pr#W+_NR~;X4lXq-32|CjPcc9vM*T7;7fr7 znLmwD);UG&^=ykv5e$kD;xho?A!(oPvmjA8+NtsA7Fll5I*y9g~KnuX`OkncV zrLRIaoo_d6Hhf2)JKCM!+G6auzIqbAaDvc%|7Rc3Ulrg-hH%6_1<<_lLZook1UF6918chY-Pjy%mp_j0Zm)IKEcEq%e0Z z!Ptb&+ySEgEtbaskuU*0606z+NAB-^7e5f;_zPX$mwe1zmhT8Ii4TBqb|)p9sJaKK^h(hmRcaDsWrsQT?mA&ce~vd zJCP9)rhD5m_+pGM?WZNJK-L|5>!8!|W9bUus6k)>vw*C5;UYyF_=|RRh?7@!1qTUR zqlxLQbcCrI1Ji`H;Ttw2D!sB}z-RG}sn!8Gujo17_3RKPsdU^>zpf*6uiP$Y!RfeL zwUS&ts3J2eFZdJ@O4K1}z1e(@(_CX;d5oVCVH^rYI6dc;c+PVzz7$3bk+(jEcRa?O z@ejBJ!aIY;v{~sSP*M|#?mLMIszsJJaeaZ+Oy+hP6P}@y%;d+^=FJ`Ls9%j#TM33O zvjz2T$f3Mf1)UpQ?6y4SCX;N}croX_5U?5e^PA==2%13naZMK%dx_OkK&q)f zwucyyoNL0)ys2E%C;E;zu_?&_%_gZ0edg?R4nk(H_isE4129qtp$|~spXJ+&CM$#y ztLvjppLODp%-wZv-ngwj3B3fp%PxCVc#CZqPuawD@utCw^xnW*#pK(5>RRZ5MYu4H zT=zyf*EWb>xY?Es3;BW zjlJ5L$?i+unJSKfg)uM<^e*6eYRlw3QF<5Up{L0OX8sKeb%E1swP1c9y*!U)=Nx{z zi3s?EVp<@BWZ33GtPYAoi`v{77znRx#szo=mc-D7zwmd(q~{hPTaT>35bl{^F#>Na{(7)?m^{jZ9L#T?98w^T^=GWaNfjU#v^Go$Q%))o1^?PX#!3Y^8d2wuz6TN1OKyo-c>(Wt{dy zB}^zp0n4B#IPuhWOHD1NLslnf1YL)B8uLwR_t~{BX3U4KS9Fk2>1hZokxK!z`NK@w zKUy&XE=rWoNd6pRqkuQl83-PFDsO?k3fh0Thp}>>z9z}hpZHp31$%KCHoB&Hbx$*= zeZ3i2ZP{C=2+ctFu__f^HASc}0RNup6=oGzf?Z1^UrE^`vPHa`<#3%|a8+Vilyb!( z2sWbqz-$Q>dfoA?A~YA^V<3N$!DS|Gyji^w?I& z=c!4Z4;M9&f*0U~j8Cwu{c{+&OgoWEO*#V8{LE(SbrnIT*+W?^Q``~O84?K1gz&Q= zrY8bK0R<9qhyEZ_I=-Xy7*FKh&AX30yr07iUMu{*ujeWN1`du*O!u*-DG^ld4WY#0 zF)nI?bNSSn>9G^-F$L?-S2>R{?!`L_>2Jx>Cg807D%-u_2TXWYXFa_yB_7w@le0rVjwcv;Zi zR&WJ_4wA)Zmjpt7iV6JxXj2N%GXM~y#2=I>pde$e46a_+42_7HnrDb!7B7vmKj6Fi zeR3%572&(xrsV9FVZn5XZd5GdccW{GSPL_s7(|X;H&-}ugPfO>6*+d}8q4o#Vc1 zo#$_nc83@tRl~@Lxj0vT%W%TT?&@D1F31+r3$t1m`o#h-8V^A269i4g6Yl3SKt}9^ z2E0NlVGo5K6OcHI<8FPmZvEu`I;fOh#=fhmICh8_?Fi~E!^P%|iqd;o+-lok(ZHL4 zoR$rGj@%{XJ^Pgub_3FY0q*|fY>*}W5PJ+lXSPCEq-@^+ZJo{#^{v^mv5ptSrQmb* z+MZql)?eQp?bj)_0q0&*xqpMo4cTCekDUy%*HuL~2J$)o=;J;)k_+tteJo|-hXmOWIFXO`>3~v~-+@L`Xr1zK+y&QxQu}icvy%omZn92A8 zxpx&X4BX5Qz?EM4_OXx$v`hq_=pRrcjThaIyZ}*LbL&Kf>#IvE5X=XHGs*l{cfg`} z%zH5q+l;22>LPH+=@D^ML|Vx->(d|*6jni!o_IR}A?5vsgy8$0dax3~wc$6V-=MG` z!X8Mypg-97> z1)goZj;Ti8nkZ%9D!?j^NR6QJi^L+p4*eszg;+-kxM1$rAj{bNt_eSxg2~AeKyR=L zb{$h>#*9-a0@!bVEz<=nV0mzBp;!C&VW*i2WwnVj;ym?Gf2VAzYI=7$d(Z%da{B?&1oliC=TE zjTE74kUeJmv!5F^Af%jdAPf*zJ9~fTiVtSVtSR3S*F$qaQ!7lmEcd4Y@`1}eIj4fpKqCaqv? zMTO&B<45iEiAo~_kZCTPI-mDmsAw6H%s>`e&vO4c%lg1n6oFIB>xWa7IhyX%1fH02 z8ypM*cTE(^eR#b5t#WJVn2dBV1Lr1LTrBlWe)g5IrKV6DqgKW*qn|Xk`W{Qa zD!PbJMh*P-7%ib02t-Mx;x?7j<3tJzfzq}ts6sI@!9kI3M^n1{CnHfAS zJR zaMsil9sr|T^%@ik@1E%}u#>rX0~&B3Kg8{lkPY@b4*6|)vzWnY?+#p)x-Vxme9->_ zN}@2Z9K7az_wkYI9Kwz%<&<$?K@_8-GeA$Fu7&r_bOd$KF2ty<;ns)ud1hrV3iH>? z{tPs=_O`sm7A^91iN*~W=xl)=pxqO34b@Tcp1<$9`~tu429IF~h|Ac?;QiSlIhugL zIWOee{ZFYZEaXRTaTvbp!}*ydgH(v)5u<6o4ieh#a2(bFHi_}|xD27x{1>qOW86p& zIGjUufQ-cpPNY6?{FxU^K49&$;FF{6k2dw({M(?acMC^c3_q40>Br39$DXK!_3MER z3dQ4K)9yNGU=!8pnvYtRGhF-iVEw-JwUDjC#z`*g6VSFCN!Fv7b3SbK!S8#_`$(fK z5lMC^!4sY}`8}0v42WA2Q_BN1TI7OxFFd4&|=|8V(XL1Pt={uUt3&Kg3qJ`s=W40h>t~o%*R$~%+TO=B?6qt z#CScCilDE2bb*$1NF&Bg$R&zmHFb2<3earKMe?WTWH}ND8Ot5)7HAwLuqe{`2ze4- zZrC)&&0F~MH1Za*;5DFoKS!T>K4wXooBc1%8yhwSgUX=RwId)GWX~E1~wWG|NTETL-UYHWm&p@z5mP%Y2K0 zxX-jPu}n_O8UuP={k{rp)AvHSGE$*nt6^grdcfE;OojmHpZAfDg+M9n#;XK ze(znwDcl>17D1A-oJAC*O;D`~h znOhnw^xY%OkbNLJtPXO-F^oYLy5U1eSaX<1%8ch0;g9VlV~48@nAAUaC}{bHHiuJ zkHnJ>u}bA3aVvlLsm^niCC5s|yyBQe5M%OcXVWF?=~2bQ(j*IY zJ9-f=a1`L5{21)ltJ>8?LxvGSMjJr{N_QR|pQd~_G0-(@uY;+z`s*$G7Hb-td)y(tKZSy0{;FWOw=DRI+Q05lU-K`V_GbWM;SC{EKsBMm619R|Zq1W?ZB5 z)mSA^x)2*nDEi?O1ci?8YeG@$SHseg^TLSM5Zh|@02NtTjRC#*UqE-2m_?&HG(FvssXk^zowdDegbsZ z+Qk!qB#HohO%n0mWearel%YNUJetAF9NnQ^*J-gnlC5`}+NxYnSH*#17vx;S>{44m4^2H?1+tnIitI9YZRJ7tIXSp!9m-`ZAYQCNH57 zfIhI9qyO^KJQ@7~6ReINi3#1>4B;H$Q3)G&B2JXyHq(ny9 zWH(-7UV^ptD7PP4P8Rxxe7|Ao3|WbrNq;xoppBOKgq4Vd@wi2 z!&{*~N2kv&oi3NqpGHAeEVT_{NPA-Gj5ES(ma^w{o+~v#lzlOsK+o2sVdC5ZR*&gNDiy)a5dM z>hhQT`1bR-X#ESq-n->yZwggjnGxXmo;xQxOj%kqAUz90v88p_(25AY4llZkS#$^4 zfq3O*Zs4FKQcIFvCe+6m^~ng7XIrm}BQisKDcuSpmFpg9fm_(EHUOv&bs^42qX{^> ztoMDHbd7_7M~9J94K5}&Us`=ffSWeq+H~5p;hgLM!v97XjwWX4gtcLO(MCzvPcbgu zKkmMEY`8D1hVj${Oe&Wwt8=f}dM8h~b(*-HNMWvdw7UZ_B;fW3!MiDsjadRFK0~Sv zR=T0A+XH6XR`V#w{AWqQK`7sPK+r(47jkjT)?C64lf<`Kptdmf{~J`{!hAO9QMmE&nQy<9Lx))hWJ1UkVmyl{@|i3 z6j6h=M2Huzh4`cx9MXLm%w#>x5S~00YsL7`YzG!S%!_%Bwdf>zNDV)|kpeUpb6UdZ zDh1ju(jT^u(}l+g{L;zwgbcdUhd0F536PWsSk&@dyK4LA^I{Sto6PGwsVK-wc-jorN@bV?um+pxe zqI_Fe%d?b$E}M@!q7*mro*BMiFgI6JBJ5upgX_s&2RDR_xC%BdBFGC|1m(T3_j7sawJO#;Ts5fZR;vcG8tjV z9!a<4Uw(+QRj<4IOXKLKbCni^jq{aTFv@E$G`Qr(P>Vc5csqzsl>jo~ATTi8Q*JYW z9Hk{=!d)GXJ4VfQ6XkjjfS1Go_6?-b%RaNjA$e&MG&-?#c~7miFhMo$`++9Qjbtbc z2O{Y8E;v*|0lEoEvYR6~9OUk*V@Alh90>?>*Ch*M<7j{mp#y z{{PNA4|5M5PF?%jd#ztB2+!%HP3A$qfVemClK1>)Ai2U2?y2)oxFkXtk7hU_4e7qh zOk${BJDyFJo{7IMfLQF>xhnBs9iP3&-cBUa+TL+4t`9k#&3;7F3Mhf)fvep5V-o1w z-te)VhruSHL`->Vv-WCmi2jVLN7HLckIQJQTgyo4m(sCvQWGRe;gL zxQIBU{dXp)p1pglJc4z!RUDX)I%m@8_sK~_Hz*y0Ug*vQVH6sb#*9vd)DoroHXq=s zze3TKjyfk7_4ooF)>#ek!#)ATwsMVM&#-kqM5FQL0`>HOJVSNl?R$y{R z89)6-D;m6Z#T>Ae%tIiTUCU3mdL)0Y)Fiz_yf66-%j)K7ak(IiyKe-J_ZHHTL=dXTVRg=~c& zPqZ((M{->wh>16b4jp_2HF+M0H^NU;f@rX?(}X5IJ_nkub66mb%&EJ4mGqJ8@pp9w zSr8|`F8v6K9{}4t{>{*FXtyo_wz_nppY0n~Gvjj5Ho7j|{|g7b(3=N>KhA=10q$;` zccnx;6&=S?Ii{e+)5EOC_X{*2?Zo$;X&efPL${#{dw@V+39QPbCeQgVVAh06eZpK0 z_Pk&1GEUWfYnSl-oj`=l_f#SOfPgUMD)kG;9W$*jPo*2kRL(n8=KMlq|@V-SE*dZ=rAv z1(Js}Alw+l0OjOa(M61~hQ?w} zmeXG}zXPq|)o0y9-0@TSf@;mX*aywM_?|8&tk-b7)_(bV5zyT^MlwkRfFzJ5VkJ8p zJVG)zkgwfyM+T3K|1{3OAXN<%=%A7Sdo}jjAFH%~kss{6Sb8NQ;iZb z4&u~lKHq3O41m!NZ&4u@?vL3oXIK<=dw$%Nj!~JgHjM=EEOq23eshkf^djh_K zIO1%3GsEh)KiSi4D12i?@*jXa?F3MW3u-yCCF%_@rPW?dyJQt{^w^7m)IC9I25VE^ zr5u$ujo&y;I619$yzu`pYp+$H8k#7(y=7+tIw@rh&*wz)O}t{H*uO<{4J69H}w?$P9YlqAr>dlfMHqO zh1@v>a5QZp@^`!XGQe3x8xge`zeN4ZmTtO4%s}tcjQGj$LG()!Pg>m7NdU1JZlVm_ zcn`qlWX&%6_8OolnmE<6hq(NdxU*EN8zEFRKk;iqc!fiSj;}Lh!6?ODCyUYCdS8U0xFb5Ps9> zJ8(k73ruscrXE27I)AOnlQKe;5dSg2Ki3ZH(jF(?uAXJeGL!Z`yEItuX!s{xUy||krcwIE~ z^)o-XQhVXKPz_xQJo1xB*FcOomO#IR&r;7NJ7=$fEgle>pkEbN8 zjW?HEa}Q5ps^GP$Hn@Cl)x5EYYZX>(sz_)Fu;gC!C{r?PNa0cm5%187Egu&joF!bg6O2Sjqr68eQ?~4y^_;ve1@Y>C7wfj>V8wWd0zlij&jfk`i zwcSs$WMsQzehR0)<*=-qN10#kn>t(uGq?gUj?Y+Bkq*I%tf0Mk^swhYY>U%^zA~_HNa{aD{hnB_{ zZgpNL5lgzk=UFm?vukwgy?`4y7rKGyvw~IP25=}|BFiY=TVm?0-ThmO29ox7FVwjw z3^L4A0=KtVHiHD)lv#K_WJT`p!p*Q`${~Pr;ep7l zxxQ_O9k{oic>+hvto`-QV~f|jjU}KIj-f@jm3PF@abA1*Ty7A8ah$W}BB^)eBc29b zjB2ycqaV?3A-MMrRYc9ps-NKV1Vgg%@M9TD-31$NLH|#CKaL%j=F&lzrOl-W)l>MG zaI!2%_6VGHONNGsNlFUo-*EEDMNwxcTbJCy5v$rI6IUsz$OS&B45ZdXcZDu7h`PIk z&O{Tpnx&rzwQNMaSc!vRrcU?>#o-_Lh$wwz<`K)uPk=TCwRRk`Z1k=8(}9c&2`9u!SMj=fE}85&M@>!mv>P(%{ilz5w9gqLN$J&UxPK zbWoui8?zWrLio@}ubOK(Jnerr{o0qR9=R8KBMU?$RF_M{zH(6$m_8DwwruyTzK#>h ztp<2B4-zRELow;QD&gPxU}CpTs6j{cB;8WSNtsP*12aP_uN4^yQ(VfrP1e1unwMh8 z(uIErBoY{pu;R`rCq5D*3c-aPg((ejQ>axy`e)wa)l>AsOnl~h(B01`d5$Cu(Q>tU zJ(KizAMIrD8qWwlvosV^xOk~)y*-q2o0$(I7>7aG7?SK#q(fs4he{`@oIOfBo)$Nw zX)9KN^z+Iq1#`SUlH#=62xRwWa9A~ZXs&=e^l)1>_WL?>5T+AsdRxsRzqhvLI;^$3 zbb=zL?&N{ZrIwYTnY{###p=FTiS1;`{fm^%BN)7wN}|sHD5otKUps&1qkEJP zi`yxP#PMWDFA?f4p7OQO$VGFMyM0ySFsOPNBSEqPUD;5DnTmf$9J8Ifz%3dF(fosG zsU?l`W_4y!884%&K$(=0KO(u8%95-X>6zIlQB9$=W5&e0!hIH$Chnf`MRJ| z(^SmtYO?_bA?g4Rhbg%#f=?H0?CBLHpGzgEmn6~3X>96mJ4Iom^`g9RtL5G43&kyC zOo59U@l;iPj^K-TGP3SJwH+-a;~j~Gt&}-sgT-Ypy4M8idVNR27>*3GAkAi%;;%j0 zAE`H17sGdkqpm0X)En9PX<0 zet~1GxcR0Lc#6nAuqUG4+!`lQ@}UmLYGsCQ*i@h9m)S{TpQ7n9 z2w+z@pt;y~+mtYAjpU9iKlNjLc}QUF=5~3FN%_6R@l(b9TQToqersvsu}a1VY58{a zh!65Nh0d%|4LwsOxbwRE6txIQAFa1o-y=#7-rH49$sQ>&;r%UVl<$42a^%url0Kbi zQ%Bj!uZD4INk8=Q+s6;RMm^8pKCW-0fJG;@Xf9Qf>>(r?!EJG)!p)@}D$wP+*0v^F zOTCF?r{vvi<-4#{NiP*JPUfNP&dN94lB!rA=3p1~!+b*6Vf5H-7=)D&iIFsL1v7-| zLUH{DzIDhR|EYF7Nxn?KF4rrG!Iwm!3yM2V+EX!8e?ZKiBjD{BEQ1*Bs%>7AH!S6n zU0G~&0P!q~c2mG%GhXg1Ucwq8po5RSaEpFJH>~pu6VXb!{PL97wl039Rjw=Xq#u(7 z6kYxd1a=x_Qe3;DiA(>{w(v;^br?>hGpX6s_Zb86Z%M>h<|3ZEpxnnA8!)eayo7aG z8a(q^b#111sk(mQ0&F#w7N;_E=bb?G_ZsF~lMw$R!irqB5HQ7jJ26#qvb~hP1B;hO z=8h6ZQkNMPWxpPzK+)Ophj(LQvdg3=hcY*TrcNLgF00A+u=4`Ry{ad;uL}Rq$-M2j*Lq_`g>8k{E?dfUCSMx((UKY~^ zye3Tsk(CB!LeLRLPCppzzXDP;T%~|Moy_`2mSYL7IJB@@2c`NuM9(Q=eiMzd1^r2! zuxD^y+HEB-qhM96KqO?jGJW|YuUIv>Af+XDr$d+O4{JgDyRRPu={+Bc11lK7-y-w= zl+s;G5^Cu%B9c$aiiY-2Al?-zqF+dVdh;+uXtP$bzX*2z8e4>DCMa=MI=3L>uprsfU8{?F&@_Av&aGYX05LhoE=|5 z>yreH<}Oa}4epj!%X_8)rc8G)H9>*=xWW8r9|-XWP;0e|Gu|Xg#uz$)UGB0A(;rM0 zH?0Y=@2uERs`${)y*eD!Gnp^<0w;QzWhbeG3G|tylMwQ@3ArkWcqw16IPJrwLy-NH z^|T{x6Ho7Y8&W$$ma%6zA0&q7<@Q5PN4{oOxckzGCJ6JNG7|=1(N0<7InbBb<8IV} z`U1%y+@LyV=L9-HBC8nEFT%|7yVT%XV!}W`-{$)d^?)h!@QSj2=6?;k{$xi%2qk(h zzR|%VYH=Lv0KNqpA*Uzf-ZRm(c}fpCwd%cHC5jbTh2i zDIZL1^qGE~g`*p4TCgw2Lww?*59D0x<_CQje~afmQQ6Rxmw~O1 zVQRMf(Z}0uv#B$8l1+)kG-JqG<=)}~-pWi0TMKhY#7(~O0Ytj3d%jtLb35bX1&vFl zf+r`B5JQ_e<1Uz0nhwB~l!P>w2nIU;j-;OAHdueQVh=&j?GfGywx1UfIN4*6Q*#{8 z3G(+n#~-2~Nt#70!~Jhb8D_CVZIfw2E@i%?CYSC+A)fBp+Qp&=aCa}9R-#U2Y?P0t z&AEJyPwXo!pCQbr#Zn*tNKJ#V%1E0dLjkLsD?cc#o=gC#(Rt-`jkHeh)2@hGf>8a5 z3?J8ta!J#2!jFNtOBYSQVZUXkX1wk*9efgrJM4W1UX#I6!nau7MbPmpBAK?{ohX?C z5_YsHE?v29hD{4$GkkOG<}bmfAyj~Y2q9wv`NV?$s#8Un!G|J}@hG$=`D{cED_^4k zcyTUp_`_ysM?x5i`vOCniBOmcT-bxF?-=i$l(Bsn&H58u&Wo5iZy{wfqx(82wie}D zE)Y&tn3+|zrAQ7{l6+@rr;jj zX%R;+RRpxYEQ`PMc!`eY;~^yD&PS1md5+Y2yl`TPeZ+GOpPW7YGJ47RO^d^6cENPn zmJy6{)e_4#{8rPUj+KeJuZSSe<*m>)%rdW4V;ZRd@#@NIjh}z$Ru&W_y|Va;L`Jc4 zUT2HU&(Yx7jd&UmVa3^~1umG*RH0(9Tec04N>q32r1@enhnYkuJp;r zyu(>@GQ_LbRltYkl@!J2+1z&{P+2GAO$n0H5`6ZJo?kmLbzSo}%C{ViA*{kAQFN^| zQ_bP`jTQ*#%#$&X2OA)8>=Q)qoriUcgz*F(q4`?)x$?eD(0-*{ntZa0kHe8XuUTfP zXkxun55~0H#91#qdp_SzAyZ)iS#+AHnPXuwI3qWz=B+<49(nI$8JrU=Mw|JaG)Nu{ zsWngG%I*5HtOZa~{JFm;wgM9tiq9}2hp(g#bBzAl6(6U4$p?<&eltFefxiMOiM}ny z+?u+PdX%*OJm6i_WUTBKn=}Pr4~Y&G%^D9EUIo1F#A;Q+CzAzY^a034U;FcPX8b{e*Pje4mFeFLZ4I|s zn3Ej8v92|2Uu9dMDB>daO&ilFQ|MNGB4}P*4>!k-z^7K&p?GH#mZAIRHMAT`raFUq ze3NME3-5`n{cgeDuy~1KL^POL&%S3u$T96tkO#U37#h4s*9luK*dD=oT|h*V>IZJJ zEyt5l>8*?gqb<2_5rCv^;I-SpX^f3vJA%UWd!m zv&)rm^CachgH_J1kij|Bl;TWCpM?4R|nN*IXi7U*ZR$|mse+p{_CN``Z`txj?tx@)Q8G9ezqDeKxr#Z82gOCz1HbkRTKK=}6Nb7EnzwB(qT3{fdNu6g2s=)!<)%Gn`)46=|~oP@5Y zg&NuX1qsop43MsJXh662Q-PP^63DA5*~pLd!W_mJA9W_F1@0{4Q(I zjV_HK)7xB8NmhB5=uh6|sH=E4vn<*8YH&gwMlyNC@K{wmU1#QFgp@v4*@|~RQuDdG zuaEbJCvEv?XY1|0eL9XJPg4ny-@1NDZwat6{iUKdA4jqBf=^g0OwQAAXK6G`tku*n z*ox-}ORTneMIHcWRE+dlnyKAx-V~>GKV#YQabp#(Ps3M*qef04ByJ}#Y`x9gt~qbCq$WeQM)*?hZVmnWMxQ;`?0Utmia;4^?_*%ryV~(a`8j3`mY2d` zxDkK5lfk`Zf`3c1I~9Xw?#wz+B}IUNrw(CTEyf@fUdz<7^NenRPCOzvLyWKxTIZ>v z)CZ!HSKyejNDw6ZT?IiQbdNijl3TxK1}BOU7*J+JnP0cE7r4EGwNhaFYi752HYvg; z?wgiugk1b{+^bHRZmf?lsT(~eV;#7f8)@H|UCWuH6)xK)A8iDwMiB-2lNf^RVLJ%Pi!v}b%LFk$o^{f|| z_c<|3LxIW@ed-GD?;9e2a{~)EB=Kr-;;O{1O59uBNnIZA&YGcG7PzQ^m&bamj(KLAMqG1#JuoBrji^kI|ttK(R+E_%(-DJ zY5sCSOmd7SR2bEEb=k#!HWk^XmW1q zH!q&(QSYd@&@KaiaYW99yODQwGEKM0a1@LrRur5%E_xdcWJsouk>7gT5OhvnZAZ_u zVj@xne}Hm+6mUyt=viZzp%Pt*on?FG*C!@BJ_F9!S-B^3vGbsB|G4}MI4469u&z|V zvg!6HgO;|x!*}p=E#P%3_R4dyUuR;(c*jv#a9U75m8TB}V7aFck?5p$$ zFFtI`$-rO!{mf!E%T5`$$U{ zxin1XUrm4O-Nhl8l(W+i_{lbJl*+_U^PKu-q}QZRdyMcGTro(JJIivCqF%4ruzd~$ zt@|I=?W1wcv&r0^5uN8P<-Cz{qy~01P$S6ZlI_UwvCP^bN<3!j7AS+RIFA#P!r{F+s02?G!M`H0wae$ z7SiReQaJX9C5&Wa+{;}LfK$uqQ*PrU#CbaGbsbt@Fi>?E-FXGL$V0P%#Sw=Y~q%d@a_`NA7}l z3~_VWVeSth7Uj#CK@MHh*?9RguJ!>ooP&@|U|+XX7CFYaA1mj@u8_$*uyVN5cypO!4gfC9_zxqVZvWxAkmx>vVZ)iBC@m(V#cLe4>pNghX~UKp z59j_hSFck<>CSqn9u$CjhJ{6fX8MJS!s-m9ll@={#bedrHG@&XF^8G(0=tx`@U+{C z{iNg_*I_D;Kd5ndVnrEi6J}TsKfmN4gLnF>x-;PDr?rqMTAQw-@eYc^xUU0)L67AM z{}ov&{R(Xc1`osf?B=_aE9Wr2+W%R~5fQX? z?VHP<;mW^Upab!dR4@B7y}R~7fkP#RFZ+|iQhP&glf8uEaEAL`fq06w+Lhxww>qfN zByoTWkO`r2^V~g#Rk)Js{hEpV*m5i=fe4vH3|@>(e|?87=z=}_owxKm34-SK6|={H@s@q23evX@0|-=8Y{m60A%M4VqfWAsY%vA(bD z>zfwZoBAZQT?K>lifsfP%F<_@ua~(MFbK4^re4vhwA1y%z%H$iKa-M@sb=NSr@*MH zxzRTZ97{h$;`Wcteo|#YS9nHA3Z={#8EWm%muJM3eP}0m_uS8DzW*Vb*Jr0zTs!!9wYuZC~xU9GVrM|9gdpMYVnXnl|9(;2ZnjbQIW zip=S->NANeXL1SH^sQB-cL8Osi$A*{#6Dlg-iu9cWcm1_MR+%v|HU>(+@DgLWLJ(b z)A!)!r3;q&)wT7WBWzW6K)TmREOB5|L+iWDi-L9@mx4B44=Ng?&f%`r+KnuVPYu~p zb{;3GxMjbwdiYhi6bupd>Yu7DEO2^Qklm)~LFj+c-ZCV>eb5HjBd6ULUW8|~$!Z{J z)o2V6xD;k&j-BC}b^)a3)rq-+s-aeDFhkwFS# zMY;7hwTX}T_hhn!6V0h)2R|{@UT%h*m$QpqbHtZxS&QTXZE^xtWBcaLPU!2kZ+vL+ z-R$~&`uOWIL8U}wo~FgWv$*qWk4R~OO_pu7vT>E7S);^I+XTQu>dsu-hkHC)O|O+} z7lBNuwaJV>tDmJ3OzGG^r?DkL&bPittS_uv?KvFs+}Kc69z40*q1A`>g2+` zrB88(EWF3QmM;6%PFjb^XRuIb>tt2Y^z+c{vVr`*K}Kw2{^hf{im!wE9sAuE3O#!Y z_Vx2*f^nUNT{kbf530US_83&H>Q$8HG%_CGaijNwA_Abvl->-A)u`yCjCRv&XhvLgqKaPrst@ZAq+72ZL zmaPTw2kI0uH5u^vZI+2tZ|OF*ri$jTo`{~)kdleaej%0KrnDuI%~ZQbzQ=OF_-pqr zL4rMg&K`DcbZvlq!Ly73D58bBX1chF-HRJ`*y93VHb_k*2Y$qJ!-?~H-}B;Q_awD{ zNe6GiY@eYmsroj9iD3NMXAXfTJ?aXFp&H}AA_rnM_MVI$(I+0g^JYV_67*21I_Yc{ z7ti{PrZQuy8fo+mdcL1P#wbqA5fzhV`R`@z^vW}OV3I_0!bvJ0>2x~_-yK`7y)Pho zT%)fd92rGPDIVF7SuEA%v1Rh~Y$0Ywb9=VpgYk<3pU{f^!tR8t1qovvQXB*w<8_y= zji**z3Uu{yy&n)Onn=6;fZZ1V;Sl4xB-=a|+P!|nXXbvFDcy0M9)P|1G;{s|z9IX( zd03+b635-RpG75nINWq(1;0%$Itrc~E47fXPP>a@Qd*E_CkX^8VJ${&)%?BvyiAG%;0%l$ zv~6sDfX#6=oW;H(b|{{sB60o~SLedBZcnzow3n=YuXtfGG`_@U)9pb$KJb6;(Zb)S zlchl6uAaz$R*QvHfVQV@@p`5ot7AXAtpBgME2@-N7=j-$zi9~v;M~u{eH@hdJO5fcd2_{?ji0w3FzN!pLT1jOIYdS`2s0m+YtM%6B~*&b9n}j zL`O)KUbz>6c}Lm=s8#T$vPr{#Oj3eviPZ#21=S6i{J}-b{7W3OGfhOgYz| z7mH*R2trO)9WX(LE;EqZBFG!G4y4K`G_BSE(p3WeIYxdCvD_{>2kxhe?Lwhwl(&iO z7)Fp`k~<75v1x4k*{_4V4!rTeUx!sNi0TDI0zLeOK%yLO;>O4uhR$RfY(JH7vkMjC z4@0Oc6m54aCS|^5qBhAQZm;3Pwd<(DsP1@Q<$}r31o2$kAWtIiq1peX0}rg}9~B^Wqy^#r5Xx^tYl?`$|CRJqj|>nkaB9qn@n z$^t5>DMLsk0xhr9flI1@eUg#qsMQb*%d`iP`%@sWszBlK=sT$Wsz5zd`@PE-s?OT) zNVeHAz5(91l%iexh_IuPNeFuz1N+y}FoS`%Z6Z*Pi0-5uq(3)6>*@e}V-*Yn9ZBRB zwY6)9k!K4Z^a7Wpph=CNppDP0^@DoI6R4?_g(cO&dzScX4`4uNO+15LEQ~knz4_sZ z?WbH09^`mOTu+Ew$^c0AiHrO{iBrmU%0F+<*UZF}0|EGH^Rbo|TDBz) zWUW;y?OQP)e`6j;*!-DLTCBQHl|w8;Ikcj$Osj`^duTXEidOF+)hG5Sg~Kp80k188 zW~0oc&v%AL>W4VkGjKcH6f)P*W~&SYvF0``msj6moa^LzN zbe*0)kZLuTE6Uzfv|)cIC|u(RG2)5Ee4Dka2o9kDH-~FWIP|Zz`KJqVQ9Pn)C|dlI zxrJyH&+Q_qCxaZlr5?Ih=0Pg!#XZ;e9gajljrK*}z@L#~0oRW-nH2gB5d`a$#Acjh<>0j#bp@4`-?VW&xNuL-Kt!Kx414FAqNM(LZHe zf?ia3&XOKOP&`Y z8&?i0*X%|TLR64xWc~;kQ5!G!;dxdb(<9LUB_!bDDbVb${0rVrBPA`4dTSR%m*JKw zni%xyd(KpN5z$uGU9h9e>7yTRi2sAN@uZ_$?%7#*n=VSym$xTQYiU z0$NwJwKwm+pu*5R6P>BCkt;FxN>msgQVh_YvzM;IM>5t!%)%EJ7CK%U)eSw&;X*4- zGLGJ#%bgBX;^rNVfF*Ihn8haPz0~+Y(3qjK@CA&t{BgFp2iVq7OkG*MkMVEH?$A{jvEVjj zub$_&DLBpA#`<+<ro1k`CYWSf6g$TyuWd*^Jp7IK-Oey2Mis4~ z=^_^^t+q)PGG04Cgo$;)35N-NWt8YCrP5#w|5uB##yK#)txn!q0(Zx}TM9qLrsla9 zq2mr^_vXgjscFl65F_xQGngwT>&Jj7V)cz2(=!yQCC%Z|QL0sf%CK2o_G6(=QjFn=v0-FIJ61+ zmAT%tRw7QBWg|!v#S~|Nx$0{|MVI+t1ta$vs9s*ZFZCyI)Cn@u)~)wwswk{kbzfFa zB=_#UryFfNU?bNs#Qr8htITb{t0#;+Iy_1hKgL^;MV+GO)a}e?B%74E0npLPBQU-l z#rrJXzSi7c#XxHNLK2T_^#<>3)g0Zmtfg6a_!H7A6!jkP>t2nFvz08CJ3IHJVq3xX z!Ek5q@@56GqC=l9gL8qZd;57olQ6yPfFjWD(EUcM2ej7frhh2QWtOz2tho|Li@DS|fm^f8jnHoe7AaZ}G-i%B|* zyCIUj^bqH(E{3r*mNVA)jkWR= z2^to7*o-5VNqoj$uRS;GOc{H5`V6h-%PrANF|RVsASJ1V5j+07A#}!l_nulZ=)Qk3 zr0&4E=xPz$ZjThV!Q=Q#AT%#RPk3DQI{hPIWuRW;k1eAH#Jxf30e(ffQ-l;L($u3HjoK5rJsMSO2>Dh+q(=gw1&PWZ7a_+(YG zQp^tAYW`53Poc2jphcx@q+`o^gc<*^65zh7cinyXRN;I^cBOO$ZULpnpm`oLiO!(G z=tYk2i^e>+F8x6xMFK0%^h~pICH?EFULn1VV<}fMvR^3ndNI4wY46HQlRPb^ywTrk zTW~k`w#}W_H>{sh%tz(8Ja4wvL|ePqYKQiEj4b$duSME%DP*jAPTW_)#O97G5)gvorG9{5EJ547mF5OUWy-raAB4 z_VpOX{kgBbhGs_oeOb9n>?4_`(cvFBeSX%rkIX=D-tTUPlwj`E44~&9`8Q?|6;idN zp4tS1ob)MH3A=bLqqrXPMWtGo%*Jq`0-bgao~waUi#*)5b*YWR6yCi?rUU9d_N>vO zy=@O+@NAD!dhf*NOl{7uAd48=Qg?Tl_-3T2eM>%y`m@t8oxgb6@z>21M-=>}iK1Ln zAJX%?J9=hs-L!a-qZYIhqh;8)?%sV})#sjZ$0gAwQGVJF1@R|p_a5W}byC3a;#DG%cw<=Vi#P|K(3?y;%S2l{B93BmVoeO%uCdatg&4Ty-`D~BDa zqf=}90uP9iCoFOfeI@JiUiZFc*U*}hrqbM$c!9nn^y%`DqSuXHTC%n+Zu1{`8~Ow< z3Ji_J*D0?edbn|S{Z$XncEv<_`?P#x{HR@yuN(WV>Ra10uM#(ZolX7z?mJUzExp(n zYV+y?f__Ey-%at4p9FQY3vr>m0>Aul+3`9NU?$c4KU1&-Q zK}0#qdyn+@H=rN`y&3KG8}F2)m?LZGX@lbkNU-pJo@9y|bOyC+1GFG!FKdezUJ7a$ z#Qpx0wFel7Sb-Hd$d41&!Dtd_`LRlM^R=~td5mgDApSk!o0;`CHaMvqz}fjeKsxdF z2SU?AliJtZWqxfB@5Y<2J-Bxa;gfYpVVJ!nc)eW!6l(P4Z#38H67xOyPR%>YQ#L1?5Qbl(&atG{nuynt)7 z86Bbysg9Q_G5=sDQusi?s=t6Av0^!A!7LoYWM=QWKzEuNMt>v1f23rSxV;Hk{d+f{ z!OYc6t<(tB%9*I$V2< zad>A*2p4|PX`@iMZwPjGcpG+ay&q2EZh9t^r(1OtWDJrn{Z-~$M~DK~8b(-FgYs?+ zc3AQmx^9M8~QS$Aq zD(~<3t*FP7hVr1NTAH|3q?mmjXYvBBIB?$@b-u@)I*gmB8`NmyCM`PO>0cqV{NWvi-Dgk7IIFh5Q-A8iV{P1I4F9`90ybw}$Vf3?-hH(s zWce*h1RoWM-$QlSf428BrJQPZtKdJ@KkgH_GnxfBD1Kk>zZev9&dS2=M4Dd$_1=Hp zwD1TN8=B;ti>gxC|NI`fp?G*USxyIpF8{B83~t7QOK^I9mtzf1V5zixi$s@&A3C|Ld!R4wI!Dg%bHbHb5fD-txrNPLuT_cXJ*Z z4!{%%zS=hrj}N(fP0iUi{!gZfOKQn7{pdF`K6>CfVR2~HD|pwpM&@m)?Gr?2pon6{v}z~|fd z3b8M=rBX_D>S%RH`2kBEqHr#u7>jn+-RN8T6G*Ru&b?;}8G1%zJg;UGYLuY6c*TQp3GmxR zuy%|eIXJNE6+L+m8j673&nJZnSZbpnIeer-N)#hA{oaRi@Xo-QCvFdx?pbb-u9S^= zLHzXyK#q<8*BM(fs{j4i0?5Y}i6=ll_Ct(*GdEUMkqlCAApDmg{lcpy7Nmx<3q>3p zFk7Yy^a|GjSt|s@WgD*$Z=Sb=<^7euJyozx#a3JAVg(=i4S^?yUK%-CPDZ4S`(&T%9(x` z&w-3DW^jFTTF_-0jbLmH?mg1+VG2^a(d#!0&05+YJQ`y z&;wK)3mgzL{8wY&kmE+BUq9D3GQMHhll9eBv*#gtC5e>})K4kvM$a z@y!vQHk?r*+_YCopH`$7pcqQWvVOq0j&}_@#-x94=|p2FV&<2D9)l&v0ZZ<2!mKc^ z!)ald#%h=@(}IW1kl7D;8rEVAS6+q!$!KZnhLA+sDB0~_*nVgx= z2NL>9xEp+Lj1MG3lxL0Y<@o48al}QF+Lg%^l^_>~OzOas_kB}zdZV2=Lj0u(cz(9! zD#%AW3H}@_3OC>#b?I0lDyFjY&8^RGWPNQmBkOUrrtqGFx!?0VT&0TIde@4LpRRr2 zhAuxJ5NIa`;)ZK+vIHkVWR@U$rUGz9xvKg5a?flePDda0gcRO>2{Cg9 z^0F+NnGa$z<^^dzG6xsQSg|z{&4&pU6CXML+3Mo1!`m>YszP|0SCppb6pGq7c;isC zju&ADfc;2GhC!?h%LVg(HL+z&co9#647-a1bBwAU>so;7*5^5`>on{FZhLmDo(&2N z_m+ghs!v^GZ(rgQ#k7f%3>zW z{n*%8zxQVk1Oii?ee@nZvhs_N-Bc;X4aEhsoK`k6`9aRt!5%Y$Uv1m}wy-|~Tghw8 zQD`SC{T_-EEc>?{Ja_cHLVKTD0z~{R_pSG1MAIYZf()b`)0x$bd~8>Eiou|Tw2&-; z{i#bU9#3<|MkUFM1i7?1VnIazR+k#JAOyL_f5mrjxJ$sEG4B&khwK?uuxFHUzgs`S zn|_)JUkRKiLlrA`1I!j@&JlAX=-ZOP*RbPp12Pu zBSkZ)z3ri~EC!*Gt5feF7~em^t>cjQ#7${dLj2DO6+8gILxm=o za?Oe1g=x4~_vi?|y<<#~QC)4A3WY`wE+Om02ZYP1{37Sh9xG>MS) z-<|FjVnlhqZ(Ue~&g+sE^s>?-3%GRSo+K6;FrlI=1FW(kNzXzo=&6$&7f+ z>}fok_pvJ30`@~S^W44FJ*r=8^x1X2qVu7C<{a(+@B^NH1gNd=xzjdS_xqABS}cL; zts0?p>OQoUunEGv~I19ROzeJ`{$MBv7iLGOaVR8_PhV7jiH`hdHv6R4)-2Zph5iS z2LDB;{w*1TXW_~v?3Ymeqf|kig;IryNKxSLL;v*%rdRN6vyw|!)6=os*|LzDAHtal z7V1TA{Mm3i8jPL=>Kbps@ts@3x^;IbH=5&TnQq0Nv5@y>JRz*9-ueL%Cs`JAVOo zsw!A2<3O$o`2tk{Rgz~FOjt<;EuL>tr36Chz$_Fmdp|2%yM@q90$|b%Y$rNR9>tMe2zbAb_+TRazgmjP}(-HuXGlZ+xIygBX7>j~K zs|aBtoeE>1P5ub8&K==~5tJx~%{jHT3=m-hL>2f0F4Y(W<}DRgPnH48H-zk7C^r1` zb4n@rSr`hyOi!%9_SO$$O%~FcfQ9vJ#ymL!$v$SGI7DSEF~Y4#{t=Y}%x=)kd|mpt z5prXP@9ORdImj$CC{0CGo~J(R`mQt$+Vm1&-NuL|HZ@YA>!d5q;F6^MdeLcISjb$4 zMFb|>eW{v%l@LOu1XAZJ*rlgZ1RWQQ+mFg!K3}1_@Cx3yh1aX$M}}_qVT{}d?%IN_ zD~X7fISwyw6{-aW2@NTBK=`Al(r72SFG z?jy$(`045Lg<1rymL5pBVp5H?&(btDdYWf~kc+*%#eEuqUV z%Rjh9X51Fzw__2sT~KjV)HWj)TjKL-e*^_3+O{V>*x)*?`PXp8t)% z!&^Otpd>+R@wiMR`j|slc#i$r>xTI5fZk8s{|d8A`6sq=VAk69?o4dmmar2r#r;NR zi52Q_sGco=!k=@U{1LaoO_27NLU1y_?Ie)&eMs1`*j#$(ID;KX6{zaK(Kx>BQ#pzI zS5C6Q1`9E8J246r6=Z-O+fxiA~&<5^%VsXHe^&KUi!s6@9?o zy^b)CUc)rpMJaZDSDx3>sX{1{mcMeJUx6}nHYh<}VVmD_E!tx!VO`%Bpc;A1!2(*K z#~evnOlx_7Z%A?D|D-v*vpd8EYLTVK-^->*3Q8Bfv?frg0VWhm;R54^)}`(!@0#3 zmU=D;%iW!wnz(8zxMnjrD2#VaPN4APpLucf?iLz@Uk}0fK^=$xO#j@_)Wi z*stMJ!hU@VurGvS+q+qRBKQe`1CaYe5Gh&XR;jxG2VZ6)F@`evv!kyaG_do45Gzha zc`?ON1q0gI_Au4P!0`OP4*qFu(k(nvM*?vy3Z(l{B7qNxBB?vQ7KsK6|4S$s;{OkM zZyru{+rE#oTFOvarYOmrA#>(&8AIkNGldG7hmy!L7nw3-iX;*WWlS;^8jPicS{0>~ zS)s76+tB+wZ@+zfzklpM_CEG;^g28**7~f^=f1D|y3X^w&a1D-$Bpxks%(yKVLMtX zMF>no-8&;TxnKh2Ki6~UP^W5F<)w8P7eaKmKJg%09{Kwl?F`lmeLHfsJ5&3XhZuR< zv^EcGTe7@gTYI`@&l=IAa(h81H?*rWs!y4Oz=Z)VRXG4C_xGM%m#L%0C=Va)DmETP z%ptvsXO1#fWlO}vkVaa!<7?_i zcq=x0_$)Yd9UZU`xqcq?Ape8YLILp>NhGW)T~YvN=EF*K1v6A?S|JDmc&L8xKV=4k zD%7>`iwwcP=w)`(!s*K1j*YsOlRy%o3)QV3=`M}uvZ7=Q7WS}a0R)ypCU{7GjQTVN z%Z`S*y{%8r+0N8%R8Wypj$s$uz-0``(djeDnwMihzDgy12pfuaE3uMKLfg=OsK(15 zVJcZNKBvItUuW9J@$jKs0I5}#$lzIB} z>9;qCa>G4DGQTr)RpEI25q~7F0`j36I(Wl#AFtdz2FQ4J0C?{~ymjdZ^!_A%on zTmR#PP<)V3kyYj68;&5zep~YCKc15)`svNNC4YJW(wB6SBtP#I_?mk2=;? zS~!3Xp-PGk?N^!X>Vp+}4R9u%Tq>Sg}!I zJ_Z%mrw?dlT*$Uy?9O@z#gNm3y_D-=l8ptGrzf3+h2b6#cspX7h2XP&%L&2tp9eGA z0;`I|nFG21GdDxG5hOPD2;A{j*lia$k8b(v2UtUh1vsk|**_a}Z7HHhgfMz(+RdH5 z_Sn{+-6jR}|H-#8I7>iP46F@~ZA_qdOcAzgXxERRWI&Vl{6R%Fm?7cH5pekI40I8` zh9q)Nhd&yH@lnX~9&g1Zo{if>w1Y>r;>O&$9R#+V2JG^tHwwrd?d+jCDqt0ct<&*9 z5k#b=sCu0R81oo3cI6L)!%5Y#N0fmO>>hm6VnqfUu3$Q+n=Zzkr0-O;ee;ZgqAoZT z6?Kqk6A?uh%xtX6N}dJXny8*=Pe&ls*O$d7Ad#%h30~^BWmdXppp9DCl=$7j2H3x0l_GG-1=_-RCCz?gJ()+)ReKbnGH!RDOBVQPg7)&uk5%9|5kh?*41R2K zRmT(1W%NU<*nRNox4sHntJa&Iq-TjcqmC{xOY`WT6Tbvdsdfmfu*jp)Mb%6fot?n$ zJOA5@8G6Bg)q?XW2dCUl*)cXKFfy`zd$kJ4^xi>pQ@fUPX#&y@19*BS6AV$UJS#zU zI}rd~)*Oq-v@sBLl;0uuz6Zcbvm0pEJ`Z2V9DqkuWCIST?sIKQE7R1BQ{78ksCf_a zdG$D-7lwxz;QIFexoMBlP0LE&f`Gh7lx6!h%f-Lp*?cgR=UNM3Qux3YYI&mMfrM)` zdy(@HnRLVQ534{xv%iDEXNn{`oj-Rc7Tq0bzHR6{zsLaRxxfxN$c#cJ#(JUY@^mBgi&;Q} znj2`dFVS~{FJ84qJ2Xjs2Vr0tjIFzWzu_dkrWPpu{P~mAjVVR_l$)CdWH|3AL~t>^C!}zi5styLr$5}X6YG7z6IcUF(83L4iZUL<5aiq%y6Byrt>1OZ zl&eQg8yT>VEJtJ2#`JDtIAJMnw{W?~Cer~cH5`<`Y{N1lO%g7VID>5n0lCMD)Q zTemaWgAtPFX@U%Rlg7L-IUODVz0%)ML#ETyREB{gLorlBHH~u(2qzfcaU7Z%Cptyy zN+&>C;SMZ}gO`Vs+(yV=(1ef>zbDKXGMiA~!;QM1w{`r|#gBWb-RXFrZI|cvwqdM| zPEWe&feA@IXgz!G>-0I9-L`f{>PaM#qIf#}mY8w`&iLv^7vRuQ?YWLOfZ@q~*zRr+ z4utNfy$bZ0bEpXFS^UFaa8r#1O7xs|&2zgDr=i5YPMZ?uuos2)C%g@8NeXu)E&FgHTk(q z`HSl#xfTx8ZE@J=^c4h`vGHImar0neTLJbv8SXw{q5LW~?f$U;V0!RoYHMH)3xKlt zfwUFqnyUc;{ynUiWuu9OwM2UQ=M{vHphEdw>zQ^Zk{-6FJoz9Vvdheg{=D~TT$`m%BecU3wWbgNo@@hW>%a~f48~~}EGfd5s6V-F zE3BGmhW8Tia1G9q7PKV4jXRBY`{v=he$bM_sm+)B_=5?RF|oXDpHFp5AkcV07Niez za)f)N`PSc4jEnel^60}HyT}Ye3kqpal!V_bFLDXS`pVf(@+h51s5hA$n$z4Y1%@aW zmIGf;eohJYx4EP@KMDzf7IsLorrK7ln6Sw4XH1Azy2K(U={*H? z(J&Y-3m040R`|h$MVNGk8z?k=k6gmgKzSC}n*Qy8SAk6nF-{a33FM%^yc8-Evm4qB zOU=tUW^5I-{J(MR4p>hLp^8EH($D=wxB|9D*i4wA8k6Ta80Z)%MbKLn#Pv7O z-9D)|16uoP;CNnreNp59RvqOpi;oNfy68d2qv!hzkz{L}S{0O)QTTgv}VCT;%10Ds(d9rQ)zX6u{0UN#UGlC?kVtMtUX0yr72c@cz=M4084Itk4?I3j7S5HD7;+@ly z6vq-248(y)v#g`5gV5@ok&0o{JAxP9-5FRSIrt|y;YW312bTq%+0!tvfm07fK3(&TYr0K~NI=f-*vRy4RT@Y_+? z2oIO1pLW1-UT))V2Y}GZp{e1}Mc09f8wxOK;0JKZ_uW$XO)J(2KRiczmHZc&v%y7@ z_Z@Y&pgx|n88hWJn++l)gnD2>+nE)}P%40yXh~C%`wEtxq{!xe6_txTtdD-^EIlm- zM$g{yfJbUIU;YJsKtn%Z4hjHEtp(ZsZ;ILrW=2=DMpM6*$LtMR@9Tj&$p?}#8JDLz?C7Can1gs#@NykDpuX8^ls)9D6xz)~QyP3XJK zrwH4tW3UmmcsUZDPtKg$ zcVClxy>HM00Jqs868iZ6imt$sqPxU=Li=x6O&)UY4xcRs^8XIV@CqXu5YyeW_&}fX zU$E1R0Mu$-JMh#0is9%@6PSUduJtOQ|L;G+wcmx+DlbLy=x=b#9kHd4(~bV?T_7-a zL13KufBXYII)nj~tQyk9Ut_V&0sJmoD#>1z!;! z%BWKyKhv5mJNOF&lb|bK!sv`&WQH4vhBGX8%`FFFEWYqK(xj1HnIZLqA>r$7|w!b5wmxUe!wMF+x%Nlsj^}f7Ob)|7;+HL)(&?VH5~|E z!$X7v3hb!=*tMvPPnk*H2iRtzzpU=xHF)^%s;$f6Yf`17xSP00ZKFfjV`MJuiWBHHW}%q2E!?5aEK_7+(a}EAwhjP>J#weD0l@n)SoA16-rb{6Zu+Fx!P1sIXU#nk zsw7--7k8{JNJYp?5|+0Dv=ro^4dLPp>FUZ2#`Cuq-&BnLoc(m~Nk3~hOsmrxS8(ji zX@qM&1><5Q0TO`_+I0#nBA`dH;CHq`h_;YjL;V-CFiU4Us^`vE{Td1y(J9!w{LuPM z!2)OW$*i`q?DoLrF@PJBcgX`rN?} z3Z$({Y0Ht%vP2Il6ZJt3eBP4l?BnK}7fjZ>G5(cI5fD*Pv;T(5-+L2D$v7CQ{)Rt@ zumQ2RwTf_m#Q-jZwQ5GV{q&vxT6_7q;#nbYy6_j8B z&pds8zbAK1fo0S^P)WQ2NfSdr7j&(<0F4k|-#L1@ zZC~KCm*dci`P`{IHOAZFDHF(;B!CJooX;0QhYx^Bkb|9LZVQe&(SVH^kAeS+6l+t1 z9EYLV*wkguR^%5p0D34CTTwc8WE5rA2q^~TUq7^x{(=lqS2;1M7brtVB%VUev@mGa z<_DzM8NoA&Cc=|EX`_KoUq!9>sb_`^9o~kyfC0b$W>d zQ0i;XL}|$5p{#1|>%VsE3&7?Mh>iY*+U5k!%Nq4$M(KY+y6Fbx@wJNFijA9QU;;dh zA+;^qH%DPd(iZgZ46vaF@|33xW}(~%VtY+005P9~|KsCJK7(E0<%`-cha13d@Wl`4 zUgcP%H%A<+=}m(sz^I)7=(A2-2HYewQxbrtL37AO*V<%3QXrTF#;={_cROXDqGAz+ zVK*3KP7zNRpem}5oWP`>7eexG27uW(w221+_M`3%O!q9p?IAP#K16=ydYtSN(gBG6(4z4^xq_2O#*Z*TINNfVS;V|Pv~ols=}r|ihztV4z~O4N*4q6; ztN{cJk|2ySKB3hV1|?gwZYCOPZk>wI2a^|Q<1q};Sqnn;4Wu!jQnf@JXrn1$4^b%GP-(EY1HKP->|+c14D z>X{}}O+GYu@6&}O4A1If*ZC6I-zd$s1RU!dINFw@F_Z4+wN{b5P)KvGoipAXXt?3% zOw>CKUG$#|*q!oXlR62@>#G1H>%o}i@QUEbGfGgLN0PXP(rd_hr8-mAL!DG$$fAu- zV1NYPASXB}ZRMCuI78oEs_#C_y65G!dkBmL>oU~Zbr+eJf&KjWTPfeM?)D5ep^vHO z#h*0YKbD_M2;H{FG_QvV&7ZbOP|q%U_d*!wXWfi@=1pkM34Ddk>id{CHuk zUpxEam&DgcvqrcR0J2Gt1sZ9gnM{JV^{*!yez}9eR^O0Sr{b?us2W?VDa%Wo^1b(s ztY@-prjSm|(HSk5B!JQ;6mhO++_W!Mk;{W~!3579RZg+Fs5-FZK=RtmAJYJ%-cQ_a zIc!Z;fOK>{9x;Kpzi>dishq@;vKCbtn&IP)jV^eI)EdSjD#{7)Y0??7TK;<<@L2SY zQRznXd>w<|U-a9Sg*24jNI8&pXnq=XfwDz)D)$8dK<$#T(->yrSlvB-$@%4PKnf3J z7DojdH7nh^{PsN9R;znYKmmo+xpmHp2>nWb8ol?*2H@_94+(CTzn+sopr(z2Ko*La z)P(1cwbHmm&P>@e5~pUqLo2U)a^z3Tc*_=*>wMyuR9%C|{A-ZV?mIl+o`DZn51gUN z_C0LNU23udb$Z=rptGN4po%hz9oz zy*znP$ogKlua|NOkJV`5CCQ^_5t*Lh3N*={+gIvJo1Mq<)KX*4D_~vpi?@IR5(={M z6VX4R;C&AfpKqw@hv!UPGHG)_3TD|2fHP@J%E0+1D z*JZ_7_P@Lb7OjjF6K5v-UT~DgTy0X`^K4TY*7VwPrA|SUYbq< ziD8oo*+JiGM(7ZI-~L zFi@wD6khu79)2R3k}vmQtn+~7s%rHBxGdjxX75ER`C-MX{CN!{zLuxgADR0xriSg5 zZAYFBmLru=T~CjZTIN5Nlog_U=8L zZVRN)`ow0yk3hbQm!!zl@z<5^Nw$O3k+qubRzVDJqljz2z|ZT&${k77lF$# zfR|q_c1|bjE@1d9lx!K-`$#hqUtVPEjV(Z2WzuImxTl{bqt(si!N4-AVP33%W=FE8tS)NFY4 z6LQK-eN(Ny%VN_D01m&w0bq*0_EYld?tYZQU?bGELH$;(J(+=(Sg&&+m3V-!9b&Zb z^rnCvQdfV|Poqm{{n?{QSJ$VA5i& z13+|(s;aIJH+tb-;10$uP98d-tHfRvhgCkNe#S-#EhaUvgZ_*#AHMCV+;g_tbV=+yF3+LWPVg=r6bp zy087D?upFn-~TKBB@YMPf@U;x&aD$i@TXmC$MRyV+0ite`;Wc>mY6ZDVk})ZfjJ1> z9(~%=>3Ya^VewjA`fbLx^>mf350Y&$!6(xwYkUD$o&QNq2-~}>ThLp|eQ4}$q1D~9 zO3tjjEgs!AUiV%PfKxOB9F0Xf0hQ!9;-MHQ&JU;_N!E zz6P$Ze<6q=sGE+}W9O|94L9)Jc#P2HKF^mh3u_QUHF&cAu*!deykt=$L={@3kWn3n zxt^8p&aWy|C=DaY`aPJYX$jOuDR5N%0C+(NvrKG3BLNW@Z4LxaDUiSYh5D?>(U3C$ z#!&8+gK^*BdRz^%@~XFIU0q+^N&&wWAE7SDqun6at`7``48d|P&!gk(ELiH9BjR*s zq{APWJ}Xesl>pa*#0^&E$0C?b^OlhU0uC7ZXr5AMa1uQ4Mw zR2WX>1vi}?huJ;BVdf2p|Bnxzx_c=bFcGu+{%*{W4g}tA;bSU?)N=rU#_v#uR{>4h zg60SQ8hp26cCq&vYISE!KrJsNKtuY&c<*kH82<_K7w4+@5}W*48LUyz~V4Hu;0La}ex} zTmh=g`VQYn!c4#W#84`0Jerj!G-d>fmOdXc?+v~Mu%cjMnQV%mw++0!lmxnq4M7^*-}taaoRGwssJ{mO%00k5ZtxLOw~zBCj#q zTubEGQz~lTNFw#5`DuQUD9k7mdGPM=p6~C$T2q(va58k0eBbA=isnHcF6lz7$lHS4 z7)FwwNOXLm=8Qjm59DF6_*N^ueIJ4W36tXD3mO6u#D)}lRz{b%Qvhez%2EWXB=tD3 zRFTUF79Vv4+I3Afngh6g{Q{uy$D~pgES)4FQzetnJ2(mX)Wj%!t`8d5#lbJ@Y<1vm zdCRVr`*fE=*?Nl3Ev`O4Ydp2zu$Mg9WZZbDoQ`F)No9ROm|dFq3CWR5%K>2F3Q3f4 z=c}krq#Rn4-ljA_%QDA_YkpGTy}G5^{FZ-~*W=xvY!;(CZ_==Q*`M;o?^fQsaT~Wc z!2b+?f)*no^!8+?kQ+(4a0{tYFxT%q>^yF4?Rg{o*ua`kNeEsU3Dn=5;+_v2PKel^|*=u=$$6u37>PCljgFmX0P=WiHD*{OEnhZMX&9Z^=C# ztE#s=?bw^Ig=+TGgKuFzP@#2!yIXO-T+f(Uop$=I_z7}OMnQr5$cp9E%0}g&2~l7B z9a_}+s}wKD2+a&2HGf~la^fmCsbHwA!{hCYFPJ1;lC4}-2jgTuG06d{YEyL9y4pS4 z`5}fDQe<(zi=g+t72ao5?4naZI0 z$|cA1V;wXzI0={i;g|%otJ|_t=?boKshdr0=gISz1NX^b#RO6Ik@tU}eYmdEBL591 zVUg|}zbDD%zd$R%(hv}jECV-kivRp2DUc$x36c4G{_}rre3yXw_q|zv{LNG2QC5@B zUbOWu;Eap$X2LnCK$@vS{C@YKPhA-)=gjJyGJ9cYq;{vyMz*X1*)rTBa-|<8yMeo+ zht9Q&29AHhW`t%WdwBi5ZQyUVoV>9G$1#7zBQ zy0H7m?J8&1%~piw1Av2}&JgI-34#UvC3AWwc7~(-j@`o=Z#uV_r>$Ez9tC+lK2K3O z&y5h38ww8_jTU7{lA^Iln%PNKik)P?$^&EPBcUJE)cm5Gu-6(^R{(HdkB-bXhcACL*6sUR&6A(~{qm8`Ch>1Jh1- zJ&o3XYe5AhLBbh?%kX-vz*tUuee`r8=b{}0$Tsi!f9w!qgE6ca^jS9g%61So`JEq- z1+V!@K%G?q$ZtgMr-3_pyPT01C7{mF>ioaEAS@DOeU1<9fi_yZC90@2z@`&L=ICg1 z%iv*D2%(-=o&pjjs%KY3Ti`%D4xPz&fhHdSD8LO`*UJFVpS?`F#o~!cm6#cu+NX^h zO{om|q5m&ch}`eblft<;g=M0kt(Ux^!ABVj{(8`fz7r`qzC+6-Uqj5PFXTXcbpk9o z9*o*#0-+@w@oH86-z}NNPa4QZae$NoMLQV=pu2*ADHxC{z+!RHNagF%oQfkKd(sJA znr=W^0#KWcd-0mp^b|n32R#O>N+Ik&2iER;Jl#iu^~W2HYORRGmcltY?`T~^R91Zv zVR!#sW@h6P2)YeRt1^31(ggW)+y?S8=vhr7p%^QrONOH3<>Z|5G2E| zr4%{Pz)AWBpmiz1>+L&jM|5)GZ9#&sU(M#$NqK{9eZd$3$eVyMO1%pRy_iVUAvV-U z2rk#;4)dS8ay%fzr*e21WcnF~q2HIAS@@bSy&L@@p&J1HskTF~FgCqmAS+$tYD3(^ zie(D(e^cJv@vlp4Vued=V*TD|!G=?V#jbqH^lP$-JwVSB?%Y}j%MXrYGG^}s30ab+ zikg|=#K`A_PsUREZ4}ho3RZEk_}38O!k6iA%}XAjeDysnX%tNM;(z9J|5V89HybV= z5olgq>kcXZlq3C7Q1fi;r~mvg8gYAvnm2d)e}3?XI{ts-+nA6ayK$2aQ%i3SD43v4 z-3usbtz3KaFTj|O+SebXJl>u{qEX~4ibh=&8xU4Q9~~>-f;J!tfdsXz;5m)RK!uSE z%n9aYmdXA!3>s{L#LZBZo&pfDGF8yZ5*Ye51_O2pgG>t2^M{abLG?yIOv@YuBWmmJ z5m4^>lO~Tnqm26o2y9O#0Hyk{LFiV-wP4;rkPm0OL&!kE`$h2{-I zPXlitRicD8SDw0sw)Vm`T^h;(ev2fxm7hM|g>tjm%W^)S-==p3JZN~AUi-7QYbN;W zuV)88U;Uhw$aw>b`wHm6Ls_6au+vF`R+0grsGf(}+y0u+-!lq$e-d;?YlElH2)M9c z2C-!aj3BuIpSt4A^syq_2XzHen9z3fp3-Z7+=Av2=z& zf(9}o$5PYK?Jy3O0l7-1AYc8bC9v0TX*z?E?cDe)eWbirVC={Q)%8;t)ySx2RKULZ zre%pFuP7v&j~jBK3HRKj?ok(LnJ!-@AqP8yL&^(Qmh=bAbdWl$Ns`gr$DVZx#@3^r zx+%kuEA&ShdT-mvSpXto8F{6RZh8rD49%Yt!#5l|l9FJG5t3XzyQ(l~^#b~GueFu9 z)qz(rnzuRArL?+`ABKtAoYlq!BbQ!LSo2`_?UvVuc}iHA^(lRg(R|x0nvj><&UvF5 znTQS>rgY$^w*QX$PW{f-!CY`o}TN|r`lV^AZ%;Bm3p5*6g6uqY^tx^fK-+@sIZ6n zH=geRaAy8B&^)8bVKZf%e0pg0e6lsN107Bs2EA8aWY2#EgNd zl(l)xA#0;XZ_{D$@7dACd7`k8Gf>H%cUvO_nd_v|HLZ^Dsd0 zG|GauYU8fVa_pT1K;q8e*<38)A)R^D=7_iuEQXFKOBfGtv_8Vv2}mW)0AjGy_=nmunEWo&ZEva4`zz$FyYVz`UB%cZ_?12LCFPuba$vY61?z zgN`2V*`h(`AiLwaY2e?;*bc*2u3yBRN?U}-_);pm)@{G)ae3HJ>;H1GHF`w++GF|* z(-m85gIY3VuHd0V`{mEARSy@Qg1UDer!u;W-$WB~n*Gqu)OS<#xaGp95aV8O?YJq1 zH9UrzTW5WT?4BsxMaJTt1~C5ifMyu3UI}I!M~Z(Uz=t1C z-E}HLnZ>>~Ny60(nIzvh*^`GErw?yeUxvA+PPY4@7e57pB~`NYC&{bO!*0m=2(K0D zCN@ngH<@x~E-?SQ;&-aZS=ao=q#U8I0g8*J$9EWA3dx>5daFF58)b?f(hHay?`0K2~R*08KXtag0Y5M=qP&+|7LKIKy4BV|t|lF5<^q?N328lU_JEGFvT74;|4`T>PL=uj*{mz)i-s zj;u0tJqtHJWr^B!`YZN83%#p3^j$WLsx_^C*oVN-?FGVZ<}Q7zc>4fLzm{z z&^O1ojC;4yoTj6bO+fcvj=NqrXly(nbu(;0|pn*<9u zAskcL?(l4$WNRC2BOrQ`k!nPFsY{h7ap4pQ(7ohLZX>M{k+uX!uv!kqoj<>;sfDrT z$#<--K<6I~E|*&qn{b6IhyrADsx3WlLH()B&XzCG+=C=>Co|0T;GCDsm;~05=H=)% z>V3&DZ}l*wQplGNjrMkJ%>XLlDg5vHlLNi2Ad9H(p9kjW(j098=f;)3LJzq@XhfD> zOiu@f2ij9Zfc*CUnUS?N5po4`YA$G%ht`5P$CG3#&%sgSSH{*aANXXvM-q%~XmlHyU zWq=F{ZdL>WG(VrwbN$fa$+r(4ysGkeXAA3Xzw|2DynoI4$@LCWm-T!F!d8JEVo<$yPiBV7AXFvY-d4Tq77Lo*NM)o zBw;&6XQ;T27>2rupu`BSa)}rJEQa;$7<=VT85SvP00KGTx06s_)%_R}o~M4VYm0)V5HpMaM=Yf-auzp3S+-EnzMaNy3rL} zJ$#>T7ZAzK9eexW{$6QMJUm@aH2RRu<5W;n6~n0$1n(^~(4F=Kx~*PNlI4K2mV-0o z@*`b2c2Gj|8{|F9QL^j98-_0JRXA$`<13;@c5^{1uNY9jlhfd>WDBaFBC8X*WkoQ~ zyMFpHGb|_u5ZuS#JK3YMoCPs;7I>gi+d+^LrOO9Ypy9e`W;UTK$=>NK)ko%y{iS&n zwvA_YLUT)|E+-+G6VI*r#E*|TnH)1!BnmnSG&Tuko-rW|?E@!rN-0UqJ+D;weAYAF z-YCPDOo`*A${jNYG&lGvfNU>j*iJ71?ZACMp%^h8 zraG`AeWaHUHcI)*Oux!z?~ItSrW+k zia9Vn$4W2}k8Y&d@&lV_Mz*zc8Nix;F2zaKz;B|J(xTMIVq#702M8y{%cMog+H?>R zG++9#QVX`~bFOn6qe%((xElRJ$zM@8%gha&r1fMWtOgI)-2Pf3Lh<@A3Hym_z9ACi zD^OA02tN4Rm6VL)+{wo3ZF|$3_c#`pKgY%$(k0B3|8`%XIv0+m<&|NZ;y%V;9^J_I z1=ax@)KxFP%S4;g@tOz<$Y{heu}_iF7B@a2n}K>`!IAXDt<9;dnkV2kMa!OTV}cNy zJ&CQ##%Afmjr#nP#&rybF;D}KHyzva%YJ?$Z(vA*@*Mqn3n+~F*^z#%7Wi9J3k)|| z^S6vA6w>Sw4#vD%It1?zP$ozzE!A^8*-p($sI9xfBbynzrrg?vSsZHf$xiTZ$=R`=^axF^gUJy;#sitB zFd}~0fIgXVcAGI}g*0Qtz_a8FRJpSFN0RjI;Nrvinlj9@7Mxj+6yBODBna|E?{kQ} z^RhK~6-KV|B7?xCaeVkLHP_uT-uXyFjMMMXjr?BUVgw>S=o;sW#~J-H7mP$8zjD4S0spy^bKmcc;zdWyvk*61*a$9&Vakm z7s*s+_ZpB;JHoojC{52Adf1q^#mb(Iyq3x-hHvU+IpAwMT3q`{Kj&H#j7`m> z&6k6&yh^`w+~U)T@#3OfWoBH**3mqnvdCJHTaCg=*hT!)dnFvcq6*{Q|UmnG*DD(YVV8Pv|wJX4r?;Wh!dSeC}cuNI&ZwRZ0FN;vr62{!$^E zqtM=5_Or2mq)fWIA-nRQmwR*mEPMN+L`GA>&^^Ph*^W}7!}{tN|IVnQcHygT7304` z$HS){q(1xAeA^3m6=R%$pTpMd+VMPAQCVg7y#;M_I!RVEM@CE$TSy8S{nj%!FuvCQ zKnI|SqYKj?F%iu`$cYb?O31tAHj5Uq=NM4pthe3J1|SVi5Pq*CA#CSCl))c6pt$$? zpR-s^1J2@hX;B6g{#r9ha!ew5_WUQY%*q{5^8a?loiAN6sE%LVewZ}o5-9B2AsvyR z)Rmn}&i;8ZdpXX3ck|*)g)r%_0dEy_9dw6PAZhFt)JZbcmwPB2TA4T=e!hciw!yPm znp{%xOcu782BnSc|D85w&E1G}6NDaUlj2J-vUI#Z-S;cI+$4kCuWUQ{+su#PuHANJ zKk)4gx@&P!i6fM67K3PR7g_UgdBoISeym*M)jZ2Z)DLkrJkQaP^M&bU+BT)ckWwsB z7<+zm{X)($#RD-(i6(eIDteyL+YJ|qP%o6{Vziy@lvoPtoLtgdMzu$1035Vt$)64LI@SP0mX+dM=y&~ z<2!X@Tjt0jHhw=#Eo2;fVs7gziq5fZ{h?T+M3VD_{9|`2blE9TR^dM}v|DV4{{E|t zfOBzyo$LkSXzq=Rz|%ur8;vo;T=vdRz}Gy>!1iXrjc`^Z=`CNpaX2j%y3TJn^s&4= zK1aF~x(TT0alkV&5`;H)m zO@aiW1BlkX^2NB5vylE&Qg<_Mt_M$AIPjC+F~cHyNNv7_R?`XYe7^T&g&Qe_Hm1@ zP8uGDWH^EsTJJq)<1cS2iv`Tf@E~z|zc5t`O-yvZYn5?tcj`3lie;8?;i6Vt$PH2W z^B8s5GW?pMmr@vdtI~k~-v2;L&ypu(TTO^Q+k07d9K${+J~YJ%2Wf5alp`L`x~|^L z0JX!3b`ZJb)kR*oe256~MGRzG7T{?08(@rz)04I`@ZU51$tv7AkZXq;giYO}KIHTI zM0Saw$fFp>YTc#?nMPdaor&E~VaHB5?{WRQ3u7%OU0n#D&ib9Sb2jmvw%n^7D^9Zq zDI8N#$w%Q0Q%|WoH6BQJCUM80yibq1VTxOXj6zMB#O?`>{|!gfR`uf@N!+8*c$H|E zd?ESwhYoE0i1nU%`J615iWesP8!N-)0vY?`z;evAxUKkhoi;RXa>0$)b#chLrzd@O z?xtlZlw?L1>G|?{7N@@YmIn?KefO2*{37=h+*8*DX`Xn_mCg?xfK=Xhs7qx`gpioi z2x#9I7MhC zuOBMY_*BQ9H{rxV>nRA4g0|Za(6Izik?kjyId~ZMu{k+0xHU*?4Us9NCBA7YQlP7` zMU@6b*_)=HQ2KS5M?V!Z4JFSH;!2c4#4UPd$H=6nd@4MLi6d{14r{Wl2a=ZUGL7sn z!5;YFUW6PIMsP@CB@SN3G7)!8P!x#{A$hfv2Jfv&tCltnhJ;HrVSV}w3{Mj|qf}i! zKXVD8mL@R%zC&P@mNl($_jKthv*p6E!ft4dx~?yz5cne*R7xh`#yIAa#R{?H8_yJ| z&%Z23|4Lfh#d_TE2#fxm^DIuG8~#c$(lg!xct zB$}X=RC!{h_!U&bFFHQMw!Q%*t{6x|mNJ*9_0=YcIaEW9)O$fD$4~r9|Gut?Poe2- z0{$3<0Yws9gXeGP7UOm9$%}BgPiM0IF661?DkQ1qKsP3X5wtl=IhoeumbJHHLT^=Y zK+FVdsE6=3@E=Kq(PB@Zr60Uu-xRvLcN%{NL);onB)>|eZ0Oi^n>lHCi$D(c8)qY~ z_PxakeqWAz!^^*ZrokU;1B-Fv*rSU}H#(+j-!~mpx5wyF1~Jw|cF0`x^gk22<;3x+ zjxa&zvi9b9%KH&VJHG9|`ejB;qU%JR?{PdP7+d2%wIyjJzAcXBGI$J){nRF4$VNaX zOFMgCy2F&pyxu}en%%D=HyFHMXw;51$u0tTodS?;TNS^Y^RT_ubqj+xI5>8Fj~G+X z!d^Rm*6SP;^$7V!5v5e+$Hedfpd=~n#Ec=5tLYFyGeaMKMJ=36nyGbV$9Mb z_)~It`qj_by|{i(KiCNgvOyiG(w9Rn_wVTBYPGCQN833tUOS=toM&px$EnyW%^H|O z9*ONb%!i!}dvHWXeervH6H{mFIqL{1Y@{b>8n+x!6K(3hgjf)eba8+Uq`w*WmKTrR z#<2@qi;mp6`Ww1$z&x}1=Bx=XFN&DuYp|1Cr^vO%DS>|+f%puX#;~mmw?Z(e(yx2o zyMq!Z_X_lGhuQI#^?ZLGlb=;MQ`!x@IT-uqAUgy>IAUx(E_Q|C~S)qGQON6yOasl!QO^%yl`u{xEU(w*r=Ml7UJsH;4p|7v*Ox8)()HV}$OhnU$ z1>?`JVb*MK(LRwf{B-EwZw!@f8q_=12pS2PeJ*ta8YOSLs`8sK4K=YYhpgZquK)Q} zEECIqd0mYioX_!BVVqBJS?iDkOrKMzXMDK z#I-M2WoGb?W2Lp!efOz-A((1Hzt!*??h+}A*LoqYJ>BijXhn7X^JeCY5T-x26=SeU zd!T2B=t8E+??=}>ny1t)Y+!ILXYJU-QDg35DfAQd$y*K{qhzMYe?qs9IdQ<2zH+jP zv+EQg>+`|GB9Y@@4Fr>taWo&m>+&+V={*6!*P;DF+Ge^jLjkOkI^^KOD7XT&X$wLP zJ^(NDwI4zLd?`DX?Vrm8pG5&80o5m|f(HKV)GALNK zOg4kW31;qjZM6aS_rnmG*+Z-NZo4rlxu?f1veGzNwk+&x`MJB{)zq?CfzIxxpDd38 zuqP>K8}wtnCMj|*@}b2-6eo3gXCp-~-8J~y$>DLV;fW|{=}JN#t>$7e9E4SJm}7VZ zx_%CTU93S{CUoX>pXXWZOvDno7n(zcPttX9(1R}BL zDDb3Ftlbb!`dD9rVtH1#Q9q6wwX&F6vWHKACKN{W?vE=ya6kjYRt6XwA}uX5(fm$H zUh9}h4P9=)$EBY>@;alTqYPLjd!z6^s_m!iN98?Km72DZS2d!jk`p^%z!k$r*p5|P z9HDcD885$qjM?LcH9W*S!9l`P6#dBuhDm-oXLQF_;j~0ZrLG(X5W56$#q2-=xXgIm zZFN=ZgG*S7=TRsAq_ZFJB#A6jFLo6)(bH;fB>-t?4HC+0lsIXDd<@s_2MLQ}Jsx99 zycOL!V@D_{q|+PGK?Cl_8X^yEqU6DXMMUT(D9T#(vy zc%~tctmOML4@jS&HGdfVK-_)E913IM+R;=i$L4Sf?>8nx=?uq9n`!Z@*I0YtD(8fk z(p4~e+#!=n=ctXc5XQZp2&1D2kKxm`haPOmB;O~yb25LH$AIemfi-kx&A_!E2KkI5 z07!4kYC4MrUFmm_hCfw5pJ&(eaO}jK6t@pHa{KPQpY0Hlg2Ys~xYbNBD8u=cMfHP} z-63(`PPc3U1X+VQt`(SuiMI$SnJv?&XVuP8_50MkNgy5uOMs*tw;BTcm7h znQq2)Vq9>nb9t!!Pi*N9av&qj$P4sXnc&v^Xm&fwTP(>lG)R+ z=p@9w0Ad({BWhlT_3M_I=6Stcyj^8B+=(kr_76sUcj`@CKlAF6HhB9v1+U+M^{-mo zWkJ9Emz85ZpGEJR0h0&v=37Qa@gz$O%Q3+k=T&J!R|;FBNP=rlPYhWAp?qVfvt7U_ z1OL5`VX3uTiBN}Vg?Huk#o$F|RW7SQtoG9%hw`GJAWVdH`wyVWuI7F&c?C!YLUfb1 zJHygCR|F2KU^yN3;@2CH;Imck7>d&YBZvNllkJc#OEyX8wP%kf_L zj=iMj5ko)~pyUMl&EjS0C6~W6G9E?<{fE{IiV?_k#)?l0JgyQN+7*{3g%ixu7faKq8)hIxdp^e4_<@O3Hx9BqcIn( zuBiqAXol>q5#=C&c7s7$T~Phivxe%1CD=&afRpi<-&wl1BAt<$= z_5;KwY(_vUTN~p;zAI&hLJgjPBIQSw7t(zSMwA1Xx~O`Ow6lMIQ+ZJM{eg{}Sd&Wf zV@L|PgK9OzsNhn_m12k3ta%)7r*(IYdHg>4DE>%n7!-@zE58>Hrg>%9?x0G59#&mw ziYwT~2EvLNcET)7EurD%Edei{htUs}Xm|y>sm(ngJ2ENI%jQROHf`--f|02Jo)9rF z@T67tCqR^G;7G0m22eK!t_$#*mj#3NmKBMRumQr1s!wp)l zUe{AG@N)sJQx2vx)Ru?j<->D_KxULyV80v>)?Lh2c2gs65nS9 zG%l!gi$fQ1RzUkN%iGp6`XCjk&v`DzdTxqf>kyuLka}7|58K`4es!+5Y{Khnz;U{Z zBokDW6kJIm)10ji<}zPTGuuuN=r$u-^Xd1GY>M>9x*qL~f>l!c_VAVaJ3imtEhe7T z^K}tYhM~_{*F>h6QfbaUE6eB>UpU3PyZ*~FiJQ|RdhrjyR6Ap5o2~_b@t>jh;|CMv z*7))V9afAER&nXAvvAYLy*RBaeqK4Mkk&JW9?}*P5 zG2qLXhOTsOpa%+Hlv*9JPsvIbd{NfUYgJW}MlT=*O`x~TSfr!B52BgLmTu^Tnu^jC z1b83?_|aE^EF}WCYmZU4kN+h5oLp5O+c6whZ~cSaOdj^7m||!1%ZRrf50h=;$A-$H z7TTqTBX?{=1tp;Q$Arq&{BfgE#D`Oup`9PUtsMQX$#xWLe{2C_$k@uWaE9jCjRTCW z3A>Y3u|#oe8Ci4Lxnwx#%qi)46NpB}N`X>$=(<;#w!Q8Q@gilVpW=C+T571rWb+y$ zc6RP$koi_6??|TzezjwZg!o^*c4a#rn|RtYKLb4oJ4E82`UOvDX4!swU};Y|mtXWd zpq(FJb?053U$Hw+R2TK{MJjq0NRkZf_{dj(f%fC89n{0VDw!C1?S@rw`4j`u)X6+J`cWvC;eZyP<*did=$1E`v*bcB$Y3J@II3H z#c7hOy&sA^-Zw`Kzoz%duGO2$Yk$EUyCpUXu`zAi9#dTX)N+BFTb?w$IfB?AR+SuG zKC0F-q9PO7-P(k8)jK~>?UA#}ZhCg)dFGLL)eBxzMCJJ}A9A|uj!X9aDiqDG%xJqL z$ot9ag1E;d*<)toqF%`J5a?T4IH!=zk^pW{>Z?TI5g2Wg1O?&~X!^butU^2;|G^s5 zbg0Z{{8>g~NHGQRuyn#@A?Zun$;7AIY->v~g~pz#Pqf}>M@+6iqahU%YUUd#v$ z^oH2+>CohXEAdRC&4JGWC36JviZ?jRy6ic1NqH6j{-}E&F;J*ocDE$KsYC2wd$gf?YV?OD|S`)q%^x} z%pP-DBl5{gQz{zV)0`AQz>IacoJywY16v);Gs@x>8xk z@J1Pw*$?)@C=iLWI>%qNC62hcU(Nu68fkxs4%naU=kl2Jj?NR_LvAvfe&T8(5U(*b zCU^$UGLl+of*u88ED>0CeSmfHzl&x)S)sbL%5s>f)7Br$05DcGGpYPEzFl~hnkv+w zAlB=^mE8c~pri~^ezxRO(8xc3K{3nz!#1HGpCKbj_&Sb_rH-YGWi;ou|82k5FqE#$ zT~(Q%m-$*k20;_B`!u6V50eK`k!>l_OK=Ap*18?rI<=O^^U|4DLs{AaX02=L6RY4?_gF512UeeF zqx9CpUr^D;q70&AQ#tW?6X zae$(a)#!|QJx*;$cIFsJl=b}lDGD#q(~=(|`H1Z_1T}K*$88PDXaf51SKP#suZ)NL zX^I!Nr(RYp4U!x_3YQ&_aHl zl7=A2O+Fak4se5l<|$oq_!H(Yu?F?GxbE6+8QqaAw-Dw>^*iy(F3#h!r}~=u zz3)+dR}*J8w=(|2vT6RKXtl@TTPd$(eQaw7IOU}sNmBwoA4{~uz_9U;k7eWRABh0I zsVRJ-js$Bze{^Y@-1E(6Tr&1Oc13zBjr@Ydq(MU@%vsgvspjP2D|X`nQwMzkjFN;eYym~ z47smBhWv2qu8&(kZu?07k>cY;DAp5fv7MTCK0SV6+2S((P~@=aXk!dIhAuWbPmX5U z4fl;_eRdl=FvR#Zk0md4*Z~;p^hdVzdJOw>(Mx;rt4VpLdu4%TMX|54# zGv}|2cLS)9To0ge;;A#6+j!#-QPp0CX-eIpcDeC#ASFd>49R1<#|)1dAMbw548;!H zJVL;hf}ZXAPoW+%P9DpuWZRm&!cv9*Q-hbgTEqa?15O}*n{DxAJrTtQ%K&qc}yMfE@qqKtTV1FbW0SPl-rHq@?MiT(%a&?tn7Ch3Dt z3qg0Nx_iNQ{Qkw`GR~SLAE`cW5A}+g{_HQ~^&vbg#hi}d%n$ZTl zep#S7-s(HS^?$MV-r-pH?;kK1viD5($S5LvZ;2E_ifq{-Xz^LWaUw6G> zYms@3bfE%g=E-Lfm(JzvGnGZqUrtv4{*M+R|Jk5hD96IZ^1Ab<QE~r99E;rV>dk-wD_-RqZ_s)FV(1}^%NyTdFci+lXF+3b<|<;}01A>U z4D-iC;YZ5x_Zi8t1LJW@!B|?Sjsw~|>Lc&C#$s?ENL?Y!3hJn$AvEucpxeHNV;mg3D-vZ34-yJ+f z(b3@{=}S-teU$z16T%@LlALn0s9-PfnKsrhgPqMsQS(3_z*}Qf_!;z1PyGO7vq)u z!W-OH(XxZ%RaU+q=h^^+5PO%Wy_@}vsy(Xplgo!?;X%npp|iLqa?_vnygvi_w#&jL`GX;Ij6hehLgm2xi%Edt`T3r$_Cm601dx@<06 zqo4YRT(M<>sr--saIG{JwewM%e9@(v)qDH=+e35RzF`UJTv$ z7**aX(7-GCCa2E!M)P350%?iT#JE=EK{F8J8qBC3F&Nj_v4ar`ecPnkpg;T^<7A&u zc+!`*f4%%l^}V^^pF!X~0tRDcWE^jKq(eO)9>G9PeK|3Uvm9tBo}+{eajtf8cESwL z5iBD26~oSV6%(d>(lv4=Q=~4!&-C%z4yT= zy7(T;uO>!1N=qX}v;zJ^9EFGQsBTLcz{uk`db=i1mGE}cU}z`)$U3-2B!O3yX4C%C zIXP!1i_IHFjuvZ5rL*kjmPR$W)>g6v`d`^K>TG)H8sCEu1 zmkRhgQHm%Ic&Q&g2K`Z)@B$@4=yvFh+^(O(9$!zA!mp1T& zE?2RT-jXkWeJ2R%=*Nn~KI$HCrO8XavK{7&ZR3TW;+HkSk+F)nFD&@(Q z6K7J56t*W8o$1Wh5YFT)xnhr|l}N%*4;7xk>+ca-9CrWE)Ws3zUG(s$`fqw!B^6&+ zk~^!eEZcd<;rOmMb7R7n-on6tiJkN?>Yyz@^dG*&CI?oOs_IpdBhs{SsWs^`6xdUEwFq(YL z!pIOvL9_TO0#EyipWLXj>}W^N@Pq}m%a&5!qwqeT8IUz=J<4;IW01{K3krnR_9=Cq`1E3m7t7FITd zy?(qOE?U_sE}-?~y~!aye^ghd8ZzeI#&?`Yp;ApqU4|GHSf~qT-(RI5A~%aj<)yq% zI@^|O9k3TZ{z{xOQ4H5iyo#u4q|y`5*qFqT=F84U5RnP@(*-XQ+e8`&MRI+ zqmz%O<~6}M*(g8o?c=VbK{hEtaYdLprb3L(VE`Ou$~0Y;nJr z1zZgHRVnW8!jd$`HO>hT_u*P`60?NSBjd)1#c?`6!kWKv8Qv5$8K z{YF$xez9IM(WV{;tiROu?uw9{Wy(?xat`0DNSVE&suGqpaNTUih<00~Ey7JE`swwJ zUGsZKx}R0p(6+y>L2htg{#4t&PEvJ7Tio!d&SC50`v@-k*IOHYcA$}n<+tH7CXTa2 zyE*#2Tb1lYd%a*Ff91m=(Bj?fc|qc*iqobLUhGEyBkCW)ZTe1I>I*bJonDRvOd%X2 zg0U98HJfbLnOEC{{e@a4M=CE5X^8F#VqS;&Q^eF;-hUHY;G+^bM;ad`a3M@jb4%GT zVlMbYXPI%`sMV#$ji9j7ZL{_CDT}@cVb@9rUP;+;^7c+N`0Z5hj*%qveL2i6;ld@qKStU@4k;)szDb$ z$v~Q0n|8${rz{_ow_Xv$_tPZP}p(o#&QOf2E+pZq8W$Ge-*k)0Di_p9%W`VXzhD6IQreGyU z{^PH*F{RCb%9?sQ6susdagBUfpE#aXaAO>()oj?EYl7V7SWC`N%I05<+g#1INj(;` z@nASit5!{+UYV+^z$}oR_XQ0P<$Pqo!ABwG=Ou)9VZHQh=1nK0RzXuq(5W_5P~A(4 zXbs`s);T$qTj@)S<^aQKv+okuXwc?K^l(T2Zz=x%tt{+ z9hJD0N4N2sE}_&StBMGBSRBod@Den7>rA7ZMfXJTjNe$57;`xXol>cqMXPNbhp#IU z5!Qz)ifT#}X;Y(wCW7~Kgd^Vh3lWO$sJ-B^zYqqRb2LSU9B>GxYr?3{i_He#W}bU- z&QIkxmDPjaMO8~NyG1tPVlIsmP3TWS?bqTS!nk?t&>D5eQ8ifj{QcuLvAypuEaPxi zx$tB@_l%e)+tNtFBma_w!M(m4x~U9xG={nRnH#scWW^T4Bg}j{kBB=yc51?{a+BN5 zG8>NvOOOKXM4Fws%k#z8x=1J1rqc~MyA+1COMP!(Qhq{eM!CR6egQk$AfGKT0C#3} zGMbgFOhon?K6UgnKqX47_(0jxIkADq0a}KS#ey64%rDto+L&}ms8wRZ?|=UkEq1|_ z9v`)0tRmUm8Ca-&+cmomTQpOx^&u(QW;AQ7c+z{0Ohv5N_I{ovG!apF)adit>%~?a zYanLeSUs=cB6(GeMV62w9MeiT+T5v9v2ZI|%#{Y)$X)`F4}W}yrYlY*=YS3(?EZ({ z^}vlOn-bxc7!hnlb)njrU8Dx5Ms7f9^8_2`%*u{p+m#6I!dI!*ln~olsUC~*(oKEB ziOW+P4amt;g72T-+$?{y#27;n{U{*;IRRAWAb=Gb=V^)Wp5>}XxR6s#kC_Gnoewn`5ZkU7$_l}vrH|8n^V z)3Mhn*@bR~aW?LJjxq`?V(3H0nW4okBcezkzY#E*w-LP;A|%UM1Py#8D1##K!epi& zWY@jPxfVqs870nP6FK`#!=Eju=`W{p(x@m2>^AzRH4dLNr~OZTf7~((X#aLiXxh`t zkLw99%JqOaioz%iOpNrSH>xf7{IvAl`o;orqdP)&6f$!;uu!#NM1Qn(RW|@QY7YZ+ z{&08xuPe!umovj@4W`)#DrXN@tLml+>{CP1D>bMhVzB%+bn**PM0;<}(HXY|>{kRW1|hFxbgvCEyDE6h3idRb-Di3^G*VoA~6E z6a{0exw!$WPydGhH86y@;FBN#)*v5?8HnixM_lu+7t^mPq982I^-VoFprJ*(ct6Pj zvkh=ZM&p@H=Gq!;94H|d8Rt1V-G}En{zc~vRHafJpZ}Y-V*^-YkWgOI0wst9yuX)m z&&y+u0}fHU_&=$UPZB{sDca9m02#Rb>kq~g7jZfCo(-iJCF%q_&u2{(CSwGKskl_C zSWQZatd_r);&{XoJ57shY0U`0;O7E?Z-*gH4 zGkv{+YfkEJvn<1^T~*Zh>xc|_{Q}bw|`4ioFkYy z!T1Ip;*GM^`y)h~ZbXLjYrNX}Ki4007(mbI1*5rVCnZ48$+b|t`{%Ho&LbDlY^EA- zm;5Iu9E;)ug3S0#%ew|IC_nX3?-H@Xw4SFK6izpq4u8rZooJ@CC;!hyAr&eOdK9_= z0jx`O@`8YPX!36!&ejc?^2MWcngd~4;A;*_$(|o*jKB?|(r9NII27f*pG}~QiFK<7 zYH3&>X=_gSDK|x&+xypcD?=WJE5ouG1Sn*YfXS2=h)N5=Qdh6aQjJe17mL%63*|+$ zbh{Dl=KKjbt9q55RT6|fJ@8=vkw+k88$=rimS?<(OSRlh5G`U8xRQ}$g@%6&!T_b{ z!s)L6>2rk{L|+XZT%7USdAacCTfRM>%HP{Qmb(!|A$8a4q(4w_5qWjHsl9311Kc?| zvYTW&OpL1B8cb~Ji5-Z(A4EE@AVdsCg?geXmj&SB5uiSx=|Q%33Y@M>2xS9e#npD! zzg~iEm(Y#-M!ezzu4MR?fyaC|qcAWxDKu;Ldx3RU($(CAmW_>eP%3lR-XTU##@K31 zG6aPvwUQyp==%Cg@E*GbR+Ub+H*Y5g6SliEN!O=cEo1C|-8L_gPm_<$gEE28=Xb ze4GbC!3%vBNQ@CMh_oSF_T@+LzlHe~Eb}w$5VNdtmnGipJ<=cU&Pe%cyb#SPk8?R8IGo zSq(zW&qZfZ^)qFc@A8ciMzY37GKG)R;ouRJDk0QgP>lVu*U3Hn$b+p=KpJS*$qsu_ zINr0ikR|H6>J2PXRyNm34fa4gf{~J&S7=!`{t8sM**Az*W7IT98GRUgc`MFz9A|!Z zh=96wfSG^Q%S1>z0(-Qp2FfpszR3lW7rSGtt!-SwO0)Qu|Q z7A{m$WNv1&C#}9=hork%eD$nuW^*Ti42aG@sZmCj1JRuNtWa`_pW9LA25NeWVcWKt zI(Id_GOfTh2u^{3Ls(aX%@CE@X2spL_F35+SR)U9ujsz{Q9HE&bMIsrLi7Sd?1dlk z2D1ocT^ERPe?$&oSYs3`zToN|{I>-&sE zxg7)(d&&Wg1fh6nZ^Z#3CUB>QVoyNQ`A$~SlB#t0^EG&jg5=sR$Inyj(*G1O3da0p z4!qpo$3WqG?WPkkBQIcBnkpA#!xzghuzhq(^|Q`*6n@G!wE0+S z8=L@qvhxvP;sfx>d$bkoMNbq5p!ejJRgj@wr>!9j@{I*2O$Me;+*zZcqr|aqh*pN$I8c@hXtn z&JPZs+x9rIvrK>`VNBtf{IS^ok?IPTLIW9#EJG!htZx&Hxkq@g^PO}vwLm|zyrW$p z7F-C^a6IbI&+E-*MG*UpCsL^H*Die~Uxw!#B-MiwW)rjohY08io`oJaP+SWGc;aSw z9HLw^0T*vQ;J~bomSteLXiyhOz6&>k?6v^tp5Kh;Ab8*-Ud(x+6J{VwA*f7@kY1CI z_qH78C~zX#XT zl1B+uH$Ku7tbxXMUs%8tCfgZ#S%I-O8tpi+hhF8xNSTDVqd$h3jR)l|pn)dWR@C;_ zDeO6|MOHxzR0K8W0Y!ZC+Ilq*_TN@ty6d^TRrl`_BkLYwH32WX5d9Sv8P4(hiPwYb zHxb0Ct{c71sx|#60Z}z4Tn}*HYu5A{n7zvSSm^??XY{~ioawi0Q({#%;wVIO$yC-( zI@|ROXlpWu$g4(B$ov14T0#t*=R zcZFN%brw(&_+7vNy&(BQM7SyFpA8}u=SjcP+x>ye;JI|@4)Q8W#8`}KOMEHN6sS43 zp7^6wFl(>6jEh15dtf%R?HRL(*K_iFcqo5sBjOsT3B552ud%4b4n$-(UN5!y0<=rt zR0a_R9jpQ#{l|c&*%Y%}d;yJc#;hbcREDb8LHHA2{Ys4cqNq7#ptBGn2@fk*L6>r< zJ+_lG|6$PsEVgx!3Nz3E7{C8Ky9v?*M!umGQfJYgFK68z0=Va}ABd~tlnLK%2vxr% zbOIVcEuLABg{vUWDqEF^j`1bdCX47aqTfe=ryK<*&rNz>>&kw-9=e;@L84%}_o<+m zxIFlQEr5y-NygwRtJn2|bCYJAl2Xv6$B?qTvg~nS8-r`*n{%CH+>pu@=g&pRlD`i2 zQJJ%mm-AaNK{SperRO#%Lgxfgqpt(E-@UyPKmEGnpPlH`E@3=7_Amw31qvB#zqcI& zf_b-qC|l96O#T?*E7g{!CMfv(ZIw0MDZ`oZp*7Z=(I6Z>OR;yx8?0;%swzP{&Sz3z zIi{&foq0WYw>bRl+ih9MT>&c`rZMhQLjq&H86O(W0H~n}cb4bY;rFG0v&AFk@UWNiwIystk{1Oi z&^g|t@V1Rtol#P8RdggT$sxx+?tv2~$U8s~8!2+8d0Kf@R{ zv<$yC-8^5Fi_-3ih36*dNnQp$1M}K%KbQw}EzU#*IMYuR&9!R;P2WeAoxg`5wV&);EA5w3{!T!ROXh~T&KeF zk7y^U>C2ZR$0%d#qPhP4B|6^2Umh5FrQ~kz>8Oq_MDuijJrF;0?e~k9GLG>XdoD?S-+P!jP>4|~b){qmEI$@RQX2L1Q|PCO<@;=}27%e;aRGSL)dl5w1c3^Ss!X|mM$-e{nQhb};qKGdzh||1 zPo-Sy%D5DV_MbyR+#FCT@omVu<=IU+X^Lb`&wum%P zU?osxz9^q&ybn%hj~wv7?Gz~fy3#^uDan6)=_ zq^S60HYJSYkIJ)*5~B~(3|aHa+W131YN0RDnk=3| zV{DU-Fhl>w*4ODMeHXr|E2!s$<`vwv7`X0XW};UaXmzAp>Uu${iRY_?hzSE-aV1^C zIg`e>R>fW18@at?LwmF41(?{g<9Gr+&be$_FE2F~A@JWH!|1ZPu=HO(#s~^7ASj3zIeP0J5wQ7)aQyn0 z@`wq0U}~0N!d?FEl5?!YqfQ9Bl$>s*oZClJX{2n@-MDvuAWD-$VwYu&g&I4grGrIL z^8sP*KN?+vD*;s&+7^4Yi_Z5lfg(=xX5soMMBbv1OU zn!`GWs3cO$7u7al#GDF!N|aexUnKCKYzSSaz0}8wdK>#UJ2&s~V8=?x*yCTiaCfP0 z2KVC{31|a;t^<3Wpp_}j9}&!gTaQ+=-!aurQX};3-xx%@wQtTd|zU*39fcyfw`&{Gh+4KD@0~Z&~P^*`x$G@IZnMZu)io^98 z0w%pNQmu#uUik$*UpLU;km1uW$lmGr;@=TT&9eX+i>vpgqW-uT?ZPwi^F-d7_}_!4 zk5MoWB3oGp1wao%ZMvC=Al9%YbHQTw6ww>P$N3Iy{YxJXlPg!Uztj~f>Haf5q98;H zuMj<|kKkx}9d-zAL;?bYoQIA=ZgZ-LMIcfsn};0(3*tR+t>)$8B!K0dIX@! z70F6p$qe&Pn@)l0wg6JY-}}8F{k0yC1p9c0vs{sG1rBzOcD zufvkU4ce`E5K~@(F7)MCt2ghjAk^aPG1Ek-&eTjnVB&f2jYZoTg8&~0NDk_dFMqa2 z6_D20HLt6c4SoE|RVS6nCf~Q$vZM_lDWK$74-{ixiJ$F-9Q8Nz3R>yGWbPc?qdM#d znNc&1%(l6Gf?U|Kbx+Xf%LBNuP?97~yh%rH`OhX49+BOTTptYiG6=M;#xRN@MXYII z8{qalBDMqiYFD6LT?9+EL|7b%F5*6>6L>ME@@5k*F~Qt4Oy>Kh=1z~-{a;{#J)Ae@ z!e#5djag)Bce~O&5N2E|aWNpmMK-#?Uu`ra#6)zGo`twPt`L(&bkNO@P*ObBr|*_2 zU4DjJwbZ=}cV&PB9{=0Nl-kacE>~2d>_xN!<^`&8y?sUD*z@x`YDzSbg5qU0+Gk)yzXYQ*hmIjow^yKWmp7~)K_j9E>H!~pVOqBF`#2ZKp&pe* zGXOJxSKN2*W zn8pRt8W#mMl(oVZ?%&li>zJp+_-+fw)+K#6>=J)G-(PjN?J;)*+!#XCxi}{ad+b1z zCGd<7nn03#;BXTBx(i&61Z2Fve^6lBt?QEnRV7~DUsa18NuV)%^ptOW&awyQ7cZtR zNRUQI9h%|;T5XG(q{vLwFjWW>V*LwPVydP<@Mr_%jU{k#&Z+THff}bnp3;S(o*ZdR3P^w(j?e?em(1`F=z`J+2Kv{oBYZ zy3;T*eOZ!+sfWys-*SGKJBYyn(q_U$Cyq@S8+Vl1FwhhKq9mAFewQa?r0iPd2^wNgY@iu?vA_sv!fctZ3}(1?!O|!hQ3qAXL70j)wC)a+ zmiON&F)rMNHV*M-ElY`d?X*)y&gjP$m zSI!u6bmOj2b9FOQl%k!~0z=PaCUQI=_G3_)S;Kj>f33w<1Kd>f1LI}>y8q<>ne`BJ z#4B##Fe=1u-g%+VCl@BWp|ab}lamCamFs2rGm^AjfFb^xz% zQR;HvpbzOaETV95D(HDc74JcD_KczU#@C=9T1AR)EOv%K3!Se?{c?`*m^kVQWboX| zdXYYs#|rk#mvrv8c$ZH|=-gN8C2!{3*t8_-?1lcUah|BF%qC-X?_Z7ez-*P9dym2Y zeMh+x;i>H|@Nf7IYx~X4Smu{G=AW>apH+SiI6cG5JcEPFZZJEHLw0Sm?d)UXz3qT% zJ>Ea=Rqswfb*CQ{2s$gfR1+hrMrhZ*apr^8?(ChYGLgM_(hE3yPXb=y=6P2YqDFDX zvjbUwjQe40oeAmv0^{E>A2gdJwv%^YLMy*eVmQYfNHc%w24OBm@M++L($vb_#+!|2 zuly`DsD!OoX|818fH{pi8`Jsyf`3=(JwH%DWb7OICZH8INMXRbxG$hV-{GN4%l$qk-5rk-2S3M(bgbn6iG%h$q<$)987YTq5 zUG_K*?D*aTwbtM)4mZigv$+=oJqde4Q!B7V7F(X+=6(*S9)?Qdu%NB>{7s72__J}x z9lv{n%Cz=2@d9lKE~=UnqSP2JhF5ME>EtOt^aImb#vUUDqZ11s_G_2h@6>V=eg(F6 z2pepBxS2FE9o{k>(j&9|vuJlTexcE}83lNdcO*dU;jnisX~bGd_G{*A3XZ%UsN0d?ErVQoqUN$@|G9>cG1>U>*mstyk;Te zgT_KVmsbMuAfn}>(j?QuDa6YU51~5B-TtfLEWdB1`#rnQ=F8r)vVn+jkD_)o3DMde zSEFTF42)8{j?MD=g0{(Px!AuEmx@oL=SgD(5%MyCLjkAUjCg0r47#GvJ{xNZdJZLs zgAy+O8-N88fKjzN%EPN%+EYf7wuwkYlwno4K`>f>ug6=aTr%x81B%JT0&CO51S*UU zS^($y;7Kx>bStfWL}1<<k(o3 zQ;xv$w(W5uyssRU<1Jo{s(Gsu)z1&O0~7W4$Pi1zR6pgu%l)?&wEh(B9_KOMGFgwl zp47TGP7KOsX)U0&mZxo}v{9#FCqfFFVief(bkS{in<~>vA9=I3Y24dvd}96i{sX@luP5g(`Hp)}4^*Ygk_-q; z3>-?B(nz{=MJ+V%t6n}c=b59fXXs;`vj16f`tqPq0u#sCb3?Ih8o#0o?t2x@qN@|V zI$v>|6)}mrw{TBRS&QNVX%i22ICNSJM>gu_ftF!k`>)A^!T~8$*<_KT|E#du5dmDl zIBs{M8&DJgYVIl?=-=ie`8M{yX<6VF{PV{eDRk1*SF%Ye@m*o1*EL`m(gL~--6yeY z*Z!zTDEwhKFh7{|$B*UDj}@@O-T7ENNJ)x&UE~?1jh-o872=tG+Ng5_!^#j8x~UDzpRUV zqjBG-hr_8csl%iEKssd0>p}f<#}X!&)yb`HdxdsHAH@bOm9I+vdm%`9UG z7~Mt0^9ZcJ?N>%%AlH>K20H9Qm`dusd0<28i#YM;Xovu^*MxY`0o?F>73=|?uY#N- zPge(0mdcjFmO|iyVd;lmWC*$xXEY01^+b~g-Q0Klpw_d2T&(F=F7>l#5MOIQvRkp< zl(OnyVaIsTL3`^x*cP8h(mRi!7+A6=-8?|*dEgfb)_r~uH9GO=V#MD`ZG@8tuMp7eH96u&H2?QTUr5*ox?R( zVwvJWw0|{Q)*JEhzDLP!_7zC_{a=|bu~Tnx5f-pR5%n{xEx5lUvu=1py*96#$wgf8 z9;GR~XQnhzXJHe<42O;GkW3x|hkk+7Py@2h7l6^82rp}i-U1YANw6jADX3rvshv%t zg*fu3U=5=M&Xs%!1%UOv|IubRVn+Rx5OGL=F;y=p-xfOP&5}T@s|hW|b{CMuGZsNt z`f2w)`6wQ*2(WFC;G-8dm<*RDJ^*EhSiuZ^8Nt%2;Iw{dN3rfR#DWv)H@ab85(I%XHShWI#$TGuJWH>kJo_{+lyT)$ z>BHC)kUaDvXcvP!NIHhBW>04Ev-jx*1qKrprKx6^{KEVs-c+ssxwK!#GX^R|x-zy7 zd7nDG-%Qqg*ImmM1TR!%E~7>ERGmu-YH1EM64gzt{531{;^NF-{ULfTU^C*q-qSmmMN+Glv1Y__60py ze9i)5>;pB?0gf^G!7;HR5WQjIWx$PZS^*lBcfX*lOQsXJQ~U~FVSxRs?`{^qR`$bh zdaw!w22{kHl09fXv7O> zRk6PUb7w9uDH`%8KzcAT>-X2uoN<9|`QiI-oUPYzqDYU5LxZ($f?tYD)2CZTlp$2S zrqDq?o6<*2%?;YYMwe}b`=Ow|6#DW*FziMEZjl8XwE#ZEYyGdcVcr)F%u)o5i>~E1 zK2`^*2oFIdp@MMbv{0F12^(WppAYBAchh7 zrW~xxSa5SrzdWGFidiFfoA&ahZ+L;tuB78Qg%?w7(6`HuiLAc8nYL>K4WrW;9R$=~xPFVr0veaxr&7} zw&aPJG)K@Lc@UH5w-&s@oRg@1k%|hX$h`Tf!p!VQ;2{{TpGSNU(nb#N3md#JfyEm} z`S5&kKfbu-DFZug!g%ob5>|K`K`;EDV;bhE>&Ow~FCB;e4kLPoE>^_Tk6p6U7uv0x?~NNFx!Vu1l#gaYh6o3`oEOj0g9!4D6TyEl`Yd~>3cVWJr< z-Rpr9_~AO|=MO0QG0Lf18wIRQ9F+q3#&!1*+m0AAoI-ru3J`DJ>ZLd$Q1^%jv(;XJ zA-}~;&oz^FK5lQo!_&P}8FR*v8}aAo6#^yQ$Y8|zU^~udSK>zBvlCPh=TI6(7BDS@ z`)tfJQ(vl`WV7r>ygE%P#F{}6IxGLs4qnM-hA*b?%l97c` zn)#G|spV^;%UhT#7$v8({=rRypkb$QHSu}8RU0KU7;3PLcDH!`roM;CMewJrp9PF-F80XJW(4OC|d=c18 zD>($Fy?bX`M$-Wj+Baj_v~}tT!sy}cg9zGC8G(>i5ROlFT4Kda6VBEFv+3)b zWL`#FlG9A1?0lxov2`@-=HizwVt-A2VUhUU2GypmyrNSbuhT+)W3MxpD+lZR3Ho=* zU4cLm{gBB}?ruDKlUzC#^f#tUEj#p16@0+r$@ENq_fS`6oml4ZxF$bx%dzo{{oEFz z%;1!iA@x_2-KXZu$LGkUGwbG+h9u%NIu#p#4r}!MD!cnY!L4zkOCh0(@`q$J*NwwF z3{70IbsN!CSsV$%Z9vR3t=y%*vC@FJufa-Q*@{W#K6Tc@PU!R?Afgo+vAuS7gX3Rx zLsnE~ov-VLN)YV0+zdk1yX=_FrQl#z6U~4j3#6eE|4-@v?;hZZ;{Wm;`p@^U2RX_)a35%b@SOr-UYhI^&j6;<*d!svhbHvC>qiWD!Vm)id{1b+;c{+^6eFvIee zKi@I(_kaJ_?|~aEH@LxS$;5s_oMZpjpNLpEs-aCwhyUYyY}hGC#77}>m&4m|4Ad(H z0RASx5UD#|?4>@SrZ6R57!H7`AmXNkxYr?Mma{_U0ZcMp*M^UFrd$x4Ngz5_$?h*C zLhYaO9wB~pAsIf2gHTNV-h)l+H@ap&7SfA(6#S6m3-sWEj*|J2oVW85T8rh-<6_#p z0P~|P=N%q65ySyEP?EYHRBel(hlVkD!yolmPYh($+`Wd^3EM{j)cmi&acLF8a`>;B zqG$0C3#*#*kjtI|mWl6SjTS&WM`sWfVp-R{zSP;9Q-sv#-vc2Eam}RCYJyCPjqW(a zFj9k!8F|dROmCQGNKGen$un!gVRf0>W*ie{&iV)Bh)`@WfNY`$e?W4cByj0zf)5Iy zyA-cqEPceDAWTuwpr&VNrx~BKh$immH}z#~NP0i0 zepx)f%$>_y-L&lkpVYMSR%GqE@^W&0b$s;Mnhvls3--VfPznl|wV;Z?N6_gKn=T|5 z2#GO40yMqZZu7W|-cDv|pt3#LPa(32Ulx1@Cqpq}aS&>y##~7TSvPQv&FvtZ@7L#7)R#{z<~U41|^InUZ-(!?xz$ zq!@^L@WpeCu&EEpdV0-;D9($J7>K6Mgo+#NM}OW@d*rMcy&h6!z*(#A*&B&=88doG zXJiIKkcG8kD-IWh0%~JR7e;>uUI4FMfu5Xuyfz3tzd@28A-pla3C}PhrMQjI!32RET#*Ls;Nu67 zC-@?$09xBuk9P0>39`uhc3SrrkGUR=El!si8MagiCO?QHAEmB~JyfjQ5_OLLnVljY zt>uSQyTRACa};XYsx%+euxaJvBT$DwjK{0xs)vW=MC9ar zj}EboQMIj*f*jw90iPcu$x< z_iL;-PI&3VcH{SqTuUA7)0Wd1{lRLiV7BZMjR?gP6yQ;p9X&IBeswR=zDW-rSV3|{ zkt9XyVjR{6@rZ$?51~4sI=_g|?D53JYr^K{pw%rs(Yy_-X-nbLYn_ocQd6%o89iZAFr164xX7>s;}u)Z`L)bxC>FYV~4d7 z?D6F?$CLRP5hqJ8I=^EaTlLB7(t7HWw>T?}oZfdhJqwpS^wT(W)U8Kh@5I}U$BiR? zJ-vhR&h{-?zjpFZIOl&ta}@(nBtf_l{bjA}xsFaNqKdV@i0sZS+*vPZES&#Iz(#Q% z`{0Lbjk?}}^IQy?ST6wnPVJjsUHzJgm*5$u^u+qL#OpFjJAh0EB+3J5qA?(yD9F#} zNX>AghuF#*tYa_9^hliYN+O2m9ADg9! zxz?#0fp{H_8~Ch>w0z#1)~!&=UnHOFt|+hgdbeq5LIf$Rp~3DM^s(n{D1lN+?)d6OlCy8QB~hu^LyR9HzlrYWU0~ofBO6 zfDEkT+6*20L7p-=Bh7O+?!Te_`;6@2jP53FIWR1?-t@vHqAe~hMc_RxpdQ!&r|Kj0 z0m{{L4YIf%J27PPcRfEDatqy@j_+Em!&3fSQi9S9|O zx62+6n3+i!8W`Bn<$v-t`Yp_r721ZSF;rt6pvZx@Qm6Ix`|qjOo$bbDY9%`nb$4#h zm5t`>KWaBFv3{1~fV855f^WMB}l*Dt!M=@>b zYx@_!1rUIf*pJG!a4X|GU0?K$Exj7qc`PsFGUwbUd9gUwZt7>Tkn+sk;fzwIv+~2v z*E!G0$?bd93XFbFbx0(o$Y@@bB;fK-%H==*M5Ms-w&`%gA^1P<-QQo4!P_T^fmpr&`9pKu7&eW@$;8P&d8PmL zpMYTaG%_Ku-ufRu6o@7iBA}|+DG>i(ulWochAjezBpeCu{cpdC%@hux{iR$9>HkW; zmLw1&nEU@v=l|vId=N7IlqY_bX8iu!A5;cEzJ2?q`f1E`f?F$jyxE2o8<&G$v6*=& z>_YCi#Y$ad9g0yfKP7~HI$=jjz<&GnsU3X%ebM6UTPzW!M&o#@|9)5CCUIPZ`yX;H zF8_Z;g$MkFedhn`KXW+@1)hZUz!EltG5_mVBlkgfvUIZkZa2GP*9Ec=U~nMSv~=|z zJgJZ?yi(WpfvB4gXEQmomP|!EvQDT5Wup-U&-qn5n{O+;A0X6m0+>RG6Gp7Prf+^j z!{n|p5`ZM`G(PjF$vOgFrzGP5zqg+So|GO^yL3COKNw3!U-*mI^|XLM^qQtx2@h%l zmrU-r4#J9ml{LqJJnxDGu#~Bc^9l=6ZPLwml3IL4oNMD@lHUuIg+- zxR}up=)8>N{H)eaKXgP4jFEsRx=W@`T2=PDu`)lOfab9q=Gua9zrTvggy^PMkWEqJ z0Ub(0FqyPQ?YJee`>EOSw_8&9?g%_N0o_XBt5vJ9Ky_^y=B5)n0hwRkfKQmD^Q0B{Udm_Znf^1zu?bdKKY}}i@F~(=buqa|BUAi*|C)86do1li)HNGe3|~G3{fV%@&!ac$ z_>{ln_K{SLNSyosN%DGuvq}P=$?)eJ}MhM7g*r(0 zmR9sN4~v@Z4j?2_;c#Mbp{k2TWbxfbEAb*|SQ25@(hUK+3rOTDaD7bSPTHQ4YdZR+ zg&-T+w}GW)gUCo$q&d@!YduKMiHq9oLoVIN1jN$`*m#byWhQSv77SrO0b$#f>r#s} zx^YN$>z^Ig@uT3nn|)dZ=7p(|vz>BOr}! zv3hv`n*ec|_#L+g?6sRn!{QCmc>Nb+0aBUX{CR)uGwh?kfP$RSh(xIU(Q*MuMRVL{ z+B{hM1&|Fr0|Rat{#UG679$|t*_m;#DQE2m$BTPFEJ=NE@++m8<`^$8m_~XdPvpAC zj0!%#$6Me6^bm@gozd*39-huutqfh$G8i!Fr!;e4p>qyWa^gXzea| z3!aMFP3b2f<3oV7UJ`jb`JEq7NIiuSX6lxJacxbG=2YGD;R!xdiL91iKp=Y$$_K8~ zy)K158|mIqQf%CW3p1nmwux?U!~EWsGam1pr>+|>xxh2$MY?HtD+bLutsgL0=;<#m zsmB7#UeEl%T^yAU`br_%6BTqVI}e}HGQXtUr+m-5;@~! z^&sugj+-(|3CZ8wvWxWmaG>_ATF0X^QfwwaSozKlAcN0)P5}i>n_vOtP&AmukOhqG zB5zBM9!#&k9P>yUn_5GV(2TQ}%^(x(-sPq%Sb?~N_1i5Iyr0YSzLvjKZ+_R^F5va8 zCo-)pb*&oj@3Y-AxwE?*#JNa4+K`!X(SrD(C*MS46hn4`fD$rA0el1Z}}xnckh zfyANBy)%7=#?CgetJk6wj#qEoGwZFG0#ha1rh6w{Ck#>L+H*Er>4GAc?y`0?t*$+u zfI4~tm3Exr@WEx+CFR@mAoRSua*5>y<0dGolLZ7N=^7Oe%&c3!=E!j9=wbV+a=Q;> zM?<`z4V~1Hc4A z4t)NQ#wnelDv|ZY;;Tnkg#6`a)iMoUKOC;FM$;cR>t12@Oj_Y6Y{0Cthcj`EV%k88 zdei=g_iA*!_tsLHN<(UCO21Y^-(|<9hBJ}A!(BnpOTf|Z_zflGew`LIzt-%m7UFLf zDIw3ZN3Z=;iqfS$<5AKR4g&0xI!B70&(EAt2Z3&Yq?R16apWIJYAlsT>IbDW*ExE) zD%m=(1cb}C=AG7l^?OuZSuE&~?mX=1p-pjUC3%2<3QlS1n&tK$!yU#yGPd7kkEJn7 z8|B-4uC!3$pFj*8Rl-Pwi7W2lAO2poI-~s)oVBaN-op#&bP=e$_6aM-4@!Tn-Rhr@ z;EWa8e)a1;Q$b~{t>5eCj*~SdQpY)`M?6oj;k$W!p%I4~4x^1JBOma3HzoFa%Jb7o zokUuF$Mw5X@;u|G;c;@YuWsM&`o5ZQBo&Bu}&=Q{Sipro5Qk5Wt&zy3e% zeR(w1ZQC}YNSO*HnL@}=h75&mo+3latYkdG6awSU)jp678M$8pxD&6~&q*RAFGd+6m} zx7J59ki*v9-;D#KL7XySDj*$M#(7pv11+p&K+AQp`&N=!dPw?Io~zq?J#SHE8yh^y z%bv9m^Q{m0lV^OkJVb6dgQ6yhEMukg)qp)V*|WPStNXP%j0qg4e+MPZ*^oC=Bsqlo z#646JwatR7`q^eM{yTG=>342;{mbCF>0`H7q`dP&1>?K`yGpt;o+T)FRfQGT(Z@q5 zWBmP5AvLMzq3x;;i|K_hI5}tO zTIil2ZY}!#D>kwtq!hYIcgX}#yW+K%d+5D66~(>ze}W$KjN^rl`GTFKYu?@09eSbn zG~9x3Cl1m{SDxe=k1azuhUdq}k`arkAiGLX(OZeFkUF6tew@5xes#5?XNSE*yy}^I zzbzVfCxU9IJ(3HxdvJciJ~!Gm_sOj+GJ1b}vRa=2p}IRmoZ|JCN>PX5NdosVZ~DS& zzB>RVT4R)LZ1D$>iY2Ru;uTOL3}1@|<}deLIh9QH3!mqQPRzuL)w!Vo{^8v8oeHZg zY~k&5HWwRYNXYB-;6W%OP*3Oay){*+F+FxFE@`z}D!R*OK){{<n(ezi=S#J>q{s*=g@U)5#%ugqqZ$Wc zGOJ2S!pqi#_A9khA2RyGq>hZdp?3)_?0peFfZunkC`pF8!eK~CgnZOBw3=)DOJeVr zDGW+yQpBG2sC)qt|6o32lvw0vxq_0`_I9Ox1^!3kHO{B@G`OMmwn%N%^lC+lNWR)j zD*@LpWY)ib8LqkPr1DRiS6zL0LR0g`*xbbA$KUom1bf2ly%WtEtaB?+>CrlFK$+(y z<=!g4gSP|$fs=O-V(GsL&*HRzlzG;pA>@J8?XYV!WU7|;Yz;g+IdBS)3 zLf*m{ljDs3@)p?e{eFqr_N#%gTDdAg2aXV#OUUKE-ZIT zQom@(y0wIvFJ{yN0~j}LysGb@wUhr)^G=@?V}mHTCun4_#-iFYfZ_Q!Dx>p#>nDbz zhPLH;Q#>e*o^5g|`88aN>g??L5?w2^IdcLMR8ziLIHq^W?q~|tD#MZK*iu07MD4;* z8IZHm1MrP~zG*cFTt0b>Hl*@=iUcjYph)@ABa-u#%Tvq~yZ3=P4babfiQDm-Bz4CkX?CHXv+L?n zMByJfb4qjU`p8ggxer_rtU5E)T)zA7QwOUDONgLR-*&dwmymZ{%sdbvPTQt-4mMb!u`euvUsT?GYCjdOVsW_y76=x z^*tnkK8+k6#(x2F{?l&{RfIsbHG!2EI zF%QzP80eiOiL!dHM3&cNNy2dDvHX%&I)P~r8vTT9_H3^=gp4KdJ^lcCSM<}$5RG*f51($C+_#XX!h=oM3SgqS6e zS2^77kkIq{djW~lxcw6R#rTchsFn@7e17v|&C#{~Ar%{V;5IV%5Ql)I90P9Pk|OEy z)KeEi0X1v}3;Qvcj1k$~tiJJ6$d^~@`p1en@tTcLe3~H=OdqX#3qnHNflz1(nW9=Q zdL7M#LF|Y=>^3BO^9lY(sWSe3E?5H5zaZ2Z@gslrNNv zHu8KoGb2r)7*z@uC6^$_r=K{P;$hKWb4W{qT}*Yv>FLvsYX>cFFxi>X9{*K*HhJX5 z9<~ZJiNShrkj%D4%^cJ$XD3!Gp!Vy@H+KPId-h(;N=ED1rook0vA}@G9q_|TCJUs6 zkKm=_`3)`h>aUz(8gd;EE}g6mqWW$y0Hxg3r6V8%cWZ39l~Y)#^-j<&>Nq;JT42Ym z49GoV1EiPM@t)v8;k|F;#UbbLxXnRNVn_y+p&g4=lb3LVoaw&0>dC7C4iz+crIY>a zJ!P(wx+Qj*C1DT^I}SX?yP-kIlF!QhrXf@1I(RH==jHSdF|f@jdmRb-szwmbobugX zPDKDBw>KC<=uQ+Qw-^c$m`1=qI8jVJU)1Xf$`Q!(h<=~_?H=(;4KAt9m=Hqt%Y?~N z(Mjke@wSn3)QAKL*Wq^my_;(|@ox!^vqe3)vZ9cPek}0+nxM(VMqtlNOZ)JQ0Bo9B zHA}GMceA$oy4+J&krgOdksDr9302m3WE?#3vI4A7FhfDZkEux>>!Gfmt2160SF1j; z7&q2)(R%18shagWy=h@e=ut=(kXA4RIuLs3?Y@*M8HJ_6sTOB!C>)Yml-%7rJjL;n;57(TB@^TX zfTofwvicqcZ^3fKoy+7NVdQ(Lv~ux*@Y@; zMaxzS|Kr`;v62q%WpEi9rS|=G-}C!63>>B!Sm^_M5-J%6i$jo zO5%lx^vXe@ie#^;obe9EC?87i0sBeVea0X<0~)tW z0P!S}nY%qlV2q)V8($uGS({~(*O`5QfBO58xvUyGD82aRerWVw7SSSiSgLK;<2H(? zw0eWfcyv%Dp2~THx8CBg>?d@!x#oGUII z#h9@5zT#{5@g(_#@gtA~5$ccP8VJLSqkdZWB)k1XuC#`-K52>T2Jl0Nn5^E>m0)@}VWFIZaJa+N-2x%c_1J&xE6_467KQ%Efm%m9kcQ;r;CS;t! zS~Fr$=9$uSOA;xA4j`RW?!~-0?Uh_@%6*$iv2Dod<3Sf#K{S{U2Ys~4N&Ya%gf1Y% zT{UKLZ-pm?loqTDH-G`#zaGYTn*<&_RX+$jButWu`kce^T}LQ*_AkcyDQ7+Ag@ag4 zr}#d`i9|NEmEfaroD^AbGrKghTYpC)98tS))Np_OBkUZ(M^;!7JvAJ#XJX>0RHTEk zPHz-+)#WO=GBb0sxoTbQH9lmZK5U-N+8^3{2XxYoK@cxA5un*m!uY=E|APtx;a(BL3crwPh_| zli18RXWsJuc0a(>ah?A0Bz6|N&>48{_cJ<2cW}knq<7R(HydrAGzx3D_`GXzoO2vy zuTvS=NeR1vv9%HYf5YbM6#e8P5=1$>|BmMq`40l1|KGg^`>`&AJ=!GHMLPV7;B9Xz zZSkH2-Yb+&=A<+{mm%#kkSlD1wCl4<$eU05B?>h4ZdG5aN@>fyW`7jg=>SfZyGjJ^ zrdR*`+bfibp4C)S`g`94!-g6k|BqjD<497XHZ7?IpSGbPufD#%+f6Je^fJR>5SFR7USwJ3z;VeV*!A%*+IcJl}nOco6a97F6O|NNg>z_)%jfb zQjF>X@&^ijKIfQcBVu%E9Mx zV)R-136Yo)EQM_E@g8_uXI87h@2Pg^E)O3*#4p}pDXKpHLj(ngG$DOsV~%I1iNSzi zI0I;5Vc0LX6CdrK`fSX*oz1|Em;etAb&bIElsaRHet+G=C!uZN>|+v&VIwJ5?z){! znc*}@isHocFic^!WN*3DQD-7zYLapbzFtXc!)Q`}sog*Ce&1ghI!L1`iW60wQSQClQJ< zAO&`ltSv|vb0sQ@8g~1~q#t;p1JmjrY?l*+eOq=*Vp^&4VTmbEtT<|)NJj4We@%({ z6Jh6<8ZSw@nuq^+v3hW^&t~a*b!i=)?faa1LOdvK-f8nWl5COv#06g>e;{95JzC~1 zd-!2G&2FE~jdaGl)VoGYo{H`JeLv~?nN1l>O$PEV+TwunY(9cn4vfu+*8i8LraQ^*s_Hf1v9Zxu$)V@}WGO zk<9R)kK0P_@%q}_k^WFFPg z7Whyn{oA9i=j8s7naWd1;n_BT`~Nd7L@)|;5zUZKWL8up`pO)6nVFbdz%-2sN-RN- z$;18aj`3^jhkmLBK%)^*6nrQF21+36{t{U2H-m6D zPr6;x=ps{CuR|rdr4nk4HC~SS5ooV@ zV`F2@1W%xb=~e5<4#$10)Jh23ZE!YB%S zNgT4rg7_Lp91+EP&L;o9mel6>R9+E-J>?&8o}bTv=F2w|(LIib95X{tVh4l8#o1do zc_ZFaM#)BbFWl>0xYs3d zXT(_ei?VTUvnr~q#)?sOSZA+XRPT6j(IN~-LPVtCJ3QjGpxm$@Zq%D|aq@+nxdtM; zL_g2}$k}jOJ^BY-O}Cpya1Cb7ZdqXMeL@MS-7 zmn+lvHiK%UjoI_jhpTifq2ik8U^c4Tz9B+org~%QMAH8Nlu=um!J5?Am@?~2LJZ0U# zdEY@K|Me#~b!9*WiaZ-L3z0oo6Y^Ndx&BA=JbV&UWBQEs%EU=b<?Lj$3Z@gYDy4Tr7a2P^@6V2pr~4V~1uO|mnAS@@l z^Ws57FA!T@QDL;g^2f2ZP2TRf`9oE3m^4V-Q7dB!QkCU~NR!bM5cBT!MZwS}N&%@03lxXp?u#s@sbHIfZ zi#vE+T*;_peqFVN7Bw&$710nJfOxV8?=9nO>2q^)g&W#nNb<6N4G8fjaNDql^lU*> zrzE0a2}N7-Hf)^ur*K5e}guz2{wrRyD?qnf?v6poGjj2|YL6tso{hfZpV?y%fjkSk$mgs9#!R;` zR9AjOEG<SVg4L16d!Nqr3O6T<~P%?y>A14mq`W4Vymu>4m))7O7l(MU0%^ipgp0S(0Af8 zhiQo7{ELv1esQ=6+ue(ZVuNz#!DXt8$Ahs8Zf#O9=5L79hoB$fYpGAUO) z8w@CMo#Hd4%m6lSXYsY!@Z>DXm39TIsjeI(DMU7_IAe2HhM<8mFl)@_>QFIq-GsQ; z($5AMc_t~&ZYa$;OPD_1^jW@o6)gh_1?$>IOuK>vFa`EY>D^z2*(_98toEbN*1s>;$$h&Sr zjGD>DZG)hrCGRdwg^Q4yh`Go|3@*@q7|z`NUx@Kl-YBvG4J0kE?$xv4*}X1B!NtU_vfs z;p^Rm@i2Udrf-j585e@X#Ic_KAtjCD!vSY2S4gRYwTyK-6ZYe;3s?9w00$Rmd=+JI zZxW}DT|-b|IS`6}0?d5z+{6kF;d6)spg<*AS9wOcm${?CVFf3O!uuC1=3Q8R0Y0l) z;$Y5d3L3|1ov6rt#G|-9bZQ(zaHfnUMeKi|Q)`s^(XKC_?2;Q1+OM|plRtzBM?QmQ z?`Np^@aduwO^vIY^%e68XRf#8-C-0o;eiBU+&KySbl&Gnb^F^;`q6>|c#3MS#klYF*;9sOos9I2iJXi3>{1k83Np#{nzlF4 zer*+X-^Ke`2 z7V(4$ORtYiqk)^CwM!f6#`3iOQ=fsn7}?3nIzMySY_qwa7YiyUznp~#Z*tI zJwi`|F2wHWeqfu3FJ5L@&lryN(75^eWYO?>s_4qI+4byn3BWz^yx$T)MmYvgZEUKt zy}~8=#{noBUj=4xnO(8@O-OIMK8!gsvJKBL%`_N7(dq19I>q2cY&xF~HBcw$<2!Qf z{|E>ps0DFnv#!9k&UnyBba0h@_`{G)Osgzxw5<}(lcY-w?N|DyWdEpMZb0}@NlzZt zRUZ`)6gV2{MdR3U1}dS0vS8bZ4hZCC0w&$>(!npj$bfTC?;1h%@p*eWHF^GjCNiaH zYyd#s2(=-7Mem|D%uxB+0xbj7!zqVtb$imO)e#;>N1dh5T(A#1N0>?G;1nBmP~1;C z)~-Mu8!kg2r`WB_g3z61ys*%{Fe<=(@Mz>s8b?}Np2tUx0DOH=X3JxcKnkN2cu^z+ z#2DFC+pIxN$yf7x*3k-anm}-^W@QDFoqxt>eci|GHn+N9~9dn3)#_7||QimZUfRA`!?y%gHk!i;! zuq9$a7%zvyD{^m=#G(XU#QKS$pZ+rN=MAQ$WIva~#Jfv|UlESlEC3sGd-I{sKn)KU z;NLrGP0}YJpS>MUr#nV>m0MNtcq{utbS==bEuppX8m8?bIIXW&E)ZlbMq$!x2D_qU zX^Z9AMc~@yoCk>gz*7XzSd7PkhL?2J#!mLck!$%|**6 z6aliH(~*B{f)E|_@k^2-GMEU_3v)NDr?D?Y#9Xrisa*KYDh-eu zr^u-v{`*)?Poi)SofYq0F3)|o@vAsrncpjVND^WgIEK);z$8&yi@nOO-{rADVCZjI zdkYY=)!y1rLm)c3d3U8n*R-bW&TKfwP_M=S4BSnIbFg6!jNU7Jf{%LX-U zVHyiaU{<~fu{-1%97fcgfCHAO01+Ne^AXxBmo^*_Q*(=?0-qyBl=iS)MT5l)&)%?v zZ=p(l+EISn2HHo;e5NqeU6Ju*w;Z+Imj)YG_RE`D;R6PH*t|^kiFvKyud(>P>YT(h z!QER!-y%UPeygv|$9m4L22HNBvo6Hv$FvlDb+I0ASVzlh0L%L6e9F^}O)mb~Ls%CY zC@dq+Xu`tE)f;!sLp{=E9I+IyiD7;C%5imi!puU2Rjp$df_;g3T>lO3P zKG!GnX6pKB$U#a$yQ1b=MPv8nt`8U`341JDoz|YpqYASecr0L3c3yI~btC8JW86}q zckM<9XN*5uizm1LwidFNNV>bb!+yT3(aGf2)&%e}mvmrlET)6W%ANh1 z-rdAuz@^Ek#?DX6RZRdJ;ndF7gsh%z{_$+NB#Y0sKijUrA;7jDiZ7iTnZIHpJwh%T zE5G2bGd^q-EF7o6w!Tv#s1p44B*__iJeY^%JS{ww7LsYSvqGDAB23}~zx7Rh(_WgU zLDvOKM11)O5mUtO0p5sHD@jJM1UfU(ShpHr@C@f4a{-jqx+)(Q7#DLNEiftB<5}uS z?1`^i|EpP#K5z&YL{qs9F{`kFM2nV^g{>YN zws#vn>z;R!Vt8~2T}>m|)*M`o!ti%1K`86T_o{xK4aPnYljz3rLOj=E(c&mptMzhK ztTEa=$Rw)zqvOZg^@@l|=&_8!XB}t0L+FK`G8}c$udX*F_Kr<)!00~|qc z>pt-!1w{7{|2sylc1mS6uhelqz0|;K(5Gd7JHqGPg;4WH14d_*EA5TeTjq2gjR*Y} zn62)qt`gECl6H_)-eK4*ejeaS-j#eiwc!(mn$faE^0E}Q*}?oP8D5$fD>^BVMeNp% zFBa-Ai>ge)F$@18tB=L|)kJ|UOdR4w-Bf&j6gc3t#MD5GWX-a0lc444{Y-hEa;3vY z%on?qHCMT96Vkpp*;3x$c~!4IR?YXf?K_$`lErs%GGJ@NE%%ztZ}#71(H0{ORexp> z`W~8xiud41zPp0*cl9zi# zMC6#SN<6{vlARs6RU%bXyXqxtiYbjKfWj?pXQ}TouZ_$)y8t>(08S?;5N3<94!@dJ zDjybUj+$_c3bYnz485i(K7I2neb{cY_|L^WW(Lq&x_8`}^3I3aJ2uchiJC1fcg{Pt zibBdL3>MwcTmTswFaZY=nR(m8ZSot({;x3|b2c3Z>1MMC3`l@7p+x>e zhXXXK5eby`E@B6zSY zRzj-J!Q@65CTa&1{^ZypMK0!bxYRG#Zv894WyHJSmp-6}c`-GU zQT=d-+x3yXgF%2qdWF)h1+MaJvAho75G&yRb<}vh-><^d>ysjLp823vY=F-|)WVKk z0>S?7Y~{*GCP9*Gct&7k-pAkvE+1)PLNNJXowl=W%NEw0|he2R=|J(J)9~QUX!i^YLFBZN7q( z4E*GHhrzwXq%$Ohl5E-gx5y9kVOY&xGWW7;SN?G&qXkXx4gDls%SJ9Zk(m8B{gzH? zSqtQRUFhzjfGW@Z{GBrp&xP= zK!{(j1i!`zhpS!>sk0-0ug#~S5*jFaLk*c-K0Kg6#g|Zg1YFZS8Xz-KTf&eJ>y80P zuQwyo?=qYZoOC$+`tYtqtI_DmD5o-8Dq66*2TI@Rxv2J>oOonNx`Z0i*2Hg)TW|** zOpkd-Xy}D}osZzlYj7+hrgJP5Q{cCq4{FZs88R^<-Me?zMmaI8EZIF7f;zN~Sn7!h zQ9glV1uk@f9XE^K#-05p@g(oQ`Kkp?osb&!#sO9tlex-eE!1+xtFpk}h|&69wLTmT zU&b(J+@d2@8chG`pHmB6j+mxh5Y)cG+S-R`I@ z%8j_KH(#`(ApuL+IX48haA;G{fMVcv_hsWn!r+awyqQ$Ne$~PkRGQr0^O*YDKu1Lw zrYY^k$al&(^X~QbFfoeqdZ1a`hE}4u)WNk`fmu1bVAMoc_Q&Z~txsK{VseA4H%E*6 zrS(&ISIWDAWOfk|*sUMt8@!SS)HK0HkQ1;9jpw_2M5PP62Y=lLFe8D16Q|Dy+Lo$Y zR*w(0eF;>_vgf^_NUGiuxO|%_vYuM!Oz=kL@Mvq-~xp=#x#Z;xC2Z{!V>OQ>$_02bs z1eRxvibBLHR=Wy(nAe~I@C{LMCB;vqLyPI~pRvbOXz zwJg2tBs%SPa|IfD;@`*gcP@9=7}_3CP-&1H>~1pAA!tiR7@KnDr^ zfI1_;wzN-TLeSaW6e}vXyoGx+c^$M*z2)Q7sSILrjZYG{j+c}8H}3B`aL=KuH+RZd z$yEoJ_;~Qja<~|VV<63Kzm}|d;DbGXT`&W&Ovs~5E}ngV-NL`-Jbz1p{`1%WEYtt$ z12R8xiiwD5&rLN&xl8|9+W*z1{m*j#XF2~?k47^5^0mCv%!1LS@*#p>YD!v)h4NPe F{tJI7qU!(v diff --git a/docs/doc/design/figures/sqlflow-arch2.dot b/docs/doc/design/figures/sqlflow-arch2.dot deleted file mode 100644 index db16cea..0000000 --- a/docs/doc/design/figures/sqlflow-arch2.dot +++ /dev/null @@ -1,50 +0,0 @@ - -digraph G { - graph [compound=true]; - subgraph cluster_1 { - label = "SQLFlow Server"; - node [style=filled,color=grey,shape=box]; - parser [label="extendedSQL parser"] - verifier [label="syntax and semantics verifier"] - parser -> verifier; - verifier -> {codegen submitter} - - subgraph cluster_10 { - label = "run locally"; - node [style=filled, shape=box]; - graph[style=dotted]; - codegen -> executor [label="TF/Pytorch code"]; - } - - subgraph cluster_11 { - label = "submit to cluster"; - node [style=filled, shape=box]; - graph[style=dotted]; - submitter; - } - } - - subgraph cluster_2 { - label = "database" - style=filled; - graph[style=dotted]; - color=blue; - database [label="MySQL, Hive, SparkSQL, ODPS...",shape=box]; - } - - subgraph cluster_3 { - label = "computing cluster" - style=filled; - graph[style=dotted]; - color=blue; - compute_cluster [label="Computation Service",shape=box]; - } - - parser -> {python_client} [color=blue, dir="both", label="SQL | table or training info"] - parser -> database [dir="both", label="SQL | table"]; - verifier -> database [dir="both", label="SQL | table schema"]; - submitter -> compute_cluster [dir="both", label="job description | training info"] - - database -> compute_cluster [color=red, style=bold, dir="both", label="load data | save model"] - database -> executor [color=red, style=bold, dir="both", label="load data | save model"] -} diff --git a/docs/doc/design/figures/sqlflow-arch2.png b/docs/doc/design/figures/sqlflow-arch2.png deleted file mode 100644 index 92d017cd212691212a137c1ce6913bcb4131c1e1..0000000000000000000000000000000000000000 GIT binary patch literal 0 HcmV?d00001 literal 96987 zcmd>mRa{ls`!^y;36dfr4bouHNC`*a(A^;2UD6;30!m1CdFbv&L8QC8yQTYG#~FV! zbNOEXFZ|57!FBdtdp+@u=LE`1i(;ZbMn^zEz!Vn~kw-wds|fxCpxy&l$h#1y!GG@9 z%Zt87DC~W*fq)>0ATA=L`0mbj5=x>%>xI8SE?p2?&ijvV?|fy@${&~8)>gzH$kVfa zSF$)_e!4kVI9_u49{v3|a>-r2I~(cm$r;Gs+uR)X&XyUuaU>YXd@=THviM@Z+8yWe zl*#4vay5xT2>X>GKB^!BBDOaI5;?-1-~UU;>BAr`i@eepgz@LnpPzoqZT|1?f*;^g zd8a}@5vvIO_b249TaEvF#al2P>7y(SG-T^P%lahJdZ%`thk`x$`v!i0nvVz?c{VOh zo8!N4&0DY+Pf&K{N)!R%zn9aIuoVNpEb;v3S;*JG5_UIK2nheZEQg4iE7D<0^1mmB ze_+@lSc3Z&2jQQM_!?hJ~nvsDj%tun<6IB_nX{kRD_I*F8o0e$x#g zF_PgsR^%>BW%zTSIZtKNP@{O-y!LaFDJKKK%5O6T-pc*iwvVKDMr9$^((5>X{=x4( z3ZnoIM)X3?0q;ML0Dc=u>1frD^v~`5wfRJ^5rm>B@F1KN|9KhfID-E_OZ%WtWr|vh z9c8~l8qHVJ&kzl*G#$!}w<53Mr2qAxesiDNyY(rhc zVJ|O-QTL;*xu&L_y1m8^d9a=9wG=Na*2c5+0@ae{TE{ITaJ}v`F=uZG|HXRwu@CXT zo?4OR4y9sLSCjl+quB@pt86*@?zozqbM9p`4ntDf?WXz)4Lao}az0Lj=d0jU zoGcaUG&Pdge7nP8*oog~G+wN`v3>9K{Xa)E+X7JtIcRE_v)}hi98zKJW^St4NbdLd z$QbDN=YQbI-0$T(oknaUeK+s(%|qGOZMP=tn{7iJr?vLHA7nlqg52wIb-Mx77O!%A@XIa?n3AMf&Tgg*hOKWHV?CC@)?|m1d$w$e-$Uv5#pg zkZV~lG=Fd!RWm+6YQ&S~qNvREDp0Jq@8xdO_Bd9mJDqhq?Bg9+&>1V#;=NThWo^D_ z9m&zCD%7&;pt(^z-gJq}RmxSQFEr~6?MddFN;G)^6ZjduQ#sXpbKd8rY_`l)DNyhK z&o+tifzW^@+ZyZ@uO++153Ysqov6CqoUd6P_HdqhC_PR=T;w#N}tsLco_zwk6Ei1*tQe!#b^!1#dwZ2k(r^H=3wdg#NT!D6yk z!vCZy>Nnn~N_x|!i@9D`OJterUnF|U3bpE7?SpLIPFXj3oDF()j3p=+sIQPQUZsP; z)fTHhYNEs`YIf|py!nnoc0`#|zNPs%m5>qlWmd9lZYZf~L(C{ojz#r?zxK16EuGua zC`J|a7Jtl!I|1EX4w6Psbp7&;`;xV*#eU$cP}}wL?!CI)Dr&l_WsM0kne)1BC}O}e z<7IlW|A^b}tEp5RrRrT2tfT88mq4n(@t1wChN!#CJ*7P21iTve`XGc_2G;cl{-O7C zJ(GAG^No7qSLN9bTNNpH>V`$wMScGD;_vwcQR93W(!p^Zs6NR`@kG~LZYvVEOUe4i zy+)v>up$y8a0?brf(#N8rQ)vWb7CizkJhX=6 zRv6z!KJ@JGCu#Qjd zGs$jZgCoyR`}j^1z4p@0&mFavpN*Sjd`65>kM-(?+`g}|U%NOZuoP&L(>6f5<2h&a zgKa?kccE(+0%7+h4f?YB^PA6)&lO+7k}NAHj5s{6-jPWST;vw}nWsi_sHmg$Sr>o5 z2UB7^oCfD}o$6&!FI~J{D}hR+>*v9Z$i}R8d5a12q6YWLX^+#HwFfGB$~{?Bxi@_R zx7(rgn;DBpQC_1DN@SsL)gZecYKtz@x5gB?BhZ`tJFiI(Vsc5I6Y;xOheGM?PL@Ii zCY3cTR}|2da@5L<8h3qANY_)6o*ngFq+QrayDx`}UOk7gZKgR0M+H^x5ig+pQ^N#5 zpceb*zCJ5mTciqoHFz?v|3KmKydN4~J!b8XsG!DWbFD9<2vLffk?qk;YKt;e;KY74 zLn9fBrz9j{JeYj;L24^iuS=SDXL&AnPj}#g1&wW|0d_3S@@oZ?)1ZalepNDJ{Y&xv zq1)yysB1Jy{b(BB`Eo?2qX7$aW$deIL1ggPIC%u?rt6b0ET`sQ9pp{@K6Hif?wYv5 zH%42N@8Aa>vY$3&iq&0U_FsV(nIbq~ZM$Oa%lIPfTOzmpRc3jK*4QK&#EhDue86g5 zZ*6eToxMPTA!og-f$}9tA7wgQ3i1ujZG^1a7P$H_0<m3q0{E~&9@fuNnFJ-oM6?;#NVPv`{?$i42aUq>XLcK~9}$#nA6h?Y9xUr+ z;b!h8x&P{zeI!IFV&&Bu1)_GAtZ}=vD~MNX@j0wR{E_0{jQ^pp@!HE zP~o?Ahu%fzW|<`3r&zwA%@U3i5qspcJ6$7q-E3`8kCTb>GueKt?@FU0YcNNF*4o>4 zOw0NFbm0No2=8t+mAl4fnO{7I=BQ?e?Xzg57n7{VtBHR9gf(3R!uwoFkqI>+dQ0_` z42N1(Ihe^u`So&svEFwKrQ@DQ1EK}COFx|wCCnovv@OD`!$zu+#tQ3p z55c=0KZ%UbDuprjo>8t!5~8cH?(^e2_HKVCizyBx|C0JeSX*`0ZI5ryUxjOtdSR)X zvfHxXy!Tm)3Sj`5=h^q=aTR4b=+Xy5?O5v?(oq#?^2_jcnXZUOUj(D5QjxOcA4HL6 z-+U~Bw%dWOK=&nbGEHw{_y=b_CJWJ&PbV1DpE&=u_bX4oJPp5iz+jOpz1*#H zn{q(Azs9k4DuPZWya=u;GS6@)YG#cfasaV$$LF>l22cD4tTOpM1B=&ml_By3W zb1TK*Q5w7S3)3+2i<2p9k0UWkNmFNNg+hE<&~5==3_c%&W~OAEUW|@sjX?g#!BGwC zol{eJe6gK89ndp1Ru%(QEULPKFmzVdN1#HKVfh z=~>4{lM0(FF*+wF>t@0@{I=mrwV&vZuJHr{Wp9j1E80}j(f8W?K?7{QR-uy zre(?x%Rgy3dH1i%Oef&OV8_99Yv`i##Pe+7=EKdVbjK^ip!Rk)A&47s$9bXURHb%o zBuD&n{hzTM(IZc5Ca;t!DLDuV%FPu#r1hjs-6nPNiV*@Es(xf3iEqDY#dXtFWT~`T z7z-*i-A=Ho%^~~3EItBR}qJ$@5BtVv}Fi;&s7b?VYnjx z1%N)%c&7%zzHe2d(IrXK8watFcpMFQO3BvuzO$Y43&Qpm4a>+0QQtztG0GAbpZlBp;%bpaF76Jz`uB43@S@(? zKltq57s*~M0^w1@*Gm&N&Bw^*6KAiyH| zu8y80#yx&7u>UHe^oUmvk`;!MW_+(#6Ra&j&o;p)_9i9@sn%gb&vA3OqqF2W-Cw{` zuo91ag>}l9E8r&~k`UJC=(N}WbXy`m2=5ut%LK{(;50a_MDY1U4Nz~pX~xjMk^nea z_^9OB>CZ5Z?)>L{wTl6~`6-cu2lKBJ`g@5xsQ;&z9(h?h=I|Ji8~S#(-i5<-klK7G z^Ycz|3l^FJT~1QmOXEp^W~{xg4gxFzVgzW5GvS3{5InkCXSYC`TT#?>GdZm7dk-_; zaclG@{xQ4R@|aGoLyiY%&CmCa)1grel^_oja{VO!AAb-Q2q+B0#r%P_MC$$P)43@L z)&>LT8OIoA?RqZVdo&d+qa;rFGeq|02ma$Xw}O>1HOFX-jG?*og= zovJ>8K8$~Eqyq$Y+$*;uz86T;Z{)S&IjyZwp4fOdNdfSI+o)x`ldo3x)2i{@h>V?Q zIZQZTzb(XRGbhUxN3p_8Ndx@3xxTqNlwqxJyxiqw-F2EUBs3rV8VEnwEDcBUTO)Zf zNiM5#Cm_XNm^(E5h$dxl+b#)UHydtmTK%UxYj*%AQ!+*9LT3XhvX=uOm+~Xw96JVF zdII_nQ|F+yc8{M6B49*kt80CzF%E#Wq^X;3&0gcB#m&{Z)3j~dxry?-SMJJ+c}_c% zP9sVx6X5wXPl;b)XPn9ln?uK+z&n! z08%z(>q!WdR@h(MpqHAcoU-D@c>hg@(|nZ4E2+1wP>wO!T6|}9E|&05 zQZZom3A&EsH7|KqtVvE@;s6fX6gc;rL&Nx{k6*nj?n&Yuqb8^RL`_=QaH5_v__Tx# zFe$gl+ImpKrO%v|1(lPX)K1jUX4XBkEr9&kMuuVzgIOE7yqE{~nr?S9IM4r zn_2Q0G}20u(eH-kMAbl6w(L(8@?-+wIbPi0`fMrG={quE?5qUa?7OAUo}Q#0M{-^} zfUPlC*LYZnJakzc`LinDsfFWzO$fOgUY95vwi7jq|{^?YO zVKALt2v-8Y`|8jzhmFBjjD%uSw5NfX9)V-r4P%6tTZMJKRavKX+oi$uzm6A!SuY$R zFFg8V+z252Z$3TJ1)#Gw382JEi-{Ks<-#j14+M@*NjNNG>7}H%B2VVLyu4Vh9E-Z0 z%GpTaqOL^^^}5>Rzgr$}!4r51yIhOKi58sC$x1E-i1Dd_-B+`q7p-qTs#lm9s^fkA z_+gZ2BiP-Rh68yIAjW+;oSK6Vkk@cT$B&+l-}FpMN@o$UdzZ#^6sfp8XMDffsrXxE z5NVk%N{eu3S(Dvbhor}EwB2@A6lSa%j+NcE3aYmKiLLh2SElXzj!*B835ZWeD&rH# z2Uf_i7I>_W4ei;UtjP07HuO=_>E&N>?U*IyNsgqm9&8D0j`CKp>boVBKcz$9j5Fe z1P3?J{|i*@G^mQ?GpA-SzFSfMhP3$)#mMF?b}d{pT;0|vm&4VG<&KE!Mn!d%ht}tS zLTzAtTYqf0Od$uu=Z``mefRB@5#LtVAUFSH*#F6GxskLS1hRtP;12TB)nsi@uc6xM z+W#8nI4Qt#PaKNwF$xOixI?9eVobPedJ?&ZvVsdVYhDP{QqXc2qVgzM*8D1#Mdc43;xJsUZ;FQ&Y(^>0}zafMh<~Rvd|FR)@07HWzN{G+&t9K>VH}3~#lyw)bdbw%OIw*2_mi2*#%Ek{ zb!QqMSbG%HF#G6udVAEUqBu456I8NuRt7 zzT2xq{?a4B)Ssb-i>$VM zag1LQrSpqoS!jx1{6*G*graG0tHt~`pV|6!w#JKYQ+oxh=>8Cf{B z6gS|)`l}hw){pFfr_$ zxdqpPakmdeN-D}*(6OO2>)MT{WbEjDv|DP;Qm@5vUe=H^7``06epVlfRB?e{ixU*} zQ|U?3{$suX&DNT)H4MSl7H%sZ%s4Ez+UnrWO&z#U{=V~9h+nUiNZody;2=Ux6@Jjl zJ%{OveZ+sLz*G29QCz3rkv}=zB>5j?vW8a{6bhw!(P&@@)E!bHH zvz*>)dN9i`9L71>AkU8wTRmu1L^eiiQz{e^006ud?B-ylbt~-B1kw3q^UXmqiSVTif8VkJk6nBgywnDMw(QLm6c^m4ZFQd zMuGaVY>cwn2sg{v>@dVo*RQW$mfpAKZ!; zrhl!Z((z6&X{TaLT_?wEl9r%DBghLT#cHeHooAGDD@v?@$=Ug+)hUgS6}O>(P#7VWj1 zG|)$$y*YT-Z9<%uWNCXtrc8$a`R%)P3j3{+@|f5f%;+Ac>4AK=)A}}!sW8ejP9g2T zg}yJMAmhJ^E=K}KklYG!(&63y{CCeGfb1J5V7slWyA3;&wn6RwwQcOX3(c|*&7mc; z4Jl^A0Vn(E8{_&j8Y<=YBpz!pW6;vgHR2aPg(J5Z1?L7x@Nw=Xz*qug|3#$l^B`y5hKXe(6A@6pXM z!)DDl8fP5Oci=bC{Ncvw>Te=KFclw24p5(#oNh`%KZdf^;Mp6gREZcC78;pwx-u1Q zm(~!TO*)MlJHs@#Z^bveyAZ9IM$SICOd3Gynq6TsK=~?e|Im3ZPAb`g!&H9Oew}>0 z)G&E`bBc3M&1t9fUDlt&A13O3`|z9^{@0J#KA>Fn8Fk0%Hn^QSHN&cP2CS>Iha@UL z+sB=qHr-mnr8S9(Ekn*3hX`iXV)BJ>i;5_23S*~)!{)Oy{jz64j*EdtUD0t`@6udU zzPlcGbK}HoxxglP_m&I#pl6wysDI_Gc0|P1YwINB7Clt>7QcuQEHmo)jWywbJ~$WN zr#^T?ndfzTbJ9)~Iu4rfWi3Sg))63eSvb#mR55GSvfajuXj_zrPZZYGj%Zk&kE+I) z(#gTHqXzF1bFkH|k-l_hOnEw9lrljR6Z@IWnx!z?LC0Zd2^Hw39f(3P*@lOm))rg zN!_-e6S6s0_*w`hB$N#OC3oR6*5j?Yz(|Zt2lVi4eaTI zjhb$gbCdg7K&FT10T*zDI`qXQu-nlQ)w42KC+$Kq3~28Jmcn&4NZ`|Pz!_P9yyyhP z7_~g*f~!VlJc8FyVbi%;$pq9;qKGxhY{{P)p&a%;Y`n|ZW%n?HlftYJP8Ko)P$%vQDAB`L>T)fzo_S3;t3dXUv^m8fjaQ$nJu8tt#gNtFcW z;{}e~9mE=l;CnM*M5q9A=gR0ed-w9vw`6Q4VGXx8mnY|IDHzT4v&tdq-{Hp(KJ2Ju zRT{e;6;WLP;=n+a$zYhYJ)9Hq`7wK^XaRLDevLfOyqqY>3dgCG*REC+gHbPS>rR*9 z$!?udFku4UzW0Pt!V09kd|HMmcP%-=c&gD#tFyoVZ^k|q6;xIn>L=cUR(P_0=t5Cb zDLytwm|C&xfK#XhYVRcA8?A@3WSn!HlPm#sWb3qU{b6MAMzg7r+t&ExY{_L2YF?CB z2$Elx~BYEgbU>`^O z1XRpaYPf&D5(9eh2+`6^E$;;>k;Sn3l%bkbTrU_;(IQ2k#(1 z_RYOJKFio)A5M5qcOoL2miN0?mGP5Y4_XP6TF;FuSE4m00aLYoTH^`~9#k(aU^g`$ z!41rH=|Cql%5?aC*!wa(Rl2^3F zLa;nG_C!ZoBu`~mf*UtvP=?o5v15$K8r+pmD zGD@(vavFfFCf@J4IAI5-5=xbg1f1R9orfEs_e+ggh|rK2Bn7v!3?dp9{E3yTnIE?^ zBZBa6z@wMpbvu+vyqqV&AsQM-*O_l_A8)2itQ?caTByT}B z)M6ARwp9vZ*R;<*hS~?fEaOsl4C46w^ZOWPjfWEp^mdNxX-0joB4`T}73x__-;TNF z>xZ((bec%hy>UWyxc@t%7=SF2i<@dMfhrP(;3f-kG@8QkY9i%tSal}>4H1M;LxCzSVkDb zN<0>ji174NW9ZM}kj+hnWuNdQ!JdOHg8gz3X1b1A^}F5i5yJ&$I6L)qe564c!3|Fa z)y%9EkHTqNJi?Qjn~N=vAGv{3__;bwo-F7N<#1s@AKRj#Vgj@y9B!4|LMB!Z3SVUj zP88R{^iDSOTrRZ@yS>-xen*gEEZCK1R8s`R2_djF+(-oIB%?O(M-T%vvGPz7-cjp& z`E(7Xk59w6fDBqRZ(T)X?C1Q;7I0n)Jf@L5tZ}Wos#%_FZjP{>{3> zX1$4D3YWS5$|MXRlVtWjx%L68kU zd%BV5r;{~}aOG%|O-}lSjWjF`K^1l}FL+N5<;!IK-A;u4&&jDKB1l}^>4O>2#1ub# z{|S^m>ouIHO>zGfIEP=IXYYR2bbHzKNkc>FQA4`Ii&EL)&6>6D(fke#^O4Gen)Sqq zUn?Jq=K21*Q(SPTPkSDEd=Lym?M@MxQ!di}+@ce7HfV8z45OiRoy44Zt;VQwZEGC}OAZe7y|6=Md>v zE&hT`cVL$}*l9tq<}Cn~Oh#y8H1z=QksO+A1$-&M{}|i3Eib>?nGYB7e*-qnm`N5GxN-NA8^#v zBmv55jp*tp$7a2ApMH@B($Gm>eDITa>h6PA^JGh>ha6~DBc0n9TSK(@qg-pBE~g7yjguUsXMiA( zTt_*fKBcUwwwuVF?96{~>$FR(etMFMQIzAk@atl$XbL!$wC4{%;d@x9yy!+0y$$p+ znq-;fA3(&~ZY9(y2b5O*a973X5$Np80bQCl>^fm1uR#Q)ERq${PK&C9V}M?*F$h_h zPAdYn!%TEqE=}zZ7-zImM9DnLZ=d+Jx%dmZfI`qI99)atE|0h^6a~Xe_|URBZ?{lx z9K~^GRzH+a5+~1s*$$=teD#&HphrNA!hEzw{EUrn4PTu+r`Zn5C@4X07aLiYptB!M z6}r3ESlz=hh0mCs?6OK^K315t5xgZ&Q)FjuHFP&zK|T0Pmr0f=^YM?Ye&1fhNDM zCx?vXluqFKNP$c|nZr`Orbz1Oy~t(mrFg2&-i>^kqTaRK!AGhPx_uX^av9ASU!?cw zku6xfO#Qw9j2f`qIbEU8mWM&-Z)y{3eA|7J4=*d{sd@8m=Ytpv_fxb^4Y|(1tzsy* zH#!{&9AmUp4*u&+*rc8}0E98*2L!+y+gPY00O%{-y-vRCdnkT=Ui1*$H?-SfvY|}4< zg4~n9$?DF9yGI7(tFOl`XzhZxoYeBs>%Q?n<-;3!P~>jiL}o_z@MEBfz{`|p0SCS) z5H?|Ws6w))S=r@p^wRcuej!bh-Pwk>a3BeMPj>$$PDXSNT&7fWdqEEU=Re2U$CfSe zw@b-DV044Vy>8=a8K8s2x{!;a&_14I0^xTKJ+b=djOIqPuAF*|JR$)*6ANY`45I{& z?0PMhe*0xxx-9UxZL8&hKlw_(!?na}hKKArp4Zu`CO$InE5Cwa)MoD`f3mA#Ma1?Y zpyM;=jL6k?%&a)3ZCd=|BfiV=NI))f9Ta*Rf>tYlp0lydbJ|G>syoNf%YX)40xFC#({tSjRt9LHAB7-QItfko5m0e6`w zsd&iDMY7xOMnD?7d5O91G9rTye)Lr0UY*D`fMrm8qBaue#G*oWh8z|a=!Qb&)9TNS zk)H1WIGB1Ma24jI;qfN&XC}M0`ks>fIZ(nARU5jKfkgi$7oWS(?Pg8j_FyYS4`Li1 z^c*FUH0yI7P{YiC5~~@GnIAR#s8HVyhts&o%d!M323G;-1gY0ImwTSwaF(e_!gTBw z#A+tTh%wqq^HWxqPGcx6nE9d#b_2DLaZbqGm>kFuy*$S|Etgyu~Wl{DmFiekFu{QGZApzHd+_npe15QRN z06~q(csh;AFTWq`%{lAo|=e@Fn1LIg6!? zvK#j%rYhwr`&_M z;YCp0u5AP4-Fqyr@v7TJJ(Cu-6T!>*4s?G~ckm8EUKFld!+lWg=`Y_i74n zLiG;qo%Bi&huk{MCm3Guw9fJmz(@+E`&NoX?G4zk1M_9is{>ITR)bYa_VY1dMEi~m zu-O&xa{s!v9jLx?)v9yC)EjGt+s?<(kfRuCO=1Nqf%tt76r1BQXsW3NOZ@8G)|^#= zdPRNgAiUkxa8ZU$E?;SH>s!UBn@h*tz{}JRNqsAtJOc z%#9EkF;4oMhxoCnBvdV8Dv_zyKBwH;n#1VRgobrojm7h=$q|eFxAvxPD_Ni#Q1!|8 zI^O_sB2DO`Zd6I5aPM*#5&9cg@ED8JZy}qiU_ls7-b-`-FIE zv4`N_gyGmQjjjY-7*UAFdUNfiQxL!IK=xbNi{we0@VId^WbL`_eB!DZCygmEnQ<7} zrP~W>*PL9XIyWC{K7IvVQy5@}MK$ZQY+S%S^!$c>kkoF7FVnOqd9y9kHT1nKQ$7}>zS-9GHk;2C`9IfR5i#_G_lSx%0>E)K+ZKxtYwbwV5!@O1h%ZDoS^2DPsQV0| zXJsiK-Xpa#`P3KDR_Uwg>(_?h4IcJcX@z>R>;`BKAJHgX`qRIEWIeF=G_XWOLf6)| zgC^z-HR=<|Q5?eyx^hr!q-=4daUacbR{_SaOiVlKPs|IRnmj3D^xzpM z`=2mQW;gs~so#HXPaj94BsFDw?C}Ys2%2cVS4s_D*d&NS((qQkH?cor8M~h_qi(mhOPc@OKp`k1zs|mduFL8N>dR||y^*f% zRNroc#kKcrtq%iftH_%^#@cS#|MHUYqgU2r>~g3=iMnl(w8bt^+l=(`1DTKs+;jm;+G~|h?6>|hPcp+S*4@<_fHY;8( z?B8k=984L;Xh~Xq2rSMvxs`t|K(K>1CYl21(+hi%2NYp15OEdNQPLb9p(B-sd)s4` z=_giIr{m($N_YRH_{skC8gGp3GAW^2V&>)6_&_3ZEjNYYz8y@Z22-ntJB0vSDX7_G zCtTHkOQD?f4Gt14l>YJh^c@Vg)lU@e4zlmNh;Wxv?1^tK#>@dc6L{Fz-w+j@RoQaP zu~73aIf^6m1@Hxu5!pXQJ2$t%KMoW80;_$=Y`fL%d9Y@{S(skIQ?j_Yk(F}7vE94= zBW3IXK;cN5QhZVFxcq^NEreqQ7gqY@LGW%K|CZ(>J`9(i`Xq(6 zyb)*#d;rw13C^{_&;CTmR_J}`(*1&osK!Tc$OuS=9~roiM%(C=j=KyU=Mw)y%?RQl zeP}y!-aNM9PAB3hZBBN48@?)RPgST}r>oJ*M>{#z&!9c6CsBWh=Z0BCk@~>A8F?(n zZ@`uUiBd>#w3+3Y3fChI4~92?YPsPK<(%G3U@_v@2Wx1V)Pt?;TQi+P zJL&rKmNs4;7utTrn(LZ&!+jvYzf%#Q^{aQRdim_231ddRFh`!8j+Vy z1uO3vuzZc%M}KY%Y2{C}6~$@$iTwp{4((Lwirr~OV)vuniM`$n3m!Ui&tHtc5D{;w zWFdXKccWx&Z`-dO;iH?c$Lzr*rfdz;GAWD=gWKh=9)Tc%>>4&R!I;2-r9@4xStMVg z{PNO_`wp0zGRk-pcf4oKBCp$<|K0tF^3zc}rp&;-1BPBPifHfAZy!qrJ0hQEMHW1p z#MNLODLSrrGZ5Vz?hN7A!^Fa$Kpi4%;RimY1BTX;SCBD!p0D_V*V3P}vQzBTl$xeV z%^(-JGQ=g0_;Q1_h2zhECWm?ye2l$#!o!}+8qaA=LZ-oNS;S5#{mnH?1hR~esc?(W zH~93M9eLz}B#%Vt8f}utn)U{5*qTn<&83|v&=U}mB7ZsiDHke`hpiw*@@Gb}!jfy& z0|}MXStGk3vnbAUmQLFxtMXUf1pzPYXEA&6$yqI>EvT6m2?*6AGRmRX(k=jexj#-F z>{vX1pVIBp@0#AM2i+mu71ezXs4C<86OQar=AiThAe_FMI+Uy8@>Ytnj(N7*1zjc%#O8auICy z;#N4cdwv}iDTpGq-6#F2dVSG){z*exY|V~m_c=xEA%o2-rcn~_T0sTWPYU??wKP(A zs6OGW%73e;IrU03t~=q8|KWq^vxr(9`Np33(G>6{9iAo+8o#-8?TCH72+$H8pxYoW zPJoe=7ZdDNfgT9T!mwXRiH3M<4wpaDal`45`t(Iqa6*w#^X;l;Z_?*V?gV>#_g@^S zEUZ*0zw*v^qKK9U1^8-3iBAknoh{WPegz~7BmIr6gIH%1gSLkH!-C4>ihiy zNkpi){bk=f<0=Ya|420f6-GKJyN9dx*G^_zTXXepw0=$4_i8w^h7WjRhE6{!>~=In z=sDHvC_IH8atR9ZybV!OmmkwV3s*UmnPEm&DY*)?eR@E@D+BiHU}bKLzl1Iu@agXN z$-{h+maA8sYDHyN=BHmSrz4f$_JHZc($DyEbo#w-0B}YCUpRUU1Gkekm-`2K>zby z00b}t#0qnuTjWg;TgQ4_+0OcZv%C8hEQQ^!1hBehlFQGeA*6E8T>q(eMiMU63P2ag z({8BcJDd0AaNJY`-p?s=TofNIXZE(PX?_Kx%ROtSL2)BST~F99=j?jW%R6E4Mbc#ugb1#=(*v-@cI`amDvoo{YMpwfo>Z6Dd7Z?mVMi+a+W%e%*aQi7&eNR1LO#Im zJgZ!&8EEJ_*fIE)1E(68p0&`Y9EkZ`9XK!0t7Qbh#F3Q8qh&i3`4;3B0y?i;TBE>1~?}$f;d0Mp3$I#v2-U3>LUAMiw5fQU1xq z!l`zs^ZP2?(>3Nsxc$J+>3lY>A8NRYKsmg{pw!*_YG}KIw#TTKY4SOdLZV~z#CzmF zdlw0pmS6E6VItC#d$;JIrw$s7;*RSDGg|-$xKj`N7#A&1{yGECfY+5Shn`4Bp{Kga z%KP`f1C|H^LQQT6VoV|X9MOEbDyqh6h0`COV5xni1~OlN*joml<$fxi7^`tt=DW$t zMb^5aI8h`IzwY}9-v^oVt88d+0UZLKf*w(#4ErOc4pJ6Wz{iz;UxY?ED)#; ztM8Kq$yeYUemVqW%BBTV!IRYr!N-Eq28euWz@@njOatwix-ds3E24P7+0Obm0jM`x z?=rg+jHeQ$2L{O}GD0&ef~VoKoZxGuTv`VLLhu1Z;s^mv@R1{uJBV-@iYkPwx$gRN zSRNt`1)|->)(GPs1;(+A0K40XT3H(zXTnQbR=WzPojIu#yPyeDw^(+otJ9K2?Q5XF zp8!60s?@M+`*hB$ihxlW-z8tOW)MDL&b)8vt23?NFh!o{db|m)u;1p7m4rT)z8)#m zn&JQ+emyeW8BAi1av)x6jO1SR)K=FlJ_29@lc`gW`d7PfWm}^#ImT1H_b0%>wPvmEXGBVRd`Vz0jHIqDzft_jU!o^PSy0W_ zNu3FXRn}UgkfKQwYTQM4EHjyW!2HRyz*Yr^dKSoq|1%m>$g+S+w;@4+jED_p9veL_ z)q#O%njb1N*&a^SHHV;rPfJ^>&4BvW6@9pjPy%1GV%5-DAR{J6<9o!80ro=+lWs^X*T ztis!`^x#08qpb(;qy>Uf^FVlW2g zJ^S@k_Z;*J=Z>yG%1W?)vje^@6kCYY0*0=mocsJO@GT#;jN zhS_NTMyrqCcOc(vBS+q}Qw^kM%ZtNwn!fZJtlGlvZc<>i62RCuolhjIOkv_1!!;TU9w9OG^hxKw^&UVJmz5%Wz< z5+$i!77=Z&QIOT;@@xG{YRJ`4he4vL0Hn8Uu<@NDImU715t z2E+Yhrg1Wx;C_}k(u?}mhmbjn%}`v+gf?*V9Z5{e%mCNqIZvv z7l|6ZT*tU}9Mf7$ymD0wz#_qx5x?>7g2;Y4MhzB;ru2airs`y{12n~ILdjXIbyHW&pK9J#`bm25NuGw@ezkncja!B(X zK1w^p=Eb{R?2Gy6**7fyv$yFgatR_A5eG=))D%luVRzs?Ly#;*_M$$78J-9>6koUu zH04;Zya?_m`~^C~vvvRUCxE-lU5cbvn10$IaPCp#k3|v`8`O^?G`gp*0=pgrDF^db zEec!t0%mbbz^%Bs0tk-PF=I&fmIE>(5xiw9Q4npTz!tdU@f=vv9%6YNP-Q$tf>M*m zu)}WXdl-m=r&o^*HR;`he)a(OK1>msl-==X}Oz3>PzJfV0Q1t)&s zGX)BF0avbLZI>ocD((*V6J?%KP#g!7sSR!51~uAI^f^+NF}iRSAM70Uq8~gOoeiL@wQuoF+|03QBDki4;2Fa&Sm2v0&;iG`}42Z6T#|d!YxVzTT8<6WD2k zt_eT|(SR#U7w{>`ce)_wThsx$aAJ>+m%s4;@%EKbRjysMvIV7-1`(0&l8{niQ%Zw0 zN=k`ziKqx0K~g%Tk?ux9N))9L1SA!Z76C~C!Mk4L=y!kKA9oDL_|7=T@7-@a&wAFH zYtFe^L#2<&!IKnR+I%ZxPc}=p3}n`A4jxcw^ay5|iCzZZ zPF%m~!ND@k873bz3-Q8uZNj(AQ*Dv=nFd8^j41!2hO@>$HDv{)e)*)^VXXMI`kdh7XPH*7S zDO|JBU>qo_sY|~U61@#=fh!~6^)mNf_XDUBBnJ74kL5-WVpldUL96WkgrR`^@h8$6ipITFQD2woCr_ z{tj^hqVs^JLJATjSAbnIc7LN=(2~$q$P*ipSt5=wU)tGHH?7mIJ>;rJL!(dlAuy7n=fW$zR@Fhttp_QSP7m+Hd}R~_)NP&Vm=XUw zE0;iD$CnL&&}4=N6h7t`IM1VRY4tYnh z7wVMP5ZPPf?5ZQl(M^S^w~jrrL;(|ZJ8>*~_t|XS1f2K|2!Y_lX`wcU{;^DjC}%@8 z3*1+vknLSs21VM}F~?2NXf;B?IQ$N*gL;5eGrK$>xMskR=Gp{|K>b8vJ7D{p6g@KKRMagD7q= zsh8b=SQM%Khx>Qx zx!FAcy&~%N!xZK#Eh~Qu6Ld=F4NhXZ5b%6?H)Dqw>@iMySum!fut^!!pWn9Ihp0eWCmch? zAB8)G^h&aAPJ&+Oal=4oW1ceZLrAfY5Uqx704 z>khDq{n~GGt_XvPne`h$6%i{YCu12oTyrK&z1No!QGwUMaJvHm_yn~8_G#Ggk$ARX zFB4H2_Fxioyq;wbPWa>*lDP5l7C(Z9{V|T5q-hbq7@DMXCB@@WcYvHBQa_e^jkx+D zk;Jg18njTaJKGp@Z|NqUPGf}aouI1RD7@H%g^n6%X6i6ei?D>syDj-RnBIlJZnCyOkSB5 z47F0?wnHfSD&R#j@cMY*EIiR|gH3tcA7jTU2G;>t9j1Q~C3JFg` zdOZ7>W%Erk&~xtaXIa*{ui8V?vd9iv94Qv;D3;fi5w^C>S8m8in}JFgdbZU_gtZ0l zk==$};2#vKrAp{zTt1O%&sO0Je7)Hc-931V%Lwt4Q7K(>p97d6?KHBvsiZcu90Hx{ zUd*h#&zA-G3ZYlc1t|CJvTpXhY|mzMVsaV<$X*|>b|OwdwA3u2NYclbG!QOpTx~zG zTv*Yw{<+=!7Q;T4^OW5 z@fA5-0I9a8uTqOO@ZXGF<>m9**ry@eUC$=DrE~xRm*{)#od@z~tTV#? zQrU8Q-Vdm^*7Fibb)BuWu4ZxwJqW);q87+q^bY?$ zuN&yKg)EVmu=-A)nRKf*15K|BA;{Tnj@n%nsJI%9jfbPZ-+^)XR%^nI!<`UWIq<#P za6-DU0<=-B&w{pJYA`F3P=0xwvkrlDS3cN_Y$Ic>jwv~E7!tK79YE`xPqn!5G~nt~ zbOp$=JA-)wFH+#0tSXY%sClh$l{;p4V}5`B&Vx4m0|?PO(gKBi7$&JKr&t|wxgehD zuOJuQ%}og5X1=&}rF1wz|5RSoTf*Qy!dhF>b zd>p$|6r-^12`Z}~iQV0~rm3xgr@|*4p{vm_{7I2-fUba?B8X>pB1PIdFeo+|sP(f( z8<`KNR$TPQZ|o;@y@v{llhi}Wa=#J>lS@V~O|eCt?nRt?1VA78oq{81UO)6PQuPkq zscv&Fvr~%xu)pbwkcd~UdpC`7PUOR#3{TJ^N~BW+Wqem$k@N7A0;1HxtSP?9aO|tB zAua32R|Ptu^0b@nh@mhNVV1$85EEq0h>p81u#(rr8n?W9!F( zh90#sLq_%U=Dj6aJuQo|Uyal*;hG*vPvaymDqG^J>P>yJw^O`S;$GLzdC z_(>-CCndi-8oxPs!R{Dw?qevKEJn}`H%Xo#>5_*dyqIJ5$+0CR_^QL$R!L5enu-DY z#JF@2An8uC*wH;R)##YdJ6U8>7?))N1M@aXgM5oLGtQ-I!B=Zv#HT1~$h%D7O% z3RObWz(G#rO)>*ehh_$N3x`He6d#t?wzDQ?F7OH^@}j;ss!%O{Tit-;FmWym_2cnt zuL|s2oJFep;_VMERC24O#nHQYk$ph)y0U8Ghup1`--7!{_@)GN zhLoIthjW3F|4oJ0dK!I?Y*%In*mq;bdow3kUFZWD!KDI`%qoDlO3DG2T-<>Fv$ z*(Qf}hXKnlM>2gXp&+f=^QsCo+ZkaJx4_e*PR+A^(uZIf$dHvi$rq?@8gn=;J@g&S zy_!)7#4enkE3=P%>OcKuEA7UH2ok{fV;HdpSnWa>`fRN@TF(|nGex*tJ^MZY`?WzZb!f-; z9-rTqN2oByD#hb9;`6xb!#@8SPS4BFUv8d6#h{OjZ7O>A@GykmLV+dXb1}Bu%=`L^ zZpPMhC}MdcEWBjIVUB={ngcn`%~_wG=cnLM#Jm$vSbJK9Zl$oB41U5Hsw@=7alQgF znh!q4Hrbn3#9wI%%?BB2BgR~_y;zsL(fxSr6KZ}co>8H(#~8w)|KQaR(n}M?K^Nhf zp==eJCXoS1=LH|w5~fXKC$&ERD5-4Ij{GmdGQfw;_tU3uS~CR07cal#Onq&I^sOn&A#$EN))%x7V7oZoJPe<0g^zyA-)=< zh8?4^TLndhpufEhI;s&6?g-qEb~R39RfC++VY~2gSd~}_b~K@4<%-T#PL0;xBy`$$ zsEB8+z<(~_A{@BHS`n$N;a@6Pu1LbSwUY4H zO`nHN{`9o4t)9kMCIY!t(GO)_B0rXB5s$5N&T-jZ&LI% zWI5Xmbs=WEJNuDS0cRoCE>@e>!cRRkerwqae#E`!xkF&FLqAZF!|1d>dJpeD)O6WR z@}y4{CGV}^Vr*IHqY{x^N>b=GSPHk88}XaAU|)c2)E0U*LkRE=M5scCkBvhGx_l0u zJ>oqBUX14KPgSC3$G;)C<~(5DdY@iDZkTLh{Cm#adwHvqAXWBn-sM}*+`V)zGKK%%obRkr^*=+%Zqg&Th zeF6$!v8ml7d;n@EaIQ)n0mGRkzvv-!yGc6OW*BvW-AdJm?Id^XKuf%@fiYp_b*m;i2H#T`{ZKIVE;S9!iv1(wU^hsaiD z7lHljGk_G>W`%nq@~LnY<)Xg&rE_rxmCBneGK}5MB35Kx(4S`>FVEsOZX6Hu1Xtlh{WDM_ zJikwpB_u)dSf7<)lunVAlr$Ym>&>wn<^%+ltm2y?Zr}&n#x}0utkC^&$Y_7D^)Ff0x!awBQ*`dJu27}f!>z}^30qI zHD7VjF8YxMD>@CxNG}4_m3O-gWZUmf>jttFxts{z!ia=Q=CNW0StJJIL*zqZJI z87R;#@m`2B>0g~Yw?uh)GP)6Q$rS*&ma}3$?1N0&5v&6%N%s{4LJoZOsS^N$1Y$EQ zCJ5Ut(M>#e=X1jiB2_Q|33H|QBp%b1sSP00UzO!<#b*z!IUzD$s#nH>#mb-Erpx{o zUy@5adqE%a=s@lmakC^}^lhY*`2M~?!Wp)-0qO%53jd}vKj_ch$DKKaC6`{=qe%2q z>fuU8&o46Zs+JBRO)NeDf9J2kb((i*68>?E{6>d`O+kl)Yfb|CJa#@~4zAz2Q)n?B z!3+THA}fX8-5Wg$boF-XAXsO}n}pJYKKv}a=uO8tUSfFytnD;K0!#b9|8*V0{QA{3 z6e?cL`%~QsK-ROLzGaA*6g2IU zo>(43x|PV2qhEo}gi)Bm^*Ab!SP#6)VOa#Lc#>h>od5Y+kcUh|A<(Axn=Il0d<>L_ zV}MOKeGBgJeM|J`7M&l#La_pkkEI~;h>(#8_X?OjMBX}FfIbK=a5afy4;$l58qOex zSgSoa*qg17T>9@vj-V*Osv{rwQn4Re7S}|cBqA`4=FfkfIRnRbkBZpQp3A?m)lXdl ztbcwxT@@OpKMT8nhaePe#R}y7^Z!HBK`%d=r1$$@S$f#WwNHrExL-!I?@ z_8~Ywv|hnNQLCDOp%x`J7KazLM3^gN8_2pig6U$0m)GTy@t(Mi?Jx4U9`+UW`)?pm zcnN-IajMIZ9vF?k&_alXf)+bzrhje|t~3qHC}_UrDy}&p9EgtJI-e}2XZ*R>|A-as zmk`0vJ(r$G^ z9%*-@4fpr1A#O8(2kA<_a5v=>Ht=Bh06r3aL=IeDlW%`s2-xmJuC_kELL3m>+!4cN zw>(@XY}IoHanx_RTmshb!D&Ik^f7-f>5aaHtEHw=QHc+?6@`X{{Kfm|$u(3yAY((o z{z9-huN1I&ECS=hoWt`s-_pldXYu^^mX2TqaEYIC9wp9UAjUx%6@cM}TG;3Aescl*ndgJFRV&vU4jI16Z?()sVVn@LPg zXvf|stjzli0-;Ej4AWCtB+yxHP=Wli=i-Jkbis;v^vg3Vmy$YxH;pkmO zsy~~V9VC$2Pvjpp8aLdng8mdRXDt&yV%t47?*X?kud#!Ni&o%vZD6<#}HuxTx7Bc zxPPYLFq!_eCi%Hc=r#f|m#0LKgiygdnep zehI$*)l`PP-|kSqe@iwN;zzlr%$q;!@8`b#`dtNP!%kATEA=>P9cr`;X6Ape6N z!UfPK$^nQQM>r2`srQ}9>!P1YAi$DNg{2COsH4f~O8hbnk|ZLqz04!RGHBgX+&lyh zaVZ`d>oBmdY`=_F4OKfB+d!ipBVLa|>m##KK1rIr%T}p?Ch)i3tKdS6;Pp^OlrYBe zK}P%G{)UR?!N&*t-H-^HAnYQ<9|S!ChrmQN$-PoAb#?N(EV(;FMgva!h%l{8MQdRT zq^kGt?IqtCv51Qq9{Rq!@ddoK__wB@7U3y7%Z9vYb8%Qf@e&U=C3&zBKk({t=U`am z{5(6s{VK>14n#s1U8l9ucpnsqKy@n_0Rww{gXh4Lq2a&b`-AZwG%J9Sqx~hxrKb!G z(VwC(eL$xPq4pjs?W|0E0apDpV5sVWzwY<>TZwkh$^ZhRsNIaWN-2Zx!w3vyu>t#< z>SX5*j%N=J?tXq>l8V@5W^cEEtd}bBqDMK>I}8<@JB9etR@sdSUo@QCK(^)P$LmW! z@Ys?};O*WC+=*Mbi*!^wxJoL4(h#c`_KQh~0lup3lh%MRDX&1p!5JxTNmi}fClJNL zGQ7w2PwcfibEzEm#Vhq2s(eBM9+M0E#hxJ-8j829qnL z!8kOxof0$raJ?X$FfMc0FM+2|@`PC55>Q-zV9^-?NjuB4iQ9;n2L* zhbo{O%u82^+DF)i^b6(L-JN`05bo!#6Y9QUcJMmIptq46+3p=Ji!F=B=8X5|){43B zrZhPYjbanTl?giR1*Ewu2l?JxK23@-TJLXxzlKOK&ojM|)Z1F|)4qtzxMJM7b5K;I z5Fg%5VVR1gEIloj!N<~bQ*~6arY|22I=)93D_zj*ULM~sU6aiM?7c@iNLzH7@qMMW zW}IsvOKiYdjn$9iJ9%y>sgdR!?a{pfI%dceO57 z?kAjD8G2#W)!L-4T$>W`9qR{5j8R5$4a?JT^%bzsSt=&W=$B_bITeCz()Z7XA%+N6 z0IEmsK~sD z-F56rO}8`kl)azxB-HLrDD8ys4?e?MATd-VYPMVyET*yBYd%V(Jf`Z|xX0~-bU)q}$3Ne@OTygxU;I54r#Zk)tJ0dYA+X?3Iyt<;8W6+35{<1Lxp; zc5%V=IJ*`CL8ze{pQ~`P{EarHyUYqj$^@)nWSXK6 zX&~9E-%arYN-Pp$j5Ek7UVfQHH?_H$V{bwOO9b(})(rR_z}`W!a)!##HDY7qO1$*K zlW9RUtC*8B(1`B_aAs~U+u;7!*5Zt?>w^kx$upO4EoKUq9C=~q7dR@xH+MSZ6o1QG zk|`EXfT~2}xcmFX=37;FU2iN#uhB@}UsuXe&gv9VP3E986*%ugolAfrT)=*MQ9aBx zGKk|%8vW$T%kRfe^;_2gz-M@yoef-cvOb3+D&R?JU+3mC!4HA7>vroAA@%f9^C6kq z#XgRVM&8Aq#GaGIwytyF)_HQ?=OH6vpPYag(pf|-U*ku0*E3x%ae7QZ5cCmjvkC1L zT^UVf#8*E=hsuc3HGZ0YRXdHU>fkJ3zMVlh0ZbOvt7c3=q+dd8UNC#znZ|Ez6Q_JD z*dzM#40xc2Qri&>MX{GHJQO?QovA9zg z8O|Mlc;;dh^UpnAjGd_-^Esx#s-cX2ah(3LM%?5R;ysOI!P9a%C>+0&uK?84FyNQ*@oru+eF0xk6?nt&u&B-Ld|gV0VGS8*)GEZ);Nkd`YN2ZCUbl zn}zLnJLo20ciqT1D96PY39sD6Y1w1W)j?Tfn92L>oxT}OtRx?N%r;YmV6GJ8)8_W+ zyZUnJHHP+LirA-!b+|r{Vfv4Qjbe-*$JN})m$-uS0n(B18B^EX#ch)`@1TAF1H*8H z{S$5n{(BQp#JhgMiDlPi|D-&SbHE^rWEg2Q&s#X3%?!3XHJ9M&xjRqF_kUzpp^Ii4oIk6?XW%z21z`WyZF4t$Bp{yMq$hT zclW||aIQ<=UwpUkS$jC+5p?`WOGUzbr2FI9#_8QRp6!=j9XY-lQhJWw9W;_PUj}0K z=d%;-osaTQxlXMTUR?WNnig;`REp&T1G_CL2{9H4yDlp4AiHBLL~B(K`;J2TDSl_Q zoTwiD$n5txUP2Bja($|USKO$nEwc*xPFO@xZ{?6X0dnb=-==?{1=HwEp zkMHo_pW}@0U}@kRBU1dfPoEI?K$jP5MPUU4YVtw-a5|3}5)t;7Y#q4M^XREt398>8 zc=hIh|EXu+voM;%622iOgL?-JbDcA_t(O8FFZSV6Gc#(5Kjd#WpJ|J*PxHkh;hTA} zq0Wec4Kb^T#@DVjsy#D6^S=J@VE@8)YvFyS);t_iAKF+ z7`gl6&NJfV=Lmf~vP(gyq}C__)Y7o<3De7RiBK(ZI=9}y}d zoKz3_D`%X1m4cqLQfiQMbKAGSdOv2|Yi?aWL78K)-gRYUWJr5a>EV%HmGrHtV{P>0 z52e_(aUTj04$1Q9P9+Gg`NR;sa5iTDYBwDBrun3eiD}Fz@Kut}i`!D{EViUVvKXHj zwP@(|a{rqu`PFTH5kg(5K->UtTrGTzALp#B>)8>q=Gnkn{u*oIEPSepHC$$1_mKMi zwHiI4dKS*H{&YT@@pOA&h0SO5cAgHF)`H5YrS&dPP2k(7JKKA0l@*ukew-m4aasYi6;OV`GefYo}h*mO}U0*yQTkeHq&W>WmkrPnSIQdu=V zF?1G%#iR|Zz0eY)I(4+#>{x%~rHl-VtH$I0PCTlthYwnKbbo&K7+Lu2EQa0OJ$`{N zWGz~<2nzylA-Ga^<3Q>3Y#TxGt*)atLNV^vc&j^$Pi8{0 ze5fu2$qx^9=HPj~?iRVPT)x^U7_Mq6f7?JjJ47hMSN4LD9{R}-ExPtW54Ptz?Mt(} z=l0H0&IyvpGST`w?k%`?T&cCXA*pyVa-_w&XU+#-{h#}(8sofQ&nvIu4 zU6jfALh7!ydU8J11>u7wmRxyDx)NtDJxSp!mN9~R^oCP(%FUF+Myr!)RFuuE(pKBunQ_?&M?Qs1Skj$D7!ilQ`X-r)GVrc&yu_s6-WUtQ{Ss zUNDN&PZb18zZvHoYQ2FSvpF;Mo`(TfT1%AIu+eqfNw7=HiZxD<%A4v^MmGCN*PH6X zPF7RBSE>o#Gtg>Y4&|WJ74D%fAWwSo{^~{x zlcgtEa4ke9Jfa@mi|HYv^(rRt&^TVqbU)Df%Hfoa+>CgkdXnT;#$%@snP#fAihQ(; zLfqPnfz{KokhQwOV8 z^n8u<*^u#f58ef)R4*=&ThZACf1aK{S+7vt)4;m*B| zbYGKq^D8gKis1HYTUZ}W@y9Ni>OJ0g3Y<_XMl;eJB?aFS8@@luIn(RHJXnx|C3vhr zedf+<0q^F9hr#PG`{HwoNX^=GQL_ng=Q(xBXu5uvGdm=`e2jUaXD^LTxL>ppjRR8g z5QoXQF?QLces6_2v>ff}6q$Q3r-wQ7$^YJ#LZw8t34KS)MF}X2WfvwDf|R^_oU`#> z_MPq*n_Yj$KhKh?s0za89f56zUd=3}s$>Bu`Yyz~qF?IZ8kP42IOnY^S6_%`jGZA^Ql-F^ZwpYI2n z&ri6eZ>hXyIH}34vG|0oQ>5qF&4Z;D!NDp;ZHX_8`N|Q(C;hcNEHxq*-uB^jM)PVP zx=k1rI)A=qOsgx3`d}AIdX-!6C2y2~+_iKa2{7|mp$?wK6G5-4zB!b}+ws(jPhVbe zk%S99nm2pa5iNN#SbMX!RpRHd7)^tb>nLi;cR0kWhuNw24b0=6j>e|W?#LgCdHdPzr4wJR}5>e`RDS2!2jn@KUsLE z@m+~gE+hx~_YVgg42pr5FZ}m6fByshB`glD#oe1FKc561=Fgu@W_mwPmeXJ09{%SK zW)ljHtoe+Ff3J@Gb9f}&SM<$FnZI}Q>o?$@q+?RTwuAm$`uA_36;T?w8tHu}|Mv&5 zNaWSv5dVLBh-elJJ#I(&hy2RFPZCW7k2rg~Rm+$pKxz`456=T?7#yLY5Tt_ca40MB z?x6bhjF>%=aA|GH2jbVTZ;;eLBfJ}im{4-rWy>OEy$4t|bwTeoLSPZFgj)#E^E!;q zVZc-^Fpv56!d~<%3w4&?Z#At4n=)@BLcBB+R|uvUBXeU|wAr-hny?YSZGbl|kOo4M zK?e95zFDq%Zs*lk4UzD2%-NbQI(m$){FEbHE|d%U-TB{^z696To*Z_s+^>3_pQ^zxlD-Yzf9ozsQePf&s&aQE(OWC>6pE z@|JmkS5?4gsf$9_Afhv$o$v2T$h`#@$=_T5JhLI^jXsLF|1E-5w%Ba6rn36;bw05Xb6x5C{2Qgp%8TF1pyN} zuNSz^U{hS1&2WDYgr^b0z`Pvo5V9RE)ky$%y;JE#z@FZlgK`ZoO$(X%t~m z8iDY-YX%T3h!B|Y+x#P}ucZCVgEp&SBG>8Y9P^g?d~h#5J$DTN5f*(?JFoKhn1mBI8`8L%n?p$6NqUSZE9d9Rvd*_z8wJ|eAK zL7*nNfSkbmI!@N#t6=ik;8iardLNkyYDPSf0+K{-yb3U|h!;*3#itjZxHP-^fKKx=}UK?lgho`*ce-f?>`99IEV|yB+vX_mjyT8$`Da(4yfniihhM_KY=ZM!D2*qPe$B*#THQlRJ7o8 z)qI3KT_XmN)%aa4C7!D2v23q@*XjsKxF6>QBSjw6dE71-r~^JTU4&;Z4p##G>viM( zA9{dapIEsUFP#HO{#^x?m41AC{6+$V?y0me;OEa1_@sql0H$y7-UTT!*V`vJBq-yP1hG z%-BKi2$VHQ>_{=rjBBvM5|CNln=t$x1P|hJ1$R9uI60BoHi<~XCRwRZRTd}szqgK# zCxyVAO{F@>%ZmXHY9hpoz9fj9az*mIA^x+YYPXW><^65?3Pc;7O=ZDf;pg($okd%zJ=*Rbg?q%_WY#SbLfQyfo`)|k{ zMC!`(hIYv|4R~FNBE3m>rUUIK@_;Ot{b6@d48zpZJu3g${=aXq z*$Ylz4JBd>H*OLr2Q2}s6ow$%JDvNT#ST$VLA+Nb=N!sG4EieSCjjN$EvR%xVB6m( z6FwvCEg#-&RhA}f_;!}a#5q9_fW8B^;djbdxfPqjvPT{ugsc!{bn8cUIY?bZ-#6{m zk!17A-TNDQz(N4lMUqXPU^^35`XrF>I+r1qX1mPycyM#<7rKxa5IuE1z!)EZBymc9zv4AHqC!O_vSRW->s*Vh#{|^2LVg=oZ&Uz?| z5#0)V=1IlH98Nrc9K;2b?jADpJ@Q#LbUafOdFpKFKB2G({s1 zAB$P%nsrtZmsZm1HJhSY=~qP1mb(a6QdvRhEz-ZAHyuelgf=;A2Jnan9b&X~_<;dz znWsBABi?WP%~wry3LEKDB*>f;dsSLm_HWdjG=`{oXz|p6F`OP9ePJj$;hd?D^>C@_ zjeB_y+H<5^#nhnxUQ2NnCItWAze)_hO1{il8I6Y=dUIfR%kR0j$&a1Eeq8El$KZ5h zdrvhO+08)*@b=SzE00gda0fp78w}Cf$hM4$Gh|B+ZAMxHl+S3zU2oQzvs4lJEy1dC z@!HKFhSB8~$4;KPsVaHky3sc)S9%^>nI+~N$_ zl7gCH;;9XgL#+fsw63L~c>6-o$M4DW7{17o>2O5VqWr&u--`&=B z9yUoWYyS6+(B`$Nm))?9ZoYm@qkG$pNcWT)9!7~x8kHy}DSH&nr;rM-WU+FDiY$l9USj{u-XL`U4 z(>7&j-Q0&}R!?0HJ%s|l^{*NUVmTTI*>5kXbaP@;jwJs8>isA}Q|FmUlyF;Im3u0^ zUfb*-fDdzA`N3mZ7{YswrfO7HpYY$ifk9Ooxtv=aOZ5Vzrz9WL$4hwD#lw1scBh;` zHs2OD=x)w)6_!d%F&@b2!@!W+yyA>foa%)?dnCUqB6S{!LfV}%Lmor)Ye1?_K>9q0 zoEE%R%U(A+A6{K(CO6e7oPtU0d{nv95RQme``!P!2|eT{w)>Y}^^=HB`I0FE{izGO zO#Lie&pjaq6{#0@U9_;-=N! z;e3yuN9%~qaOqPsQ2vrtfKD+U%pS71#q#_A1|8x!SnV~L(Z!6o-$7Dhe~`H{@jMZl zpW3g){M1l>w_sfS+#5PCLI>W39k4s^R|f0R`(SxJ7cO&>AK?Z|b?Tb_y`H)}vSJmk zTTmn)*~r#wa|c9Su0%Lx%3vvq;p z9DB()`M;dO8&?)D`p5KzKEZwm5_xHpws3|y$TqtN z>~A9b5E)7O@N?Eb0h-ef;RiUgxSdtNBQ{;h%?K5M`Lr0*FL}3!hthv`JzXyu98r>j zbH{{L`N?EDQTgRT=yRG5!EU^({RL_pS9@Q?c3K}SSG$pL8hh;)#|AJR;=wPaUgv}4 z_qB%}pwBz87`Tl%E78n(6qUU`eZHSm{^7-z@|C*J2~|Vaxav7(cVKk=2$FRvCkpde zB-pwmkrLUgk6h$|nzJ4lTq;jHX-J5eY){lmg22-fMk}}G`gD;0PDFf2*&lz&g_9}wwn?Q3T#E5Nfj-thKFmeEH2LZt)f z$mf0>uKpOmtBXx71`(ZbNKAbFYCKR@kowL8mgtF%1t1cBA@70g>UGkCy~Ri0>(lN- z3BJKQvv2B8YZ+_&z@q{<7W)&_k|2Y*w}H4IDaWaxmVSWVzE*uRFH~kR0YbwIQe^&f zwqoqgmzqI=J5R-2JgyX>ennVCU)VvUwZ!?CSMW+tLeo0wQy~uAkLn5md*7Kdb%xfH z$ti?HB)E7vs(M=pp|309JePiX$QZ6GCJT>?a2IOrcO3jkWr|3IZ=La)0Lv+4Kns*| z&3>I%NT8-q7Ml)W>I$J2?0OYQHu1pgtw1NYnO+Xzi=bzqqD}Gcd zxeJ9LqumI1wsLa}a43cB>h7;16M+eX^8c);d_TDDTebb#cT)5ibo5MdTt0ZrYY&c8 zl&-*Vfa{U?6jVN;l9c4dCptQSbX49bQ35ta6$q__vZca*VhdN+dA*(O!j6Q2WJXI; zou;Z->$umCgYyP62xC2xx9k;ZpUl$1JvbhE_F}$xd=jR<0(24*G7-C^?@N)#Y%>43 zl{qAk7ME|-L*Z<|7GEo_Czr3H$bU}m+i4+(592(0&%wO@)M1V42sYse2qn7#qp(N( zUl7(%MZY3f?Q+?48odV`k|BxJZvRpsdbD}t`)4;gTG^)d@p`^4WQ2)lQMFou{^NEB z%PxHO{ho<6sI1Qc2_@^rXp}M-o9k1RMb;YnW`{9F@51`4@vsC7fbJT~G2ncCnKSzu zC|3fkub%RS&X!hU8>U!w?&caNXTz7I*q92`^fCFfAP$o+`W#+tgm~%H0s=(Kw1k;{ z4keGch*~r8m>q;0%kAZnvN7UfJ9d*UuJjDLloGPr ztFru3qoX)sns~A0kRaEISzQg*%{N=BORlx+nvz}^cS)u+zVEcN58qr6m~YOsUeete zr^?x#^>E9};mD{pA?l)Iy2ThIEzWS~cyTyXRYeT+xEL`<0#8WVHP2pN3x)tP!NGwH ze~t&{ucq>MEZ^BH95PQ4ZfDXBzZO7==ysWpDV=f21E6a)6tO6=!Om;bDuGbZUkJkO zp2tpbW!V`|Cv9^|qw}@j)*)mQ%$HKC%C0&}%L{v$6WAof@}5vs-zUC6 zW0QKCpvHOu{RN(aNe<*gEk}V};HXz`By)<+G_( z$ov#<%X2&MW1NArf=NuuD~D4rx~@f&j-?yJ4EptE&f36kovZlBnE0JBPBu*PbxH#$ z&;t*I*OW}=<8k2BEE%&bdVr}SJnpzz4n@(i?;e4gB>*qf)m@wIqodTrX}5d+Df9yA z*s_6wZi@Ew_%p;`q6|V?$+$!Bcu&02nW$jCA;qPo$l&^D|DP2QdY2R+-X3i&3q56& ztnPD&_m98W4mwM6S~lPt3da()*a~X&CS_?*GHyCeIDUGm;WwBMSQT|a*Tphq)-Pfn z-@|B4K}~fwRgErkq8HO)9>2Cvuz2=#^1Sc81f-{YbN;p?EnyzKNA|W25Dh2N&V5! z>Y~h-OqcQs;zYOBd3{X_G9%vm z9hx_u_QC1$Dsj*5_>~f;Eki!c`CKU*KZg;W^9?k9{I|_N0@YhgyF7n*}T61AjdXfe1}e)O29m2nKxlD?9ZdoQ4h zG)pmdOIYlpG%b8|#;Of-9DjuU2dE-xJ^Lb$qYUJg3E z3LvqEt^-BU?0Z_kOiC`ahTRM>Z70Vg@jEJl&Me-V9yV9x<4tS>pk`VkUeiyYb-Wf} zUzTWg16CJ~na6J*Wvb6c7Oo>TAZDF0O-yUntwC!{2zzg5b#*SVbX%*6z_fkQJY78| z*u+!nDLNuurq>|Vyv3lPH0z-d5Y;LPyj0h)xG%X5ldMarqcqwEr z^T7ov(gpeV+UfWdeeC)&OE6{D8N)g)+mX?zM_>D)9lI69@an#Bi_?(Z08XPxC^f@e z9+i)V&vfL{-Ybrr;w{Ei_G(9u2XuD6Td!~DM8Ub{Bw+ z{0}z9SFH|V(bk=Rgqx^BUV4k_a%Y)ibd+@Kd7JT?vWkF7RnT%iuQkE#vk6=usCM{x zts^A)VT_~y2JHl*#Pn0H-taW2Y`SUm{jw7pw>bPU@d86=KUw!X_q~RyM+fXj*fnd0 zcj+md3OOq8jiXFsl+|k--!s+tx!Y_mE?;EY0G{Nx)bkQwzVqgYj(Au<643kH*;C+b zUB924tf~F!RLdZl6MyJHm66jr>xqY4!19bEXnCJrv zL1+FYGVyFzdnYeO#e|tAY{z{r5x<}WGyOy%>qr+(Ci3ltz6=`!jdT?9{a3E}^T_I% zKDj45aG@kRO)rva#aLKfx`n07Dqy<#{snU0p;znckZC?IkxdWAw;+>?tp7Yc=hu4N z13XfaEJEacyWC`+5IyqQLVht-uD*_}1OVbLeZw8Q=7ov+COqPpDtR0U zADVeHih$xp+QW z-jbryJ2%LH<`7xU-ga$>!@%%XvjMHG-bOQL0G4vCM6|=aC610%haMjz?`SjG%Bklc z^w617pDj-fn~>QjR^2!@WGG0VOsniPhQJuj#j~@|NNOYIeGEIfN@GQ-Oy53QBC>sZ zTDRBW`=?PSRUNS&x!WDTdTMFPctU9(w*A!E%J}%)&k)!u%BG#dc8p=Xd&;K!+;O>e zK1?7PTb-NtslFY1%O5K|sSbK}rM#vGCJh{nQ>LD4v1ik#O#MtIo+^`^!25=7CLMp_ z7*%1~_<81;uJPTAx65K>sS;WlAL8^7#8R&|E~HlLJrXrGI8#AlaHt2=33Q?A;;Sn~ zW+|s;_ix_oGb`OT;Wv|-ty2`<>c}QslKUXx%-B`C67OSWu6gYJaV`piGh@WtBWLm} zf9)*_Sv0Q~g)4u-S1re{84m-bS@IN)#7P=N^vl&0c*t5GG6vDDSk=6+s~Xg?{}85| zoy#AD!+SG6nxntGl!sw7-s%do6#4m8o!8LU#kWWLx}^fG4kgN8CkR$TMmROkbUF>S z#9obOEONJC7V_k@2y&IRDhv~lJ3BHEyy@iEOl1^5ed@`)Mm(chG10~g^^3ooMT1JP z>$i;d%Oy8JH1wkT;}bv?65Tbo*`hy@;rgrgIt}6U+J7#p76uh_I>Ga9ap|G%f*v$(@f%=dU*~iLFt-h7?MF7&$vGoWOasz(%q?cHrCyl(1OD=Pb zrN>U%e0~)#qWO$1;bG92yqyKL_@#-w4D6AG+jEeZx$>rVrMA!l@vfB~eU6+%tChY) z0D-d7T#-XN+2(1&D|+;+IzaB9SGntb*N#jpjDR{jSyyEy;I&=|`K`|Aop8lrG`KAF!V)5qcWU20*VCj2Lj784b5~RD- zlF{V{lcxL;Frf={KZfLJeFEtdXyXmlBYr{Q%h1d;{m`j_^=lPoDZsb3HliBj#6!@z zQ_tV#QiCdkA*q5{5H2WwFQ(9Z5+&~)8E162y_7fHkZ`DqWQZX$0Nse^7)By-0L`6F z_p2ShdY);r2<57eAmi2`bF)v;Py$B7X ziU~S%1Y{)w2m3oBdIroRB)K!qGIu8DcNV zr6KjwteSoHR!i=~pP$qgdE7B2U0RAa9ga{UB%Y)dukjEVbQ-F?q(q*ts6Q@3BS64Y z87>B9GTnfXewow$0%gh^2oB#e7LK5`M)k14oYqVRTch^L>y|88YFRcH%I z!PhJ?;03^7r@jn{aKH}g4gd|R$)Lb*^mU2gH;W*-MuXZce~gI%p(GRhn;HmS3DLrc zd;Av{QlbNBSujkm%ap;pLH$4YkDF0FI2Hz&5c%4#+K8`nUj zCPc0N@kj;lA>{J
    hua1e@+R67qid-A||Z{muXtG;OwdP21RXff{;02*Hal=*!Q zXyZ-6Zp;IFJKy4h5xc{13zWoj}!fcMtzE?07b>Z?%zVY=IU{=wn)B_s z_(**uT~XF|g5!V#jv0)H`sXx;99MGyoJBZ!z+V9sVJB<={Ol zB*9O;7ax-yF4zmD46c2`Q(*e&!BT+;(o^k&i|~u;-|Ffpzx}vL?*OoZWXLcFm2r~U zb>cu^h0~C#v^v?}_IM0ExisOrRV z5K>l@Wor;4q1Ga+AlV98B|tr6vgvYXkTp5$95bXNIR?D&1cw8;E)EPNNURZtrt!X0 zkxGeBkodP7@_p(Xll&e^$#dAjwM}P#3HLY=aMC zv>gs` z-~n)a)*F{#K2N$!M)0(mQ>fi)jzyV@65!mc(8SpdLk@xIUIGB^NTb}I{V8HMnM3HC z>s*#^pt5!JfjT|C`2pLfH0@BEcBGdSkei#_U#{d=aWIw;zb@So` z*4TxrAwHqz2FP}F)W_;*Yr3yk((gUfhms6k{;=M?cMABsr_V!s`uQnG@3bPghv@%Q>#4=ntC95YC-$cX`weG`TJIh54KVCkm;C^5kR{Mtu(2aCWVGTVJ-BUI zF8@veTx)BUpoIzSb(kPuk)p&{s`{6~MAs_WQwHO6Mjjp6C* zb?;194FKRjPBimuEwvR2buM`g9Tw0|hu-pJO8>rDn|8VWvxPnl1b=Gr9FSgn(NX74 zm?s(cR{>9>WDZd|@{KZm*?u3O0~%DD_SfC0F$ZL&^Ai>U?Y32g2_WvJu|^1V2P23D0~a%l|2W1ynoEC<_^{1zJ zeT!_V03`*~Cb8B|!8Xb}-e!&&(Xwhv*rYDc>{O2_{3VnjIuCeyxqf-wd9-b1`r0c( zIsMMtQSxcRPtTKBrT?8CSr9jPBCB*sp#LWojUC>FwfTh{;SA5t_-71k^Tv8eQ6mV6)=`_9w<0{w?M8+GzUH>VI)HHr@D|F^8cwSr4vZ`f%E+&IM1QVH$6spE}3%v-SD%L<# zp%W=_jV$Bcx!2vg^H;z=Ti?nTRXg%SO@g&2j(7OympILYBl{4!37~-@x8}n+cffb`1LVO*&}qw9lgbSw zqK8vNAJ%&4%-q~2_lvUc9&?{kpV?WZPu|U^FWFh$p2PY7{<846Ml$ZfO0J)!cCkE1@R~;Lo_g)zl@xxu$sGqB z3XDf9Zi32}{!4d&DS8M}SwoI_B0P4mI$>InVwgeALoeK$|6 zN<(Q#-P6co6pCux8!)x>*Et{LryO}Le?7~;9{T2^BkDCO;!v#!JCe^pKpI!G5dx*$ z^O@gx)F;^-b&&8jCQ6ebUMV$M7P$CW_T>6%{s79X!xW@~*=_r1`PVL&$v#butVP$B zVm7LYsd>T{1}e2|Gba+XOi60jl)vgV0ghLegrbkDazL{TcO_mNSStqr;FdbefF>&Y z-@KXby`~g*$H(@AM?Y#5vLa$%)#1h!Mt`EHQW53+3??^Ne77;2*-fgDy=(Ex1$={# z5L-CC+aB4#rAZwcFSm-Zwxe^yfAIK#X{j!QbD}75Jvz4OAMBe6U8V3hSy;?KOn*QZ z8%Yp^ob_(?CmrKIDG`dFwyz!>T>?S7T?Ct7OSDa|dU5OQfFF7BY9V(7EV}i`niNzQN!5^UR@*(>#Ku~^%bgrfy7*8DIOCZJU1!L(W0h292 zLpz52LH!^28jmSaqN_*j7>O=UULQoqu7M%J#d00eMwqx%YT7LKn=v=wg{W#~j?sz- zpWyYWh|Z8YIMh3CE@xRbgW!7wA5lQ6$<2PVn%8~_esrs-LCr1+@5l>D>qZbs*bi>f zj7Qd)qZ2xiJw6uI>klB**=Q`Kg{*>ofvuGw?A@_s~?sE0D2|C??;qUHx}qK4MVf z^e<}UDm=48<_v_o33$Qo9}=VUF8C$5gId<;xB1dcNJtSy-<#m{RbOQGRyhSjZ6dgO zOimqauYpbBFD*}9{QA^n33PdH0X(&=hF#}p2hP2}y9*B@((3}^Q?q7oa7|?sVGEcA zfe6snWF|@#$gC3g2SA|k&mNi?(Lca8Zu431Cpfvr%5)oyUsac;!I01FfIKW2#0rdq zAa11s_=IF28XEa7>(@r^#=@26e2Z*)jS<{ZCGcbG4kv;n(OF?&bca}bhc;G7LG>hm z9k{?*M!sn1)qshWvkFL}>TS>KgoV-wAhP4>?@%85B&*U!`%Af^xH|sN;-7)6>QP6! zBupkwybYga4TV#J<%it8CuI1;ee4bW;8`U~!s;;yHa0mPO#xTFWVdGz*;v?20940WoUTX4>ncIl>K;h`s6h|%5`kF^QkiWfd+0?dPP{nQ*7rA6u*IFA zH5y(6_G&I@Fmn*I1k$?gMd{fW{)S9j@9Z!V9j*Ym^95untd_6~=PL>yN;&{YXN_&h zm}K=RDVEI(=xY^Q%Qy}4!!F)l;roozLSO2MLX)HAc!iNPDC|hjQsvZ|hUbxihu)fw zPb(K9UWC(FqqnYPw zxQo{VjIFBEv?c_E2kRyTG7vp8^dm@kzC(1pDX6Pe_+bIUQ$Q)S?)>@3k~=*F$0IjX zNqym?9$I-Y`06;0>%@Koqb2FupvkrH;2Fi>Di1T!dgqC8O+ zO<_cx0HJu~3#!{f4i%VBSFizMk)nNlR6(sqT56+ZH(1kbg1`L1@`97U++qUCUSGG6 zn{t|ytt%^`mrnFxTl;hszmnDpImsdA{Gn)H-U=k$Js{H9`iQlDl!RYfCZpj2D;Wzv zoht+ltRx9o2uXN`I;?YCnl1h~vk;68$#_hGp2ZZ{^&f0w90zN<1Vve`xo{p0Wqs>` zf(bwf$Fh!jVwE@}l#&4TlxsEZwEuKpjnn7{=dcwR3pDaOcN&ieYXqk`nogqM3?GAQ z(GzfWZh!(GI<$z_g#Vt}|3fHG1J>PQXXwOFspVpWr)VAkOPz;+^$V#3f*rDlS${y# z7{hCIwW@UGTZj#n7thD@l^6rTgMKP9q?x-1ZlyI7_Uc*C-Kk;|G&ET9C@WR>fyRLz zBtk05Z$yM>xv9HASp$`uI1>gyf>y`}5MckVK@_b~ zGpTZ%^oZHZ7B5HGyP*its{(3OQB+D32TH0r(uLB{@{?)}Q67}CylVQ`W(ZBH%O6=% zWSHh$Ap;`&ttrl55kB5(8oGSDb;}<^PQ$!0=Qv>Y$g-7YFhXCsc9EXO8kG>XSJEkG z?UZ}o@LcdFxen4tx{84vywNg0X;AU{H;yJ4{d9$>1h+uc_m@#pDqU;n3Ibo`D)6CC zUYB<~@Hgv#)GHqX57-v9mzR?^DZ;OtU!KZkEfn^it;cRPpIr0I`-!HC{)Lm6U^~xi z2MyPykZ#n9rifna$3?a3CAMB=Km2AyGKDhlqb54XJR5GKzwS@3@JnG06RHiFS|qhj z-mQKB#dZkY0?h};P9OIafGXx~glLik*r$nRN1QfGNOi;JCvfpKi(l%;)wMR`-zUvw zEl1VUpNxBzQ)TesJ}DVZIjY71$Y()v<`sjomhr9&(YmJ!^n5UK$v|u?f4L7b?)B@q zff~$o4mo`zp@bl#OGARb3=l%ae)mYIcQH`S`~GZQnA+=E9L$Rl)`?!cP)e}{dI(y& zKf9zrCX-Z-{?>-nz1^xx<~2=f4IvDgIxVH$Yh{{cjf4!?!*JN)cn9WSIm71{$GYM4OwJJVW^_`o?VMbZ^2Yu6`0o(wOwCpZvd zfmA?!30JQPH$hOcTSDlWSzP8Kh+{>5AIJ;`Lo%WeGR~AwkU2k}=d40w!+D7ciKY}! zIeY%IzWB$w4U7hr{Eq^uPb=BgebkN+2b0`)AGwz8sIFfDbq$hpct*o7Cci66P#&?D zC^v+hz%(HCcAfVXZbPN(!jY!uUR`JA(@VqB*FH5W^fL+MoG3fu-oF!L5MzVf;_X|8 z*Q%nxx%^}b@Vobmhv7!Mi_-djVX%qH#GU|tn>a=nuhBCr?95io4|V5~^=<-V*EC~1 zL8v*b$$vKYWY6*&9!_&W(h46#JbnD)!RP_B=0Sh_OY!ZiWx+IO{8=Zm4Sa1e2MyNk ze7)yw`m~#k5?%X{{(#&@djE3s6R+de=&MRh+EgWcemzgUosFixUoHik@S{=mkQ!T< zb0>GjJAiYc?U0pr?gRL5(n;@6fU!vxM~Tu%UFAe@2GGIo<&y2)1G-Y73|gJ=X!(av zl8>Le)?^Qt@%)Y;75-3u_oxxXpX><+5U}9e$Ho>&_~S}GM5TUB;y>iBqlkVP!NVwp|$0LY0P zmR(MCHj%Zr)$~hEf_R;qLxeOM%F<@z_ei1yn7R_)&%7xVwt44(C43+*EatT{U*uUCl&f<%dZS5p>WIO6!YKFwL ziO`mn+tT{FB@ksT;RFc3F~Gck`Ds2aaaR8^;rHfp^d*q0^8F`Y$`fmO%kos$ z1lSYFIou?ks4(nsA|k#C<;#oT5vj!}-Bn?yVVb{Q9C0A%*d$&=urlp^Kt^+$kg`QN zQNHS#WGQDSotwh`C2AXUygc?52v>|j7Vc%+&{-Ocaxo_aR}G+d$COm< zR15M(fCVoNIlr7Ay;tp!BRYg!O!suM#P%}xmRuB8+{zcb{@@t7lyAJwuASh?i&yi= znyXz?2*|a(5{{lmMP#L?-?8N$5Lt5Q&{|d$*@@w5_FxJU3+Wr$BCU&xXARxjUYmKP zoJ;*b3%ter?~TBPCTiJQvbWWqDjc_Kb+|{BGLGugQE7GP{nNp{kdmH%TG_4REf|?C z>6VZ-84s)Or@IQVY_@MjGakR(`D_|yH2Npw{-FC8pxPvHzfJOz2t>`~>Ay*`Czz0_ zQuk|AO}{jqXy2H)fr`sRYk1dk$nXhg-qmAY|G=jg>ki_Ne3g}2qPp3!sWcH~o_xB? zr&iZbPAl-hkI0}RdvPh<>8crM!33lxv~zVDYOTYq$@6SYmmC&7ea9g8*TwTF@gxaX z@uR~~HH1vHb1%Ll9kw`J#cN6 zDJ#LrtNfgUx&7O}H;SymXh+PvPT=*899cs-awho=9KYxM_9RafArvwACAbm9If>cF zN7AHOY7q^{S{LTVb351h& z6|~=h0Bm<`^fI2=J~d5NjN)gapIWKp9@HM&%oT8`6EQ?jYAB5=X>ZKsMVm7QIqKmg zg7%N14&2!o^7JW^D_jM$`DgQK(ipwy7K_byG}>;{34tO9M# zCrk1GW6!Vr(~!?uSnj>PVmdwqw!`uq54=S0LBZXWK>^D#sZy#!hf#_LEy$k6>nBf5 z`0#LAuzA#r3O>zppI<&8eZR;yCb_#JD#x&7H1#&&T6X?vGV>qkF4h=Iar7xIEzyFUcOBzVMW zKb)96v&2q+R|WEI-B-(a@n&-fQ1Yt=1&6q9XE26nzeZm%*uy zQ(9*vV6r+2L3)c-!%dLi$Rf~@H!&BV9k3|-7hIP^pCpI(>V8>kKmi{ zbfu>P5Lb^Iz-gh&m~yb~_@R*FIx#=GHM{LLp)Y8&A6#Rzy4lO5>-eD=y z7do5nxXY=R)|%gJUw6{S*5lbfIuwo;q1Ewa`O{6+r)8G9I)f36Kg1YsroZsWPtZ0# zhe15~x9B1T>aU0qTKL;UzfQL{EWO3sJ~HN5?LDdR{8aAGesB1#eqtd<#v}L|pO<@C z=eWI9N|*cZt#F*R2;OX24wjGOEwtO+RN zd8V1m38{#rVMKzH;2KH=1{`s~rQguJ9pwtn%wI?nY^&uUZgHN-3rUycWHd|Laq=T)XBiBwH~ z`RKaTG8`>e;e_FAI)i{+Cg#)#3nEjsP2TjWiz8bL`EDLnvs2a3e4HLs)_>cBnyKS= z(jumhOxYgy28pRW^;mmk|MNk&6D?bPo42}jN-hCLiz*{p8Esu#-_N2uC9EuS~2Sy=havlHQ{?Ob@`;u z653K8y)U=+WKR*4^|HmnwVJ43c;L)Ayc)Yhp>W<@e*ZH``YYpf1RHTZ7Nu!;ZE*{YrkKvX)scrj1(;P<@ zwny;;5mUcS-!b~>GtV+PxC)u&B$>?ncJ3%P%Jh4+OlvaJ2S3Z?TIRhf3s~71ojo46 z6fTQ-L=rIQcjLSWs$&oBuYx;pXZ~4?@2FXdQT}5*IpiEU$K;TLWFHgvyVed9Zf@Yw zZXYRAj;tJEn;evvtIItz(3n(UD;(4J)Mmz5$6$rb)KgSF<@YGp%A>_cJv3R>E|fT2 zDK1_1yN~=ultzdoE$T zcIuqUDH1=>+eyWCwARM1SEG<7G6--Ra-XVyA>p;<{|aAYY%i6r`1G7n`dn>m|GgkQ zxA&gE{~|U#0M}enr5;xK&PIK|3N@s{dLq|NwAoa$u;WkYQ-}VxDs`QFooWBC9??a2 z)!W&1-jwkDS4$;VYgG3DCPxpBUxxyjgE_U_M409}^4-TQ5&zYk7zKL=n$K6=!m(G{ zvhA+hZU)H1PiZqyqIT%~8C!3jP@-oXaQ^(c~Fwf*5XZ&;>{r0RvjY5xK8bnQ~ih z>Zv~;W$9H{ebxMX3SlIMQO>cf|D3nt?;xFp_GGqs4K*2`=$-*J4Vkqk0OT(B@GenW zE&;zh{8*Dz^qWpgHrdt2$D~1@y4QB=iYC3qiPs&d1XlWkDNpylEI(lS!GTaxfBD%n zxcF`1V0!R#M_`$;MTJF41?lwO$W7Q8HK$`(tHhx_!0Q@&5W)rJxVZvCA;u70`b_SN zI4N|s7Ka@{^rc>Tyhfg!zyeT*%=<>bFb^cA^n7X1K^eJC(PP+!cOt!D}6877F}qQs}A#TobkKLY4`@CZA0znMr2ZBdtpiYMzEN` zPh@Fbt=^r*gmP)^31~jd|8yO>b~{ImTsb24NAW*j4}Ro7_^*@vl4J3*m`T=~>5V3X zZJ~jvABMp;8-7VKA7BP^qJ~)r$R-eJbk&N=DaL_+eVySOk791VZZ`ThUCewC*{e>D zPuFpLwsg?Ztoa73SYSj}>sRn;tSwq~W?L#Bnx$(l8V-wr_xZiF`^ZDA4X{85HPf(K zi`U$&f4*$rfHmp$gGQdCR^`;16?(ZT9f?Wq!=&ML=UtrIjY3_GV!=$+4K2JQH-HN& zd@_`dG9B40zy+CTmO;p|j}YPuT2BW!9>}(mH4LVsO%A0L1M7I=Ykpu=G@B*^W_6H2 z8$tpC=NjaktvK>^;%E3?dw~C|B+xS8%Vip_yj5P5sd~C3^E`hBO@?6mWDdZ4?Cyu4 zpkT#ivkDPe2UG=;4EaVK09B2#Tt19Oi49Rx+0L%x7@_zlWU2=ty4nQltPLAA$-jDU zDuGk@2^W>GzboL;Yt}jlw;vVi;Pen6kaF}MFyi>A&;tN7*cCC3)2Bt1!&rB_3+)0p zU5lIqWC$cA3xBG=8DQ3^!F~5!3l}n%|4lq2gv&c2<``7Jp+Q^~_a44uQeCj{CwN#) zkPfqJ&E+eD?5n1&-UxOvMD*m=Gm?L=_&vq7e)-oq~{6aY1c**pebpqsf4A zTdfeW;Ot`)P^~cN8Z*t^Yq?w_tz8%gJ+ejQ8eJ)34(x>5OqaWgRJf0ew7C-MixfKF zloGcdJCss<+@M}>?)9d)FUdr``H8OrE(Syp>K*p!dx;Wz4SgQw5SpGxmL$QN~EUBAhYeW}XFF_Bwl<6ZDSSpdNId%-Gg zuyPCdJ_Uy$e@qg~+m}a* z_gI7XTh}Xjg-Uy-=zr41-@6Z!M1}+u7~9%QgWck|OWi@nK)S0TD21w-bGSF?4myuR z{&F#8pSl8mxV{H**@IzGK4p&WA}JBhabm+}#c&`Vu};Qko0!9Tjs z#rO(UA_J5>(gKq8DbfNpo{CL7PAwg_dVTg{MQC%t$b`ybIIZp|!R1?=mT(uw)X zdm)v$rh*^CHg#Uqb}eP{=00x?kDKaxE1NHxs~Pk7+f@4i-xWYec8$-Y1|Rvm={uqz z#ImgE?t<1!Yw8`tbXh2t5)_K0sPsJ2d3loLuW|u#15X{+;D7$X7BrENv-+s7HpiiR zmwrdE-P{b&hdLVu?T0Vkj?#N@7wM#eQ;L+=T!}^;X)zyK=oGU z9q_`xp+Kx{RXu*CtYnL@0|gAf?|o~vynd5@chfBl1abj&dL;t{HR^}g-$`vk(=%26 zQ*$Ok8|I7KBn~QOazrLoy$_dVz|pq`5MmyP+xVs9lwK0|{mZ+bbe_tpvCM+8fbw@a z)c|JmB1;?pY|@q-z_-t1i$s{vasPml&;emN_Ck6fR+|0%H@BeW1e;$CiH6lzO#`iD zP6G+{5EZ?Dj2SaB*r%2=-wxtkygp#v9*}H4!YkjTDyJZbDlgJPF4ftycAwT0i!A`4 zWKW%7MQatJim|NHW?yP6WMhR0c+GqtS*p$2X;adUFN(QT9|*y>M{nI!3OU-8bAgtj zv+kCjeOd_b|j2AvuDOd!_qVK`jT?V|>BioCQ&+J%z z{ew^EWVE1ghrhlk+(ckBc)Ak#oLhZ>#E5s1C%phKkSn=3QDYYlDOBVd4Tv#0wZ zhOox}se^01rE^DCKc`$afqj}%(@?rqJDMCpkk|1&-89}}^LMpx0NR3ZyNk6%B#!bk0| zz*-3hwo`&b%2`CCAjPPh3F(41{C~a!Q3A1|KENr|yc~NpnqT(lKO3;PHBkQbK2}^+ zUxR6~hVhLpBR^B`r5~0_IadDCj&}jHi$=Az?=8AjEci9E|HqFADHeDGi|$3LNnT>> zm-5bCASb81y0UP7&mcMTmJZQAj2ZU>!N2GMr8p3YUDVrG76U>yn2-=9#lo^16C`xi z9{w900noexUH-YYhTIXr(a(*bbD_y2|4(7_&qq#$Xa~J3hI*)g%BE53g@LFC`fR>2 znM>3{>whm3^mip~=+B9pcvsB-fB*2`pP=N0h=eAS>56#TqS^cGug2}~`BfC=oyxZ? zNpbnpd9`od-6WGG-7@#fe{1tU2Ip8}uwBzTKT}OTq85)NK+I8Vipyx}-O$w=G`Grzeez+TY85i3D>t8U@ zo-xCOvhNS7y5}F0%%U615+0x%WKsqTWCfDNgC-Py{%>r>XUHi-CL{Bj4`#arU;+V< zMxRf73)aE{R<7_vByYSdc#HO*V73spa-^}WaRx|Geqf4E(E<|x{uiWHs>?Nx@3tz` zAF2BZRDuwXu2<>@yg_^hgE%r1cQ)?x=7giU|r4_?0;N@w)rC z9bhxpz<=t34)&6kYANG4Z_@%Ha{;i)y7U9qSpRRHI;d4i0xn@*5bc0jtAqpvgL6-MdWf{;MG+z~e1z`CtzOzGGHBueZB$K%&yrzc&f6oobV2}iM zQmB}E0PsTQ&~W7dRP&W!(xZt)LeK)CU1}RS!blBdq_!ww=i0VJYY2?A6sCEai zBPYKb%Uj(=2e%22V-Rni2n9cshFbIx=QgISEoP=bO1!@j1K_SxUI)C5DtdV!y`U>Y zJSzU|b9;0ccUNmGL-PCO2LfPaImRkYWF5fgOuD;S7qgCJ);?JI5hD+{e{`9-vW7hi zNYxDJa7s|jGn)vQ-1*4#1SrV}DK^s@wDl!wfykg3XBCetxpWA))dI}P01&Ng4<@t) zK*x<8aquH@=icA@fIL%Ra&)94eukjeY*L}~8P57uM!!P3sEG0e6sV=HWkTT4j=~5x`fv_xi}68!PX9!gd87f0 zA>Rz+IjWpFD|#@OKx>^+K2ToNe;l0vGwN5#8goA;?P{m+=TB3B7Z|eJHas*xM*sW) zQZdY(fd-32zy;3&b!5l981P4>C}x4c2p=|N-tz6|NLo2xqsALvbZugDjy@NB!7Hn~ zf!qk<(&?4u*;FBQD0vWMu&3KnO_0&oeFKDoTlVJDlpQ_1@f?CK{?QeKNUzNx&%Wp ztpLTFs?zE)(Bs8j2gdjTkajVrj~zS=i2WFP*}7fux;dhv4z-rAu0hUB&@aRTxn`HA zz>6}dB?~c?4nT^?`0aQV)Yo%4GAQc-M*7RtIT+D|ip6M;>pm_}I*lu@0ns@Z6wl<2 zg_ddv=d+p_w^bZgXo7-HeF|?~BjCFT)bXJu2z$$6O4qx<*mFg;mzTmwwAwt#I z_iV`q&ko^WdFGqTYw%Y8fLW6VlBY}qez+rW{L(GutPUX4ljZh?(_Jhf^pLnyAyz&i z%qJAmj*Gm$#C*cR6aguBU#ETh<>^+N)L6ck6`Lurbm+SpTs_C5SH9ThdjcB250j72 zf`0CTZ`cJ@e*snbJB<<*;+^OSnrcSykMMiJOt_wxrU`y38ng&AAvQvnf_DYnr;txF z+!)>n0^$;Z&MOHl4ojofufD{4y+jklCC`{>;8Rlm3ctW^g)1Wk?;w&--g>h45?vzW z^Ct*b=9@d+xJh+C`Z3tVQPgRx1GJCxI?{7L%6&f^I`yuKV+B(Hz)EY$%273Y=-80? z^*Ag0`a#{w?`xWn&SW3#+Gb7L1&?#x{0yl$<}FAzE`*eT&{uxSJ-JI`eQnIiGA9~S zL%lWN-~nbTm*8QrDu&+{K3e*r`c-g?oQUq zep~YVn=+KdV>TfW>p>=+EPde#PKbu9#drx3*B?fi85mJmvq{&>^tz;LCW_ZmdLNdM zx_$VWnK2NVV4(-J@M9??7o`_-cG;+Clj$QLZ$G3BTVppOmPb~k@&`delf4*Bj#L#6 zIFr9$8;F1SE^*^P8&~i4AdfKtMkHqSFPR@vMzIJVAcBm91r^*kk#s7RUif zuh5!zbBe#X)aEuV@y`MWy^B@BTTR=L6$)0qio>juWb4A8GN(8`UeNHl?4wo`b8*+D zd~O4G#ikB_dOu|y1T0y$W3eE@z@4v~6E!ddgF{{rwK1dJw-^~O&u)JjWOmGj&^1z~ zzZ2wSQchgkj0B3hrthE31hYfAOY66atUd3g&eDbpv@eqGZg<)>797;Awe^})a8P~v zZ8x-J(t{Rpw4pTCCIXr*Ey9QcxAXT;8SmXe!99>TXz(qC3xOaLvV*45_f37OMmoES z=dZe@#n_m5;o2k2&zL*H0>X)WX|Uq#^oA`5c5Bk?05z3$&$vI!e0lcM&uqhd4AIkTf@h%$6$MoA>*B`o4Js`svpV zSYrMNP{*H#QXxxf0s(*JBbkU_-NJ;iT(OjE7fZo``t(`Zfuj*pCaJ_hWu zK9aLG0^k@ln-z`0h|DYUOV5|a4smSfN$^CHc;kn%?}D$j7V5M)N{<8cbIVO698+Cy ztBkmNbH$~GYenRf1<}!J<5m3*OTzt1!o*0r?<=}eytv8V-bJn%ynr#v9m_?2R|`(S zd-vh#$TKP8pzcTB%T)lbXb2irrnDAJ6yEv#-H(!2<`LE+lrXc})&iNGoV1gsWJH;C zkd?zew2dTDfYK(Rcp@Hh7x&VjvyPC@Q9^AB7!BqD-TUDY=|bPB_+Kw!ata zGuf8zB1m6efPMP<=;z$}mQQti-pL&Mn!nVZL+8dWDMdyev8 zh3sldOs@&KJL)KRwo3?Q>5eUB$Mo9>1~?vI3-Z6uw=%Kla7=@1low=u=F)uTp)^I8 z``p_yte6}(ZK(6u$A?x2F$QIb7N0IHdC57BLNHkH5Z1Ulkf9<&Q#qCMRwd$4AkpG|&Y`NP@Q)|T8=9s8#3&A6*I zXU0qIt?+A#W9JVV{;R-6Zab6)!a9DY+Dwg-C6FrXSS_lG5jjg0Brsj@vB(Uk21IY|zL^WMk;UvEd8|tz53fBM6cWBQUn`N_98=#UvHem#H7E~l<5)mL zj<>wMY2qBkoXj4M3>Wj|TkP$s5`Oa|fCn|22%r3cA^c2qc^z-!(Vi}=GG=*!zb<~0 z;dZ+8E&2L!Cq1@?aLF=y30;ib>Em4frFG-WvxFV@skL*)C6u)K_2 zyRZ6(>yIuGu+zhl6W~@n^nTyEYF5(9q)WKdod(6Uj+FaE*T{;w`>B0CREaxi>gSqt zF$ogpS=G*e`(8LwaT1JtjZ=bYaG3gVzgIGG-1zds=)>X??T|U%M5(GopkKlxLhcv7 zPs1@o-gXCHU0_!h9=-fF-=Gk?+`+HqaojY($`D$#t^zmJ&BqF3AQ z;i|L4`aGNcF~z9i+q9a67qsw)sCa1)p~@sJ?qOdf#BDz1O?&!Hfqa&ZuQsr^Y-8xd zi+TZS!RI9RKQ7eEQ{1MVrH{DVU3P)0PTmrRO*%D?rs()G`N*&7cqv2tZQ?BIMUGB~ zyLiPMGIl!DE27J~ckSJWdPmB)?X%gxw79*Wu`fLNFwME^jesDQNFe{VM`BoBz_nvN zc^HBthtIk;$+@*cS-;W|(%E31(MK#Y)Kj9)4Sq$ln`H zR6-7?3PdWVZ&MS9Uw zD_`EH-B;I}yfW!{sz%}|j*UbDS>;mvgTHyJbe{IUFn@Ssw`U+r=h|7yfl4s+RQP%M zhg{tKtptT6MQ66l?*uwqQ3AVoUkSvpw=rdoBhEsLDcyyfwqhAfc6H$)FZ@)wJ_mLI zi8pRxs^2#-O-iD$;m>FMn#NgN3PgKQh9Qa?HROY%s`uS9_VJbY>T2urFdVj{HW9RK zWtIdC!Ol>Ik&(~MD_@bt%&y- zKk0<={9(M<48OxhW9K_FrnY|huwadT+v9)3Jb-wd8>92~-VLlE-~N!?H5wJW-4UMp zu$kS89faZE%PNv%=7@8SykGfjob40QJq|=YNzX8%(NV38Dql+^tk*Z|n{+%DvWc2v zc0{?K&9yW$E{tON<`CL3yu;Bsj0DAnb(jN*BC%d=k6pWJj`z8!XQh&zy|S0SH;MgX z4@d94`BW@JQZG1A)K z{J%EX5H-t3282`kA)vEy5=9ufGfuxjwzUi!3utmP5b)|P;H6eE<4>(p5%8KJnVft} zyCvDl8EZE8x`RD?L2$8@c>2XYx%I9&9SJsK@ORiOSJpOu+cPXi<|rHwC!2neDSJnlOip=)fDUyXfJvw#e4hz6~&Bm(UY`!-m5FFu-^*IoZPxqi0r)J?KTsum#zWtUXs}I)~oNhZ zcpm%^H#=`cAhXn1FYCVdhD@t}3D8M@B!)=~CZ%>Lm8` zVR2)LQsfAikaUIOxj!XJ)oe%XqpWt!as6DGZp%G-gquD@xo${Ve|#@u+Vg^X9pP1q zdkPsjV?cM5;o$V+cH|Gm>s7qH0oOjU?Z(KzXU5(cTFcQtlVV`nW34=}8^94{a@5>d9DZgv9@oh-F5MljfThRPF1;CQKViLpy^J9y%wy!K|RxulRE|KPzE{4sa!xcrXj0ZJb} zr@1!{p??kuF&4UE1-jG;VMV$RYz9Z<23`FyG~(AA^t25R{m2l7r!6F(_PWr6-19dU zk5gsQmRdGm-_HK9LC>5m&#)r)9Dj9LeK2oIbw9gGVI9okB+!834yYD66qKl*j z66cahixA5R9|Hg`uJ!)&hf9~Al<^)`ae5}Y^k(3nQx6%Sb3J%wNU@(wc3|FhH}Yp`T-+{et<emx8A>9#Mdpf zm7#&N!H-D$LAvQE6GL$X!@c=uYqIZ;J$)Ni=ugFy=7HpE? zW3AzzQqR@4x6F9T#Y3+6(Z(Y+p*-V!mV3KCF5DhS=k|uyH|hPSm;;?e6#ct*2QX zskM>%zL?D03vp=VE@?K%zbN-I6T@2?-(T_gVHgK>P-TtDb#pi)j;(%i9dd}{UANQC zHQh2_BSs%TabeMY_=4)bIeM@fOUHTWfwd+w$9; z;_|$zQ+m6Xv7!vy*CkhZ+kQAvRL1aCy~Q+31_z7!!HZ^!Uz+UAN5U%~4{yHP7CH_X zPCGZ*iyi)Zi9?RMO}77t{ancpA{w5%u~G2f9K_gADnuZOD`iVqW1{1SssTf zyb6}0(mOth(4fc2=IUye`K~%gv7amBRQ>S{{qG3M4SOS&E%i+b>l%fkuW;DGy(NMI zx{*Xf%GBx-az0^P$Hc?;-%k-p>J(iLSW`Lr_#02{(q6?Mm9o4p9t$IayT1xQaos$? zE7RQj8DjEfDO9wkxnZ@WSR_3QGpvz6nywu@cI0ybiKVwyBo zl~3Jl{$X2_EX`qQwUfPWCoYkd%tlQC&vqYk`plshDIlM+9ys-uJfdcYFOy`_JCkz| zh~d9y{MIDw$m!6G2Ky+N)N6zVe-3uL ze^=niGe_HY!B9nh6HbDAxTBxJEzq5krhpg9bc-w-dXBU?jqj!Ge2#(2mpYS**eOm;G@@HoX&cLAQ$c})U&S{esXP{qWejsOT z{ZjB)xAbV{i7lHBa}W@TPdcU z2gGXqmwWG;vJDn~VyOKFzl8qE2tnz|*jf%L&f`SRLanaXa%$!^wpVVHE8*S5Kwzo6GheH`I;+8q2ZjGPA{xsU$)^w0uKB43n5QOUr^EH#=PW>~S*)^%9lvHYbI|A&C3QQ}18 zEvRu#|IJTz+Jg{&&UTq znuSK`h4=$C5^N8ApAr}#!vubReX|14?7jHB6L%rxh(jQKT=1K{o~u~F#AR2$2@J>sW>C8yp>!k>Uv`*L}cdeC792(KB9 zw;m>dqp@CdCaBP7EVq90gu(g=E(M@U&WXLE_vks(pX|!qC4Vr-!#WH<_rGP_WJk<) zb(j+~3Az3Sw`n224VZH^GhaEsy|cS>v)%ah{Prs4&cGddM;^S;hwZ!b{?RowCipRW zNtXC!-t9z&wbVG>^Q!RA^{>0Vg%{3BeZk9IJ*3mKw;g?tbkNw!y+uy2#j2u(k$DYU zw_TIF2{uO^E;3TBD*UIP{C`Znby!th)b%ZhaOm!mP6ectZb4G%kQSt+8$`OhrKG!& z?(XhJy1U_9=YF2|eZTAS5A%4=K6|e<*PLVg#tMqrDgYDJZHtWi!Y#Xwm;)3z{&Nu+ zmGp3G{%C7HBgH)~Z+yoZ^IA0=RDzcMLT1*!Mah2wq`6l3> z)BsYh1qSDCu_8L6~

    #nE32RufBV2PQ6P=%xuJYrona|7r}{91+v z2yPv?+S$;>yg9V;Bt{8t<9XIixdw_Ov|O^ejeWKfC}fdT?q}E;V4sj$cb=dEL-E>Q z6KO}x1V`=4^I`V$X4Q^$+e7mqPGSKWga9EH?B1dDy>Gp-^B!O@25X3qUk#0aB?doe9OG0$N{VLH%GBF`p$EOiqA#ebzSh zl?<32k9xuJ)yX6voF7z9#m4&=@FZI!9p%0c~98YbJ*MkJSJdYx;w(#l;v z%F*{=OJKxOWz`YVmvEW_j3smEH1|h9YdZFO0QNaRn9d<&Bv1vJ)Sn;nRKYhXN3qH< z2GtC(lPb?7V6bALOS*W!m{~|W<$8XYn+GTwqs*|RV)L+UwX^czIQPNG-*U0|(2b{& zNa^Anst$4hqwaYqJ)M1!Rdj4B*5j+(&x0tdm~X$@Yrg=0jE)m$X#>RVzIcu{`NA0w z+@lT)M#)3wji;vZW6MXBLN-B6{3bvqIAew1$FS^D=tg~RVH`Ar)WBxkIU-3j^9q9kKdG&9xm>&)=G zcVb7inS{$Du#f^X=JXf%B8*Bjue}QxlctZ~gFLPScKMI>z@J>zY%`@s0P9q&=EM26 zQ$xC3tt2%!_?51K`tnRv(RRyboaO!dS}mOZy0tj#BMt6Cu2bn5bP z_gU-W{k52MjAVrw`(-nHjB$nAPMdfk-}V>7j=D%yYPR~^jf)hwOJxf>J%h^+$1<(h zG~F-f5wfnQcVdG)SNLbAu2;H=lMm~R@4{G<4baZ(AT>`_*YmNJM<#)&vei#BEm!Jn z-^{S%j#^MsPaNEy@0m*FDHTw_K4MdkrUP=})mpE^Js&b&$Rl6Xb6?&%JlM9S;qseIaxnoZc6pq*LG?k zi+5Vh$aXj}&4x#j4Y7<%eID|>urjFFG%rsTrjI}D9`m)i*f=QUyqL_ham3dCda27i zxC+bb7Y0Gw$MnbCzaIjr3R4#I%}X{z$?bgvTW?%}G{*c-AgyU5ENLV%3<~4XK8)KN zn=#vTI*Nyw`-mA+k=Yfl1#<(>Z!urq?Z90$k^-=f+bakR^<3o0L@~WIi-kiEDXyc5 zK{zJR)l_!jvFkK+4I<3=VINUw#hrSbuDt9xIrGr~Yzwb7#D7~Ee=_^sytHFj9zPzb z$s6D23O-78E45kubZ|xqnmzz&Ed@DjzRf<5j~nFAtXmq;)t~3*PtQ{6?#WJU^wP2| zdDM|C)xRO^aysjQ24VR!x=~Vtmd$ej7;<-2S=BwQne1pd=!wHfo>@#e+VNl2@%8W@ z;={|jk55~fJS*=<0;$h`$;|4K7EIuB67-R%FnxjQE4>5q@KAZ1MB@HwLrSs~ zq&nPy3P`e8U^9kvnl)dO0CwQ5*%iQgun$eZiVA|HffQIXdXR+$$B3o7R=%Bfyih$Ru~pvKfv>6S z1-7eT2#3f|4B?8$Ax_nm&wv6@^DoK>ARHv$Ffze@m3qzXUBI+EWim-9z}U+p2$a+x zQ=&kAowxi8^eLc%;=c)!dwM5UTw6SyAC5Mmbr9e0?`GV7;WJ<=#p!6L7F><06qAa8g+2K# z&TbB3T9jjf>`!Hj0mQ+*=vk>qxe6rU_z?2PIDuoQ`Aj>9h16ui?p5By1!(Mg$Z(a+ z!$C}F(hDnF6NN6A^hBUCv#6zHMjBXTc$Xu~dst=h-n+mYALiXcxse7-|DZqx|1V=$ zSApTK9dU?hCBzO3hNoOrXDRoz9Xa_;xfq=Z%U9l%)}74cuP*t7h)fvmPb%7T2j?UA z0{bnH3yl9>e-YP3I3;C*gGxFo$wARGIf)#Izg{c>WLkL`kmt_@r-;O3h z>@Vf}DA^zm0vacR}QHf-~Y#;=}D zbLN$o)sb1a`^CoH{EV1hWE9{=>&kn6S)@6-X#)}+h&(}=q#FOXiZ{7O!=xy2y0D!R zKn$*1T&r(Q|3d80ym*fY;^1_FaBZ;z+>XLjtf!ss!Oz_$Ximj3);6PW{pEfOLLZT> z>_%bQgtzffB@Z+qVT5weM2v>?BUm!j*o#$Pc?TIXHl2JuoH8qNa76S9eoHhR5z0+= zrF|@>TLzI~51m8hP6*}JPJ<^aHrP%RRGZ&vq>NyK77|*ToRV}P{b8GLcc}VI4i(EY z#bFf=@8bOp*x2X5@27gH=0FcH^P;Wsn`HxU_Y0G|RNP;SQNsGZX1zZ!0+L^@aY^?* zbAS_d`a+=;jYT;s2U|y|WZ1la75R)U_r6VSz&Mn1`ujGg;u7U37py@jr^-p$ax1+7 zDKtOg5cs4D6bqh@fXRTzu$7wGAzkDw#31ZLHwYov@>1bk-2hs_SH=YL~81{$yu?dDgmNsq&vu3E@SXv%Kue%oM9I$f|= zUz2K7f<3~!MWWjHtcE>=koc<+T(zhma?b+|57PulZ@TFiztFZNhdg;b(6CnLNhnjn zdW-6U9Jg|SnGv*uGcyrVdrft$?mPSs95xfGorY|HcqJ1W|B-=05#;@*<#x#G02Gg` zvz&}1sz;dAY`x$uEq_&}DQLcKu2KxKb>Dp^!}^hnjM<4-Sy2ECKLgckk^FncT9m zur*21jXiZCr}TI00q)_YH$0_rIDs_j}6*N-tEa|T0s&_HXb`Jti|OfZWpm?6O`y5^ zuv(i?VOCXr-IT)gw%N&j-7F4!vhmpsqwTAj5w_f{Dz4toW;x)lsgG=3ML%?aneqf{ zyPnuHDBcY4t0ZJGV32M3A=^}9rSswY4P@Qc{<)Ug;g=F$L?rQ2nc9;G+8(A#J(Sr{ zprT|ni3M=_C)-j*BYa1#EWP;`-0`aK>6Uf(B>H)yd<51`DR6!S(!}j>d;VhW=5-H^ z=^=AvwzDxlw^xMX8$IY;K_wn~?zWGti1qXy3`OIvum$WQ{WUF)gY&eH_e=P|0}RE+ zzmk~b1(V(d5#%>G{kBhqi_uV4udnU$&bf9@k}2Z?kePayU!;t@nH`<@x&nU11yoOd z4n@1j`f_YtIBU-Buxo<@pOC$|`K%K5SwoE{f4wRF49SL~qrK@Oy1Eie*^6`8w)mz> z4ESp1&ED`Ao~B2O2hd4>YYo^hLp!Jc@(Z71shgaiBtVYCrNXTZL?T-Nn6g`K%Za7L z?c7WagYlAN0|0I|W59Md{oD=^W5$-{_*~SyZpL2!hpHQ=o*$6GD8ypT1~?_4J|ouS z{Q2slJP7dw={VpqXfR7Zt^j%UjlgM2*UrLhb#|Zy<*>(f-S=;40$YaT(Pz~`TZ3TiaO5yj$hi2-437g>%L_S z9%BI-?-~Gyld$hFi`$oPW1h+X+;q<%LS%1@iFZ#E;x-{WMBcg;x}z#XE|$K5coY6< zVqNFJr$rmz^M-ag|G*yQ>zK#0^*H~b>bPg^WBdpfH>_Vn$#RBNV{~+_KSd%y%-$bf z_niioeu#JC^H@)*>TLSj2iBZuH*jZU{AM8{*}DwWT=)C54lkY8!%Gn8$a6oX{M7cb zFdC@P8BgV8*2z*j<7pxG&?=f9_p%PF54xxhAC^A`62A{57R-K}Tq8|CA^h4X`gOi1 z88CDOKY%X+n;RaRyB>=hB?uDOXM;{>f#A|^*Mxf&ZmV7Vf{5I)dh0f5whEYk^j63U z^KbUx<-a>-r^H^^i8AJQZvd#R{CmNfC0_Tyk)cnEYI1yi5_!w0xmk^&mRkPC^vsJl z(u-#CH*S-yt)UT}+j)KlOG1QU_Ed*gVvJsa!~=oo@YG~}_$3pdixhLCmAjTCi1F2g zdnSg<*~zs1af|fd5459VkgQ#}S4|=pS-ZBK21F-dvOWZw_R0YSgZAn* z=QZtVA2U^{#n;kq>9jkC@aj%}hahFB4?xQahc~-*Mx);~PkwZwyt-J0#LFvsXs`y%B|%u%4qPa%>S0t$FHM@`BMmZzJKOB*Ki}|`D@2IAp(jx zf^1p4s}FP%I9R;BN4|&)nZjKH^9Re7{{S6j-2fv^9=m^RPGsjsVFef&2?csQe{)+X zjFq?bCV^^2Kk_7K&FeD;uMu#sI_8ahjFH|XILYWxlQtz(B2kqcG3Dc1FaZIzrvT~d zz)J5-p4}KquvexAZa(bmPG5ZjX>(gEWj^NoP9EuoI!6g|_Y2*@uKvaS$PfU-Z955G zU~GbJlw!|<&$lDlrgZ>%BnwU380T7!(HZdI-S7wZ=QIQ9ygJMAs8KO_=1P-SpsSF` z;VDs%S19vJU7pGv=f3sj5#|x$b!S$g+}`eD>B%ViB?=?_$2e`5-SG^8p5hA`Jp+Dz zhK}=)HavLpKdd($t?2EmJ;Dg2ILqfFO>t|@^xW;AUimfaz3Ad!rlGFMv5;Hbx27lc5=Fcw=pdXcM&LZZ3~{^@rE*wq37P4u>ONyw*8(FY->VFBSw zZWyi?CB|!?sV0i8lwU!6F(dfY6WA1X%ZT)`bV_D`aLRoQ@nub@?cmKjEoRB63~=@r z^bxP4E3D#w^cxlJuWeJAm#M|W5`BVnQIQ0go65v&h5sNFae`7q6)b@d12uz-&HfTV zhqViiSv+rp!NO#kJROXoeetT*}8nSzUoS(4UO0EZzoeNw0BYIMyH; z{yt4jpsDyMr@bt}?aA!k&T~B|h7PeV`cSH!St1 z0_ls#a8bQV9-~iYg^}K^mBejzkzG+wHtK?wLuohZB9r=_qkKqRxs_!t!doAD%2S?i zx5HhwIXabO@tb4}Eh~Gwzo)78U7MMzQev*Cwbz_uJp{guMH{114D)ScG#5V9^{}|n z9yYtqeHx)5IAx|hsEa#k+_=X$wjx(TTnqQX=(ZR%pn)6Z|C3QrG$S}*Z>n+}WI^eS zAsF^o3ru2$4?y(RPZ0x$!+ae zb9R_1JRkRUygzScuGD~Vai3w11*!Vh)6MhR_a6@y72ghv29=tdiS0Z%1<38xW`{Fb zNxP6Avwe!Fdg~Sj>POiG;)w#+daaROie?GYTB`)^&O4PIiQI?}_z@x3*eYxhgVsgAU#q<|9HqP6eTXdC41K6#_~^J zXEx3`7y{2b!g;M0n<~wh=0tWzYNFI1>Kf))8e5mks1$p7H`5?jJk3kr^CE^gTDVt5 zO2zPT`Ra7^IgAdRD5p-d7&%%ErbXk}E9|7i(ib43&MUxQk9v&YjZp0qhK^m#pT$Bz zw_9N?!gXJWn2NTiYgen~rY%bF`0bq|O<>Afo+_DnT5C~GiZ(dx6GAw>e}u>HGCS>N zcO|1B87b>)@7v7cQcs3&^@$5ocwKVTj^Dls`8knT(8cxDBmH=PmvEm>`8`cts+FcgdO*jH&8lRDk=K)95>%>w1 zb}fj76f!)D`ayqafE}TTWiOm+pW;uwo}CY5)O<)&ozY@GpEc6bCQBhy#iY5}ZAKikwbK46_YV(yK<-Xk*H7MT~>X*yZss3xZQ#^BGpd)kr!^Qd4LW zL>P9Nfr-6;Xg)VJy}e5wMwL?M^FgH z94g$;!soUz{|^h`;H)lV>%5?SP5mNdyDd|GT0nl#fSDhYKhEcnP2t};lfP>i24w=8 zai~4&KhA;65u23xoWDA%=WX8|UtV$ArnwLH!w12Sdyq200&O$5chFSTfI&Ax;aKr| z7OE6Jh*TixURcryAxyG&&aHqB2kXio72a^bbj4uw9>ektLy~~I*GKSn7f=_(LtJJ- zA#6V_=+o|8XeeW4^~1c35+Kp@(&OW{`%X|r;pUHNEHdPDR3<>On|>;yXWn(f_9d(` z^|RIlBs0=I6^DYEzMuZwh%BnT<~u=WS(1%BCJ)6XJUUg|?hn4LtXsI7DB-Sv_pnr8 zsj&9B)9pNkbW@U{@cHhd5m@xRlqAL+=cLNA4*MlHiN8Slm-|dZcH|S^x*K(S>Zm`? z9RYflnlZTq_m^dfs|qZnD@s#rB+6Yfi!p~MA0v>orijL7ju_TQHA+O?YLwWQ_s zF%l;0oAKYIm1%-FPG$NLRdFWa?*A@HYH^+isrm&`hGyyOy@QAlC4~?Kohv2E~7T zjma(u9yi?1kI%?8DdSOqLlSEZjgO!i-X^)6=%1O`jo%DUS$0NasF3133M7w`<(i|m zwo=%AKdu}tvba6zWA`uXkS6*yVfS7ju@pr#UgCUj<41T#a!jQ>r3TldF0sRZ~quWA+T)H=M{bWc9fO!sy;qkK{t`B~gzl0?g6W zkqqYThPS=n=<()PECtB$)qRu<9v(+BjHOak)##1uz3sL9s3P{ob>eM}^RLj(CYy}+ ziCe2%Qg222Z7xo!i2@z_J3Nl_e3|`8KKij8Gt*_6A7CDlb${(Fwnn1f zY@4Aw*TUR0RsT^kTAoYph@v@M>m_yY@VA0V+2mpN*6kS@!+R`X5owsaXkn31|U#g$k&wc0EO0MR9U1cnKiD2rC=#J+VM!D+9c5Q3Bo)Yxq zb!KZ5bqM2nJ6)dy53vFZn}Vim@Cz~=oofc4VlgWSa?Cp(!b)#(O1>{@5Nwv8rM?qZ zzZ{u6n5b$x@7WoDE1%4=n#$@Uh%CL2&qXCa`f+^X|KZhUW(!HUV=zzub=RlaaLeCR z!&CwHM9wr-2St^`NM!gGiERpV| z6z7^X-PO(LVX!;cgx+drwVd1g)Gyg=2 zt{t(ALC`t&Ri8VbVQ<7qPYu0~LYFI*<-i8#d9{rHy>j1?s?>_hrETd)gP@JNZWDtb z3-FO+WToRa_Lit(Lz0P-Tynw-ztLmz{MMGBaM$8po@sdc&Hmp< z9HiOwXnwqE78GG|f14}@$Mc`Kl?PLfCWVl+Bb^eyS;fC(ynLP(HbrI?q_QKLjGgws zBlo?f$RdkZqHUGr%xtkHli2(+pM(!nA$Ip`u-lD)3u68Cq@;ojJ&=Kw7KzEkn~~GA zpzQ7w8)omP*Eyea8;z4=0)%q0B7~<_V+~NT+Z_|!7x2oxItLZ}Mm04%kENR4DZbIo$5XWA(R@bF%~(K*k*``lFRUx%D(UwDB(QF4vv zyIxac5W;{_kx=TNgg{tNroMvki0}25V%3|p8Ouh;;G|ncy|YPLqtM8X0a_`&Q;K9L ziuM)1N%+yf)0oAJSf-*;`mNY}>yssj6VN+V(N-_}0u@Bi!lDbjFIxA-|8KI8o(ut6 zp|WWb*MqWa7dlulS)%Yn)ye}_RTd`}j`_tG9Jq%Xu)yQ2lVIidxeLw{&kIV#sY(zc zDW)^&7D2AAwNa*neMKQu({j7b?fMV%{D7^fP%m}7`gOk)DrTNMP&(pfK4ldhDmjC6 z??cdO;nvBkQcQl`B%Zzr2e}7*??)Y-9S^JfiL@u2X(!WLDWwd^#XGN^3`#XR$DMsG-w|LFvH#9y_e`ynT>)LO%Ug&sq@-zgk&gd2hr2 zYW40VDQzYT)81RGrt5R$_mXMH%<1aUW?IA_*gfK7I_^^dSDREM=^rT63>U-I zgmw(&>4W))nI`CO@5f9YS-x(C4&36(^cxI?;GiiFRJ>bpnwBQgw|P{3OlX1+KPD<( zzHFb5^ERW=1wHZeP0EUt>F7H6)5~%*r0w8uW+};^T&^BgJgu%C&b>D1&quz+AAH3& znI%jN^0ZJ7`Ch-GzF))tR3ka%|$tx zmjj;_p}VSoIoYl7{PBQYt?S?8e?T@TK*S~)ba7;I`ACpucI-X-WT3Z&DN=JWZsfQK z{$!_$G9pFDlETvtp3G_0>mUg?oqEkf^?pb4pchAoCV$ij%ulC>Ucy;)t5)9uCEsMK z$DK)X58ZSraLrQqTtNW}8rr;P6cJ?kBikW7(nheN9giZnb=$*k2}m_zKby&m4dxVN z4Z=)7>)GAEtnAWph10GUDY#V^<3FJlm^UvIgm7`W%lY1OhL#wh8y0hX`6;a-uM0>j zRnUlxlKMTy1|a0K0L__dsG`UNaD}GRWH^l|SO2ns}nI1n?lTlUbrr1!@@@`nx~RG9w&{E3RN+)TUTh`xExx;vNVn&{6xXd z3}4=0oqR~85^ALJMhxInawoZH8HFwkNU|Uz3hBD=g?O@9dKa7#=+`!Y!G`Nr{12`Xd}42_Iztv-^|t{YWl*Kw5Y8Npk+*Ni5cA-gYX;vKXofC~Rj@8l2IscXXH?r8Zk5>y8K(6(n!pEVpx#2)E`-u#!UXPiSg#F0 zVo?5f`Y0R9nyjRj$~6wGpu!C;Kb`kM^QRes#*Jm1Z5 zfeE;u1z--Tuqw&%9wa>gO8E=~qJe$Vj5L+Uwcq+mdnhVX6d%R#Y78DXJ|1c*jx z-NU1Dn?SH;4CR?bvN^8#;@7;w*BYfw`u@jmf~ncqave0PHV1Cq%%CzZ=>CH{P#)=f zaEg*iWm~v9Bm{4fsvpE~aNY9kzB>!btQ8PKp|KJB?;W@>o~vEgXJET1m!eDIU{wQ` zsiT~v_sa8Lq!>Dq%VR1Q{7eBooS^m|@|x@p575_}XpZ^P#yVgl2h7eo;uO@nzbLTYV?*VxZ$m z2L&{3a$Q1YJy?`(IZ!?f4A47ypZdVVNqhwbq69j&c3jVdCU^2~1(w>z-Dv#klcWlb zaoQG~lp@q=|Gyqw9g8r0JxVY0{tI+ZBikWF}53TV%A8STkR_V z6@^Y90>8g;_wF{CI#EAJ;i?nVEBA}^NY{5j=0tR1@68TYkSFYZrMtfXKt%@tD?iRI zXLRUJGNE6gZ9e@q4)gfBpMHzQ4=@9=j_$(KEK(W_O*8a*g(tbV`pIqBdUfyR=RP1M zNxDh@_}v@b!RJab?zo8fdmyolFuGdp-?u+9zM${Rx-s~suYm72=4Re51MSGSEr!xv z5I(viJ=OJM?7kV9sgy6zbmt$XLU`*>k59z6UfUnN$us{9AkzJgxC!`vBg|4?0=5R1 z#Gm0ebOQC{Jf(a(JFCzc=g$1GCDq7+Tm#EaOv5(yEhtj8zmST;bhI_mW{8HtxPvss zxU2JS(E$-clsJ06mZg_CA&?q9jGM;B&TZt_8h|WSU3)LzkAW@?ja<@Azt{MlZGk}Y zLuh9<&1ElNkNCcJW*i>Z4|CyJd9c4kKFh-ogZ3{F-{iD<*I+0c)$T zYU0Wr|Kav6rWYmlC<&wIS+4I7K|BJ|L00c70T&j)j7xlBD+qRc(gvXA3c+~0rD_36 zS}qVO%L{VrVhPLM)reTqgpBa6KLQ$?h(s^p+6?vk*Y^fIV&Q%mSUby9=fke~Uyo!F zDZcycQh;18IY@(nocUsR`?$}GcY-WP8QFZ-h?~Ld@ba?~y;_-H1h`GA#w3VKg7U(N zgR|XjZ^a%0`8wYQ`13VwGE5@QNat!oMUF#F!V&?7oVNoyPvhZqn{S&g#!tTHJIG8Q z?nvMKAR9FhWnu6?`w0se{eeQ$iGz-Z+JKU5zZ7lJ#%Q0d&@JzCB<3A zEeR^4#u9c4(g$O-*KR4JF#EEa=Qw$zT1^b`-h7Tqe(8PuiR!G(77EJ5; zt!*wR>2gdmtLYifr{|S5?Bz%;Ix%Jk{_sZX_;oAIy=14jOHlFEX2gJ51q(~GcIGY0 z+pOzMMJ9OUc#s0eS~|wx-sWU~oEXa`3XW}U0KhFg8GffymsAS1-o%FZv|Lr z>}C@h<+sm4*b8^sz?gIB9!#jz>`&_rkY`il0B~a{3#Vc^zWbqiVo}~2SiIMp-wG_k zH*0kyf3&%L5B=0sUpU46dA|4j>N9#I_TmJ!UL86(fn3qZcTS|~*ay%r%7u}I-q#CRlBGHA`%W~P6|z5>c| zE)aUGxiwicSEwQd3K|l`)x(p2u+4b?6uX5G=Vo>;BCA6O^$b107W;zm5rdD+1e!N{ z5X4kcR;&HRd53f+9>q?eh0a}VVA-G*(-J=u?~kLuMKzoOEX!y1@sV_*q)ilg!fW=# z>o|3z#E8(+xJdt0Fd}a#NAVf4+RKIan+2gZ6IPmcFMFoV$SL9o+CWP|AF`aj=PvIspwq^Q+!3qrR;aOR6@S5n=>(W61%rNI5}E%YpZ>1BA>1#JF$6v zF80ULle+NdNY`yIPOzG`OnxJ31`jZ^euPlHY%}vVMmF$covf0B#?%)r^XW-AjqsJx zXJ%iyHtHo%Nw_${BSNI7?oYC8aNvv;HZDYM@sU*`KmKCdX?iqqMHT+NuCK@Mv^F#! zTZ-S(l(00gnPD0(`#OGl$Cto5DbY&qBrDDRErk-bG`z?+u^R&3tPAQ>6Zgl4(VDWZ~2ArjC@_9x6V6XS@bFT0Kpi>jLW3(bp8k+$J@B zRewQnD~+YDFQwr*@5QL7XwG{u9~;9ScZnt(^>oBevxqV zjC$Qc{(RHp6SD${+}gc4ADc|p8c9C(JA7$ywegF<4Y&Jubc-U+Z^xGPBA0|=L~4X= zB4^~o_V2YP+YU-=7R6ETnDypiWl1KFAY-nJ(6wWKRUK4yCikHtg)oOs!3x40_T6mz zkYg#mFfB>zaHB~il;hie;OnIuC5YAW(eikNHRz}D!& zBp^yu1m$|TkNbmdxbpku(7I04k%-U;oU5Zpqy=Z14vH0VxEw=jF?k&|{?qD$c85o7 zFT)90=@tLZN?R~$$b*v*V)G2(qU<$D&-vuUDOvl|*xLuQ7)xZ&&AoCJQ@`yi9!|-X z=1=2b1<@#xlf8W{_1{c5m%{Qj%F-m51E|;i>vaz~F0-%o&P=2_HJAwmf}&k zI<-Cl4kLmU=6P=cBJ(L)TY*-OX;0ZBSm5Ys5Hp%m>zkBF1cTWd4;+bpXS9pq!%EfUqi!t0 z!(3``kd;8&7x2;uDcpjvK6JOe&O0$t7+%g$iLJ3@Whjg&&+&VFYQf?&-sqL1jFr?C z!u_;9en#a`hOnG-x%m?nBE?OONQ#PSE0}I^ju=)WMnc9N!hx6>o%2fL6J!xew9@44 z7m^Hec^ReEcp^6tZ)%e9CKLP0s|@09V;WMk-N@nE1g@O-@CHa=^dRNWo-;elUbUE@ z&zs?wm3@q3P!OgIraXf zO%6Mi@?|mGp@6>14C`?wYOYG`D~i>OpMKlf)AMq5d3)D75q-4CEM!kc&9fGgte7*Y z`l$5NHk53>@IP(p7^7jYwcy%3cD6{aT$6{VWLe`^L;6&Bc@EIl-CO_uz1S-{e?oZ~h_iSZzB{ zp&G=-F)_a*x!!yt;#Z26CzEprxk-MV!~(fYPDdwyqEM}Y?3lokZE4Z|@C)~hZVC)}XFM$sc|XUTyC671-K*IUZ$%`PA8prihg%7a-0 z+bhZt`fESJ=%X3+9zHSs1h%dx&!-8u3d|3)LwrLmEZyHSQ-8p^0^KVwoPL4b)|bA7 zy_EfBuQx+O>FvBzC0+Ok{iNRg6P;*PM5Lfj8+{2UDI!#hYla!Q8Dj5p(z>x}6!iqK z&cuzU)?ZV|g#OTSiM}xCn~^MJU~V1w&4~BM<3#wptKJlEMh2$hr^asI99=>kLFFo3g;PG) zVi^mvZd)?tr5NaX>(_D+veHm0=z<+TCgM-)U*o3&KpM8y2H!6lY?dmWa~DvD7z3?< z7gSVAQMm{p!kBm%wASeRTKu5mIeh>4!@hze-ST>^>|X;#uyAz;+U;Iev;6HasxP#iBfv~mM@&ReC-7a;TcJ7FPLW_ikxqG zx@0jJQ`yo!6Z2ygO}y4PI6o7(g2uuD z5c`z_G4mT{q7a1G?I}hwS_d$VSST!Qp07#*E%2sLp1ntp7VfXUXy&+Hrh1;yvY?x% zUO4>6g@}4p5aG^)qqk_YD^qw=T1iqYy%*XLgqT6eyhvTJ-lipWTZ}PsUyOc$xeL*5 zG5Bh6V11r3@BS{@47_5Qh9>h^A{PZC|UyvNRmL3n$WK zbSecB-lmACMTS zph>A&C2@^1GC-&lIy@{?ZKhgCNfuV-;DI(`+&7L4eIIZS`z6Ui`ELXn1iI!L(7__^ z9tUu1oC*Kqx) zj5L`95Lwa%RJd{YsrMUpqQR4XZ)sU=@c$MYO+c*l&aO#p2R~Q-c7SpXgmN);!*JDK z-iks=uML2id;plcf59`#UgV{mFY9y!O6Ro`Ek2->HwR2_MqC5uK+qB?yHK~%kp8Lc zAaK?4`7vC%*GR|1`I*H8AEvP!sM}o|3RIhlU!Qd0cbIQ^49*4g>8hhJfT8CkP6G_6I-t zn`k0K+-!C-xDx54@48DzXg0|T@g_?d?;sItU&dj;CywP)A_Ye`&8}3=JZ;d}O$J(P z*ZaM~OYpz(fBYHI8p*c?_HR! zkWtW86^J!NTRYx?_O7O4yeT0o42j>}2LjgaEhb6VJYdv!4&!-;56e$@)^o=MW%u$a z)h+7tK#g{vArT?|N_)VB5Ge~tW-U5c>bQ!#>%rwjf*%L&KY0EB{V^&6 zYvSD>|0^OuC=zORVmp&3@c?wxpfcQcu-p9KzxNvU6;%XR-oFgS*{(<)FbGAyrCKbd zKLhCzb0CW(=ua3h^ZlOy0vJq5?3=rNy|jNBMBO6CNWeXbt{kcH3@+y^j-t$G>e#IR ziO>KK2=p1DhA!X762wDpr;WABrIZSiu>2+pU?Skk1#~+2XTKyF{@+XQIKz&Ka&9H? zhl{)jl24~YEdY4<831OlSl1yn`47hOKX1E9ay8lNPJ_RjplQ~nx~>CqktKWy{1^+s zi#D$p|7q|4az_G$!oe7JiVDA;H)$>!ehTbOSa{_JVbXajRZ*?ak1NlFU?MM;jt{vx zJi2yesZz4Q@#;q0xgUF98AnFpU6B1=S=|-!;PRBR0IK94(D`)810c31dxvd+jy@_+Z#vP1^k>Oe|HVdtjUE|JjY#3ID~py% zbvYr}w>dQ8||eE#!*)|JOChI z8c#B5a(p$p&WHHl8&5T!tP@MOh3wq3_0H_wAK&8;*iQM4Rd)h9Y8A$W&{i!M zFy1ZzjD8xi_idTX>(0}vKtyD0J6Z;EL~6}7HrLa%W?dG6%eQDrO!lZ zd^`0T)`B)7*+1y>GavcnRINLwT}CZib%`*i*!)}Qz50eisK*`jzRN(5r3(PywUeQS zFfh&d3&I8F@YHl+lJ)!u55SzwyswU@I;wJ*X8`IMT%0skum&4ESNq9~ngMF^_dkK= zV3#ci9eB=b`l!Z*RAR%FJWi;E%rDcl&lg&#{~45y^kylKs+R6;{ zuyxLPLBH-AGNMU7D$7`Yej5mV%63FxffG&*a^4}T%#dZ(O=w*8677Af{-h8c#s8T? zSx9cCHQLs;!4uCyYnbEr?IS(C)@$Ry`!(qCbZdMM8g}=))*uYiFQUbRc;SKKxaM!& zhl1uQxIe<1dhP!+d*xFShiMBZdUw{#KXdU%3wDxy?xsrbiczmaOJ!|=@qZo=@=ch>J-{(@4jAbO+h-)s zA`ier&NA7j^@ftF8{9$)&!E-%9n{E$V}ZtK`1TkswZlC7vy{Qi`oNeL_7#+a-Qz@{ zfA;5b9lwPYB^PYbd?e3MOH~fAiI!PF%~?F)XC~RUiaRaOkEQ%4tuOUT2$nzWNf7id zfOYZ^41ExT`zxnO@8p3gEd$jlxMo&ASb^&aS_6nIfPH$7bp#P@&@rgJk>g?s6v2m} zIz&7a7918`C`7!(_=SECeFy)tlYX=#TFiWPBt2ra>yGXuDTL}`Da2Uu^ksWM$PoSD zb3HI@iZ&o_b|U*@$kl?-GT&=x=ue*UxD;JAxRv2J4(i?bwM{MiDP;@b zR!8035acuAUi=~2OMP}2NP?C%i)^J3|BLOpDBY?;NfPrx2|WSICZ zPD%iD<`p`5Hv@@9GmgNBm+O?Zx|3fwkl3;h90-$td)(~?665rNW@>!xSkw(k#*g2` z*eFr757g!jAhV#O20Z^9 zr*dx4`4P3{4;nX#FyMGeFLd>cQq^{*Ap{BDXMX_^>fn5n=a=QfRXR)0SPp2T@&ZzV zbyOd1GILDu^I~w!AgImyzQuoys~w93brxG6$aqir4GuvW0mK<=50gF4OF~jW4RV>j z4NfipS-HsCI#Ej%;<%i4^t)^<;b<*UXe?Q2cFvPm9%vJt)&rxVrsNnwPBs+uR;j=r z`DPAWqK?P5%U{Sd6_^>1;1k@`#9EIhx4xBmt|pHy z(yjX;xU=`{B^=QS^#qfffKyqh5O3;VkB0ll)Z*+=0;QuAq>;JrqlO<|cK>pX#tI3i zhdLhfj59;I{=eehI;yHKY8O_d8>EqLkP-x?Lt45Um6DR~@JI<#(%p@MbO|UZC?QBU z0@96ierux`zxTcWedAsS!@)T0v-e(e&9&ByXFfAH3AY=auxjgOco|qnf!0Kn%1}!d zV~`M$6JVyY0`LMt@q``9bdTO=4`v*KnkmOXkr#eDFbqNnLwwDZlt-ZUPc_wysmyP` zaQEH+jLE;MG%88FUmPv;C|ao-D8BY+D=B(Ig{VIj7YG;AA(6WpHQAH1#k=Oae@+jw zcE(aENcPyl-N%j*()5@$0O>6|1h1aYg7zrgjTJ^8$Q*k_bL3a^RzcyL+GxO5I}ai= z&VH`HCi=tuNIE@VM)GAM^o)aqq3tS%QwXc%_)^LAJpv*67^tUnrH=XzD0Xhfj6kKO zn2%>)SCqFWy+Tx{nF?pl>z#*TZ+djF+sq7}h**3rrH(*@E!$l^i(ZwqE1Vd?Q zvj@|!!}JYVwFi1Z7}cP*Rw`Jz^8tQ2L&&uJ=^nRNL4#$+uEU_?318L1Sahf79x}3H z`7e~}$kAg$k^R}W!Vv~a5|^6&B3>J%I@_+Pw|Eni{nZGUwfma;DGb( zsz=D?)XHso%uE+-WFU@mC?!qjo2i0EwD_13VLp4Q^+LSPGiRSD zmv&bAMfVHTNt&IH9RZy97_}@;yr3>(oVb^60JJNF8*}ZK!`)&KSBA@Kz2-2qT(NPe+67fSA6i;x(sQR2D1UUlaXVz%A zs54>#GDsG=R4$J0^gQq1SvP`rhl}tY)7i?8t9y}w1+B~*%l1j9!X>%i|fepA* z8ldP>T@3l0^p^2+h#u~}gG4-)FAO=rfIAYI|Ixa_&byaqq|U8vXi5pf=W{ECrs z7coQ7^-(I=-8U1QJpn2yr&f{(Bb$5Spkmq{EEGS-4LfobJ*i+fb!7nuBaYZY|5G=~ z_Yq;3ruN=XsRcmXYK*aQ&lZRep&pcIsVt)#GR)74v(2QaXF-uIV`0xmjM^sX zcv9!qj5#|#r!wihEagJrX*z8@?`92bya%KXtU7(Bk~5OlvV^oyn`Z~$lwgGdU?|aQ1?>$b59tT(WF$llwIf5 z6LWXu)#{bF&U;U+&aT(k!(68f9F4R>hPi66gF9bUs^ufM(JY6ymw+TUw&7&c}0#$R-qbvw<0sTmrD6iJyiW2k#{ z;8MS*9ar9cu!FOdeU~AGCv*Qw;>3-cWFQ6xD3&NWhP!pezQ*cU6nrB#&n0%XMAoIz zhEGkuqnZDfEw$@r#0$;9``$6Q3egl%b6n(fZ|@mLI?OWOyb{98959o!@~-G3T$8K1 z>qmZq1^kcpt}(;0dvDVnj=0F}%3DkNP|(W)Q~)7I+@~@vEXol``FQFg#RUF7scN($sj6qLwRR{|yF-|40@ldTt9zhm03~gEDFYs+2{r&$C0Hn|Sh-QC2_G!> zH0Nrdfq1*4%Zo>C5(OuA+x1(};&b_FtC2q-tk zUwtN>@}BM;?OHs&o9*qdBN71Ns6}FHuCck7W`*7TZDcP5#7DTrpznopC~E`9wjKU^ zMq#IfE+n|Wf700ifPlG-^zrwm^<5>9rTJ!g%9wLTX?QI~NEptT#>RU(o;7IEm=6I>mR*{TQ%VV9_=;lK{i+Cj-(>=VqxHUr<}|GL(i`BRw4mVFT?o_- z#fQuBHJET1@4%q+L2*E+r|_~yE;L+ydFbdVVk)$QUJQ7uu>gC@_9P>b;g>MFa6VaM z;!iS(fnJY)O~S?1tWZq`Q-6uB4RPg94gakbl(?U|#PJm&KF7a;T#zeH07OzZCtFDN zpWD@52e^KI`xK@B=$0;k*)GSj6gXG-y-Rx12mtK#I4}dsZ`Jiv8x1`wmCEoW4CxzE zvz?B9$UDp+blKoOH?STI4`DF@>>d)`RVdrC;U5dtXx{<36%w;tntSP*DBok}N;I$j zoE*S@L+<)+HI(;AXLD?7&_-s*S)MDD_(i71b;+}(SJba^Id}WLHds!F4nO+5Sr=Z1 z3L`)VH0lU+5dnPnWXWtbAVwVO=K*M;Bxrg*{opBD zhrASndhT#n@&s-2q*A?xG=wk29m^*(h5#u(>9aLnxzpFkgZmsS$Q>)1r!c)xf|3P> z!x;a0yHJTSK(f^Qi6HRp6uy(!%$;?2Rp%gqDg}uEl$~sw8m>W$UOieJK%#{JQnXLec*S-wm4O7y&UieU%iceZ~CUYC3-zTJoFB$FNT=PmFOA+G2QCh5EX3=urQ{2BNQKFxXGeJ=Uy%*)vS{W~K) z?Bx%Y6@aw=yrRYz{^yoINBCU~3goiA{(8Y9{d;wQUnr%3`0V{&tNuqCj)XvkC!7Qa zNdA2kz+rt#Z^x-p_<2_RUWz^XAQI;^w5w)`D1Esb{XMzbKo;aL}GD zc?P(9x=@+TqmTib>sR97?@P;|;)SD$4R14m2KFIHQImv<_rO+Tp~Gd}v$^u&&$D{$ zATymWECdg<+5J5?&@V7P=r2UOhM+kZ@EQKQ8B+J$g_ryyr`{#^`9;N-|4d>8Aa)f{ zbGt45U;V>|y@iJMKfiIo66iERr|$drK)~43B+uYk$vmWcz;fT5)4w1tTQZuO%VkA3K{rd@&A1qpi}|-9I%i- zWk$6F@>`sEumImOaE?z{Bzo1VYGnom4-bj`ako? z2nSSCq6W1$gX1zv@LLVRFA*RVtbJC-7-g0KLV5Y?OQ7X1_oIv?#@LlBeA+ViMAe`OG(g`5cXLtgN z#!=A3(Nu3%ClXMhHY3yES$sc-aR0z<8+G1N4hP9k8jeIzM$$sK^vxHm3dBwMi3C7P z$4pkhto>e9f~xsF1kv8C2JJvcq1L`cH|FdCEMbffka2?XXIKM#$Af*qDs81}Rq77~ zyr$zp@@h;0Q2YTNX6tsWlH)_rShGksR~^B;O+}7%pl}&cuRDb6i#d>!b-vm6m%1y) zr+{DE66`|&4}t${Ex&t;(z$Va^;^@S+vg|VKS*_5r~O}lZImc)<9u#dl9j4CR(~qc zwL=DEZlw!~bgW-W5EQ&%05(uD{awO%;I><)zO-#JM^?5?wrefWfnrM4AXkt+bq2JK znn=y+1_(zVuf3pf9)5t-VpRVac)AcWYEkS_02b_P@j3g=ZXbpv_y|NSofuyir>xvZo92A&ydksXzs{lBMMNg}F@{ZKb8?Vda79wS>^JUE@lCVR7 zQ6UWBMQ;bJ@Iw$hmxIJN6)K>@={*1>g#opOuB+_`8yIUJSpqA&Thb@x13IF4ux0N8 zM$!=wu@*TTp8f@4n*UZO95Vh#48pmQCWNoIt-73>?uxeK!4xe9Le13or(C#W<0z* zP$kLfe46I$z4JBDUmojJ@~EijRatNs2>b7W%GW?4+Ht6YlN3h_`9Xr~czl?s2gsTo zCLY;MLcj*|V-V4%IC@iR_#$=V@T3-0(=hMEl+v`5;q4X+Gj#X5SgCEuhPHQ!`x6HA z)Y?P=B2@+=jwNTD^)?I*MGje@R#L|x97)`GZ~T?(N9ijtPX7?FuHJu}c_A``hK>-z z8WIdLZk5Iyf}peK^gNkKv@&}45Qw{5fDnG*$DVhfg0u@ToL4F#ZV*#B1m3WJPUiOg z<+On7u=Mr?%b-2z^sIy&h|V8)_K4fR*4VSQCJ4prRa`cx&P_!Nz~H;S{QQAGRltMi z{e2#b1yI!v*Ssw!)&&VXM5gsX({pipdJKf<*T5!R0f?*w?=#QAw!pgmz`XOjJ3|j0 ze7wEr3^HO2aX&q03y3=tBIZA)ZoNWw3d+G+;vb6oyigEIIbcNJvA^ltu*0$_-lG`g z^+K!orjutc*c;8EoI!kZ z^8#b#7nx6_UTQ~1JMx^l*yu^hQe*?t-|<74GV|hZiJ!-a9#;+-@b5^?y0sfF4Ml2& zMASjeFT+3<)~NrX`7KHXIOHSGPOK~+sWdKDFkL1qyJSp#`I9$PmB%ftb;fn>AFR;Y zwXoS?pQ<$i6>8>b`g&xo=(CAN-hL1oc|gc>a)@!g3QFHDWd@OkeFmT#rO`cr{qeUK z+I|TE6NAD9XoY-Oh{04{Z%V*bv{GhXp}7fo>Ez>TqzWs#LX$sWg!h^?$z0zTi4i)M zmYQi5w}%Q~s6h3{h1;}VLOsvEnNK0&ke0R*U z-G+{Nhw#}}_?tCQ0VgwV#Im-}?l!zJ`KFhr0SBf)-xCC?bmmznfRm4yaK-EPxyPta z+fuzNQhheI*OiTa1|e6yQOR>Z`^hcLF^D7m15)Aj?ur{i^)B?Q`;1AwoIK!wz6XRA;V@=d>)2mCh0N81iFzeO^9z+zX?Xm~K+sw=F?9wqtm)BC1qsv_#iyJvR- za47V)e|tHwb*iWPfsD~Ry<9trpW0Uu#gJuMeu%c3l--`iX{LF0@Xg-!#0&ChyF(rJE#6@z^nO2Z%AQZh5{!S4 z)@Ufmi}Ws4XeK?3JT#YiETlc2_dBKl#PV$ {^9Agy4$~ztcxQytxxHf}1=f3hk zFyIfsZGP+I*NA}IGTe-G9=C4z(fUFt?JfWummnoN0^kW`A!viQnB#U17xMt&*2(az zG*Expl<1)?MnwK25sKd1l81HMD-3T*Q4shBRc!msOJq=dK4HJ5iKQ0btHO5QO!N~= z1fVNh;a4$a*^LcE&Af&kg7B7Ax$H|?@LY!>r}J~-P!#G-Q%-j!AOp&Rk(c)2xHL%P zex3D&^yBdYdIInzIxlmpRoQukw+dotUrf4g6XDB{zFw;PhK=!&Av&{c6(t1gJ*IkvG5SN8A=7?}fNaH32V8Ck|cN9?T{GkI< z<;NJNZFR2lX+^1DmS#!Z6eBt38v&MLFJK|WJcHMf{`B<0&k=~;ODj%>jFmS^5*BwmEW02l zONYFa*vO~ILN@j&puHdVp{v`t+uqAgvAo?$O;UP6-nK6@NA2Wus1t-nj3Be0t!e!#nhhq}1Sx?fQPOr3rHWjB)8y5qt_I9#2X-STXU zaFc?sez-wDZ&~_@15tg{wFbBtVtX+)wj8|tVgyMQ5nfnX;ddUE!QI+^12%Ji>n+iz zIQQs@8sn9TCXC>p4u6*w%csr~&o(=6p*X{yPlbGp2X>Fr;)!j(-*n#q=Uj&SEp&Tt za#;vN)50_Q+(xK?J9~+?1uKhP!1yuYQ1t5aKby!Gn#~cV5+tt(Mq@^LZgg=fRIh7q z8CJ8(71NMfU25kEnt0iSBO2@x_6!GSuhj9 z<^eAUfH%kP%) z-jC0=Ij2j!sh(6@PN}^~xQcHhUI7Y+X|2GtgRYotuFf5n)Zs^V@0YI3`wQ&JCcLGuOE z1V3x#r1KYsmq1y8-jT%0|Nu<)k3%5pY6V`1N)}$2FbFCwqW()jF{}o zxA(rf1LEZ?R#XRJvpPm6WAFDlcN1Tc2Tzo*z2B$WRamtlv;;kcbbZ(KJ9 z-1k*CKCK6CX$+~m7S*rhs81=E#fS;kPG@{wlQR*J-EDVoS@6^qO z(B>>`UR{Uz8MV;C)N6ZpJI|qKUuIv=yC*5NQBlIfe{g@$Q{gP0@5!!4{Og*98Q)2q zU9$U89z3iE-m%uYulwu<`YNlOgo~@|B%f(`HK{B1>*Q^`DfVX?R%kG@)PMBoQ4#4F z%dkn~N=fLpmrFkn*?_KdF;zP2wNG6WtU2MY)t4(;)ywi{6mA|QtC%-t=E_^=1o-4Mc-7b%(|$vZZK2g|vd8F&MZlNaU!r=;Qpr@V+fA{3^^3*^ zR}Jk4c5YzWASZm|_pMSzlQCs5tYwpI>@G0)y7AZ|C}>c5SGh>b@y)c@N^+6H$l%Ix z)Lzks(#ZPgQYWXAnTpg@Nxy1k;}0`_g{r!kGPbGRp+>aB_53y2*SrlUGNtz=CZgkt zYD$xyrJmWZb-Ub7IBXm=TREh6sJAM?&#YTMzfydgb&P6k5MSkqpyHf(=EfE%*q*w% zNFXsJ8&{h+giQJL3LG+}k_p_=%zFLE7iWtjgV~no$)UZ32eKf2*D<8?i*QA8kt~?KS%U2e<*^u@jbXSL zojktf8=?%VD{~COhccV!(M{Fr`&la3knn!}M_CJ3icly!wH!At*JHr7`!G**$Z^bw zs#>|qCz9^+D(L!rzbg$V>SBUzoCBmG>$^^5+VOR8c|QkmF^Z=mWeCV>DD^@K`$P1v z+|`hNKLmW88VY_&0u%jdQN1YXw`dYrfA0MGlStSy4y7}tQ6Cs)P zvh=VP7gkqQwK_3|{ka;t_XlbbH1(UEH{kysbfGeM%TUc22eH3j6BGlgm2x;>|07(< zt8nVK4b;D&{QdohIFw2sn>$5&sliCNO-T^`{1GMw?&8l_D&K`{{)#&^K~nV3Bhs9 zagxOUeu{__3U+?sin~3}r4B&Tft_^)#w=-t>Hm9ffV2nbMA8e3S3ZwKGjcdBtM$b{ z#{28Bs6cA?nIHbS%~KKZCU7y{hiSY|3}Ttzpu)cYZ*l|Is6Tp?b^G>hd=-W&(e#IK z7#J9>@sIIm7v5dImB=#uLo~I5sEp>zPcB|fg#6+GyA*frjZ>F5d46j$^w$V-n2nu! z%!0~HWLhbaM>&r>xl#YR%Kt(a^kn7a zL;ohx?8>*gcH(uUlL!H>DQ1^59|rMt0Rq_xNs==(4G``F6xV@p8OuE5T8Xu z?Jb+yp9M`RZnU?$hmUfwK5Rf7i~WhMhxK@QyCmh7iR__^&IfXj-DP}Fi=gA(7Iwl? zQlURKBR0ich9{~j}X)^m_JnLmSLB3M)?f^08;gIe7wr9 zVeMEZ&_rm#5;4@EnwN*{sQBWS?_;6zZ91?E0B{9%v%+qz^ySeL-&t}*>DNnNo@#2^ zUHsZI2|6h{>N8oqxV*Jm`8!R+gcK(Y*MQ}TC(V8qLb_jG2g6&iDHR;=u~v`w&VNbK zm56((`g;|qu-SgOYx<-CKlMWpLm~!Q;`a5-qe^Ne2be4&Xy@iq0Wt;U8pWIoBPg%} z)x1P8-pkGLAzsZKLHSBC8rB`#^bvKPEacT<;1P`po{vk@mqi5(?k1(R($=}zKh&+S`2-~jz=k}Q?IJ7U=lC$|FL z>N>!Yu*g5kRlYn*5n;MdZu3vWmzOm0xV(5i0$glBa)^#JkBuCy2-d^E@UXeOy!>2| z8;S=B=R?2OdbPaKKJu2S&g~<6g`Wxx4_Bm_B%f@~S~E$=%RNuH4e9b0SK)*4YBq$*^<8d* zwq>?RAPNmiDIdtGzpVfM2w6=B+9eiLmRD9-yc(_Bgh5-5wa*M`=r*U8x}V{GYq`iF z0s$L0Xg6C<6V9!Q7So>Uktlfnc7!U>ujJ4WUWFOFiELtbzH*k;NSO&l`a}ELqW`zq zkdJwFzq3x?tFrYKXZ=d?X7Zq&p)d%_kCzjgu5+TJ!aXL;9}&6jx@n~P#0OxRpNy7J z?Mhv0X(WT~{)DMbJ_)Jtq}A5U8*ly%tG%5*(>Oz$!3K}*<@=%7jAH4gfZ#pMvFi(} zTG2os`lSkCi=;ckkSMrGc{<1M(3+N;1s<;qbgcHubs2b%jEnE)Q1}W`Mo_{MG1b~? zJ=_+$)4HvFP>y_GIjVDOB(~nqx-?=K;CE&#bh7cX!|pV%g2FqjmA%F&gz8ppYOOJ&-6+eEi_86is$zM$mr~lmTqZgBkF6Z^ z?<4^Z%LIetOv>GH2h;tjC+je$rr5i1>MTE1qpPbc$XDR=ZSfL&8pkP6%n(#WiXUwF z?@?fk{7MUFaxl;4i_7r))xsjs99Ur7ykWQV?1wP4IiRbw1e$rq$04hgK+C8*r=p^Q zR>A%eZ+dd;r#geKf#^iS9<9em9_hb?^^O;AS$d|Et7U9&XX+E^LHw%nwr9Daa5)+q zM-ufYGnmowqH^aynW-?hz>cM{eqf*!0>wJ-JMfo_N7@mp-{IKK>A7|Hd6Z`aB{ zN3}$7TMmvg2tV{2sj)Lb(ZM(GuX)4;?POw7zXM?yBmL` zY}9-)H$ZhilXLefptxKKN@ednHWEmqxzJU*1^T>M(xysNL-}@kHV*;4)h%5guX9WX zcfVlhrl!3LFkN?%-#e(by|MwlYzp@Hxt*(H&tAqmye6qXi71_oHmqI&PvW z7w75n&&S&Ws*;~yz6*|3oOr06TtK~zP`3+>n_?0V>o(@)EuM7edxO_$sOA@Ysi##1I_^Py*Pd61Sq|^a}YPJ!oq%AV^rNn z52H1+Foxxe8#Or6)v41B4*OGTa{-Afz-g^A%m=(EJwL$W!f5mueqib@FM0(Q30{lX zbD2lKTQ!tIcoH=5kkBHXdRqIj#ufWCN#REkv$m>k?>|G6F=rryAqMvm2 zS|{e+4_5v8wqIU**1mQ{g582FpylzFu)0u-3dvywS17l?3R1Igg03^vnQ?{NrHu8| z%Qv%Yj}+!Lb<>6?XjEM{k-FH_KjVWc$2a>8J|{GuXEHwye>e=efJT3NBWObshBgK;NamXL&vB%Djf7Dl(2M z7eI|wDaaxo+bC^P)KGYh3v2zhrql6YC5^X0hqJvaIoJbVe<<>4Gst67B0wSbNMO2-7y zt(%_j`tbKqI3Y@TK)_y*Sv5b;tUf$0y-0eygD7H6s(mpileM(6Lg&jTnse`jTEMJ<0uN-N zyN`Qm;nc~+CnYZqmh?h+0(Ksk60AfCi0wzGl{B~ks0WqH*fxg;$a9UlKdT4ikG?RYA(Q)oaI_;o{YDrb z$}0mhp{eXx`ia7Yh2>@eLuCe>`c+&+1_VZVR8>@UFg5wiV2Q)xet;}@CA9g0)Z{op z?#@?<7X-CeKU27f3)Ym90|8#x%Ml7N1hi(|J;$+Boh0s|osWcv-Far+95*q@Ean+7 za>ufy&brIM>|`4NU-ccGCqSdD0iiloi~2k++Ijjo+tHy(1kfB;!N)~DNZ z9T0p5ONVDvqbx%9YG_QZ+5AtjFkx#|b*`>ecTIZst-h{qmg&&Z{PfzLPa<@oq6Krl zZb#o?)Z0TRMP-vg1RqEG8FYZ9@R{E%k3Oi8%&wOPo9I8=5&QWGocz zoYmOP_PPf56JI!6uwwikR#%J1D&~n?W?hF!6pn1%1rjn>WnRlIV~aU7y;7ioq%0Bw z1(vucmg71M1y2h+cEH|<=zF@XAzZ0&xZNgv_EKbyGycL1?~i@jYLzHFwL0O_@-4eI z4KG{hDBppcWYxKYwX153EqUjD6qcgy6YE^N)Qelg!5<9eM-gTyrU^mJq$Ba1kF3DP zQ@k)RWLAmTez%i&zJ3ts%_dLwey>eSVRp74cdf}&uZO)$t6*Emb zKU;ru_SgnqBKPBwK>7yfhEeCb#RPhP8u>AsvBpkt@1D6yG%6<~IT+j)NgSh?eH%bTM5&slDoB;h5--eg+ON4;Oi zs`y$a(bH)A2kcsx6HIr!@emI@ObG!5`OVg)E-xx)h(8YUf8;Z(oYcw%3Lw-5#U z!5p(0sU>rVskSD>a=XrB=VExr1mkML&mbi=oS~P=dhBys6^>r}s#=}2U{*o*_<^CM z6;>v6*~DX+_c9&D<#ZR5EV2a@j`?(1*!#rQ(%hV}vC0z{#zp{aOcJh8$R);OO;6te z+`~Y+22u+;XRGCzSU8?NfpiGDtX0UqQL$c6RkTPr(kmWqK87Y4_7>7$;9Meh=V0e= z8(infPZw!8rf&YBivRY?4H6hF+F}*oMDJ^vt4u?lUo(5VGoKi8?Tu242Kpxm$l>YyX(^)1x(1H! zUHHM${$S$?4m?ra+83b%wk^yNEr0CEgzWxS+gcw(nx80>u2Jur+Z>}ircA?xb(QuV zP3c}sd2elpR88hUo?~cJz=hmB*?w(R;)@~eAm`&Vkwl!7J&QCP)GNWy)hJ=>t}Upy z74*%H$WC3qP-S6F^W~9pfi`3Ii$kOUoMAcZk;)Ii9VkD-)mg9iDYnR5Wpo%R`Qgye zWN&h)P~)+`O2tG1b{v3u+a`+>ya?>6y#}kvf}NEi07&1l^y#xOJ|m?frTOpIUZ$}S zJw7x%XMU2yP1^!m*zayoAktMbO5`oNBhWeRiTsQMq=!H{_ppowdU7S4TAzC zz60wp3z)yBB5}yr`E`CVDY@$DIXNGpP%JDMIN>%ECOxmS$`HloqPd*pJlV~PY ztDVKKmTFH4k!fW_e9}S^VzXIPV0Ab3Nf4?0spU~665ccFN^g%bVmq0zykz)VE@^S9 z9EkqaDEg4^C^dGs5iu_3(tOf}!kZ=qto+a%*rRxE`6o1oh(*k!-VwrfGBqgX46z^J zsx#lbtZ797A{^4W)32=SGyK7vC zu+#bJmA0vPVD1rjz9%M(y$QWVksJ9?1TRG@j;nfiV`(I-=7kMR^^Nl4){n(9!??Se zw!O|Nu8rooK2=Y~tY?Mf!UBIqm&hOs)Ulod0rm_5K*CnBd_bgzr<_W>)3lTEAj@i4 z0&W?>X5<#v1nvTZKdHQXigl{BzT=pzA7yu6d7gHf&;#st-fr}JLfpHH)BCMAe#I0m z)WG4Wp4FJ_YzE)oYrBy0Ai{lbWKWUJaI&#e#rvuo)4?J^&7uHFKxPN(ov%Z+u}w$J z=cV=TXo9@`^)AC@S=iYGsfz**wOQ%kZGO+&6}TK!?0i}L?>}M@_Pg<<{I%l=9bb2T4YWrJ_5d)WvN%Ci&MB5Gs)J?B8uVdUB zX^yN?&j*gpFF~Myz&G7@u@BmzfK~DXrr=J`9jw~DQi7Wtpbtj#V7|+w_VMXrVlk## z-P--`?ZrEZQh8P@CsnBI?uHJ+CIQKvLo6Eix^KF52%Rw2n0Hcjm9t9uMof0^jh6HY zG5xYA!KMw~zi_8{p6@jT!u38sp3;2Dc_k$ERJnQxRUlC+?vzh>e`BdK*J}DxAbr=s%8g6g4;)KIR~rTK?+o$rnH&S+Hw zdUF6hRtdlcUjPT#_2a?nMw6BR475S1^94YlP@xH0MQGH-hT+gi4ujfvinpf#&^LT? z_ygQE)fw=7)gVT(-)B`W8fP_4s)6xA$adBE#5;DG^^Xd_fEMFT+D+c4Rsbh#Gj<99 z#$>A=wP;lQt8c`WG_6-UZVru2$`T5JW|BolQy_d6xCVf{XKTqnFeW{^5rt0PSOc$x z0R3U2L&uFL`L`{Jy9QP8ZSnABw+)Js*z#uD%Y zK;kM0QIr933L;Aaj+?R-T_bJm{0gjzK^F)-wyTuByavL7PWo&s<;%?mC0>UBl1vI; zYd719a15FycC3c+ufG9s_)im8_j4dP-rlPf#u)%!aSrF3uttqQ4UE&X51RU@^8-Ww zU~GNV@^$x0Ny)CkO_t>H0$of0h{f#Gk<60Hldihzl&*dx@;7PAZ>d*?hK0r!ysl(s z`D(GTBK-~oTNr_^6ZCnnyk6rbff}DkPUrf|Jm(j<0+91Qjln&Zh zm%7+EFmq336mj=J{_g+snh%5LCa+U!G z9!+cjd&v4As(3aaP@g5dIiwvpM}&U)30!7~CA-2FcijbX4e$yff{?2HfyBy}E9-94 z&Hgr8d3$y#{gJx>zMNQBwNE7{x(p(o6)WRLnZl<_ifbT#fD9KC2BHDHHRNUbnj>!Q zktxn)L*83xEBiZrgBt~TRy0Fp8k&pN-e0|n5P-{GEeLOd#$7uov9J1Wo7%;4ngvy5 zJ^A-S1>3^+>tN?)GG0WhM?83)o9QqLqItdqH+&2%P7mW!kAv_KaQzD*bm$kseZ7^j z>L+9M770h~PBO34XI|K#jDy-3-?0O{{Qm8Mo}O=0nDZ9XM|IDi;x>dT@~aF_0I2I& zbM)ZS4FD(R^5~q%b(v2vL(EWK!4Dpg%{9`H*5)*N%4+*shl1Q8#CrN!dljlDo6=}j zctH0({Icu2I=tSVRJyTmLqmwUUK>j~U+!ec{UN~gDI}2?ko^Hbn+Vmz5@{0c&c$AO zE+_Niv8K*`;SvBiP4&4>OXz~Va$CS{QQ19 z`0Bnno*6R8Z2IZ<1S8{}uhm2i43wN%OyuXzwD|L-e$v#aSL*5SHwK~h7pL>QO(@Ue z&qMaI783hm_o^~`vbu|AgY2Wt(|gNoeyExeCcM@^+q>^np}d!A7N*w`W9iNXf?jVv zE;pTz>6Pp4i5FB0{SNp6X`n#!x1=F_jv)8|Q9?5>4h^A>RiT@2DYn7ft$(67*DS{K ztF;rT4GGr=nFLW=6$Y%vMK{LXS9z!RUg*37v`3Tav#95u{cyXkgiLjG-4wdjGX?ui zo^)Jyn7$XxqU;J0u`b^iuinGBafhR>?xe%Q&7RD8cy}nNE6I$_Em1z+X#GLi-Xom+ zQC6)#d2DHDf|xnY^yY)1dH~T4*E)Sr-_I1C?|`yf1=T$uN~xj^LSEEzve-0)St`rZ zm%>2RXI?gK@Hs!*tDidkLa&%=1=-`iNMkNav$D38v*3<=PD5XZM#+xwG{j9GM$h z;HV-Ix+{vx!xSZsH+nNu4|iV0^JUn=zUXPLId6TOXa1TKXtHp5q=A)!NAqT#U5{zr z_b5b1U#+kp3&lSa#Ly8J)IXi)i=$2y=8>QDqos+3=y)pJnfCie%;101iM$llrXazj zs)q~aWE=LF@F3>;_gV@hs_)O znQ(V8Z+7X2Vy(fhu9QT^ZO}ylO>w2-|)TrKgZE&}V?3Dy2Cj?;- zjpB#0EkJlSw)`{TG_!J;??jm#a?l-(YnuB?(UevDMYkH0xyPxnMuF9;>9m>rjBi(z zM>4$*zHn=5Y`IP>`6A~l@*1_jmLP2Kp`wA{eu0{5iMhoCl@|(T+Kyu>I$r&7IXx;? z%IV4no3C{fj+HxJBA%%ewd0%PWqPuE7REh!zkRG?)NeNM+Pz{idG}@G8mmIrxKzpvG25lLUxG2BbN1qHV_`z~XC(7o`00GRIuSzAc{o3??z><^q7RArF5 zzuelOx1@Xl?YPCoVTNbrdRH{)*)Ite+>g$x(r(NpOO19E#C!^%g zSR4`tiQ0X#N9X9QcykPG6(_l?IKQt;--er*DQpJ5$_$^HOdBys&RPS`A-XkREHlEz zk#e&=UajBxV&HMfA^{z0{PYX1OpdO_eJ06=OKepzYP{s#br4%P!v z#9jE`lLx-9mIpMSuY|?y9{><1(BOw6Ag}({^OaKICS(O)gV5iPYYMS->_Xwt<0JG% z&IiM8+`V<2bP47Ho_jQ2{z16M$6`_>$txgP*wa*flBD~HFs1KezW{L;{UVlve5RPc z;bkC}LjR5n;M*ww*yND*_tF4e>aNVp{rhDiIDoO<%!WMh+cE`f4YmVbyVv1$KGWYL zq`eP3syAq2PXDY%K_H7AJr;}LKdKB_1Hi#seWvusD1p>I90*u&C^+#SwQmM$|Nr(? b4}7%BKOiZi7=&H{|70W-?iGs}J^TLvsqpK+ diff --git a/docs/doc/design/pipe.html b/docs/doc/design/pipe.html deleted file mode 100644 index b3e343f..0000000 --- a/docs/doc/design/pipe.html +++ /dev/null @@ -1,383 +0,0 @@ - - - - - - - - - Piping Responses - Sqlflow - - - - - - - - - - - - - - - - - -

    - - - diff --git a/docs/doc/design/sql_parser.html b/docs/doc/design/sql_parser.html deleted file mode 100644 index fe0db48..0000000 --- a/docs/doc/design/sql_parser.html +++ /dev/null @@ -1,364 +0,0 @@ - - - - - - - - - Extended SQL Parser Design - Sqlflow - - - - - - - - - - - - - - - - - -
    -
    - Sqlflow - -
    - - -
    -
    -

    This site uses Just the Docs, a documentation theme for Jekyll.

    -
    -
    -
    -
    -
    - -
    -
    - - Search -
    -
    -
    - - - - -
    -
    -
    - - - - - -
    -

    Extended SQL Parser Design

    - -

    This documentation explains the technical decision made in building a SQL -parser in Go. It is used to parsed the extended SELECT syntax of SQL that -integrates TensorFlow Estimators.

    - - - -

    Lexer and Parser Generator

    - -

    In 2001, when I was in graduate school, I defined an extended SQL syntax for -querying distributed relational databases, as part of the course project of -Distributed RDBMS by Prof. Li-zhu Zhou. I wrote the parser using bison (a -modern version of yacc) and flex (a modern version of -lex). yacc and lex generate C code; -bison and flex generate C++ code. However, this time, I’d use Go.

    - -

    I surveyed goyacc, a -standard Go tool. The usage is very similar to that of yacc and bison. -However, the Go toolchain doesn’t provide a tool like lex/flex.

    - -

    Google revealed golex, which is out of -maintenance.

    - -

    The Medium -post -recommends Ragel, which is a C++ -program and could generate Go lexer; however, it lacks documents.

    - -

    Handwritten Lexer and Parser

    - -

    Some documents, including this -one -recommends handwriting lexers. However, it doesn’t explain how to write the -parser.

    - -

    GoAcademy always provides high-quality tech blog posts. This -one is from the -author of InfluxDB. However, I -stopped at where it explains wrapping a SQL statement as a string by an -io.Reader, because it is obvious that we should keep the string as a string so -that that token strings could refer to the same memory storage of the SQL -statement.

    - -

    Following a link in the above GoAcademy post, I found Rob Pike’s excellent talk -on how to write a lexer in Go in 2011. Many works after that change Rob’s -implementation somehow but always lead to longer and less comprehensible -codebases.

    - -

    The Choice

    - -

    Therefore, I wrote the lexer and parser both following Rob Pike’s idea. After -few days work, I realized that:

    - -
      -
    1. I should borrow the idea from Rob to represent SQL statements as strings, but -not io.Reader as other work do,
      2. -
    2. but no need to use channels and goroutines at all, and
      3. -
    3. it is technically intractable to write a SQL lexer/parser manually.
      4. -
    - -

    So, I switched to write a lexer manually, and to generate the parser using -goyacc. During my work, I referred to this -example and -the official yacc -manual for details -about operator association and precedence.

    - - - - -
    -
    -
    -
    - - - diff --git a/docs/doc/design/syntax.html b/docs/doc/design/syntax.html deleted file mode 100644 index 2ce44b3..0000000 --- a/docs/doc/design/syntax.html +++ /dev/null @@ -1,457 +0,0 @@ - - - - - - - - - SQLFlow: Design Doc - Sqlflow - - - - - - - - - - - - - - - - - -
    -
    - Sqlflow - -
    - - -
    -
    -

    This site uses Just the Docs, a documentation theme for Jekyll.

    -
    -
    -
    -
    -
    - -
    -
    - - Search -
    -
    -
    - - - - -
    -
    -
    - - - - - -
    -

    SQLFlow: Design Doc

    - -

    What is SQLFlow

    - -

    SQLFlow is a bridge that connects a SQL engine, for example, MySQL, Hive, SparkSQL, Oracle, or SQL Server, and TensorFlow and other machine learning toolkits. SQLFlow extends the SQL syntax to enable model training and inference.

    - - - -

    We could write simple machine learning prediction (or scoring) algorithms in SQL using operators like DOT_PRODUCT. However, this requires copy-n-pasting model parameters from the training program into SQL statements.

    - -

    Some proprietary SQL engines provide extensions to support machine learning.

    - -

    Microsoft SQL Server

    - -

    Microsoft SQL Server has the machine learning service that runs machine learning programs in R or Python as an external script:

    - -
    CREATE PROCEDURE generate_linear_model
-AS
-BEGIN
-    EXEC sp_execute_external_script
-    @language = N'R'
-    , @script = N'lrmodel <- rxLinMod(formula = distance ~ speed, data = CarsData);
-        trained_model <- data.frame(payload = as.raw(serialize(lrmodel, connection=NULL)));'
-    , @input_data_1 = N'SELECT [speed], [distance] FROM CarSpeed'
-    , @input_data_1_name = N'CarsData'
-    , @output_data_1_name = N'trained_model'
-    WITH RESULT SETS ((model varbinary(max)));
-END;
-
    - -

    A challenge to the users is that they need to know not only SQL but also R or Python, and they must be capable of writing machine learning programs in R or Python.

    - -

    Teradata SQL for DL

    - -

    Teradata also provides a RESTful service, which is callable from the extended SQL SELECT syntax.

    - -
    SELECT * FROM deep_learning_scorer(
-  ON (SELECT * FROM cc_data LIMIT 100)
-  URL('http://localhost:8000/api/v1/request')
-  ModelName('cc')
-  ModelVersion('1')
-  RequestType('predict')
-  columns('v1', 'v2', ..., 'amount')
-)
-
    - -

    The above syntax couples the deployment of the service (the URL in the above SQL statement) with the algorithm.

    - -

    Google BigQuery

    - -

    Google BigQuery enables machine learning in SQL by introducing the CREATE MODEL statement.

    - -
    CREATE MODEL dataset.model_name
-  OPTIONS(model_type='linear_reg', input_label_cols=['input_label'])
-AS SELECT * FROM input_table;
-
    - -

    Currently, BigQuery only supports two simple models: linear regression and logistic regression.

    - -

    Design Goal

    - -

    None of the above meets our requirement.

    - -

    First of all, we want to build an open source software. Also, we want it to be extensible:

    - -
      -
    • -

      We want it extensible to many SQL engines, instead of targeting any one of them. Therefore, we don’t want to build our syntax extension on top of user-defined functions (UDFs); otherwise, we’d have to implement them for each SQL engine.

      -
      • -
    • -

      We want the system extensible to support sophisticated machine learning models and toolkits, including TensorFlow for deep learning and xgboost for trees.

      -
      • -
    - -

    Another challenge is that we want SQLFlow to be flexible enough to configure and run cutting-edge algorithms, including specifying feature crosses. At the same time, we want SQLFlow easy to learn – at least, no Python or R code embedded in the SQL statements, and integrate hyperparameter estimation.

    - -

    We understand that a key to address the above challenges is the syntax of the SQL extension. To craft a highly-effective and easy-to-learn syntax, we need user feedback and fast iteration. Therefore, we’d start from a prototype that supports only MySQL and TensorFlow. We plan to support more SQL engines and machine learning toolkits later.

    - -

    Design Decisions

    - -

    As the beginning of the iteration, we propose an extension to the SQL SELECT statement. We are not going a new statement way like that BigQuery provides CREATE MODEL, because we want to maintain a loose couple between our system and the underlying SQL engine, and we cannot create the new data type for the SQL engine, like CREATE MODEL requires.

    - -

    We highly appreciate the work of TensorFlow Estimator, a high-level API for deep learning. The basic idea behind Estimator is to implement each deep learning model, and related training/testing/evaluating algorithms as a Python class derived from tf.estimator.Estimator. As we want to keep our SQL syntax simple, we would make the system extensible by calling estimators contributed by machine learning experts and written in Python.

    - -

    The SQL syntax must allow users to set Estimator attributes (parameters of the Python class’ constructor, and those of train, evaluate, or predict). Users can choose to use default values. We have a plan to integrate our hyperparameter estimation research into the system to optimize the default values.

    - -

    Though estimators derived from tf.estimator.Estimator run algorithms as TensorFlow graphs; SQLFlow doesn’t restrict that the underlying machine learning toolkit has to be TensorFlow. Indeed, as long as an estimator provides methods of train, evaluate, and predict, SQLFlow doesn’t care if it calls TensorFlow or xgboost. Precisely, what SQLFlow expect is an interface like the following:

    - -
    class AnEstimatorClass:
-  __init__(self, **kwargs)
-  train(self, **kwargs)
-  evaluate(self, **kwargs)
-  predict(self, **kwargs)
-
    - -

    We also want to reuse the feature columns API from Estimator, which allows users to columns of tables in a SQL engine to features to the model.

    - -

    Extended SQL Syntax

    - -

    Again, just as the beginning of the iteration, we propose the syntax for training as

    - -
    SELECT * FROM kaggle_credit_fraud_training_data
-LIMIT 1000
-TRAIN DNNClassifier       /* a pre-defined TensorFlow estimator, tf.estimator.DNNClassifier */
-WITH layers=[100, 200],   /* a parameter of the Estimator class constructor */
-     train.batch_size = 8 /* a parameter of the Estimator.train method */
-COLUMN *,                 /* all columns as raw features */
-       cross(v1, v9, v28) /* plus a derived (crossed) column */
-LABEL class
-INTO sqlflow_models.my_model_table;      /* saves trained model parameters and features into a table */
-
    - -

    We see the redundancy of * in two clauses: SELECT and COLUMN. The following alternative can avoid the redundancy, but cannot specify the label.

    - -
    SELECT *                  /* raw features or the label? */
-       corss(v1, v9, v28) /* derived featuers */
-FROM kaggle_credit_fraud_training_data
-
    - -

    Please be aware that we save the trained models into tables, instead of a variable maintained by the underlying SQL engine. To invent a new variable type to hold trained models, we’d make our system tightly integrated with the SQL engine, and harms the extensibility to other engines.

    - -

    The result table should include the following information:

    - -
      -
    1. The estimator name, e.g., DNNClassifier in this case.
      2. -
    2. Estimator attributes, e.g., layer and train.batch_size.
      3. -
    3. The feature mapping, e.g., * and cross(v1, v9, v28).
      4. -
    - -

    Similarly, to infer the class (fraud or regular), we could

    - -
    SELECT * FROM kaggle_credit_fraud_development_data
-PREDICT kaggle_credit_fraud_development_data.class
-USING sqlflow_models.my_model_table;
-
    - -

    System Architecture

    - -

    A Conceptual Overview

    - -

    In the prototype, we use the following architecture:

    - -
    SQL statement -> our SQL parser --standard SQL-> MySQL
-                                \-extended SQL-> code generator -> execution engine
-
    - -

    In the prototype, the code generator generates a Python program that trains or predicts. In either case,

    - -
      -
    1. it retrieves the data from MySQL via MySQL Connector Python API,
      2. -
    2. optionally, retrieves the model from MySQL,
      3. -
    3. trains the model or predicts using the trained model by calling the user specified TensorFlow estimator,
      4. -
    4. and writes the trained model or prediction results into a table.
      5. -
    - -

    Working with Jupyter Notebook and Kubernetes

    - -

    The following figures shows the system components and their runtime environment. The left part shows how to run the system on a PC/laptop, the right part shows how to run it on a Kubernetes cluster.

    - -

    - - - - -
    -
    -
    -
    - - - diff --git a/docs/doc/design/walkthrough.html b/docs/doc/design/walkthrough.html deleted file mode 100644 index e10f61d..0000000 --- a/docs/doc/design/walkthrough.html +++ /dev/null @@ -1,371 +0,0 @@ - - - - - - - - - SQLFlow: Code Walkthrough - Sqlflow - - - - - - - - - - - - - - - - - -
    -
    - Sqlflow - -
    - - -
    -
    -

    This site uses Just the Docs, a documentation theme for Jekyll.

    -
    -
    -
    -
    -
    - -
    -
    - - Search -
    -
    -
    - - - - -
    -
    -
    - - - - - -
    -

    SQLFlow: Code Walkthrough

    - -

    User Experience

    - -

    SQLFlow allows users to write SQL programs with extended syntax in Jupyter Notebook or a command-line tool.

    - -

    The following SQL statements train a TensorFlow model named DNNClassifier, which is a Python class derived from tf.estimator.Estimator:

    - -
    
-SELECT * FROM a_table TRAIN DNNClassifier WITH learning_rate=0.01 INTO sqlflow_models.my_model;
-
    - -

    And the following statement uses the trained model for prediction.

    - -
    SELECT * FROM b_table PREDICT b_table.predicted_label USING sqlflow_models.my_model;
-
    - -

    Please be aware that the part in the above statements before the extended keyword TRAIN and PREDICT is a standard SQL statement. This feature simplifies the implementation of the SQLFlow system.

    - -

    System Implementation

    - -

    If a SQL statement is of the standard syntax, SQLFlow throws it to the SQL engine and redirects the output to the user; otherwise, SQLFlow translates the statement of extended syntax into a Python program. Currently, it generates a program that throws the standard-syntax part of SELECT to MySQL, reads the results in the train-loop of a TensorFlow program. We will talk about how to extend SQLFlow to connect more SQL engines like Oracle, Hive, and SparkSQL, and generates more types of machine learning programs that calls distributed TensorFlow, PyTorch, and xgboost later. Before that, let us explain the system components.

    - -

    SQLFlow as a gRPC Server

    - -

    SQLFlow is a gRPC server, which can connect with multiple clients. A typical client is pysqlflow, the SQLFlow plugin for Jupyter Notebook server. Another once is a text-based client /cmd/sqlflowserver/main.go.

    - -
    Jupyter Notebook          ---(SQL statements)-->       SQLFlow gRPC server
-(SQLFlow magic command)   <--(a stream of messages)--
-
    - -

    The protobuf definition of the gRPC service is at /server/proto/sqlflow.proto. The return of the method SQLFlow.Run is a stream of Reponses, where each represents either a table header, a row, or a log message. The header and rows are usually from a standard SQL statement, for example, SELECT or DESCRIBE, and the log messages are usually from the running of a generated Python program.

    - -

    SQLFlow in the gRPC Server

    - -

    Once the SQLFlow server receives a batch of SQL statements via a gRPC call, it runs the following steps for each statement:

    - -
      -
    1. the parser to generate parsing result,
      2. -
    2. the verifier to verify the semantics given the parsing result,
      3. -
    3. the code generator to generate a Python program, or the submitter, from the parsing result,
      4. -
    4. the executor that runs the submitter locally.
      5. -
    - -

    Step 3. and 4. are only for a SQL statement of extended syntax; otherwise, SQLFlow server proxies the standard-syntax statement to the SQL engine.

    - -

    The executor calls Go’s standard package that captures the stdout and stderr from the submitter process and passing the result back to the gRPC client. Therefore, it is the responsibility of the submitter to print log messages to its stderr and stdout.

    - -

    Minimal Viable Product

    - -

    In the minimal viable product (MVP) of SQLFlow, the code generator generates a Python program consists of two parts:

    - -
      -
    1. throw the standard SELECT part in the extended-syntax statement to MySQL via ODBC, and
      2. -
    2. a loop that reads outputs from the run of the SELECT statement and trains the model (or, using a trained model to predict).
      3. -
    - -

    The training part calls TensorFlow to update the parameters of the model specified in the TRAIN clause.

    - -

    Extensibility

    - -

    By writing more code generators, we could extend SQLFlow to support more SQL engines, e.g., Hive and Oracle, and use machine learning toolkits, e.g., PyTorch and xgboost, in addition to TensorFlow, on various computing platforms. You are welcome to add more code generators such as

    - -
      -
    • codegen_distributed_tf.go to generate a submitter program similar to the MVP but runs a distributed TensorFlow training job.
      • -
    • codegen_kubernetes_tf.go to launch a distributed TensorFlow job on a Kubernetes cluster, other than running locally, in the same container as where SQLFlow gRPC server resides.
      • -
    • codegen_gcloud_pytorch.go to launch a submitter that calls PyTorch instead of TensorFlow for training on the Google Cloud.
      • -
    - -

    Job Management

    - -

    The current design of the gRPC interface assumes that the connection between the client, e.g., the Jupyter Notebook, and the SQLFlow server keeps alive during the running of the training program. This assumption is reasonable because even if the user closes her/his Web browser and disconnect to the Jupyter Notebook server, the connection between Jupyter to SQLFlow server might keep alive. However, this might not be robust enough if the Jupyter Notebook server runs on a user’s laptop and gets killed. In such a case, the gRPC server cannot stream the messages back to the client and would cause the failure of the submitter.

    - -

    A solution is to change the gRPC interface of SQLFlow server to have a method that files a job and returns immediately, and another method to get a batch of recent messages given a job ID. We will make a design for that soon.

    - - - - -
    -
    -
    -
    - - - diff --git a/docs/feed.xml b/docs/feed.xml index d8c18cf..963dd98 100644 --- a/docs/feed.xml +++ b/docs/feed.xml @@ -1 +1 @@ -Jekyll2019-04-18T16:37:07+08:00/feed.xmlSQLFLOW +Jekyll2019-04-20T11:30:34+08:00/feed.xmlSQLFLOW \ No newline at end of file diff --git a/docs/index.html b/docs/index.html index 48ee8b8..4498a1e 100644 --- a/docs/index.html +++ b/docs/index.html @@ -942,7 +942,7 @@

    Extends SQL to support AI. Extract knowledge from Data.

    sqlflow> 
    -
    - Sqlflow - -
    - - -
    -
    -

    This site uses Just the Docs, a documentation theme for Jekyll.

    -
    -
    -
    -
    -
    - -
    -
    - - Search -
    -
    -
    - - - - -
    -
    -
    - - - - - -
    -

    Piping Responses

    - -

    Streaming Responses

    - -

    As described in the overall design, a SQLFlow job could be a standard or an extended SQL statemnt, where an extended SQL statement will be translated into a Python program. Therefore, each job might generate up to the following data streams:

    - -
      -
    1. standard output, where each element is a line of text,
      2. -
    2. standard error, where each element is a line of text,
      3. -
    3. data rows, where the first element consists of fields name/types, and each of the rest is a row of data,
      4. -
    4. status, where the element could be pending, failed, and succeeded.
      5. -
    - -

    To create good user experience, we need to pipe these responses from SQLFlow jobs to Jupyter Notebook in real-time.

    - -

    Stages in the Pipe

    - -

    The pipe that streams outputs from SQLFlow jobs to the Jupyter Notebook consists of the following stages:

    - -
    Web browser 
- ↑
- | HTTP
- ↓
-Jupyter Notebook server
- ↑
- | ZeroMQ streams: Shell, IOPub, stdin, Controls, Heartbeat
- ↓
-iPython kernel
- ↑
- | IPython magic command framework
- ↓
-SQLFlow magic command for Jupyter
- ↑
- | gRPC
- ↓
-SQLFlow server
- ↑
- | Go channels
- ↓
-SQLFlow job manager (Go functions)
- ↑
- | IPC with Go's standard library
- ↓ 
-SQLFlow jobs
-
    - -

    In the above figure, from the SQLFlow magic command to the bottom layer are our work.

    - -

    Streaming

    - -

    We have two alternative ideas: multiple streams and a multiplexing stream. -We decided to use a multiplexing stream because we had a unsuccessful trial with the multiple streams idea: we make the job writes to various Go channels and forward each Go channel to a streaming gRPC call, as the following:

    - -

    Multiple streams

    - -

    The above figure shows that there are multiple streams between the Jupyter Notebook server and Jupyter kernels. According to the document, there are five: Shell, IOPub, stdin, Control, and Heartbeat. These streams are ZeroMQ streams. We don’t use ZeroMQ, but we can take the idea of having multiple parallel streams in the pipe.

    - -
    service SQLFlow {
-    rpc File(string sql) returns (int id) {}
-
-    rpc ReadStdout(int id) returns (stream string) {}
-    rpc ReadStderr(int id) returns (stream string) {}
-    rpc ReadData(int id) returns (stream Row) {}
-    rpc ReadStatus(int id) returns (stream int) {}
-}
-
    - -

    However, we realized that if the user doesn’t call any one of the SQLFlow.Read... call, there would be no forwarding from the Go channel to Jupyter, thus the job would block forever at writing.

    - -

    A Multiplexing Stream

    - -

    Another idea is multiplexing all streams into one. For example, we can have only one ZeroMQ stream, where each element is a polymorphic type – could be a text string or a data row.

    - -
    service SQLFlow {
-    rpc Run(string sql) returns (stream Response) {}
-}
-
-// Only one of the following fields should be set.
-message Response {
-    oneof record {
-        repeated string head = 1;             // Column names.
-        repeated google.protobuf.Any row = 2; // Cells in a row.
-        string log = 3;                       // A line from stderr or stdout.
-    }
-}
-
    - - - - -
    -
    -
    -

    - +

    Easy to Learn

    @@ -951,7 +951,7 @@

    Easy to Learn

    -->

    Manipulate data and running AI with SQL

    - +

    Work with Many Database Management Systems

    @@ -960,7 +960,7 @@

    Work with Many Database Management Systems

    -->

    MySQL, Hive, Alibaba ODPS, Oracle and you name it!

    - +

    Support Many Machine Learning Toolkits

    @@ -969,7 +969,7 @@

    Support Many Machine Learning Toolkits

    -->

    TensorFlow, PyTorch, xgboost, and more!

    - +
    diff --git a/docs/pages/about.html b/docs/pages/about.html index 97f1d2d..f9a3302 100644 --- a/docs/pages/about.html +++ b/docs/pages/about.html @@ -1,353 +1,348 @@ - - - - - - About Us - SQLFLOW - + + + + About Us - SQLFLOW + + + + + + + - - - - - - + -
    -
    - SQLFLOW - -
    - - +
    +
    + SQLFLOW + +
    + + +
    +
    +

    This site uses Just the Docs, a documentation theme for Jekyll.

    +
    +
    +
    +
    +
    + +
    +
    + + Search
    -
    -

    This site uses Just the Docs, a documentation theme for Jekyll.

    -
    +
    +
    + + + +
    -
    -
    -
    - -
    -
    - - Search -
    -
    -
    - - - - -
    -
    -
    - - - -
    -

    About

    -

    Add info about the product.

    +
    +
    + + + +
    +

    About

    +

    Add info about the product.

    -

    Add info about the team.

    +

    Add info about the team.

    -

    Add info about the contact.

    +

    Add info about the contact.

    - -
    -
    +
    +
    +
    - - \ No newline at end of file + diff --git a/docs/robots.txt b/docs/robots.txt index d297064..e087884 100644 --- a/docs/robots.txt +++ b/docs/robots.txt @@ -1 +1 @@ -Sitemap: http://localhost:4000/sitemap.xml +Sitemap: /sitemap.xml diff --git a/docs/sitemap.xml b/docs/sitemap.xml index adac8c6..273a020 100644 --- a/docs/sitemap.xml +++ b/docs/sitemap.xml @@ -1,7 +1,7 @@ -http://localhost:4000/pages/about.html +/pages/about.html /doc_index/contribute.html @@ -13,6 +13,9 @@ / +/sqlflow/doc/quickstart.html + + /sqlflow/doc/build-tensorflow.html @@ -31,9 +34,6 @@ /sqlflow/doc/pipe.html -/sqlflow/doc/quickstart.html - - /sqlflow/doc/sql_parser.html @@ -74,30 +74,30 @@ /gohive/git.apache.org/thrift.git/lib/go/thrift/ -2019-04-18T15:54:36+08:00 +2019-04-20T11:20:34+08:00 /gohive/git.apache.org/thrift/lib/go/thrift/ -2019-04-18T15:54:36+08:00 +2019-04-20T11:20:34+08:00 /gohive/hiveserver2/ -2019-04-18T15:54:36+08:00 +2019-04-20T11:20:34+08:00 /gohive/ -2019-04-12T10:17:18+08:00 +2019-04-20T11:08:15+08:00 /gohive/service-rpc/gen-go/tcliservice/ -2019-04-12T10:17:18+08:00 +2019-04-20T11:08:15+08:00 /gohive/thrift/ -2019-04-18T15:54:36+08:00 +2019-04-20T11:20:34+08:00 /gohive/vendor/git.apache.org/thrift.git/lib/go/thrift/ -2019-04-12T10:17:18+08:00 +2019-04-20T11:08:15+08:00 diff --git a/docs/sqlflow/doc/contribute/build-tensorflow.md b/docs/sqlflow/doc/contribute/build-tensorflow.md deleted file mode 100644 index a4329cb..0000000 --- a/docs/sqlflow/doc/contribute/build-tensorflow.md +++ /dev/null @@ -1,140 +0,0 @@ ---- -parent: Contribute -nav_order: 1 ---- - -# Build TensorFlow from Source Code using Docker - -To contribute to TensorFlow, we need to build TensorFlow from source code. The [official guide](https://www.tensorflow.org/install/source) is great. However, it interleaves the native building process and that using Docker and makes it confusing because packages needed by the former are not by the latter. Also, we found some useful tricks to start Docker containers in practices that are not in the official guide. Hence this document. - ----- - -## Build the Pip Package in TensorFlow Development Container - -On either Mac or Linux, or any other OS, we don't have to install and configure the building tools; instead, we can use a Docker image where all tools have been installed and properly configured. - -1. Get the Docker image containing all the building tools: - - ```bash - docker pull tensorflow/tensorflow:latest-devel - ``` - -1. Then, let's get the source code. On any OS, please install git using the native package manager. For example, on Ubuntu, please - - ```bash - sudo apt-get install git - ``` - - or, on Mac, - - ```bash - brew install git - ``` - - Then, use the git just installed, let's clone tensorflow source code: - - ```bash - git clone --recursive https://github.com/tensorflow/tensorflow - cd tensorflow - ``` - - By default, we will be on the `master` branch. Feel free to do you change in your feature branches, or switch to a release branch, for example: - - ```bash - git checkout v1.11.0 - git checkout -b v1.11.0 - ``` - -1. Then, let us start a Docker container running the `tensorflow/tensorflow:latest-devel` image: - - ```bash - docker run --rm -it \ - -w /tensorflow \ - -v $PWD:/tensorflow \ - -v $HOME/.cache:/root/.cache \ - -e "HOST_PERMS=$(id -u):$(id -g)" \ - tensorflow/tensorflow:latest-devel \ - /bin/bash - ``` - - - `-w /tensorflow` brings us to the `/tensorflow` directory in the container once after we start it. - - `-v $PWD:/tensorflow` maps the current directory, which is the just cloned TensorFlow source directory on the host, to `/tensorflow` in the container. - - `-v $HOME/.cache:/root/.cache` maps the Bazel temporary directory on the host into the container, so the intermediate files generated by Bazel running in the container are actually saved on the host. This allows us to interrupt the container during the build and restart it later to resume the building. - - `e "HOST_PERMS=$(id -u):$(id -g)"` passes the user identification on the host into the container as an environment variable. We can reset the mode of files generated in the container to this user identity. - - -1. From now on, we will be working in the container. Let us first configure the building: - - ```bash - ./configure - ``` - - Usually, I would simply choose all the default options by hitting enter all the way down. - -1. Build the pip package builder: - - ```bash - bazel build --config=opt //tensorflow/tools/pip_package:build_pip_package - ``` - -1. Build the pip package and save it into `/tensorflow`: - - ```bash - ./bazel-bin/tensorflow/tools/pip_package/build_pip_package /tensorflow - ``` - -1. Change the generated wheel file into the usual mode as on the host: - - ```bash - chown $HOST_PERMS /mnt/tensorflow-*.whl - ``` - -# Install the Pip Pacakge in TensorFlow Development Container - -1. Let's now try the new pip package. First, we need to uninstall the current tensorflow pip package, then we install the newly built one: - - ```bash - pip uninstall tensorflow - pip install /tensorflow/tensorflow-*.whl - ``` - -1. Now, we can verify if the new package works. First, we need to switch to a directory out from `/tensorflow`, so we don't import from the source directory: - - ```bash - cd /tmp # other directories also work - ``` - - then, we can try import the newly installed tensorflow package and verify it: - - ```bash - python - >>> import tensorflow - >>> print(tensorflow.__version__) - ``` - -1. Now, let us quit from Python and from the Docker container. We should see the `tensorflow-*.whl` file on the host in the current directory. - - -## Install the Pip Package in a Clean Python Package - -After we quit from the development container, we should see the wheel file in the TensorFlow source directory on the host. Now, we can start a Python container and install the Pip package in it. - -1. Start the Python container - - ```bash - docker run --rm -it -v $PWD:/tensorflow python:2.7 bash - ``` - -1. Install the pip pacakge in the container - - ```bash - pip install /tensorflow/tensorflow*.whl - ``` - -1. Try TensorFlow by starting Python in the container - - ```bash - python - >>> import tensorflow as tf - >>> print(tf.__version__) - ``` diff --git a/docs/sqlflow/doc/contribute/build.md b/docs/sqlflow/doc/contribute/build.md deleted file mode 100644 index 597bede..0000000 --- a/docs/sqlflow/doc/contribute/build.md +++ /dev/null @@ -1,133 +0,0 @@ ---- -parent: Contribute -nav_order: 0 ---- - -# Canonical Development Environment - -Referring to [this example](https://github.com/sql-machine-learning/canonicalize-go-python-grpc-dev-env), -we create a canonical development environment for Go and Python programmers using Docker. - -## Editing on Host - -When we use this Docker image for daily development work, the source code relies -on the host computer instead of the container. The source code includes this repo -and all its dependencies, for example, the Go package `google.golang.org/grpc`. -Code-on-the-host allows us to run our favorite editors (Emacs, VIM, Eclipse, and more) -on the host. Please free to rely on editors add-ons to analyze the source code -for auto-completion. - -## Building in Container - -We build a Docker image that contains development tools: - -1. The Python interpreter -1. The Go compiler -1. The protobuf compiler -1. The protobuf to Go compiler extension -1. The protobuf to Python compiler extension - -Because this repo contains Go code, please make sure that you have the directory -structure required by Go. On my laptop computer, I have - -```bash -export GOPATH=$HOME/go -``` - -You could have your `$GOPATH` pointing to any directory you like. - -Given `$GOPATH$` set, we could git clone the source code of our project by running: - -```bash -go get github.com/sql-machine-learning/sqlflow -``` - -Change the directory to our project root, and we can use `go get` to retrieve -and update Go dependencies. - -```bash -cd $GOPATH/src/github.com/sql-machine-learning/sqlflow -go get -u -t ./... -``` - -Note `-t` instructs get to also download the packages required to build -the tests for the specified packages. - -As all Git users would do, we run `git pull` from time to time to sync up with -others' work. If somebody added new dependencies, we might need to run `go -u ./...` -after `git pull` to update dependencies. - -To build this project, we need the protobuf compiler, Go compiler, Python interpreter, -gRPC extension to the protobuf compiler. To ease the installation and configuration -of these tools, we provided a Dockerfile to install them into a Docker image. -To build the Docker image: - -```bash -docker build -t sqlflow:dev -f Dockerfile.dev . -``` - -## Development - -### Build and Test - -We build and test the project inside the docker container. - -To run the container, we need to map the `$GOPATH` directory on the host into the -`/go` directory in the container, because the Dockerfile configures `/go` as -the `$GOPATH` in the container: - -```bash -docker run --rm -it -v $GOPATH:/go \ - -w /go/src/github.com/sql-machine-learning/sqlflow \ - sqlflow:dev bash -``` - -Inside the Docker container, start a MySQL server in the background - -``` -service mysql start& -``` - -run all the tests as - -``` -go generate ./... -go install ./... -go test -v ./... -``` - -where `go generate` invokes the `protoc` command to translate `server/sqlflow.proto` -into `server/sqlflow.pb.go` and `go test -v` builds and run unit tests. - - -### Release - -The above build process currently generates two binary files in -`$GOPATH/bin` on the host. To package them into a Docker image, -please run - -```bash -docker build -t sqlflow -f ./Dockerfile $GOPATH/bin -``` - -To publish the released Docker image to our official DockerHub -```bash -docker tag sqlflow sqlflow/sqlflow:latest -docker push sqlflow/sqlflow:latest -``` - -## Demo: Command line Prompt - -The demo requires a MySQL server instance with populated data. If we don't, we could -follow [example/datasets/README.md](/example/datasets/README.md) to start one on the host. -After setting up MySQL, run the following inside the Docker container - -```bash -go run cmd/demo/demo.go --db_user root --db_password root --db_address host.docker.internal:3306 -``` - -You should be able to see the following prompt - -``` -sqlflow> -``` diff --git a/docs/sqlflow/doc/contribute/close_producer_from_consumer.md b/docs/sqlflow/doc/contribute/close_producer_from_consumer.md deleted file mode 100644 index a598521..0000000 --- a/docs/sqlflow/doc/contribute/close_producer_from_consumer.md +++ /dev/null @@ -1,166 +0,0 @@ ---- -parent: Contribute -nav_order: 3 ---- - -# Closing the producer goroutine from the consumer - -The producer-and-consumer pattern is well used in Go concurrent programming. When -the consumer stops, we want to gracefully stop the producer as well. - -## Problem - -When a gRPC server receives a streaming request, it usually calls a -[function that returns a channel](https://talks.golang.org/2012/concurrency.slide#25), -reads the result from that channel and send the result to the client one by one. - -Take the following code for instance: upon receiving a request, the main goroutine -`Service` calls `launchJob`. `launchJob` starts a separate goroutine as an anonymous -function call and returns a channel. In the anonymous function, items will be sent to -channel. And `Service` on the otherside of the channel will reads from it. - -```go -func Service(req *Request, stream *StreamResponse) error { - result := launchJob(req.Content) - for r := range result { - if e := stream.Send(result); e != nil { - // should we signal the running goroutine so it will stop sending? - return e - } - } -} - -func launchJob(content string) chan Item { - c := make(chan Item) - - go func() { - defer close(c) - - acquireScarceResources() - defer releaseScarceResources() - - ... - // if stream.Send(result) returns an error and the Service returns, this will be blocked - c <- Item{} - ... - }() - - return c -} -``` - -There is a major problem in this implementation. As pointed out by the comment, -if the `Send` in `Service` returns an error, the `Service` function will return, -leaving the anonymous function being blocked on `c <- Item{}` forever. - -This problem is important because the leaking goroutine usually owns scarce system -resources such as network connection and memory. - - -## Solution: pipeline explicit cancellation - -Inspired by this [blog post](https://blog.golang.org/pipelines) section -*Explicit cancellation*, we can signal the cancellation via closing on a separate -channel. And we can follow the terminology as `io.Pipe`. - -```go -package sql - -import ( - "errors" -) - -var ErrClosedPipe = errors.New("pipe: write on closed pipe") - -// pipe follows the design at https://blog.golang.org/pipelines -// - wrCh: chan for piping data -// - done: chan for signaling Close from Reader to Writer -type pipe struct { - wrCh chan interface{} - done chan struct{} -} - -// PipeReader reads real data -type PipeReader struct { - p *pipe -} - -// PipeWriter writes real data -type PipeWriter struct { - p *pipe -} - -// Pipe creates a synchronous in-memory pipe. -// -// It is safe to call Read and Write in parallel with each other or with Close. -// Parallel calls to Read and parallel calls to Write are also safe: -// the individual calls will be gated sequentially. -func Pipe() (*PipeReader, *PipeWriter) { - p := &pipe{ - wrCh: make(chan interface{}), - done: make(chan struct{})} - return &PipeReader{p}, &PipeWriter{p} -} - -// Close closes the reader; subsequent writes to the -func (r *PipeReader) Close() { - close(r.p.done) -} - -// ReadAll returns the data chan. The caller should -// use it as `for r := range pr.ReadAll()` -func (r *PipeReader) ReadAll() chan interface{} { - return r.p.wrCh -} - -// Close closes the writer; subsequent ReadAll from the -// read half of the pipe will return a closed channel. -func (w *PipeWriter) Close() { - close(w.p.wrCh) -} - -// Write writes the item to the underlying data stream. -// It returns ErrClosedPipe when the data stream is closed. -func (w *PipeWriter) Write(item interface{}) error { - select { - case w.p.wrCh <- item: - return nil - case <-w.p.done: - return ErrClosedPipe - } -} -``` - -And the consumer and producer be can implemented as - -```go -func Service(req *Request, stream *StreamResponse) error { - pr := launchJob(req.Content) - defer pr.Close() - for r := range pr.ReadAll() { - if e := stream.Send(r); e != nil { - return e - } - } -} - -func launchJob(content string) PipeReader { - pr, pw := Pipe() - go func() { - defer pw.Close() - - if err := pw.Write(Item{}); err != nil { - return - } - } - return pr -} -``` - -## Further Reading - -1. [Google Form: Channel send timeout](https://groups.google.com/forum/#!topic/golang-nuts/Oth9CmJPoqo) -2. [Go by Example: Timeouts](https://gobyexample.com/timeouts) -3. [Google I/O 2013 - Advanced Go Concurrency Patterns](https://www.youtube.com/watch?v=QDDwwePbDtw&t=111s) -4. [Go Concurrency Patterns Talk](https://talks.golang.org/2012/concurrency.slide) -5. [Go Concurrency Patterns: Pipelines and cancellation](https://blog.golang.org/pipelines) diff --git a/docs/sqlflow/doc/contribute/figures/sqlflow-arch2.png b/docs/sqlflow/doc/contribute/figures/sqlflow-arch2.png deleted file mode 100644 index 92d017cd212691212a137c1ce6913bcb4131c1e1..0000000000000000000000000000000000000000 GIT binary patch literal 0 HcmV?d00001 literal 96987 zcmd>mRa{ls`!^y;36dfr4bouHNC`*a(A^;2UD6;30!m1CdFbv&L8QC8yQTYG#~FV! zbNOEXFZ|57!FBdtdp+@u=LE`1i(;ZbMn^zEz!Vn~kw-wds|fxCpxy&l$h#1y!GG@9 z%Zt87DC~W*fq)>0ATA=L`0mbj5=x>%>xI8SE?p2?&ijvV?|fy@${&~8)>gzH$kVfa zSF$)_e!4kVI9_u49{v3|a>-r2I~(cm$r;Gs+uR)X&XyUuaU>YXd@=THviM@Z+8yWe zl*#4vay5xT2>X>GKB^!BBDOaI5;?-1-~UU;>BAr`i@eepgz@LnpPzoqZT|1?f*;^g zd8a}@5vvIO_b249TaEvF#al2P>7y(SG-T^P%lahJdZ%`thk`x$`v!i0nvVz?c{VOh zo8!N4&0DY+Pf&K{N)!R%zn9aIuoVNpEb;v3S;*JG5_UIK2nheZEQg4iE7D<0^1mmB ze_+@lSc3Z&2jQQM_!?hJ~nvsDj%tun<6IB_nX{kRD_I*F8o0e$x#g zF_PgsR^%>BW%zTSIZtKNP@{O-y!LaFDJKKK%5O6T-pc*iwvVKDMr9$^((5>X{=x4( z3ZnoIM)X3?0q;ML0Dc=u>1frD^v~`5wfRJ^5rm>B@F1KN|9KhfID-E_OZ%WtWr|vh z9c8~l8qHVJ&kzl*G#$!}w<53Mr2qAxesiDNyY(rhc zVJ|O-QTL;*xu&L_y1m8^d9a=9wG=Na*2c5+0@ae{TE{ITaJ}v`F=uZG|HXRwu@CXT zo?4OR4y9sLSCjl+quB@pt86*@?zozqbM9p`4ntDf?WXz)4Lao}az0Lj=d0jU zoGcaUG&Pdge7nP8*oog~G+wN`v3>9K{Xa)E+X7JtIcRE_v)}hi98zKJW^St4NbdLd z$QbDN=YQbI-0$T(oknaUeK+s(%|qGOZMP=tn{7iJr?vLHA7nlqg52wIb-Mx77O!%A@XIa?n3AMf&Tgg*hOKWHV?CC@)?|m1d$w$e-$Uv5#pg zkZV~lG=Fd!RWm+6YQ&S~qNvREDp0Jq@8xdO_Bd9mJDqhq?Bg9+&>1V#;=NThWo^D_ z9m&zCD%7&;pt(^z-gJq}RmxSQFEr~6?MddFN;G)^6ZjduQ#sXpbKd8rY_`l)DNyhK z&o+tifzW^@+ZyZ@uO++153Ysqov6CqoUd6P_HdqhC_PR=T;w#N}tsLco_zwk6Ei1*tQe!#b^!1#dwZ2k(r^H=3wdg#NT!D6yk z!vCZy>Nnn~N_x|!i@9D`OJterUnF|U3bpE7?SpLIPFXj3oDF()j3p=+sIQPQUZsP; z)fTHhYNEs`YIf|py!nnoc0`#|zNPs%m5>qlWmd9lZYZf~L(C{ojz#r?zxK16EuGua zC`J|a7Jtl!I|1EX4w6Psbp7&;`;xV*#eU$cP}}wL?!CI)Dr&l_WsM0kne)1BC}O}e z<7IlW|A^b}tEp5RrRrT2tfT88mq4n(@t1wChN!#CJ*7P21iTve`XGc_2G;cl{-O7C zJ(GAG^No7qSLN9bTNNpH>V`$wMScGD;_vwcQR93W(!p^Zs6NR`@kG~LZYvVEOUe4i zy+)v>up$y8a0?brf(#N8rQ)vWb7CizkJhX=6 zRv6z!KJ@JGCu#Qjd zGs$jZgCoyR`}j^1z4p@0&mFavpN*Sjd`65>kM-(?+`g}|U%NOZuoP&L(>6f5<2h&a zgKa?kccE(+0%7+h4f?YB^PA6)&lO+7k}NAHj5s{6-jPWST;vw}nWsi_sHmg$Sr>o5 z2UB7^oCfD}o$6&!FI~J{D}hR+>*v9Z$i}R8d5a12q6YWLX^+#HwFfGB$~{?Bxi@_R zx7(rgn;DBpQC_1DN@SsL)gZecYKtz@x5gB?BhZ`tJFiI(Vsc5I6Y;xOheGM?PL@Ii zCY3cTR}|2da@5L<8h3qANY_)6o*ngFq+QrayDx`}UOk7gZKgR0M+H^x5ig+pQ^N#5 zpceb*zCJ5mTciqoHFz?v|3KmKydN4~J!b8XsG!DWbFD9<2vLffk?qk;YKt;e;KY74 zLn9fBrz9j{JeYj;L24^iuS=SDXL&AnPj}#g1&wW|0d_3S@@oZ?)1ZalepNDJ{Y&xv zq1)yysB1Jy{b(BB`Eo?2qX7$aW$deIL1ggPIC%u?rt6b0ET`sQ9pp{@K6Hif?wYv5 zH%42N@8Aa>vY$3&iq&0U_FsV(nIbq~ZM$Oa%lIPfTOzmpRc3jK*4QK&#EhDue86g5 zZ*6eToxMPTA!og-f$}9tA7wgQ3i1ujZG^1a7P$H_0<m3q0{E~&9@fuNnFJ-oM6?;#NVPv`{?$i42aUq>XLcK~9}$#nA6h?Y9xUr+ z;b!h8x&P{zeI!IFV&&Bu1)_GAtZ}=vD~MNX@j0wR{E_0{jQ^pp@!HE zP~o?Ahu%fzW|<`3r&zwA%@U3i5qspcJ6$7q-E3`8kCTb>GueKt?@FU0YcNNF*4o>4 zOw0NFbm0No2=8t+mAl4fnO{7I=BQ?e?Xzg57n7{VtBHR9gf(3R!uwoFkqI>+dQ0_` z42N1(Ihe^u`So&svEFwKrQ@DQ1EK}COFx|wCCnovv@OD`!$zu+#tQ3p z55c=0KZ%UbDuprjo>8t!5~8cH?(^e2_HKVCizyBx|C0JeSX*`0ZI5ryUxjOtdSR)X zvfHxXy!Tm)3Sj`5=h^q=aTR4b=+Xy5?O5v?(oq#?^2_jcnXZUOUj(D5QjxOcA4HL6 z-+U~Bw%dWOK=&nbGEHw{_y=b_CJWJ&PbV1DpE&=u_bX4oJPp5iz+jOpz1*#H zn{q(Azs9k4DuPZWya=u;GS6@)YG#cfasaV$$LF>l22cD4tTOpM1B=&ml_By3W zb1TK*Q5w7S3)3+2i<2p9k0UWkNmFNNg+hE<&~5==3_c%&W~OAEUW|@sjX?g#!BGwC zol{eJe6gK89ndp1Ru%(QEULPKFmzVdN1#HKVfh z=~>4{lM0(FF*+wF>t@0@{I=mrwV&vZuJHr{Wp9j1E80}j(f8W?K?7{QR-uy zre(?x%Rgy3dH1i%Oef&OV8_99Yv`i##Pe+7=EKdVbjK^ip!Rk)A&47s$9bXURHb%o zBuD&n{hzTM(IZc5Ca;t!DLDuV%FPu#r1hjs-6nPNiV*@Es(xf3iEqDY#dXtFWT~`T z7z-*i-A=Ho%^~~3EItBR}qJ$@5BtVv}Fi;&s7b?VYnjx z1%N)%c&7%zzHe2d(IrXK8watFcpMFQO3BvuzO$Y43&Qpm4a>+0QQtztG0GAbpZlBp;%bpaF76Jz`uB43@S@(? zKltq57s*~M0^w1@*Gm&N&Bw^*6KAiyH| zu8y80#yx&7u>UHe^oUmvk`;!MW_+(#6Ra&j&o;p)_9i9@sn%gb&vA3OqqF2W-Cw{` zuo91ag>}l9E8r&~k`UJC=(N}WbXy`m2=5ut%LK{(;50a_MDY1U4Nz~pX~xjMk^nea z_^9OB>CZ5Z?)>L{wTl6~`6-cu2lKBJ`g@5xsQ;&z9(h?h=I|Ji8~S#(-i5<-klK7G z^Ycz|3l^FJT~1QmOXEp^W~{xg4gxFzVgzW5GvS3{5InkCXSYC`TT#?>GdZm7dk-_; zaclG@{xQ4R@|aGoLyiY%&CmCa)1grel^_oja{VO!AAb-Q2q+B0#r%P_MC$$P)43@L z)&>LT8OIoA?RqZVdo&d+qa;rFGeq|02ma$Xw}O>1HOFX-jG?*og= zovJ>8K8$~Eqyq$Y+$*;uz86T;Z{)S&IjyZwp4fOdNdfSI+o)x`ldo3x)2i{@h>V?Q zIZQZTzb(XRGbhUxN3p_8Ndx@3xxTqNlwqxJyxiqw-F2EUBs3rV8VEnwEDcBUTO)Zf zNiM5#Cm_XNm^(E5h$dxl+b#)UHydtmTK%UxYj*%AQ!+*9LT3XhvX=uOm+~Xw96JVF zdII_nQ|F+yc8{M6B49*kt80CzF%E#Wq^X;3&0gcB#m&{Z)3j~dxry?-SMJJ+c}_c% zP9sVx6X5wXPl;b)XPn9ln?uK+z&n! z08%z(>q!WdR@h(MpqHAcoU-D@c>hg@(|nZ4E2+1wP>wO!T6|}9E|&05 zQZZom3A&EsH7|KqtVvE@;s6fX6gc;rL&Nx{k6*nj?n&Yuqb8^RL`_=QaH5_v__Tx# zFe$gl+ImpKrO%v|1(lPX)K1jUX4XBkEr9&kMuuVzgIOE7yqE{~nr?S9IM4r zn_2Q0G}20u(eH-kMAbl6w(L(8@?-+wIbPi0`fMrG={quE?5qUa?7OAUo}Q#0M{-^} zfUPlC*LYZnJakzc`LinDsfFWzO$fOgUY95vwi7jq|{^?YO zVKALt2v-8Y`|8jzhmFBjjD%uSw5NfX9)V-r4P%6tTZMJKRavKX+oi$uzm6A!SuY$R zFFg8V+z252Z$3TJ1)#Gw382JEi-{Ks<-#j14+M@*NjNNG>7}H%B2VVLyu4Vh9E-Z0 z%GpTaqOL^^^}5>Rzgr$}!4r51yIhOKi58sC$x1E-i1Dd_-B+`q7p-qTs#lm9s^fkA z_+gZ2BiP-Rh68yIAjW+;oSK6Vkk@cT$B&+l-}FpMN@o$UdzZ#^6sfp8XMDffsrXxE z5NVk%N{eu3S(Dvbhor}EwB2@A6lSa%j+NcE3aYmKiLLh2SElXzj!*B835ZWeD&rH# z2Uf_i7I>_W4ei;UtjP07HuO=_>E&N>?U*IyNsgqm9&8D0j`CKp>boVBKcz$9j5Fe z1P3?J{|i*@G^mQ?GpA-SzFSfMhP3$)#mMF?b}d{pT;0|vm&4VG<&KE!Mn!d%ht}tS zLTzAtTYqf0Od$uu=Z``mefRB@5#LtVAUFSH*#F6GxskLS1hRtP;12TB)nsi@uc6xM z+W#8nI4Qt#PaKNwF$xOixI?9eVobPedJ?&ZvVsdVYhDP{QqXc2qVgzM*8D1#Mdc43;xJsUZ;FQ&Y(^>0}zafMh<~Rvd|FR)@07HWzN{G+&t9K>VH}3~#lyw)bdbw%OIw*2_mi2*#%Ek{ zb!QqMSbG%HF#G6udVAEUqBu456I8NuRt7 zzT2xq{?a4B)Ssb-i>$VM zag1LQrSpqoS!jx1{6*G*graG0tHt~`pV|6!w#JKYQ+oxh=>8Cf{B z6gS|)`l}hw){pFfr_$ zxdqpPakmdeN-D}*(6OO2>)MT{WbEjDv|DP;Qm@5vUe=H^7``06epVlfRB?e{ixU*} zQ|U?3{$suX&DNT)H4MSl7H%sZ%s4Ez+UnrWO&z#U{=V~9h+nUiNZody;2=Ux6@Jjl zJ%{OveZ+sLz*G29QCz3rkv}=zB>5j?vW8a{6bhw!(P&@@)E!bHH zvz*>)dN9i`9L71>AkU8wTRmu1L^eiiQz{e^006ud?B-ylbt~-B1kw3q^UXmqiSVTif8VkJk6nBgywnDMw(QLm6c^m4ZFQd zMuGaVY>cwn2sg{v>@dVo*RQW$mfpAKZ!; zrhl!Z((z6&X{TaLT_?wEl9r%DBghLT#cHeHooAGDD@v?@$=Ug+)hUgS6}O>(P#7VWj1 zG|)$$y*YT-Z9<%uWNCXtrc8$a`R%)P3j3{+@|f5f%;+Ac>4AK=)A}}!sW8ejP9g2T zg}yJMAmhJ^E=K}KklYG!(&63y{CCeGfb1J5V7slWyA3;&wn6RwwQcOX3(c|*&7mc; z4Jl^A0Vn(E8{_&j8Y<=YBpz!pW6;vgHR2aPg(J5Z1?L7x@Nw=Xz*qug|3#$l^B`y5hKXe(6A@6pXM z!)DDl8fP5Oci=bC{Ncvw>Te=KFclw24p5(#oNh`%KZdf^;Mp6gREZcC78;pwx-u1Q zm(~!TO*)MlJHs@#Z^bveyAZ9IM$SICOd3Gynq6TsK=~?e|Im3ZPAb`g!&H9Oew}>0 z)G&E`bBc3M&1t9fUDlt&A13O3`|z9^{@0J#KA>Fn8Fk0%Hn^QSHN&cP2CS>Iha@UL z+sB=qHr-mnr8S9(Ekn*3hX`iXV)BJ>i;5_23S*~)!{)Oy{jz64j*EdtUD0t`@6udU zzPlcGbK}HoxxglP_m&I#pl6wysDI_Gc0|P1YwINB7Clt>7QcuQEHmo)jWywbJ~$WN zr#^T?ndfzTbJ9)~Iu4rfWi3Sg))63eSvb#mR55GSvfajuXj_zrPZZYGj%Zk&kE+I) z(#gTHqXzF1bFkH|k-l_hOnEw9lrljR6Z@IWnx!z?LC0Zd2^Hw39f(3P*@lOm))rg zN!_-e6S6s0_*w`hB$N#OC3oR6*5j?Yz(|Zt2lVi4eaTI zjhb$gbCdg7K&FT10T*zDI`qXQu-nlQ)w42KC+$Kq3~28Jmcn&4NZ`|Pz!_P9yyyhP z7_~g*f~!VlJc8FyVbi%;$pq9;qKGxhY{{P)p&a%;Y`n|ZW%n?HlftYJP8Ko)P$%vQDAB`L>T)fzo_S3;t3dXUv^m8fjaQ$nJu8tt#gNtFcW z;{}e~9mE=l;CnM*M5q9A=gR0ed-w9vw`6Q4VGXx8mnY|IDHzT4v&tdq-{Hp(KJ2Ju zRT{e;6;WLP;=n+a$zYhYJ)9Hq`7wK^XaRLDevLfOyqqY>3dgCG*REC+gHbPS>rR*9 z$!?udFku4UzW0Pt!V09kd|HMmcP%-=c&gD#tFyoVZ^k|q6;xIn>L=cUR(P_0=t5Cb zDLytwm|C&xfK#XhYVRcA8?A@3WSn!HlPm#sWb3qU{b6MAMzg7r+t&ExY{_L2YF?CB z2$Elx~BYEgbU>`^O z1XRpaYPf&D5(9eh2+`6^E$;;>k;Sn3l%bkbTrU_;(IQ2k#(1 z_RYOJKFio)A5M5qcOoL2miN0?mGP5Y4_XP6TF;FuSE4m00aLYoTH^`~9#k(aU^g`$ z!41rH=|Cql%5?aC*!wa(Rl2^3F zLa;nG_C!ZoBu`~mf*UtvP=?o5v15$K8r+pmD zGD@(vavFfFCf@J4IAI5-5=xbg1f1R9orfEs_e+ggh|rK2Bn7v!3?dp9{E3yTnIE?^ zBZBa6z@wMpbvu+vyqqV&AsQM-*O_l_A8)2itQ?caTByT}B z)M6ARwp9vZ*R;<*hS~?fEaOsl4C46w^ZOWPjfWEp^mdNxX-0joB4`T}73x__-;TNF z>xZ((bec%hy>UWyxc@t%7=SF2i<@dMfhrP(;3f-kG@8QkY9i%tSal}>4H1M;LxCzSVkDb zN<0>ji174NW9ZM}kj+hnWuNdQ!JdOHg8gz3X1b1A^}F5i5yJ&$I6L)qe564c!3|Fa z)y%9EkHTqNJi?Qjn~N=vAGv{3__;bwo-F7N<#1s@AKRj#Vgj@y9B!4|LMB!Z3SVUj zP88R{^iDSOTrRZ@yS>-xen*gEEZCK1R8s`R2_djF+(-oIB%?O(M-T%vvGPz7-cjp& z`E(7Xk59w6fDBqRZ(T)X?C1Q;7I0n)Jf@L5tZ}Wos#%_FZjP{>{3> zX1$4D3YWS5$|MXRlVtWjx%L68kU zd%BV5r;{~}aOG%|O-}lSjWjF`K^1l}FL+N5<;!IK-A;u4&&jDKB1l}^>4O>2#1ub# z{|S^m>ouIHO>zGfIEP=IXYYR2bbHzKNkc>FQA4`Ii&EL)&6>6D(fke#^O4Gen)Sqq zUn?Jq=K21*Q(SPTPkSDEd=Lym?M@MxQ!di}+@ce7HfV8z45OiRoy44Zt;VQwZEGC}OAZe7y|6=Md>v zE&hT`cVL$}*l9tq<}Cn~Oh#y8H1z=QksO+A1$-&M{}|i3Eib>?nGYB7e*-qnm`N5GxN-NA8^#v zBmv55jp*tp$7a2ApMH@B($Gm>eDITa>h6PA^JGh>ha6~DBc0n9TSK(@qg-pBE~g7yjguUsXMiA( zTt_*fKBcUwwwuVF?96{~>$FR(etMFMQIzAk@atl$XbL!$wC4{%;d@x9yy!+0y$$p+ znq-;fA3(&~ZY9(y2b5O*a973X5$Np80bQCl>^fm1uR#Q)ERq${PK&C9V}M?*F$h_h zPAdYn!%TEqE=}zZ7-zImM9DnLZ=d+Jx%dmZfI`qI99)atE|0h^6a~Xe_|URBZ?{lx z9K~^GRzH+a5+~1s*$$=teD#&HphrNA!hEzw{EUrn4PTu+r`Zn5C@4X07aLiYptB!M z6}r3ESlz=hh0mCs?6OK^K315t5xgZ&Q)FjuHFP&zK|T0Pmr0f=^YM?Ye&1fhNDM zCx?vXluqFKNP$c|nZr`Orbz1Oy~t(mrFg2&-i>^kqTaRK!AGhPx_uX^av9ASU!?cw zku6xfO#Qw9j2f`qIbEU8mWM&-Z)y{3eA|7J4=*d{sd@8m=Ytpv_fxb^4Y|(1tzsy* zH#!{&9AmUp4*u&+*rc8}0E98*2L!+y+gPY00O%{-y-vRCdnkT=Ui1*$H?-SfvY|}4< zg4~n9$?DF9yGI7(tFOl`XzhZxoYeBs>%Q?n<-;3!P~>jiL}o_z@MEBfz{`|p0SCS) z5H?|Ws6w))S=r@p^wRcuej!bh-Pwk>a3BeMPj>$$PDXSNT&7fWdqEEU=Re2U$CfSe zw@b-DV044Vy>8=a8K8s2x{!;a&_14I0^xTKJ+b=djOIqPuAF*|JR$)*6ANY`45I{& z?0PMhe*0xxx-9UxZL8&hKlw_(!?na}hKKArp4Zu`CO$InE5Cwa)MoD`f3mA#Ma1?Y zpyM;=jL6k?%&a)3ZCd=|BfiV=NI))f9Ta*Rf>tYlp0lydbJ|G>syoNf%YX)40xFC#({tSjRt9LHAB7-QItfko5m0e6`w zsd&iDMY7xOMnD?7d5O91G9rTye)Lr0UY*D`fMrm8qBaue#G*oWh8z|a=!Qb&)9TNS zk)H1WIGB1Ma24jI;qfN&XC}M0`ks>fIZ(nARU5jKfkgi$7oWS(?Pg8j_FyYS4`Li1 z^c*FUH0yI7P{YiC5~~@GnIAR#s8HVyhts&o%d!M323G;-1gY0ImwTSwaF(e_!gTBw z#A+tTh%wqq^HWxqPGcx6nE9d#b_2DLaZbqGm>kFuy*$S|Etgyu~Wl{DmFiekFu{QGZApzHd+_npe15QRN z06~q(csh;AFTWq`%{lAo|=e@Fn1LIg6!? zvK#j%rYhwr`&_M z;YCp0u5AP4-Fqyr@v7TJJ(Cu-6T!>*4s?G~ckm8EUKFld!+lWg=`Y_i74n zLiG;qo%Bi&huk{MCm3Guw9fJmz(@+E`&NoX?G4zk1M_9is{>ITR)bYa_VY1dMEi~m zu-O&xa{s!v9jLx?)v9yC)EjGt+s?<(kfRuCO=1Nqf%tt76r1BQXsW3NOZ@8G)|^#= zdPRNgAiUkxa8ZU$E?;SH>s!UBn@h*tz{}JRNqsAtJOc z%#9EkF;4oMhxoCnBvdV8Dv_zyKBwH;n#1VRgobrojm7h=$q|eFxAvxPD_Ni#Q1!|8 zI^O_sB2DO`Zd6I5aPM*#5&9cg@ED8JZy}qiU_ls7-b-`-FIE zv4`N_gyGmQjjjY-7*UAFdUNfiQxL!IK=xbNi{we0@VId^WbL`_eB!DZCygmEnQ<7} zrP~W>*PL9XIyWC{K7IvVQy5@}MK$ZQY+S%S^!$c>kkoF7FVnOqd9y9kHT1nKQ$7}>zS-9GHk;2C`9IfR5i#_G_lSx%0>E)K+ZKxtYwbwV5!@O1h%ZDoS^2DPsQV0| zXJsiK-Xpa#`P3KDR_Uwg>(_?h4IcJcX@z>R>;`BKAJHgX`qRIEWIeF=G_XWOLf6)| zgC^z-HR=<|Q5?eyx^hr!q-=4daUacbR{_SaOiVlKPs|IRnmj3D^xzpM z`=2mQW;gs~so#HXPaj94BsFDw?C}Ys2%2cVS4s_D*d&NS((qQkH?cor8M~h_qi(mhOPc@OKp`k1zs|mduFL8N>dR||y^*f% zRNroc#kKcrtq%iftH_%^#@cS#|MHUYqgU2r>~g3=iMnl(w8bt^+l=(`1DTKs+;jm;+G~|h?6>|hPcp+S*4@<_fHY;8( z?B8k=984L;Xh~Xq2rSMvxs`t|K(K>1CYl21(+hi%2NYp15OEdNQPLb9p(B-sd)s4` z=_giIr{m($N_YRH_{skC8gGp3GAW^2V&>)6_&_3ZEjNYYz8y@Z22-ntJB0vSDX7_G zCtTHkOQD?f4Gt14l>YJh^c@Vg)lU@e4zlmNh;Wxv?1^tK#>@dc6L{Fz-w+j@RoQaP zu~73aIf^6m1@Hxu5!pXQJ2$t%KMoW80;_$=Y`fL%d9Y@{S(skIQ?j_Yk(F}7vE94= zBW3IXK;cN5QhZVFxcq^NEreqQ7gqY@LGW%K|CZ(>J`9(i`Xq(6 zyb)*#d;rw13C^{_&;CTmR_J}`(*1&osK!Tc$OuS=9~roiM%(C=j=KyU=Mw)y%?RQl zeP}y!-aNM9PAB3hZBBN48@?)RPgST}r>oJ*M>{#z&!9c6CsBWh=Z0BCk@~>A8F?(n zZ@`uUiBd>#w3+3Y3fChI4~92?YPsPK<(%G3U@_v@2Wx1V)Pt?;TQi+P zJL&rKmNs4;7utTrn(LZ&!+jvYzf%#Q^{aQRdim_231ddRFh`!8j+Vy z1uO3vuzZc%M}KY%Y2{C}6~$@$iTwp{4((Lwirr~OV)vuniM`$n3m!Ui&tHtc5D{;w zWFdXKccWx&Z`-dO;iH?c$Lzr*rfdz;GAWD=gWKh=9)Tc%>>4&R!I;2-r9@4xStMVg z{PNO_`wp0zGRk-pcf4oKBCp$<|K0tF^3zc}rp&;-1BPBPifHfAZy!qrJ0hQEMHW1p z#MNLODLSrrGZ5Vz?hN7A!^Fa$Kpi4%;RimY1BTX;SCBD!p0D_V*V3P}vQzBTl$xeV z%^(-JGQ=g0_;Q1_h2zhECWm?ye2l$#!o!}+8qaA=LZ-oNS;S5#{mnH?1hR~esc?(W zH~93M9eLz}B#%Vt8f}utn)U{5*qTn<&83|v&=U}mB7ZsiDHke`hpiw*@@Gb}!jfy& z0|}MXStGk3vnbAUmQLFxtMXUf1pzPYXEA&6$yqI>EvT6m2?*6AGRmRX(k=jexj#-F z>{vX1pVIBp@0#AM2i+mu71ezXs4C<86OQar=AiThAe_FMI+Uy8@>Ytnj(N7*1zjc%#O8auICy z;#N4cdwv}iDTpGq-6#F2dVSG){z*exY|V~m_c=xEA%o2-rcn~_T0sTWPYU??wKP(A zs6OGW%73e;IrU03t~=q8|KWq^vxr(9`Np33(G>6{9iAo+8o#-8?TCH72+$H8pxYoW zPJoe=7ZdDNfgT9T!mwXRiH3M<4wpaDal`45`t(Iqa6*w#^X;l;Z_?*V?gV>#_g@^S zEUZ*0zw*v^qKK9U1^8-3iBAknoh{WPegz~7BmIr6gIH%1gSLkH!-C4>ihiy zNkpi){bk=f<0=Ya|420f6-GKJyN9dx*G^_zTXXepw0=$4_i8w^h7WjRhE6{!>~=In z=sDHvC_IH8atR9ZybV!OmmkwV3s*UmnPEm&DY*)?eR@E@D+BiHU}bKLzl1Iu@agXN z$-{h+maA8sYDHyN=BHmSrz4f$_JHZc($DyEbo#w-0B}YCUpRUU1Gkekm-`2K>zby z00b}t#0qnuTjWg;TgQ4_+0OcZv%C8hEQQ^!1hBehlFQGeA*6E8T>q(eMiMU63P2ag z({8BcJDd0AaNJY`-p?s=TofNIXZE(PX?_Kx%ROtSL2)BST~F99=j?jW%R6E4Mbc#ugb1#=(*v-@cI`amDvoo{YMpwfo>Z6Dd7Z?mVMi+a+W%e%*aQi7&eNR1LO#Im zJgZ!&8EEJ_*fIE)1E(68p0&`Y9EkZ`9XK!0t7Qbh#F3Q8qh&i3`4;3B0y?i;TBE>1~?}$f;d0Mp3$I#v2-U3>LUAMiw5fQU1xq z!l`zs^ZP2?(>3Nsxc$J+>3lY>A8NRYKsmg{pw!*_YG}KIw#TTKY4SOdLZV~z#CzmF zdlw0pmS6E6VItC#d$;JIrw$s7;*RSDGg|-$xKj`N7#A&1{yGECfY+5Shn`4Bp{Kga z%KP`f1C|H^LQQT6VoV|X9MOEbDyqh6h0`COV5xni1~OlN*joml<$fxi7^`tt=DW$t zMb^5aI8h`IzwY}9-v^oVt88d+0UZLKf*w(#4ErOc4pJ6Wz{iz;UxY?ED)#; ztM8Kq$yeYUemVqW%BBTV!IRYr!N-Eq28euWz@@njOatwix-ds3E24P7+0Obm0jM`x z?=rg+jHeQ$2L{O}GD0&ef~VoKoZxGuTv`VLLhu1Z;s^mv@R1{uJBV-@iYkPwx$gRN zSRNt`1)|->)(GPs1;(+A0K40XT3H(zXTnQbR=WzPojIu#yPyeDw^(+otJ9K2?Q5XF zp8!60s?@M+`*hB$ihxlW-z8tOW)MDL&b)8vt23?NFh!o{db|m)u;1p7m4rT)z8)#m zn&JQ+emyeW8BAi1av)x6jO1SR)K=FlJ_29@lc`gW`d7PfWm}^#ImT1H_b0%>wPvmEXGBVRd`Vz0jHIqDzft_jU!o^PSy0W_ zNu3FXRn}UgkfKQwYTQM4EHjyW!2HRyz*Yr^dKSoq|1%m>$g+S+w;@4+jED_p9veL_ z)q#O%njb1N*&a^SHHV;rPfJ^>&4BvW6@9pjPy%1GV%5-DAR{J6<9o!80ro=+lWs^X*T ztis!`^x#08qpb(;qy>Uf^FVlW2g zJ^S@k_Z;*J=Z>yG%1W?)vje^@6kCYY0*0=mocsJO@GT#;jN zhS_NTMyrqCcOc(vBS+q}Qw^kM%ZtNwn!fZJtlGlvZc<>i62RCuolhjIOkv_1!!;TU9w9OG^hxKw^&UVJmz5%Wz< z5+$i!77=Z&QIOT;@@xG{YRJ`4he4vL0Hn8Uu<@NDImU715t z2E+Yhrg1Wx;C_}k(u?}mhmbjn%}`v+gf?*V9Z5{e%mCNqIZvv z7l|6ZT*tU}9Mf7$ymD0wz#_qx5x?>7g2;Y4MhzB;ru2airs`y{12n~ILdjXIbyHW&pK9J#`bm25NuGw@ezkncja!B(X zK1w^p=Eb{R?2Gy6**7fyv$yFgatR_A5eG=))D%luVRzs?Ly#;*_M$$78J-9>6koUu zH04;Zya?_m`~^C~vvvRUCxE-lU5cbvn10$IaPCp#k3|v`8`O^?G`gp*0=pgrDF^db zEec!t0%mbbz^%Bs0tk-PF=I&fmIE>(5xiw9Q4npTz!tdU@f=vv9%6YNP-Q$tf>M*m zu)}WXdl-m=r&o^*HR;`he)a(OK1>msl-==X}Oz3>PzJfV0Q1t)&s zGX)BF0avbLZI>ocD((*V6J?%KP#g!7sSR!51~uAI^f^+NF}iRSAM70Uq8~gOoeiL@wQuoF+|03QBDki4;2Fa&Sm2v0&;iG`}42Z6T#|d!YxVzTT8<6WD2k zt_eT|(SR#U7w{>`ce)_wThsx$aAJ>+m%s4;@%EKbRjysMvIV7-1`(0&l8{niQ%Zw0 zN=k`ziKqx0K~g%Tk?ux9N))9L1SA!Z76C~C!Mk4L=y!kKA9oDL_|7=T@7-@a&wAFH zYtFe^L#2<&!IKnR+I%ZxPc}=p3}n`A4jxcw^ay5|iCzZZ zPF%m~!ND@k873bz3-Q8uZNj(AQ*Dv=nFd8^j41!2hO@>$HDv{)e)*)^VXXMI`kdh7XPH*7S zDO|JBU>qo_sY|~U61@#=fh!~6^)mNf_XDUBBnJ74kL5-WVpldUL96WkgrR`^@h8$6ipITFQD2woCr_ z{tj^hqVs^JLJATjSAbnIc7LN=(2~$q$P*ipSt5=wU)tGHH?7mIJ>;rJL!(dlAuy7n=fW$zR@Fhttp_QSP7m+Hd}R~_)NP&Vm=XUw zE0;iD$CnL&&}4=N6h7t`IM1VRY4tYnh z7wVMP5ZPPf?5ZQl(M^S^w~jrrL;(|ZJ8>*~_t|XS1f2K|2!Y_lX`wcU{;^DjC}%@8 z3*1+vknLSs21VM}F~?2NXf;B?IQ$N*gL;5eGrK$>xMskR=Gp{|K>b8vJ7D{p6g@KKRMagD7q= zsh8b=SQM%Khx>Qx zx!FAcy&~%N!xZK#Eh~Qu6Ld=F4NhXZ5b%6?H)Dqw>@iMySum!fut^!!pWn9Ihp0eWCmch? zAB8)G^h&aAPJ&+Oal=4oW1ceZLrAfY5Uqx704 z>khDq{n~GGt_XvPne`h$6%i{YCu12oTyrK&z1No!QGwUMaJvHm_yn~8_G#Ggk$ARX zFB4H2_Fxioyq;wbPWa>*lDP5l7C(Z9{V|T5q-hbq7@DMXCB@@WcYvHBQa_e^jkx+D zk;Jg18njTaJKGp@Z|NqUPGf}aouI1RD7@H%g^n6%X6i6ei?D>syDj-RnBIlJZnCyOkSB5 z47F0?wnHfSD&R#j@cMY*EIiR|gH3tcA7jTU2G;>t9j1Q~C3JFg` zdOZ7>W%Erk&~xtaXIa*{ui8V?vd9iv94Qv;D3;fi5w^C>S8m8in}JFgdbZU_gtZ0l zk==$};2#vKrAp{zTt1O%&sO0Je7)Hc-931V%Lwt4Q7K(>p97d6?KHBvsiZcu90Hx{ zUd*h#&zA-G3ZYlc1t|CJvTpXhY|mzMVsaV<$X*|>b|OwdwA3u2NYclbG!QOpTx~zG zTv*Yw{<+=!7Q;T4^OW5 z@fA5-0I9a8uTqOO@ZXGF<>m9**ry@eUC$=DrE~xRm*{)#od@z~tTV#? zQrU8Q-Vdm^*7Fibb)BuWu4ZxwJqW);q87+q^bY?$ zuN&yKg)EVmu=-A)nRKf*15K|BA;{Tnj@n%nsJI%9jfbPZ-+^)XR%^nI!<`UWIq<#P za6-DU0<=-B&w{pJYA`F3P=0xwvkrlDS3cN_Y$Ic>jwv~E7!tK79YE`xPqn!5G~nt~ zbOp$=JA-)wFH+#0tSXY%sClh$l{;p4V}5`B&Vx4m0|?PO(gKBi7$&JKr&t|wxgehD zuOJuQ%}og5X1=&}rF1wz|5RSoTf*Qy!dhF>b zd>p$|6r-^12`Z}~iQV0~rm3xgr@|*4p{vm_{7I2-fUba?B8X>pB1PIdFeo+|sP(f( z8<`KNR$TPQZ|o;@y@v{llhi}Wa=#J>lS@V~O|eCt?nRt?1VA78oq{81UO)6PQuPkq zscv&Fvr~%xu)pbwkcd~UdpC`7PUOR#3{TJ^N~BW+Wqem$k@N7A0;1HxtSP?9aO|tB zAua32R|Ptu^0b@nh@mhNVV1$85EEq0h>p81u#(rr8n?W9!F( zh90#sLq_%U=Dj6aJuQo|Uyal*;hG*vPvaymDqG^J>P>yJw^O`S;$GLzdC z_(>-CCndi-8oxPs!R{Dw?qevKEJn}`H%Xo#>5_*dyqIJ5$+0CR_^QL$R!L5enu-DY z#JF@2An8uC*wH;R)##YdJ6U8>7?))N1M@aXgM5oLGtQ-I!B=Zv#HT1~$h%D7O% z3RObWz(G#rO)>*ehh_$N3x`He6d#t?wzDQ?F7OH^@}j;ss!%O{Tit-;FmWym_2cnt zuL|s2oJFep;_VMERC24O#nHQYk$ph)y0U8Ghup1`--7!{_@)GN zhLoIthjW3F|4oJ0dK!I?Y*%In*mq;bdow3kUFZWD!KDI`%qoDlO3DG2T-<>Fv$ z*(Qf}hXKnlM>2gXp&+f=^QsCo+ZkaJx4_e*PR+A^(uZIf$dHvi$rq?@8gn=;J@g&S zy_!)7#4enkE3=P%>OcKuEA7UH2ok{fV;HdpSnWa>`fRN@TF(|nGex*tJ^MZY`?WzZb!f-; z9-rTqN2oByD#hb9;`6xb!#@8SPS4BFUv8d6#h{OjZ7O>A@GykmLV+dXb1}Bu%=`L^ zZpPMhC}MdcEWBjIVUB={ngcn`%~_wG=cnLM#Jm$vSbJK9Zl$oB41U5Hsw@=7alQgF znh!q4Hrbn3#9wI%%?BB2BgR~_y;zsL(fxSr6KZ}co>8H(#~8w)|KQaR(n}M?K^Nhf zp==eJCXoS1=LH|w5~fXKC$&ERD5-4Ij{GmdGQfw;_tU3uS~CR07cal#Onq&I^sOn&A#$EN))%x7V7oZoJPe<0g^zyA-)=< zh8?4^TLndhpufEhI;s&6?g-qEb~R39RfC++VY~2gSd~}_b~K@4<%-T#PL0;xBy`$$ zsEB8+z<(~_A{@BHS`n$N;a@6Pu1LbSwUY4H zO`nHN{`9o4t)9kMCIY!t(GO)_B0rXB5s$5N&T-jZ&LI% zWI5Xmbs=WEJNuDS0cRoCE>@e>!cRRkerwqae#E`!xkF&FLqAZF!|1d>dJpeD)O6WR z@}y4{CGV}^Vr*IHqY{x^N>b=GSPHk88}XaAU|)c2)E0U*LkRE=M5scCkBvhGx_l0u zJ>oqBUX14KPgSC3$G;)C<~(5DdY@iDZkTLh{Cm#adwHvqAXWBn-sM}*+`V)zGKK%%obRkr^*=+%Zqg&Th zeF6$!v8ml7d;n@EaIQ)n0mGRkzvv-!yGc6OW*BvW-AdJm?Id^XKuf%@fiYp_b*m;i2H#T`{ZKIVE;S9!iv1(wU^hsaiD z7lHljGk_G>W`%nq@~LnY<)Xg&rE_rxmCBneGK}5MB35Kx(4S`>FVEsOZX6Hu1Xtlh{WDM_ zJikwpB_u)dSf7<)lunVAlr$Ym>&>wn<^%+ltm2y?Zr}&n#x}0utkC^&$Y_7D^)Ff0x!awBQ*`dJu27}f!>z}^30qI zHD7VjF8YxMD>@CxNG}4_m3O-gWZUmf>jttFxts{z!ia=Q=CNW0StJJIL*zqZJI z87R;#@m`2B>0g~Yw?uh)GP)6Q$rS*&ma}3$?1N0&5v&6%N%s{4LJoZOsS^N$1Y$EQ zCJ5Ut(M>#e=X1jiB2_Q|33H|QBp%b1sSP00UzO!<#b*z!IUzD$s#nH>#mb-Erpx{o zUy@5adqE%a=s@lmakC^}^lhY*`2M~?!Wp)-0qO%53jd}vKj_ch$DKKaC6`{=qe%2q z>fuU8&o46Zs+JBRO)NeDf9J2kb((i*68>?E{6>d`O+kl)Yfb|CJa#@~4zAz2Q)n?B z!3+THA}fX8-5Wg$boF-XAXsO}n}pJYKKv}a=uO8tUSfFytnD;K0!#b9|8*V0{QA{3 z6e?cL`%~QsK-ROLzGaA*6g2IU zo>(43x|PV2qhEo}gi)Bm^*Ab!SP#6)VOa#Lc#>h>od5Y+kcUh|A<(Axn=Il0d<>L_ zV}MOKeGBgJeM|J`7M&l#La_pkkEI~;h>(#8_X?OjMBX}FfIbK=a5afy4;$l58qOex zSgSoa*qg17T>9@vj-V*Osv{rwQn4Re7S}|cBqA`4=FfkfIRnRbkBZpQp3A?m)lXdl ztbcwxT@@OpKMT8nhaePe#R}y7^Z!HBK`%d=r1$$@S$f#WwNHrExL-!I?@ z_8~Ywv|hnNQLCDOp%x`J7KazLM3^gN8_2pig6U$0m)GTy@t(Mi?Jx4U9`+UW`)?pm zcnN-IajMIZ9vF?k&_alXf)+bzrhje|t~3qHC}_UrDy}&p9EgtJI-e}2XZ*R>|A-as zmk`0vJ(r$G^ z9%*-@4fpr1A#O8(2kA<_a5v=>Ht=Bh06r3aL=IeDlW%`s2-xmJuC_kELL3m>+!4cN zw>(@XY}IoHanx_RTmshb!D&Ik^f7-f>5aaHtEHw=QHc+?6@`X{{Kfm|$u(3yAY((o z{z9-huN1I&ECS=hoWt`s-_pldXYu^^mX2TqaEYIC9wp9UAjUx%6@cM}TG;3Aescl*ndgJFRV&vU4jI16Z?()sVVn@LPg zXvf|stjzli0-;Ej4AWCtB+yxHP=Wli=i-Jkbis;v^vg3Vmy$YxH;pkmO zsy~~V9VC$2Pvjpp8aLdng8mdRXDt&yV%t47?*X?kud#!Ni&o%vZD6<#}HuxTx7Bc zxPPYLFq!_eCi%Hc=r#f|m#0LKgiygdnep zehI$*)l`PP-|kSqe@iwN;zzlr%$q;!@8`b#`dtNP!%kATEA=>P9cr`;X6Ape6N z!UfPK$^nQQM>r2`srQ}9>!P1YAi$DNg{2COsH4f~O8hbnk|ZLqz04!RGHBgX+&lyh zaVZ`d>oBmdY`=_F4OKfB+d!ipBVLa|>m##KK1rIr%T}p?Ch)i3tKdS6;Pp^OlrYBe zK}P%G{)UR?!N&*t-H-^HAnYQ<9|S!ChrmQN$-PoAb#?N(EV(;FMgva!h%l{8MQdRT zq^kGt?IqtCv51Qq9{Rq!@ddoK__wB@7U3y7%Z9vYb8%Qf@e&U=C3&zBKk({t=U`am z{5(6s{VK>14n#s1U8l9ucpnsqKy@n_0Rww{gXh4Lq2a&b`-AZwG%J9Sqx~hxrKb!G z(VwC(eL$xPq4pjs?W|0E0apDpV5sVWzwY<>TZwkh$^ZhRsNIaWN-2Zx!w3vyu>t#< z>SX5*j%N=J?tXq>l8V@5W^cEEtd}bBqDMK>I}8<@JB9etR@sdSUo@QCK(^)P$LmW! z@Ys?};O*WC+=*Mbi*!^wxJoL4(h#c`_KQh~0lup3lh%MRDX&1p!5JxTNmi}fClJNL zGQ7w2PwcfibEzEm#Vhq2s(eBM9+M0E#hxJ-8j829qnL z!8kOxof0$raJ?X$FfMc0FM+2|@`PC55>Q-zV9^-?NjuB4iQ9;n2L* zhbo{O%u82^+DF)i^b6(L-JN`05bo!#6Y9QUcJMmIptq46+3p=Ji!F=B=8X5|){43B zrZhPYjbanTl?giR1*Ewu2l?JxK23@-TJLXxzlKOK&ojM|)Z1F|)4qtzxMJM7b5K;I z5Fg%5VVR1gEIloj!N<~bQ*~6arY|22I=)93D_zj*ULM~sU6aiM?7c@iNLzH7@qMMW zW}IsvOKiYdjn$9iJ9%y>sgdR!?a{pfI%dceO57 z?kAjD8G2#W)!L-4T$>W`9qR{5j8R5$4a?JT^%bzsSt=&W=$B_bITeCz()Z7XA%+N6 z0IEmsK~sD z-F56rO}8`kl)azxB-HLrDD8ys4?e?MATd-VYPMVyET*yBYd%V(Jf`Z|xX0~-bU)q}$3Ne@OTygxU;I54r#Zk)tJ0dYA+X?3Iyt<;8W6+35{<1Lxp; zc5%V=IJ*`CL8ze{pQ~`P{EarHyUYqj$^@)nWSXK6 zX&~9E-%arYN-Pp$j5Ek7UVfQHH?_H$V{bwOO9b(})(rR_z}`W!a)!##HDY7qO1$*K zlW9RUtC*8B(1`B_aAs~U+u;7!*5Zt?>w^kx$upO4EoKUq9C=~q7dR@xH+MSZ6o1QG zk|`EXfT~2}xcmFX=37;FU2iN#uhB@}UsuXe&gv9VP3E986*%ugolAfrT)=*MQ9aBx zGKk|%8vW$T%kRfe^;_2gz-M@yoef-cvOb3+D&R?JU+3mC!4HA7>vroAA@%f9^C6kq z#XgRVM&8Aq#GaGIwytyF)_HQ?=OH6vpPYag(pf|-U*ku0*E3x%ae7QZ5cCmjvkC1L zT^UVf#8*E=hsuc3HGZ0YRXdHU>fkJ3zMVlh0ZbOvt7c3=q+dd8UNC#znZ|Ez6Q_JD z*dzM#40xc2Qri&>MX{GHJQO?QovA9zg z8O|Mlc;;dh^UpnAjGd_-^Esx#s-cX2ah(3LM%?5R;ysOI!P9a%C>+0&uK?84FyNQ*@oru+eF0xk6?nt&u&B-Ld|gV0VGS8*)GEZ);Nkd`YN2ZCUbl zn}zLnJLo20ciqT1D96PY39sD6Y1w1W)j?Tfn92L>oxT}OtRx?N%r;YmV6GJ8)8_W+ zyZUnJHHP+LirA-!b+|r{Vfv4Qjbe-*$JN})m$-uS0n(B18B^EX#ch)`@1TAF1H*8H z{S$5n{(BQp#JhgMiDlPi|D-&SbHE^rWEg2Q&s#X3%?!3XHJ9M&xjRqF_kUzpp^Ii4oIk6?XW%z21z`WyZF4t$Bp{yMq$hT zclW||aIQ<=UwpUkS$jC+5p?`WOGUzbr2FI9#_8QRp6!=j9XY-lQhJWw9W;_PUj}0K z=d%;-osaTQxlXMTUR?WNnig;`REp&T1G_CL2{9H4yDlp4AiHBLL~B(K`;J2TDSl_Q zoTwiD$n5txUP2Bja($|USKO$nEwc*xPFO@xZ{?6X0dnb=-==?{1=HwEp zkMHo_pW}@0U}@kRBU1dfPoEI?K$jP5MPUU4YVtw-a5|3}5)t;7Y#q4M^XREt398>8 zc=hIh|EXu+voM;%622iOgL?-JbDcA_t(O8FFZSV6Gc#(5Kjd#WpJ|J*PxHkh;hTA} zq0Wec4Kb^T#@DVjsy#D6^S=J@VE@8)YvFyS);t_iAKF+ z7`gl6&NJfV=Lmf~vP(gyq}C__)Y7o<3De7RiBK(ZI=9}y}d zoKz3_D`%X1m4cqLQfiQMbKAGSdOv2|Yi?aWL78K)-gRYUWJr5a>EV%HmGrHtV{P>0 z52e_(aUTj04$1Q9P9+Gg`NR;sa5iTDYBwDBrun3eiD}Fz@Kut}i`!D{EViUVvKXHj zwP@(|a{rqu`PFTH5kg(5K->UtTrGTzALp#B>)8>q=Gnkn{u*oIEPSepHC$$1_mKMi zwHiI4dKS*H{&YT@@pOA&h0SO5cAgHF)`H5YrS&dPP2k(7JKKA0l@*ukew-m4aasYi6;OV`GefYo}h*mO}U0*yQTkeHq&W>WmkrPnSIQdu=V zF?1G%#iR|Zz0eY)I(4+#>{x%~rHl-VtH$I0PCTlthYwnKbbo&K7+Lu2EQa0OJ$`{N zWGz~<2nzylA-Ga^<3Q>3Y#TxGt*)atLNV^vc&j^$Pi8{0 ze5fu2$qx^9=HPj~?iRVPT)x^U7_Mq6f7?JjJ47hMSN4LD9{R}-ExPtW54Ptz?Mt(} z=l0H0&IyvpGST`w?k%`?T&cCXA*pyVa-_w&XU+#-{h#}(8sofQ&nvIu4 zU6jfALh7!ydU8J11>u7wmRxyDx)NtDJxSp!mN9~R^oCP(%FUF+Myr!)RFuuE(pKBunQ_?&M?Qs1Skj$D7!ilQ`X-r)GVrc&yu_s6-WUtQ{Ss zUNDN&PZb18zZvHoYQ2FSvpF;Mo`(TfT1%AIu+eqfNw7=HiZxD<%A4v^MmGCN*PH6X zPF7RBSE>o#Gtg>Y4&|WJ74D%fAWwSo{^~{x zlcgtEa4ke9Jfa@mi|HYv^(rRt&^TVqbU)Df%Hfoa+>CgkdXnT;#$%@snP#fAihQ(; zLfqPnfz{KokhQwOV8 z^n8u<*^u#f58ef)R4*=&ThZACf1aK{S+7vt)4;m*B| zbYGKq^D8gKis1HYTUZ}W@y9Ni>OJ0g3Y<_XMl;eJB?aFS8@@luIn(RHJXnx|C3vhr zedf+<0q^F9hr#PG`{HwoNX^=GQL_ng=Q(xBXu5uvGdm=`e2jUaXD^LTxL>ppjRR8g z5QoXQF?QLces6_2v>ff}6q$Q3r-wQ7$^YJ#LZw8t34KS)MF}X2WfvwDf|R^_oU`#> z_MPq*n_Yj$KhKh?s0za89f56zUd=3}s$>Bu`Yyz~qF?IZ8kP42IOnY^S6_%`jGZA^Ql-F^ZwpYI2n z&ri6eZ>hXyIH}34vG|0oQ>5qF&4Z;D!NDp;ZHX_8`N|Q(C;hcNEHxq*-uB^jM)PVP zx=k1rI)A=qOsgx3`d}AIdX-!6C2y2~+_iKa2{7|mp$?wK6G5-4zB!b}+ws(jPhVbe zk%S99nm2pa5iNN#SbMX!RpRHd7)^tb>nLi;cR0kWhuNw24b0=6j>e|W?#LgCdHdPzr4wJR}5>e`RDS2!2jn@KUsLE z@m+~gE+hx~_YVgg42pr5FZ}m6fByshB`glD#oe1FKc561=Fgu@W_mwPmeXJ09{%SK zW)ljHtoe+Ff3J@Gb9f}&SM<$FnZI}Q>o?$@q+?RTwuAm$`uA_36;T?w8tHu}|Mv&5 zNaWSv5dVLBh-elJJ#I(&hy2RFPZCW7k2rg~Rm+$pKxz`456=T?7#yLY5Tt_ca40MB z?x6bhjF>%=aA|GH2jbVTZ;;eLBfJ}im{4-rWy>OEy$4t|bwTeoLSPZFgj)#E^E!;q zVZc-^Fpv56!d~<%3w4&?Z#At4n=)@BLcBB+R|uvUBXeU|wAr-hny?YSZGbl|kOo4M zK?e95zFDq%Zs*lk4UzD2%-NbQI(m$){FEbHE|d%U-TB{^z696To*Z_s+^>3_pQ^zxlD-Yzf9ozsQePf&s&aQE(OWC>6pE z@|JmkS5?4gsf$9_Afhv$o$v2T$h`#@$=_T5JhLI^jXsLF|1E-5w%Ba6rn36;bw05Xb6x5C{2Qgp%8TF1pyN} zuNSz^U{hS1&2WDYgr^b0z`Pvo5V9RE)ky$%y;JE#z@FZlgK`ZoO$(X%t~m z8iDY-YX%T3h!B|Y+x#P}ucZCVgEp&SBG>8Y9P^g?d~h#5J$DTN5f*(?JFoKhn1mBI8`8L%n?p$6NqUSZE9d9Rvd*_z8wJ|eAK zL7*nNfSkbmI!@N#t6=ik;8iardLNkyYDPSf0+K{-yb3U|h!;*3#itjZxHP-^fKKx=}UK?lgho`*ce-f?>`99IEV|yB+vX_mjyT8$`Da(4yfniihhM_KY=ZM!D2*qPe$B*#THQlRJ7o8 z)qI3KT_XmN)%aa4C7!D2v23q@*XjsKxF6>QBSjw6dE71-r~^JTU4&;Z4p##G>viM( zA9{dapIEsUFP#HO{#^x?m41AC{6+$V?y0me;OEa1_@sql0H$y7-UTT!*V`vJBq-yP1hG z%-BKi2$VHQ>_{=rjBBvM5|CNln=t$x1P|hJ1$R9uI60BoHi<~XCRwRZRTd}szqgK# zCxyVAO{F@>%ZmXHY9hpoz9fj9az*mIA^x+YYPXW><^65?3Pc;7O=ZDf;pg($okd%zJ=*Rbg?q%_WY#SbLfQyfo`)|k{ zMC!`(hIYv|4R~FNBE3m>rUUIK@_;Ot{b6@d48zpZJu3g${=aXq z*$Ylz4JBd>H*OLr2Q2}s6ow$%JDvNT#ST$VLA+Nb=N!sG4EieSCjjN$EvR%xVB6m( z6FwvCEg#-&RhA}f_;!}a#5q9_fW8B^;djbdxfPqjvPT{ugsc!{bn8cUIY?bZ-#6{m zk!17A-TNDQz(N4lMUqXPU^^35`XrF>I+r1qX1mPycyM#<7rKxa5IuE1z!)EZBymc9zv4AHqC!O_vSRW->s*Vh#{|^2LVg=oZ&Uz?| z5#0)V=1IlH98Nrc9K;2b?jADpJ@Q#LbUafOdFpKFKB2G({s1 zAB$P%nsrtZmsZm1HJhSY=~qP1mb(a6QdvRhEz-ZAHyuelgf=;A2Jnan9b&X~_<;dz znWsBABi?WP%~wry3LEKDB*>f;dsSLm_HWdjG=`{oXz|p6F`OP9ePJj$;hd?D^>C@_ zjeB_y+H<5^#nhnxUQ2NnCItWAze)_hO1{il8I6Y=dUIfR%kR0j$&a1Eeq8El$KZ5h zdrvhO+08)*@b=SzE00gda0fp78w}Cf$hM4$Gh|B+ZAMxHl+S3zU2oQzvs4lJEy1dC z@!HKFhSB8~$4;KPsVaHky3sc)S9%^>nI+~N$_ zl7gCH;;9XgL#+fsw63L~c>6-o$M4DW7{17o>2O5VqWr&u--`&=B z9yUoWYyS6+(B`$Nm))?9ZoYm@qkG$pNcWT)9!7~x8kHy}DSH&nr;rM-WU+FDiY$l9USj{u-XL`U4 z(>7&j-Q0&}R!?0HJ%s|l^{*NUVmTTI*>5kXbaP@;jwJs8>isA}Q|FmUlyF;Im3u0^ zUfb*-fDdzA`N3mZ7{YswrfO7HpYY$ifk9Ooxtv=aOZ5Vzrz9WL$4hwD#lw1scBh;` zHs2OD=x)w)6_!d%F&@b2!@!W+yyA>foa%)?dnCUqB6S{!LfV}%Lmor)Ye1?_K>9q0 zoEE%R%U(A+A6{K(CO6e7oPtU0d{nv95RQme``!P!2|eT{w)>Y}^^=HB`I0FE{izGO zO#Lie&pjaq6{#0@U9_;-=N! z;e3yuN9%~qaOqPsQ2vrtfKD+U%pS71#q#_A1|8x!SnV~L(Z!6o-$7Dhe~`H{@jMZl zpW3g){M1l>w_sfS+#5PCLI>W39k4s^R|f0R`(SxJ7cO&>AK?Z|b?Tb_y`H)}vSJmk zTTmn)*~r#wa|c9Su0%Lx%3vvq;p z9DB()`M;dO8&?)D`p5KzKEZwm5_xHpws3|y$TqtN z>~A9b5E)7O@N?Eb0h-ef;RiUgxSdtNBQ{;h%?K5M`Lr0*FL}3!hthv`JzXyu98r>j zbH{{L`N?EDQTgRT=yRG5!EU^({RL_pS9@Q?c3K}SSG$pL8hh;)#|AJR;=wPaUgv}4 z_qB%}pwBz87`Tl%E78n(6qUU`eZHSm{^7-z@|C*J2~|Vaxav7(cVKk=2$FRvCkpde zB-pwmkrLUgk6h$|nzJ4lTq;jHX-J5eY){lmg22-fMk}}G`gD;0PDFf2*&lz&g_9}wwn?Q3T#E5Nfj-thKFmeEH2LZt)f z$mf0>uKpOmtBXx71`(ZbNKAbFYCKR@kowL8mgtF%1t1cBA@70g>UGkCy~Ri0>(lN- z3BJKQvv2B8YZ+_&z@q{<7W)&_k|2Y*w}H4IDaWaxmVSWVzE*uRFH~kR0YbwIQe^&f zwqoqgmzqI=J5R-2JgyX>ennVCU)VvUwZ!?CSMW+tLeo0wQy~uAkLn5md*7Kdb%xfH z$ti?HB)E7vs(M=pp|309JePiX$QZ6GCJT>?a2IOrcO3jkWr|3IZ=La)0Lv+4Kns*| z&3>I%NT8-q7Ml)W>I$J2?0OYQHu1pgtw1NYnO+Xzi=bzqqD}Gcd zxeJ9LqumI1wsLa}a43cB>h7;16M+eX^8c);d_TDDTebb#cT)5ibo5MdTt0ZrYY&c8 zl&-*Vfa{U?6jVN;l9c4dCptQSbX49bQ35ta6$q__vZca*VhdN+dA*(O!j6Q2WJXI; zou;Z->$umCgYyP62xC2xx9k;ZpUl$1JvbhE_F}$xd=jR<0(24*G7-C^?@N)#Y%>43 zl{qAk7ME|-L*Z<|7GEo_Czr3H$bU}m+i4+(592(0&%wO@)M1V42sYse2qn7#qp(N( zUl7(%MZY3f?Q+?48odV`k|BxJZvRpsdbD}t`)4;gTG^)d@p`^4WQ2)lQMFou{^NEB z%PxHO{ho<6sI1Qc2_@^rXp}M-o9k1RMb;YnW`{9F@51`4@vsC7fbJT~G2ncCnKSzu zC|3fkub%RS&X!hU8>U!w?&caNXTz7I*q92`^fCFfAP$o+`W#+tgm~%H0s=(Kw1k;{ z4keGch*~r8m>q;0%kAZnvN7UfJ9d*UuJjDLloGPr ztFru3qoX)sns~A0kRaEISzQg*%{N=BORlx+nvz}^cS)u+zVEcN58qr6m~YOsUeete zr^?x#^>E9};mD{pA?l)Iy2ThIEzWS~cyTyXRYeT+xEL`<0#8WVHP2pN3x)tP!NGwH ze~t&{ucq>MEZ^BH95PQ4ZfDXBzZO7==ysWpDV=f21E6a)6tO6=!Om;bDuGbZUkJkO zp2tpbW!V`|Cv9^|qw}@j)*)mQ%$HKC%C0&}%L{v$6WAof@}5vs-zUC6 zW0QKCpvHOu{RN(aNe<*gEk}V};HXz`By)<+G_( z$ov#<%X2&MW1NArf=NuuD~D4rx~@f&j-?yJ4EptE&f36kovZlBnE0JBPBu*PbxH#$ z&;t*I*OW}=<8k2BEE%&bdVr}SJnpzz4n@(i?;e4gB>*qf)m@wIqodTrX}5d+Df9yA z*s_6wZi@Ew_%p;`q6|V?$+$!Bcu&02nW$jCA;qPo$l&^D|DP2QdY2R+-X3i&3q56& ztnPD&_m98W4mwM6S~lPt3da()*a~X&CS_?*GHyCeIDUGm;WwBMSQT|a*Tphq)-Pfn z-@|B4K}~fwRgErkq8HO)9>2Cvuz2=#^1Sc81f-{YbN;p?EnyzKNA|W25Dh2N&V5! z>Y~h-OqcQs;zYOBd3{X_G9%vm z9hx_u_QC1$Dsj*5_>~f;Eki!c`CKU*KZg;W^9?k9{I|_N0@YhgyF7n*}T61AjdXfe1}e)O29m2nKxlD?9ZdoQ4h zG)pmdOIYlpG%b8|#;Of-9DjuU2dE-xJ^Lb$qYUJg3E z3LvqEt^-BU?0Z_kOiC`ahTRM>Z70Vg@jEJl&Me-V9yV9x<4tS>pk`VkUeiyYb-Wf} zUzTWg16CJ~na6J*Wvb6c7Oo>TAZDF0O-yUntwC!{2zzg5b#*SVbX%*6z_fkQJY78| z*u+!nDLNuurq>|Vyv3lPH0z-d5Y;LPyj0h)xG%X5ldMarqcqwEr z^T7ov(gpeV+UfWdeeC)&OE6{D8N)g)+mX?zM_>D)9lI69@an#Bi_?(Z08XPxC^f@e z9+i)V&vfL{-Ybrr;w{Ei_G(9u2XuD6Td!~DM8Ub{Bw+ z{0}z9SFH|V(bk=Rgqx^BUV4k_a%Y)ibd+@Kd7JT?vWkF7RnT%iuQkE#vk6=usCM{x zts^A)VT_~y2JHl*#Pn0H-taW2Y`SUm{jw7pw>bPU@d86=KUw!X_q~RyM+fXj*fnd0 zcj+md3OOq8jiXFsl+|k--!s+tx!Y_mE?;EY0G{Nx)bkQwzVqgYj(Au<643kH*;C+b zUB924tf~F!RLdZl6MyJHm66jr>xqY4!19bEXnCJrv zL1+FYGVyFzdnYeO#e|tAY{z{r5x<}WGyOy%>qr+(Ci3ltz6=`!jdT?9{a3E}^T_I% zKDj45aG@kRO)rva#aLKfx`n07Dqy<#{snU0p;znckZC?IkxdWAw;+>?tp7Yc=hu4N z13XfaEJEacyWC`+5IyqQLVht-uD*_}1OVbLeZw8Q=7ov+COqPpDtR0U zADVeHih$xp+QW z-jbryJ2%LH<`7xU-ga$>!@%%XvjMHG-bOQL0G4vCM6|=aC610%haMjz?`SjG%Bklc z^w617pDj-fn~>QjR^2!@WGG0VOsniPhQJuj#j~@|NNOYIeGEIfN@GQ-Oy53QBC>sZ zTDRBW`=?PSRUNS&x!WDTdTMFPctU9(w*A!E%J}%)&k)!u%BG#dc8p=Xd&;K!+;O>e zK1?7PTb-NtslFY1%O5K|sSbK}rM#vGCJh{nQ>LD4v1ik#O#MtIo+^`^!25=7CLMp_ z7*%1~_<81;uJPTAx65K>sS;WlAL8^7#8R&|E~HlLJrXrGI8#AlaHt2=33Q?A;;Sn~ zW+|s;_ix_oGb`OT;Wv|-ty2`<>c}QslKUXx%-B`C67OSWu6gYJaV`piGh@WtBWLm} zf9)*_Sv0Q~g)4u-S1re{84m-bS@IN)#7P=N^vl&0c*t5GG6vDDSk=6+s~Xg?{}85| zoy#AD!+SG6nxntGl!sw7-s%do6#4m8o!8LU#kWWLx}^fG4kgN8CkR$TMmROkbUF>S z#9obOEONJC7V_k@2y&IRDhv~lJ3BHEyy@iEOl1^5ed@`)Mm(chG10~g^^3ooMT1JP z>$i;d%Oy8JH1wkT;}bv?65Tbo*`hy@;rgrgIt}6U+J7#p76uh_I>Ga9ap|G%f*v$(@f%=dU*~iLFt-h7?MF7&$vGoWOasz(%q?cHrCyl(1OD=Pb zrN>U%e0~)#qWO$1;bG92yqyKL_@#-w4D6AG+jEeZx$>rVrMA!l@vfB~eU6+%tChY) z0D-d7T#-XN+2(1&D|+;+IzaB9SGntb*N#jpjDR{jSyyEy;I&=|`K`|Aop8lrG`KAF!V)5qcWU20*VCj2Lj784b5~RD- zlF{V{lcxL;Frf={KZfLJeFEtdXyXmlBYr{Q%h1d;{m`j_^=lPoDZsb3HliBj#6!@z zQ_tV#QiCdkA*q5{5H2WwFQ(9Z5+&~)8E162y_7fHkZ`DqWQZX$0Nse^7)By-0L`6F z_p2ShdY);r2<57eAmi2`bF)v;Py$B7X ziU~S%1Y{)w2m3oBdIroRB)K!qGIu8DcNV zr6KjwteSoHR!i=~pP$qgdE7B2U0RAa9ga{UB%Y)dukjEVbQ-F?q(q*ts6Q@3BS64Y z87>B9GTnfXewow$0%gh^2oB#e7LK5`M)k14oYqVRTch^L>y|88YFRcH%I z!PhJ?;03^7r@jn{aKH}g4gd|R$)Lb*^mU2gH;W*-MuXZce~gI%p(GRhn;HmS3DLrc zd;Av{QlbNBSujkm%ap;pLH$4YkDF0FI2Hz&5c%4#+K8`nUj zCPc0N@kj;lA>{J
    hua1e@+R67qid-A||Z{muXtG;OwdP21RXff{;02*Hal=*!Q zXyZ-6Zp;IFJKy4h5xc{13zWoj}!fcMtzE?07b>Z?%zVY=IU{=wn)B_s z_(**uT~XF|g5!V#jv0)H`sXx;99MGyoJBZ!z+V9sVJB<={Ol zB*9O;7ax-yF4zmD46c2`Q(*e&!BT+;(o^k&i|~u;-|Ffpzx}vL?*OoZWXLcFm2r~U zb>cu^h0~C#v^v?}_IM0ExisOrRV z5K>l@Wor;4q1Ga+AlV98B|tr6vgvYXkTp5$95bXNIR?D&1cw8;E)EPNNURZtrt!X0 zkxGeBkodP7@_p(Xll&e^$#dAjwM}P#3HLY=aMC zv>gs` z-~n)a)*F{#K2N$!M)0(mQ>fi)jzyV@65!mc(8SpdLk@xIUIGB^NTb}I{V8HMnM3HC z>s*#^pt5!JfjT|C`2pLfH0@BEcBGdSkei#_U#{d=aWIw;zb@So` z*4TxrAwHqz2FP}F)W_;*Yr3yk((gUfhms6k{;=M?cMABsr_V!s`uQnG@3bPghv@%Q>#4=ntC95YC-$cX`weG`TJIh54KVCkm;C^5kR{Mtu(2aCWVGTVJ-BUI zF8@veTx)BUpoIzSb(kPuk)p&{s`{6~MAs_WQwHO6Mjjp6C* zb?;194FKRjPBimuEwvR2buM`g9Tw0|hu-pJO8>rDn|8VWvxPnl1b=Gr9FSgn(NX74 zm?s(cR{>9>WDZd|@{KZm*?u3O0~%DD_SfC0F$ZL&^Ai>U?Y32g2_WvJu|^1V2P23D0~a%l|2W1ynoEC<_^{1zJ zeT!_V03`*~Cb8B|!8Xb}-e!&&(Xwhv*rYDc>{O2_{3VnjIuCeyxqf-wd9-b1`r0c( zIsMMtQSxcRPtTKBrT?8CSr9jPBCB*sp#LWojUC>FwfTh{;SA5t_-71k^Tv8eQ6mV6)=`_9w<0{w?M8+GzUH>VI)HHr@D|F^8cwSr4vZ`f%E+&IM1QVH$6spE}3%v-SD%L<# zp%W=_jV$Bcx!2vg^H;z=Ti?nTRXg%SO@g&2j(7OympILYBl{4!37~-@x8}n+cffb`1LVO*&}qw9lgbSw zqK8vNAJ%&4%-q~2_lvUc9&?{kpV?WZPu|U^FWFh$p2PY7{<846Ml$ZfO0J)!cCkE1@R~;Lo_g)zl@xxu$sGqB z3XDf9Zi32}{!4d&DS8M}SwoI_B0P4mI$>InVwgeALoeK$|6 zN<(Q#-P6co6pCux8!)x>*Et{LryO}Le?7~;9{T2^BkDCO;!v#!JCe^pKpI!G5dx*$ z^O@gx)F;^-b&&8jCQ6ebUMV$M7P$CW_T>6%{s79X!xW@~*=_r1`PVL&$v#butVP$B zVm7LYsd>T{1}e2|Gba+XOi60jl)vgV0ghLegrbkDazL{TcO_mNSStqr;FdbefF>&Y z-@KXby`~g*$H(@AM?Y#5vLa$%)#1h!Mt`EHQW53+3??^Ne77;2*-fgDy=(Ex1$={# z5L-CC+aB4#rAZwcFSm-Zwxe^yfAIK#X{j!QbD}75Jvz4OAMBe6U8V3hSy;?KOn*QZ z8%Yp^ob_(?CmrKIDG`dFwyz!>T>?S7T?Ct7OSDa|dU5OQfFF7BY9V(7EV}i`niNzQN!5^UR@*(>#Ku~^%bgrfy7*8DIOCZJU1!L(W0h292 zLpz52LH!^28jmSaqN_*j7>O=UULQoqu7M%J#d00eMwqx%YT7LKn=v=wg{W#~j?sz- zpWyYWh|Z8YIMh3CE@xRbgW!7wA5lQ6$<2PVn%8~_esrs-LCr1+@5l>D>qZbs*bi>f zj7Qd)qZ2xiJw6uI>klB**=Q`Kg{*>ofvuGw?A@_s~?sE0D2|C??;qUHx}qK4MVf z^e<}UDm=48<_v_o33$Qo9}=VUF8C$5gId<;xB1dcNJtSy-<#m{RbOQGRyhSjZ6dgO zOimqauYpbBFD*}9{QA^n33PdH0X(&=hF#}p2hP2}y9*B@((3}^Q?q7oa7|?sVGEcA zfe6snWF|@#$gC3g2SA|k&mNi?(Lca8Zu431Cpfvr%5)oyUsac;!I01FfIKW2#0rdq zAa11s_=IF28XEa7>(@r^#=@26e2Z*)jS<{ZCGcbG4kv;n(OF?&bca}bhc;G7LG>hm z9k{?*M!sn1)qshWvkFL}>TS>KgoV-wAhP4>?@%85B&*U!`%Af^xH|sN;-7)6>QP6! zBupkwybYga4TV#J<%it8CuI1;ee4bW;8`U~!s;;yHa0mPO#xTFWVdGz*;v?20940WoUTX4>ncIl>K;h`s6h|%5`kF^QkiWfd+0?dPP{nQ*7rA6u*IFA zH5y(6_G&I@Fmn*I1k$?gMd{fW{)S9j@9Z!V9j*Ym^95untd_6~=PL>yN;&{YXN_&h zm}K=RDVEI(=xY^Q%Qy}4!!F)l;roozLSO2MLX)HAc!iNPDC|hjQsvZ|hUbxihu)fw zPb(K9UWC(FqqnYPw zxQo{VjIFBEv?c_E2kRyTG7vp8^dm@kzC(1pDX6Pe_+bIUQ$Q)S?)>@3k~=*F$0IjX zNqym?9$I-Y`06;0>%@Koqb2FupvkrH;2Fi>Di1T!dgqC8O+ zO<_cx0HJu~3#!{f4i%VBSFizMk)nNlR6(sqT56+ZH(1kbg1`L1@`97U++qUCUSGG6 zn{t|ytt%^`mrnFxTl;hszmnDpImsdA{Gn)H-U=k$Js{H9`iQlDl!RYfCZpj2D;Wzv zoht+ltRx9o2uXN`I;?YCnl1h~vk;68$#_hGp2ZZ{^&f0w90zN<1Vve`xo{p0Wqs>` zf(bwf$Fh!jVwE@}l#&4TlxsEZwEuKpjnn7{=dcwR3pDaOcN&ieYXqk`nogqM3?GAQ z(GzfWZh!(GI<$z_g#Vt}|3fHG1J>PQXXwOFspVpWr)VAkOPz;+^$V#3f*rDlS${y# z7{hCIwW@UGTZj#n7thD@l^6rTgMKP9q?x-1ZlyI7_Uc*C-Kk;|G&ET9C@WR>fyRLz zBtk05Z$yM>xv9HASp$`uI1>gyf>y`}5MckVK@_b~ zGpTZ%^oZHZ7B5HGyP*its{(3OQB+D32TH0r(uLB{@{?)}Q67}CylVQ`W(ZBH%O6=% zWSHh$Ap;`&ttrl55kB5(8oGSDb;}<^PQ$!0=Qv>Y$g-7YFhXCsc9EXO8kG>XSJEkG z?UZ}o@LcdFxen4tx{84vywNg0X;AU{H;yJ4{d9$>1h+uc_m@#pDqU;n3Ibo`D)6CC zUYB<~@Hgv#)GHqX57-v9mzR?^DZ;OtU!KZkEfn^it;cRPpIr0I`-!HC{)Lm6U^~xi z2MyPykZ#n9rifna$3?a3CAMB=Km2AyGKDhlqb54XJR5GKzwS@3@JnG06RHiFS|qhj z-mQKB#dZkY0?h};P9OIafGXx~glLik*r$nRN1QfGNOi;JCvfpKi(l%;)wMR`-zUvw zEl1VUpNxBzQ)TesJ}DVZIjY71$Y()v<`sjomhr9&(YmJ!^n5UK$v|u?f4L7b?)B@q zff~$o4mo`zp@bl#OGARb3=l%ae)mYIcQH`S`~GZQnA+=E9L$Rl)`?!cP)e}{dI(y& zKf9zrCX-Z-{?>-nz1^xx<~2=f4IvDgIxVH$Yh{{cjf4!?!*JN)cn9WSIm71{$GYM4OwJJVW^_`o?VMbZ^2Yu6`0o(wOwCpZvd zfmA?!30JQPH$hOcTSDlWSzP8Kh+{>5AIJ;`Lo%WeGR~AwkU2k}=d40w!+D7ciKY}! zIeY%IzWB$w4U7hr{Eq^uPb=BgebkN+2b0`)AGwz8sIFfDbq$hpct*o7Cci66P#&?D zC^v+hz%(HCcAfVXZbPN(!jY!uUR`JA(@VqB*FH5W^fL+MoG3fu-oF!L5MzVf;_X|8 z*Q%nxx%^}b@Vobmhv7!Mi_-djVX%qH#GU|tn>a=nuhBCr?95io4|V5~^=<-V*EC~1 zL8v*b$$vKYWY6*&9!_&W(h46#JbnD)!RP_B=0Sh_OY!ZiWx+IO{8=Zm4Sa1e2MyNk ze7)yw`m~#k5?%X{{(#&@djE3s6R+de=&MRh+EgWcemzgUosFixUoHik@S{=mkQ!T< zb0>GjJAiYc?U0pr?gRL5(n;@6fU!vxM~Tu%UFAe@2GGIo<&y2)1G-Y73|gJ=X!(av zl8>Le)?^Qt@%)Y;75-3u_oxxXpX><+5U}9e$Ho>&_~S}GM5TUB;y>iBqlkVP!NVwp|$0LY0P zmR(MCHj%Zr)$~hEf_R;qLxeOM%F<@z_ei1yn7R_)&%7xVwt44(C43+*EatT{U*uUCl&f<%dZS5p>WIO6!YKFwL ziO`mn+tT{FB@ksT;RFc3F~Gck`Ds2aaaR8^;rHfp^d*q0^8F`Y$`fmO%kos$ z1lSYFIou?ks4(nsA|k#C<;#oT5vj!}-Bn?yVVb{Q9C0A%*d$&=urlp^Kt^+$kg`QN zQNHS#WGQDSotwh`C2AXUygc?52v>|j7Vc%+&{-Ocaxo_aR}G+d$COm< zR15M(fCVoNIlr7Ay;tp!BRYg!O!suM#P%}xmRuB8+{zcb{@@t7lyAJwuASh?i&yi= znyXz?2*|a(5{{lmMP#L?-?8N$5Lt5Q&{|d$*@@w5_FxJU3+Wr$BCU&xXARxjUYmKP zoJ;*b3%ter?~TBPCTiJQvbWWqDjc_Kb+|{BGLGugQE7GP{nNp{kdmH%TG_4REf|?C z>6VZ-84s)Or@IQVY_@MjGakR(`D_|yH2Npw{-FC8pxPvHzfJOz2t>`~>Ay*`Czz0_ zQuk|AO}{jqXy2H)fr`sRYk1dk$nXhg-qmAY|G=jg>ki_Ne3g}2qPp3!sWcH~o_xB? zr&iZbPAl-hkI0}RdvPh<>8crM!33lxv~zVDYOTYq$@6SYmmC&7ea9g8*TwTF@gxaX z@uR~~HH1vHb1%Ll9kw`J#cN6 zDJ#LrtNfgUx&7O}H;SymXh+PvPT=*899cs-awho=9KYxM_9RafArvwACAbm9If>cF zN7AHOY7q^{S{LTVb351h& z6|~=h0Bm<`^fI2=J~d5NjN)gapIWKp9@HM&%oT8`6EQ?jYAB5=X>ZKsMVm7QIqKmg zg7%N14&2!o^7JW^D_jM$`DgQK(ipwy7K_byG}>;{34tO9M# zCrk1GW6!Vr(~!?uSnj>PVmdwqw!`uq54=S0LBZXWK>^D#sZy#!hf#_LEy$k6>nBf5 z`0#LAuzA#r3O>zppI<&8eZR;yCb_#JD#x&7H1#&&T6X?vGV>qkF4h=Iar7xIEzyFUcOBzVMW zKb)96v&2q+R|WEI-B-(a@n&-fQ1Yt=1&6q9XE26nzeZm%*uy zQ(9*vV6r+2L3)c-!%dLi$Rf~@H!&BV9k3|-7hIP^pCpI(>V8>kKmi{ zbfu>P5Lb^Iz-gh&m~yb~_@R*FIx#=GHM{LLp)Y8&A6#Rzy4lO5>-eD=y z7do5nxXY=R)|%gJUw6{S*5lbfIuwo;q1Ewa`O{6+r)8G9I)f36Kg1YsroZsWPtZ0# zhe15~x9B1T>aU0qTKL;UzfQL{EWO3sJ~HN5?LDdR{8aAGesB1#eqtd<#v}L|pO<@C z=eWI9N|*cZt#F*R2;OX24wjGOEwtO+RN zd8V1m38{#rVMKzH;2KH=1{`s~rQguJ9pwtn%wI?nY^&uUZgHN-3rUycWHd|Laq=T)XBiBwH~ z`RKaTG8`>e;e_FAI)i{+Cg#)#3nEjsP2TjWiz8bL`EDLnvs2a3e4HLs)_>cBnyKS= z(jumhOxYgy28pRW^;mmk|MNk&6D?bPo42}jN-hCLiz*{p8Esu#-_N2uC9EuS~2Sy=havlHQ{?Ob@`;u z653K8y)U=+WKR*4^|HmnwVJ43c;L)Ayc)Yhp>W<@e*ZH``YYpf1RHTZ7Nu!;ZE*{YrkKvX)scrj1(;P<@ zwny;;5mUcS-!b~>GtV+PxC)u&B$>?ncJ3%P%Jh4+OlvaJ2S3Z?TIRhf3s~71ojo46 z6fTQ-L=rIQcjLSWs$&oBuYx;pXZ~4?@2FXdQT}5*IpiEU$K;TLWFHgvyVed9Zf@Yw zZXYRAj;tJEn;evvtIItz(3n(UD;(4J)Mmz5$6$rb)KgSF<@YGp%A>_cJv3R>E|fT2 zDK1_1yN~=ultzdoE$T zcIuqUDH1=>+eyWCwARM1SEG<7G6--Ra-XVyA>p;<{|aAYY%i6r`1G7n`dn>m|GgkQ zxA&gE{~|U#0M}enr5;xK&PIK|3N@s{dLq|NwAoa$u;WkYQ-}VxDs`QFooWBC9??a2 z)!W&1-jwkDS4$;VYgG3DCPxpBUxxyjgE_U_M409}^4-TQ5&zYk7zKL=n$K6=!m(G{ zvhA+hZU)H1PiZqyqIT%~8C!3jP@-oXaQ^(c~Fwf*5XZ&;>{r0RvjY5xK8bnQ~ih z>Zv~;W$9H{ebxMX3SlIMQO>cf|D3nt?;xFp_GGqs4K*2`=$-*J4Vkqk0OT(B@GenW zE&;zh{8*Dz^qWpgHrdt2$D~1@y4QB=iYC3qiPs&d1XlWkDNpylEI(lS!GTaxfBD%n zxcF`1V0!R#M_`$;MTJF41?lwO$W7Q8HK$`(tHhx_!0Q@&5W)rJxVZvCA;u70`b_SN zI4N|s7Ka@{^rc>Tyhfg!zyeT*%=<>bFb^cA^n7X1K^eJC(PP+!cOt!D}6877F}qQs}A#TobkKLY4`@CZA0znMr2ZBdtpiYMzEN` zPh@Fbt=^r*gmP)^31~jd|8yO>b~{ImTsb24NAW*j4}Ro7_^*@vl4J3*m`T=~>5V3X zZJ~jvABMp;8-7VKA7BP^qJ~)r$R-eJbk&N=DaL_+eVySOk791VZZ`ThUCewC*{e>D zPuFpLwsg?Ztoa73SYSj}>sRn;tSwq~W?L#Bnx$(l8V-wr_xZiF`^ZDA4X{85HPf(K zi`U$&f4*$rfHmp$gGQdCR^`;16?(ZT9f?Wq!=&ML=UtrIjY3_GV!=$+4K2JQH-HN& zd@_`dG9B40zy+CTmO;p|j}YPuT2BW!9>}(mH4LVsO%A0L1M7I=Ykpu=G@B*^W_6H2 z8$tpC=NjaktvK>^;%E3?dw~C|B+xS8%Vip_yj5P5sd~C3^E`hBO@?6mWDdZ4?Cyu4 zpkT#ivkDPe2UG=;4EaVK09B2#Tt19Oi49Rx+0L%x7@_zlWU2=ty4nQltPLAA$-jDU zDuGk@2^W>GzboL;Yt}jlw;vVi;Pen6kaF}MFyi>A&;tN7*cCC3)2Bt1!&rB_3+)0p zU5lIqWC$cA3xBG=8DQ3^!F~5!3l}n%|4lq2gv&c2<``7Jp+Q^~_a44uQeCj{CwN#) zkPfqJ&E+eD?5n1&-UxOvMD*m=Gm?L=_&vq7e)-oq~{6aY1c**pebpqsf4A zTdfeW;Ot`)P^~cN8Z*t^Yq?w_tz8%gJ+ejQ8eJ)34(x>5OqaWgRJf0ew7C-MixfKF zloGcdJCss<+@M}>?)9d)FUdr``H8OrE(Syp>K*p!dx;Wz4SgQw5SpGxmL$QN~EUBAhYeW}XFF_Bwl<6ZDSSpdNId%-Gg zuyPCdJ_Uy$e@qg~+m}a* z_gI7XTh}Xjg-Uy-=zr41-@6Z!M1}+u7~9%QgWck|OWi@nK)S0TD21w-bGSF?4myuR z{&F#8pSl8mxV{H**@IzGK4p&WA}JBhabm+}#c&`Vu};Qko0!9Tjs z#rO(UA_J5>(gKq8DbfNpo{CL7PAwg_dVTg{MQC%t$b`ybIIZp|!R1?=mT(uw)X zdm)v$rh*^CHg#Uqb}eP{=00x?kDKaxE1NHxs~Pk7+f@4i-xWYec8$-Y1|Rvm={uqz z#ImgE?t<1!Yw8`tbXh2t5)_K0sPsJ2d3loLuW|u#15X{+;D7$X7BrENv-+s7HpiiR zmwrdE-P{b&hdLVu?T0Vkj?#N@7wM#eQ;L+=T!}^;X)zyK=oGU z9q_`xp+Kx{RXu*CtYnL@0|gAf?|o~vynd5@chfBl1abj&dL;t{HR^}g-$`vk(=%26 zQ*$Ok8|I7KBn~QOazrLoy$_dVz|pq`5MmyP+xVs9lwK0|{mZ+bbe_tpvCM+8fbw@a z)c|JmB1;?pY|@q-z_-t1i$s{vasPml&;emN_Ck6fR+|0%H@BeW1e;$CiH6lzO#`iD zP6G+{5EZ?Dj2SaB*r%2=-wxtkygp#v9*}H4!YkjTDyJZbDlgJPF4ftycAwT0i!A`4 zWKW%7MQatJim|NHW?yP6WMhR0c+GqtS*p$2X;adUFN(QT9|*y>M{nI!3OU-8bAgtj zv+kCjeOd_b|j2AvuDOd!_qVK`jT?V|>BioCQ&+J%z z{ew^EWVE1ghrhlk+(ckBc)Ak#oLhZ>#E5s1C%phKkSn=3QDYYlDOBVd4Tv#0wZ zhOox}se^01rE^DCKc`$afqj}%(@?rqJDMCpkk|1&-89}}^LMpx0NR3ZyNk6%B#!bk0| zz*-3hwo`&b%2`CCAjPPh3F(41{C~a!Q3A1|KENr|yc~NpnqT(lKO3;PHBkQbK2}^+ zUxR6~hVhLpBR^B`r5~0_IadDCj&}jHi$=Az?=8AjEci9E|HqFADHeDGi|$3LNnT>> zm-5bCASb81y0UP7&mcMTmJZQAj2ZU>!N2GMr8p3YUDVrG76U>yn2-=9#lo^16C`xi z9{w900noexUH-YYhTIXr(a(*bbD_y2|4(7_&qq#$Xa~J3hI*)g%BE53g@LFC`fR>2 znM>3{>whm3^mip~=+B9pcvsB-fB*2`pP=N0h=eAS>56#TqS^cGug2}~`BfC=oyxZ? zNpbnpd9`od-6WGG-7@#fe{1tU2Ip8}uwBzTKT}OTq85)NK+I8Vipyx}-O$w=G`Grzeez+TY85i3D>t8U@ zo-xCOvhNS7y5}F0%%U615+0x%WKsqTWCfDNgC-Py{%>r>XUHi-CL{Bj4`#arU;+V< zMxRf73)aE{R<7_vByYSdc#HO*V73spa-^}WaRx|Geqf4E(E<|x{uiWHs>?Nx@3tz` zAF2BZRDuwXu2<>@yg_^hgE%r1cQ)?x=7giU|r4_?0;N@w)rC z9bhxpz<=t34)&6kYANG4Z_@%Ha{;i)y7U9qSpRRHI;d4i0xn@*5bc0jtAqpvgL6-MdWf{;MG+z~e1z`CtzOzGGHBueZB$K%&yrzc&f6oobV2}iM zQmB}E0PsTQ&~W7dRP&W!(xZt)LeK)CU1}RS!blBdq_!ww=i0VJYY2?A6sCEai zBPYKb%Uj(=2e%22V-Rni2n9cshFbIx=QgISEoP=bO1!@j1K_SxUI)C5DtdV!y`U>Y zJSzU|b9;0ccUNmGL-PCO2LfPaImRkYWF5fgOuD;S7qgCJ);?JI5hD+{e{`9-vW7hi zNYxDJa7s|jGn)vQ-1*4#1SrV}DK^s@wDl!wfykg3XBCetxpWA))dI}P01&Ng4<@t) zK*x<8aquH@=icA@fIL%Ra&)94eukjeY*L}~8P57uM!!P3sEG0e6sV=HWkTT4j=~5x`fv_xi}68!PX9!gd87f0 zA>Rz+IjWpFD|#@OKx>^+K2ToNe;l0vGwN5#8goA;?P{m+=TB3B7Z|eJHas*xM*sW) zQZdY(fd-32zy;3&b!5l981P4>C}x4c2p=|N-tz6|NLo2xqsALvbZugDjy@NB!7Hn~ zf!qk<(&?4u*;FBQD0vWMu&3KnO_0&oeFKDoTlVJDlpQ_1@f?CK{?QeKNUzNx&%Wp ztpLTFs?zE)(Bs8j2gdjTkajVrj~zS=i2WFP*}7fux;dhv4z-rAu0hUB&@aRTxn`HA zz>6}dB?~c?4nT^?`0aQV)Yo%4GAQc-M*7RtIT+D|ip6M;>pm_}I*lu@0ns@Z6wl<2 zg_ddv=d+p_w^bZgXo7-HeF|?~BjCFT)bXJu2z$$6O4qx<*mFg;mzTmwwAwt#I z_iV`q&ko^WdFGqTYw%Y8fLW6VlBY}qez+rW{L(GutPUX4ljZh?(_Jhf^pLnyAyz&i z%qJAmj*Gm$#C*cR6aguBU#ETh<>^+N)L6ck6`Lurbm+SpTs_C5SH9ThdjcB250j72 zf`0CTZ`cJ@e*snbJB<<*;+^OSnrcSykMMiJOt_wxrU`y38ng&AAvQvnf_DYnr;txF z+!)>n0^$;Z&MOHl4ojofufD{4y+jklCC`{>;8Rlm3ctW^g)1Wk?;w&--g>h45?vzW z^Ct*b=9@d+xJh+C`Z3tVQPgRx1GJCxI?{7L%6&f^I`yuKV+B(Hz)EY$%273Y=-80? z^*Ag0`a#{w?`xWn&SW3#+Gb7L1&?#x{0yl$<}FAzE`*eT&{uxSJ-JI`eQnIiGA9~S zL%lWN-~nbTm*8QrDu&+{K3e*r`c-g?oQUq zep~YVn=+KdV>TfW>p>=+EPde#PKbu9#drx3*B?fi85mJmvq{&>^tz;LCW_ZmdLNdM zx_$VWnK2NVV4(-J@M9??7o`_-cG;+Clj$QLZ$G3BTVppOmPb~k@&`delf4*Bj#L#6 zIFr9$8;F1SE^*^P8&~i4AdfKtMkHqSFPR@vMzIJVAcBm91r^*kk#s7RUif zuh5!zbBe#X)aEuV@y`MWy^B@BTTR=L6$)0qio>juWb4A8GN(8`UeNHl?4wo`b8*+D zd~O4G#ikB_dOu|y1T0y$W3eE@z@4v~6E!ddgF{{rwK1dJw-^~O&u)JjWOmGj&^1z~ zzZ2wSQchgkj0B3hrthE31hYfAOY66atUd3g&eDbpv@eqGZg<)>797;Awe^})a8P~v zZ8x-J(t{Rpw4pTCCIXr*Ey9QcxAXT;8SmXe!99>TXz(qC3xOaLvV*45_f37OMmoES z=dZe@#n_m5;o2k2&zL*H0>X)WX|Uq#^oA`5c5Bk?05z3$&$vI!e0lcM&uqhd4AIkTf@h%$6$MoA>*B`o4Js`svpV zSYrMNP{*H#QXxxf0s(*JBbkU_-NJ;iT(OjE7fZo``t(`Zfuj*pCaJ_hWu zK9aLG0^k@ln-z`0h|DYUOV5|a4smSfN$^CHc;kn%?}D$j7V5M)N{<8cbIVO698+Cy ztBkmNbH$~GYenRf1<}!J<5m3*OTzt1!o*0r?<=}eytv8V-bJn%ynr#v9m_?2R|`(S zd-vh#$TKP8pzcTB%T)lbXb2irrnDAJ6yEv#-H(!2<`LE+lrXc})&iNGoV1gsWJH;C zkd?zew2dTDfYK(Rcp@Hh7x&VjvyPC@Q9^AB7!BqD-TUDY=|bPB_+Kw!ata zGuf8zB1m6efPMP<=;z$}mQQti-pL&Mn!nVZL+8dWDMdyev8 zh3sldOs@&KJL)KRwo3?Q>5eUB$Mo9>1~?vI3-Z6uw=%Kla7=@1low=u=F)uTp)^I8 z``p_yte6}(ZK(6u$A?x2F$QIb7N0IHdC57BLNHkH5Z1Ulkf9<&Q#qCMRwd$4AkpG|&Y`NP@Q)|T8=9s8#3&A6*I zXU0qIt?+A#W9JVV{;R-6Zab6)!a9DY+Dwg-C6FrXSS_lG5jjg0Brsj@vB(Uk21IY|zL^WMk;UvEd8|tz53fBM6cWBQUn`N_98=#UvHem#H7E~l<5)mL zj<>wMY2qBkoXj4M3>Wj|TkP$s5`Oa|fCn|22%r3cA^c2qc^z-!(Vi}=GG=*!zb<~0 z;dZ+8E&2L!Cq1@?aLF=y30;ib>Em4frFG-WvxFV@skL*)C6u)K_2 zyRZ6(>yIuGu+zhl6W~@n^nTyEYF5(9q)WKdod(6Uj+FaE*T{;w`>B0CREaxi>gSqt zF$ogpS=G*e`(8LwaT1JtjZ=bYaG3gVzgIGG-1zds=)>X??T|U%M5(GopkKlxLhcv7 zPs1@o-gXCHU0_!h9=-fF-=Gk?+`+HqaojY($`D$#t^zmJ&BqF3AQ z;i|L4`aGNcF~z9i+q9a67qsw)sCa1)p~@sJ?qOdf#BDz1O?&!Hfqa&ZuQsr^Y-8xd zi+TZS!RI9RKQ7eEQ{1MVrH{DVU3P)0PTmrRO*%D?rs()G`N*&7cqv2tZQ?BIMUGB~ zyLiPMGIl!DE27J~ckSJWdPmB)?X%gxw79*Wu`fLNFwME^jesDQNFe{VM`BoBz_nvN zc^HBthtIk;$+@*cS-;W|(%E31(MK#Y)Kj9)4Sq$ln`H zR6-7?3PdWVZ&MS9Uw zD_`EH-B;I}yfW!{sz%}|j*UbDS>;mvgTHyJbe{IUFn@Ssw`U+r=h|7yfl4s+RQP%M zhg{tKtptT6MQ66l?*uwqQ3AVoUkSvpw=rdoBhEsLDcyyfwqhAfc6H$)FZ@)wJ_mLI zi8pRxs^2#-O-iD$;m>FMn#NgN3PgKQh9Qa?HROY%s`uS9_VJbY>T2urFdVj{HW9RK zWtIdC!Ol>Ik&(~MD_@bt%&y- zKk0<={9(M<48OxhW9K_FrnY|huwadT+v9)3Jb-wd8>92~-VLlE-~N!?H5wJW-4UMp zu$kS89faZE%PNv%=7@8SykGfjob40QJq|=YNzX8%(NV38Dql+^tk*Z|n{+%DvWc2v zc0{?K&9yW$E{tON<`CL3yu;Bsj0DAnb(jN*BC%d=k6pWJj`z8!XQh&zy|S0SH;MgX z4@d94`BW@JQZG1A)K z{J%EX5H-t3282`kA)vEy5=9ufGfuxjwzUi!3utmP5b)|P;H6eE<4>(p5%8KJnVft} zyCvDl8EZE8x`RD?L2$8@c>2XYx%I9&9SJsK@ORiOSJpOu+cPXi<|rHwC!2neDSJnlOip=)fDUyXfJvw#e4hz6~&Bm(UY`!-m5FFu-^*IoZPxqi0r)J?KTsum#zWtUXs}I)~oNhZ zcpm%^H#=`cAhXn1FYCVdhD@t}3D8M@B!)=~CZ%>Lm8` zVR2)LQsfAikaUIOxj!XJ)oe%XqpWt!as6DGZp%G-gquD@xo${Ve|#@u+Vg^X9pP1q zdkPsjV?cM5;o$V+cH|Gm>s7qH0oOjU?Z(KzXU5(cTFcQtlVV`nW34=}8^94{a@5>d9DZgv9@oh-F5MljfThRPF1;CQKViLpy^J9y%wy!K|RxulRE|KPzE{4sa!xcrXj0ZJb} zr@1!{p??kuF&4UE1-jG;VMV$RYz9Z<23`FyG~(AA^t25R{m2l7r!6F(_PWr6-19dU zk5gsQmRdGm-_HK9LC>5m&#)r)9Dj9LeK2oIbw9gGVI9okB+!834yYD66qKl*j z66cahixA5R9|Hg`uJ!)&hf9~Al<^)`ae5}Y^k(3nQx6%Sb3J%wNU@(wc3|FhH}Yp`T-+{et<emx8A>9#Mdpf zm7#&N!H-D$LAvQE6GL$X!@c=uYqIZ;J$)Ni=ugFy=7HpE? zW3AzzQqR@4x6F9T#Y3+6(Z(Y+p*-V!mV3KCF5DhS=k|uyH|hPSm;;?e6#ct*2QX zskM>%zL?D03vp=VE@?K%zbN-I6T@2?-(T_gVHgK>P-TtDb#pi)j;(%i9dd}{UANQC zHQh2_BSs%TabeMY_=4)bIeM@fOUHTWfwd+w$9; z;_|$zQ+m6Xv7!vy*CkhZ+kQAvRL1aCy~Q+31_z7!!HZ^!Uz+UAN5U%~4{yHP7CH_X zPCGZ*iyi)Zi9?RMO}77t{ancpA{w5%u~G2f9K_gADnuZOD`iVqW1{1SssTf zyb6}0(mOth(4fc2=IUye`K~%gv7amBRQ>S{{qG3M4SOS&E%i+b>l%fkuW;DGy(NMI zx{*Xf%GBx-az0^P$Hc?;-%k-p>J(iLSW`Lr_#02{(q6?Mm9o4p9t$IayT1xQaos$? zE7RQj8DjEfDO9wkxnZ@WSR_3QGpvz6nywu@cI0ybiKVwyBo zl~3Jl{$X2_EX`qQwUfPWCoYkd%tlQC&vqYk`plshDIlM+9ys-uJfdcYFOy`_JCkz| zh~d9y{MIDw$m!6G2Ky+N)N6zVe-3uL ze^=niGe_HY!B9nh6HbDAxTBxJEzq5krhpg9bc-w-dXBU?jqj!Ge2#(2mpYS**eOm;G@@HoX&cLAQ$c})U&S{esXP{qWejsOT z{ZjB)xAbV{i7lHBa}W@TPdcU z2gGXqmwWG;vJDn~VyOKFzl8qE2tnz|*jf%L&f`SRLanaXa%$!^wpVVHE8*S5Kwzo6GheH`I;+8q2ZjGPA{xsU$)^w0uKB43n5QOUr^EH#=PW>~S*)^%9lvHYbI|A&C3QQ}18 zEvRu#|IJTz+Jg{&&UTq znuSK`h4=$C5^N8ApAr}#!vubReX|14?7jHB6L%rxh(jQKT=1K{o~u~F#AR2$2@J>sW>C8yp>!k>Uv`*L}cdeC792(KB9 zw;m>dqp@CdCaBP7EVq90gu(g=E(M@U&WXLE_vks(pX|!qC4Vr-!#WH<_rGP_WJk<) zb(j+~3Az3Sw`n224VZH^GhaEsy|cS>v)%ah{Prs4&cGddM;^S;hwZ!b{?RowCipRW zNtXC!-t9z&wbVG>^Q!RA^{>0Vg%{3BeZk9IJ*3mKw;g?tbkNw!y+uy2#j2u(k$DYU zw_TIF2{uO^E;3TBD*UIP{C`Znby!th)b%ZhaOm!mP6ectZb4G%kQSt+8$`OhrKG!& z?(XhJy1U_9=YF2|eZTAS5A%4=K6|e<*PLVg#tMqrDgYDJZHtWi!Y#Xwm;)3z{&Nu+ zmGp3G{%C7HBgH)~Z+yoZ^IA0=RDzcMLT1*!Mah2wq`6l3> z)BsYh1qSDCu_8L6~

    #nE32RufBV2PQ6P=%xuJYrona|7r}{91+v z2yPv?+S$;>yg9V;Bt{8t<9XIixdw_Ov|O^ejeWKfC}fdT?q}E;V4sj$cb=dEL-E>Q z6KO}x1V`=4^I`V$X4Q^$+e7mqPGSKWga9EH?B1dDy>Gp-^B!O@25X3qUk#0aB?doe9OG0$N{VLH%GBF`p$EOiqA#ebzSh zl?<32k9xuJ)yX6voF7z9#m4&=@FZI!9p%0c~98YbJ*MkJSJdYx;w(#l;v z%F*{=OJKxOWz`YVmvEW_j3smEH1|h9YdZFO0QNaRn9d<&Bv1vJ)Sn;nRKYhXN3qH< z2GtC(lPb?7V6bALOS*W!m{~|W<$8XYn+GTwqs*|RV)L+UwX^czIQPNG-*U0|(2b{& zNa^Anst$4hqwaYqJ)M1!Rdj4B*5j+(&x0tdm~X$@Yrg=0jE)m$X#>RVzIcu{`NA0w z+@lT)M#)3wji;vZW6MXBLN-B6{3bvqIAew1$FS^D=tg~RVH`Ar)WBxkIU-3j^9q9kKdG&9xm>&)=G zcVb7inS{$Du#f^X=JXf%B8*Bjue}QxlctZ~gFLPScKMI>z@J>zY%`@s0P9q&=EM26 zQ$xC3tt2%!_?51K`tnRv(RRyboaO!dS}mOZy0tj#BMt6Cu2bn5bP z_gU-W{k52MjAVrw`(-nHjB$nAPMdfk-}V>7j=D%yYPR~^jf)hwOJxf>J%h^+$1<(h zG~F-f5wfnQcVdG)SNLbAu2;H=lMm~R@4{G<4baZ(AT>`_*YmNJM<#)&vei#BEm!Jn z-^{S%j#^MsPaNEy@0m*FDHTw_K4MdkrUP=})mpE^Js&b&$Rl6Xb6?&%JlM9S;qseIaxnoZc6pq*LG?k zi+5Vh$aXj}&4x#j4Y7<%eID|>urjFFG%rsTrjI}D9`m)i*f=QUyqL_ham3dCda27i zxC+bb7Y0Gw$MnbCzaIjr3R4#I%}X{z$?bgvTW?%}G{*c-AgyU5ENLV%3<~4XK8)KN zn=#vTI*Nyw`-mA+k=Yfl1#<(>Z!urq?Z90$k^-=f+bakR^<3o0L@~WIi-kiEDXyc5 zK{zJR)l_!jvFkK+4I<3=VINUw#hrSbuDt9xIrGr~Yzwb7#D7~Ee=_^sytHFj9zPzb z$s6D23O-78E45kubZ|xqnmzz&Ed@DjzRf<5j~nFAtXmq;)t~3*PtQ{6?#WJU^wP2| zdDM|C)xRO^aysjQ24VR!x=~Vtmd$ej7;<-2S=BwQne1pd=!wHfo>@#e+VNl2@%8W@ z;={|jk55~fJS*=<0;$h`$;|4K7EIuB67-R%FnxjQE4>5q@KAZ1MB@HwLrSs~ zq&nPy3P`e8U^9kvnl)dO0CwQ5*%iQgun$eZiVA|HffQIXdXR+$$B3o7R=%Bfyih$Ru~pvKfv>6S z1-7eT2#3f|4B?8$Ax_nm&wv6@^DoK>ARHv$Ffze@m3qzXUBI+EWim-9z}U+p2$a+x zQ=&kAowxi8^eLc%;=c)!dwM5UTw6SyAC5Mmbr9e0?`GV7;WJ<=#p!6L7F><06qAa8g+2K# z&TbB3T9jjf>`!Hj0mQ+*=vk>qxe6rU_z?2PIDuoQ`Aj>9h16ui?p5By1!(Mg$Z(a+ z!$C}F(hDnF6NN6A^hBUCv#6zHMjBXTc$Xu~dst=h-n+mYALiXcxse7-|DZqx|1V=$ zSApTK9dU?hCBzO3hNoOrXDRoz9Xa_;xfq=Z%U9l%)}74cuP*t7h)fvmPb%7T2j?UA z0{bnH3yl9>e-YP3I3;C*gGxFo$wARGIf)#Izg{c>WLkL`kmt_@r-;O3h z>@Vf}DA^zm0vacR}QHf-~Y#;=}D zbLN$o)sb1a`^CoH{EV1hWE9{=>&kn6S)@6-X#)}+h&(}=q#FOXiZ{7O!=xy2y0D!R zKn$*1T&r(Q|3d80ym*fY;^1_FaBZ;z+>XLjtf!ss!Oz_$Ximj3);6PW{pEfOLLZT> z>_%bQgtzffB@Z+qVT5weM2v>?BUm!j*o#$Pc?TIXHl2JuoH8qNa76S9eoHhR5z0+= zrF|@>TLzI~51m8hP6*}JPJ<^aHrP%RRGZ&vq>NyK77|*ToRV}P{b8GLcc}VI4i(EY z#bFf=@8bOp*x2X5@27gH=0FcH^P;Wsn`HxU_Y0G|RNP;SQNsGZX1zZ!0+L^@aY^?* zbAS_d`a+=;jYT;s2U|y|WZ1la75R)U_r6VSz&Mn1`ujGg;u7U37py@jr^-p$ax1+7 zDKtOg5cs4D6bqh@fXRTzu$7wGAzkDw#31ZLHwYov@>1bk-2hs_SH=YL~81{$yu?dDgmNsq&vu3E@SXv%Kue%oM9I$f|= zUz2K7f<3~!MWWjHtcE>=koc<+T(zhma?b+|57PulZ@TFiztFZNhdg;b(6CnLNhnjn zdW-6U9Jg|SnGv*uGcyrVdrft$?mPSs95xfGorY|HcqJ1W|B-=05#;@*<#x#G02Gg` zvz&}1sz;dAY`x$uEq_&}DQLcKu2KxKb>Dp^!}^hnjM<4-Sy2ECKLgckk^FncT9m zur*21jXiZCr}TI00q)_YH$0_rIDs_j}6*N-tEa|T0s&_HXb`Jti|OfZWpm?6O`y5^ zuv(i?VOCXr-IT)gw%N&j-7F4!vhmpsqwTAj5w_f{Dz4toW;x)lsgG=3ML%?aneqf{ zyPnuHDBcY4t0ZJGV32M3A=^}9rSswY4P@Qc{<)Ug;g=F$L?rQ2nc9;G+8(A#J(Sr{ zprT|ni3M=_C)-j*BYa1#EWP;`-0`aK>6Uf(B>H)yd<51`DR6!S(!}j>d;VhW=5-H^ z=^=AvwzDxlw^xMX8$IY;K_wn~?zWGti1qXy3`OIvum$WQ{WUF)gY&eH_e=P|0}RE+ zzmk~b1(V(d5#%>G{kBhqi_uV4udnU$&bf9@k}2Z?kePayU!;t@nH`<@x&nU11yoOd z4n@1j`f_YtIBU-Buxo<@pOC$|`K%K5SwoE{f4wRF49SL~qrK@Oy1Eie*^6`8w)mz> z4ESp1&ED`Ao~B2O2hd4>YYo^hLp!Jc@(Z71shgaiBtVYCrNXTZL?T-Nn6g`K%Za7L z?c7WagYlAN0|0I|W59Md{oD=^W5$-{_*~SyZpL2!hpHQ=o*$6GD8ypT1~?_4J|ouS z{Q2slJP7dw={VpqXfR7Zt^j%UjlgM2*UrLhb#|Zy<*>(f-S=;40$YaT(Pz~`TZ3TiaO5yj$hi2-437g>%L_S z9%BI-?-~Gyld$hFi`$oPW1h+X+;q<%LS%1@iFZ#E;x-{WMBcg;x}z#XE|$K5coY6< zVqNFJr$rmz^M-ag|G*yQ>zK#0^*H~b>bPg^WBdpfH>_Vn$#RBNV{~+_KSd%y%-$bf z_niioeu#JC^H@)*>TLSj2iBZuH*jZU{AM8{*}DwWT=)C54lkY8!%Gn8$a6oX{M7cb zFdC@P8BgV8*2z*j<7pxG&?=f9_p%PF54xxhAC^A`62A{57R-K}Tq8|CA^h4X`gOi1 z88CDOKY%X+n;RaRyB>=hB?uDOXM;{>f#A|^*Mxf&ZmV7Vf{5I)dh0f5whEYk^j63U z^KbUx<-a>-r^H^^i8AJQZvd#R{CmNfC0_Tyk)cnEYI1yi5_!w0xmk^&mRkPC^vsJl z(u-#CH*S-yt)UT}+j)KlOG1QU_Ed*gVvJsa!~=oo@YG~}_$3pdixhLCmAjTCi1F2g zdnSg<*~zs1af|fd5459VkgQ#}S4|=pS-ZBK21F-dvOWZw_R0YSgZAn* z=QZtVA2U^{#n;kq>9jkC@aj%}hahFB4?xQahc~-*Mx);~PkwZwyt-J0#LFvsXs`y%B|%u%4qPa%>S0t$FHM@`BMmZzJKOB*Ki}|`D@2IAp(jx zf^1p4s}FP%I9R;BN4|&)nZjKH^9Re7{{S6j-2fv^9=m^RPGsjsVFef&2?csQe{)+X zjFq?bCV^^2Kk_7K&FeD;uMu#sI_8ahjFH|XILYWxlQtz(B2kqcG3Dc1FaZIzrvT~d zz)J5-p4}KquvexAZa(bmPG5ZjX>(gEWj^NoP9EuoI!6g|_Y2*@uKvaS$PfU-Z955G zU~GbJlw!|<&$lDlrgZ>%BnwU380T7!(HZdI-S7wZ=QIQ9ygJMAs8KO_=1P-SpsSF` z;VDs%S19vJU7pGv=f3sj5#|x$b!S$g+}`eD>B%ViB?=?_$2e`5-SG^8p5hA`Jp+Dz zhK}=)HavLpKdd($t?2EmJ;Dg2ILqfFO>t|@^xW;AUimfaz3Ad!rlGFMv5;Hbx27lc5=Fcw=pdXcM&LZZ3~{^@rE*wq37P4u>ONyw*8(FY->VFBSw zZWyi?CB|!?sV0i8lwU!6F(dfY6WA1X%ZT)`bV_D`aLRoQ@nub@?cmKjEoRB63~=@r z^bxP4E3D#w^cxlJuWeJAm#M|W5`BVnQIQ0go65v&h5sNFae`7q6)b@d12uz-&HfTV zhqViiSv+rp!NO#kJROXoeetT*}8nSzUoS(4UO0EZzoeNw0BYIMyH; z{yt4jpsDyMr@bt}?aA!k&T~B|h7PeV`cSH!St1 z0_ls#a8bQV9-~iYg^}K^mBejzkzG+wHtK?wLuohZB9r=_qkKqRxs_!t!doAD%2S?i zx5HhwIXabO@tb4}Eh~Gwzo)78U7MMzQev*Cwbz_uJp{guMH{114D)ScG#5V9^{}|n z9yYtqeHx)5IAx|hsEa#k+_=X$wjx(TTnqQX=(ZR%pn)6Z|C3QrG$S}*Z>n+}WI^eS zAsF^o3ru2$4?y(RPZ0x$!+ae zb9R_1JRkRUygzScuGD~Vai3w11*!Vh)6MhR_a6@y72ghv29=tdiS0Z%1<38xW`{Fb zNxP6Avwe!Fdg~Sj>POiG;)w#+daaROie?GYTB`)^&O4PIiQI?}_z@x3*eYxhgVsgAU#q<|9HqP6eTXdC41K6#_~^J zXEx3`7y{2b!g;M0n<~wh=0tWzYNFI1>Kf))8e5mks1$p7H`5?jJk3kr^CE^gTDVt5 zO2zPT`Ra7^IgAdRD5p-d7&%%ErbXk}E9|7i(ib43&MUxQk9v&YjZp0qhK^m#pT$Bz zw_9N?!gXJWn2NTiYgen~rY%bF`0bq|O<>Afo+_DnT5C~GiZ(dx6GAw>e}u>HGCS>N zcO|1B87b>)@7v7cQcs3&^@$5ocwKVTj^Dls`8knT(8cxDBmH=PmvEm>`8`cts+FcgdO*jH&8lRDk=K)95>%>w1 zb}fj76f!)D`ayqafE}TTWiOm+pW;uwo}CY5)O<)&ozY@GpEc6bCQBhy#iY5}ZAKikwbK46_YV(yK<-Xk*H7MT~>X*yZss3xZQ#^BGpd)kr!^Qd4LW zL>P9Nfr-6;Xg)VJy}e5wMwL?M^FgH z94g$;!soUz{|^h`;H)lV>%5?SP5mNdyDd|GT0nl#fSDhYKhEcnP2t};lfP>i24w=8 zai~4&KhA;65u23xoWDA%=WX8|UtV$ArnwLH!w12Sdyq200&O$5chFSTfI&Ax;aKr| z7OE6Jh*TixURcryAxyG&&aHqB2kXio72a^bbj4uw9>ektLy~~I*GKSn7f=_(LtJJ- zA#6V_=+o|8XeeW4^~1c35+Kp@(&OW{`%X|r;pUHNEHdPDR3<>On|>;yXWn(f_9d(` z^|RIlBs0=I6^DYEzMuZwh%BnT<~u=WS(1%BCJ)6XJUUg|?hn4LtXsI7DB-Sv_pnr8 zsj&9B)9pNkbW@U{@cHhd5m@xRlqAL+=cLNA4*MlHiN8Slm-|dZcH|S^x*K(S>Zm`? z9RYflnlZTq_m^dfs|qZnD@s#rB+6Yfi!p~MA0v>orijL7ju_TQHA+O?YLwWQ_s zF%l;0oAKYIm1%-FPG$NLRdFWa?*A@HYH^+isrm&`hGyyOy@QAlC4~?Kohv2E~7T zjma(u9yi?1kI%?8DdSOqLlSEZjgO!i-X^)6=%1O`jo%DUS$0NasF3133M7w`<(i|m zwo=%AKdu}tvba6zWA`uXkS6*yVfS7ju@pr#UgCUj<41T#a!jQ>r3TldF0sRZ~quWA+T)H=M{bWc9fO!sy;qkK{t`B~gzl0?g6W zkqqYThPS=n=<()PECtB$)qRu<9v(+BjHOak)##1uz3sL9s3P{ob>eM}^RLj(CYy}+ ziCe2%Qg222Z7xo!i2@z_J3Nl_e3|`8KKij8Gt*_6A7CDlb${(Fwnn1f zY@4Aw*TUR0RsT^kTAoYph@v@M>m_yY@VA0V+2mpN*6kS@!+R`X5owsaXkn31|U#g$k&wc0EO0MR9U1cnKiD2rC=#J+VM!D+9c5Q3Bo)Yxq zb!KZ5bqM2nJ6)dy53vFZn}Vim@Cz~=oofc4VlgWSa?Cp(!b)#(O1>{@5Nwv8rM?qZ zzZ{u6n5b$x@7WoDE1%4=n#$@Uh%CL2&qXCa`f+^X|KZhUW(!HUV=zzub=RlaaLeCR z!&CwHM9wr-2St^`NM!gGiERpV| z6z7^X-PO(LVX!;cgx+drwVd1g)Gyg=2 zt{t(ALC`t&Ri8VbVQ<7qPYu0~LYFI*<-i8#d9{rHy>j1?s?>_hrETd)gP@JNZWDtb z3-FO+WToRa_Lit(Lz0P-Tynw-ztLmz{MMGBaM$8po@sdc&Hmp< z9HiOwXnwqE78GG|f14}@$Mc`Kl?PLfCWVl+Bb^eyS;fC(ynLP(HbrI?q_QKLjGgws zBlo?f$RdkZqHUGr%xtkHli2(+pM(!nA$Ip`u-lD)3u68Cq@;ojJ&=Kw7KzEkn~~GA zpzQ7w8)omP*Eyea8;z4=0)%q0B7~<_V+~NT+Z_|!7x2oxItLZ}Mm04%kENR4DZbIo$5XWA(R@bF%~(K*k*``lFRUx%D(UwDB(QF4vv zyIxac5W;{_kx=TNgg{tNroMvki0}25V%3|p8Ouh;;G|ncy|YPLqtM8X0a_`&Q;K9L ziuM)1N%+yf)0oAJSf-*;`mNY}>yssj6VN+V(N-_}0u@Bi!lDbjFIxA-|8KI8o(ut6 zp|WWb*MqWa7dlulS)%Yn)ye}_RTd`}j`_tG9Jq%Xu)yQ2lVIidxeLw{&kIV#sY(zc zDW)^&7D2AAwNa*neMKQu({j7b?fMV%{D7^fP%m}7`gOk)DrTNMP&(pfK4ldhDmjC6 z??cdO;nvBkQcQl`B%Zzr2e}7*??)Y-9S^JfiL@u2X(!WLDWwd^#XGN^3`#XR$DMsG-w|LFvH#9y_e`ynT>)LO%Ug&sq@-zgk&gd2hr2 zYW40VDQzYT)81RGrt5R$_mXMH%<1aUW?IA_*gfK7I_^^dSDREM=^rT63>U-I zgmw(&>4W))nI`CO@5f9YS-x(C4&36(^cxI?;GiiFRJ>bpnwBQgw|P{3OlX1+KPD<( zzHFb5^ERW=1wHZeP0EUt>F7H6)5~%*r0w8uW+};^T&^BgJgu%C&b>D1&quz+AAH3& znI%jN^0ZJ7`Ch-GzF))tR3ka%|$tx zmjj;_p}VSoIoYl7{PBQYt?S?8e?T@TK*S~)ba7;I`ACpucI-X-WT3Z&DN=JWZsfQK z{$!_$G9pFDlETvtp3G_0>mUg?oqEkf^?pb4pchAoCV$ij%ulC>Ucy;)t5)9uCEsMK z$DK)X58ZSraLrQqTtNW}8rr;P6cJ?kBikW7(nheN9giZnb=$*k2}m_zKby&m4dxVN z4Z=)7>)GAEtnAWph10GUDY#V^<3FJlm^UvIgm7`W%lY1OhL#wh8y0hX`6;a-uM0>j zRnUlxlKMTy1|a0K0L__dsG`UNaD}GRWH^l|SO2ns}nI1n?lTlUbrr1!@@@`nx~RG9w&{E3RN+)TUTh`xExx;vNVn&{6xXd z3}4=0oqR~85^ALJMhxInawoZH8HFwkNU|Uz3hBD=g?O@9dKa7#=+`!Y!G`Nr{12`Xd}42_Iztv-^|t{YWl*Kw5Y8Npk+*Ni5cA-gYX;vKXofC~Rj@8l2IscXXH?r8Zk5>y8K(6(n!pEVpx#2)E`-u#!UXPiSg#F0 zVo?5f`Y0R9nyjRj$~6wGpu!C;Kb`kM^QRes#*Jm1Z5 zfeE;u1z--Tuqw&%9wa>gO8E=~qJe$Vj5L+Uwcq+mdnhVX6d%R#Y78DXJ|1c*jx z-NU1Dn?SH;4CR?bvN^8#;@7;w*BYfw`u@jmf~ncqave0PHV1Cq%%CzZ=>CH{P#)=f zaEg*iWm~v9Bm{4fsvpE~aNY9kzB>!btQ8PKp|KJB?;W@>o~vEgXJET1m!eDIU{wQ` zsiT~v_sa8Lq!>Dq%VR1Q{7eBooS^m|@|x@p575_}XpZ^P#yVgl2h7eo;uO@nzbLTYV?*VxZ$m z2L&{3a$Q1YJy?`(IZ!?f4A47ypZdVVNqhwbq69j&c3jVdCU^2~1(w>z-Dv#klcWlb zaoQG~lp@q=|Gyqw9g8r0JxVY0{tI+ZBikWF}53TV%A8STkR_V z6@^Y90>8g;_wF{CI#EAJ;i?nVEBA}^NY{5j=0tR1@68TYkSFYZrMtfXKt%@tD?iRI zXLRUJGNE6gZ9e@q4)gfBpMHzQ4=@9=j_$(KEK(W_O*8a*g(tbV`pIqBdUfyR=RP1M zNxDh@_}v@b!RJab?zo8fdmyolFuGdp-?u+9zM${Rx-s~suYm72=4Re51MSGSEr!xv z5I(viJ=OJM?7kV9sgy6zbmt$XLU`*>k59z6UfUnN$us{9AkzJgxC!`vBg|4?0=5R1 z#Gm0ebOQC{Jf(a(JFCzc=g$1GCDq7+Tm#EaOv5(yEhtj8zmST;bhI_mW{8HtxPvss zxU2JS(E$-clsJ06mZg_CA&?q9jGM;B&TZt_8h|WSU3)LzkAW@?ja<@Azt{MlZGk}Y zLuh9<&1ElNkNCcJW*i>Z4|CyJd9c4kKFh-ogZ3{F-{iD<*I+0c)$T zYU0Wr|Kav6rWYmlC<&wIS+4I7K|BJ|L00c70T&j)j7xlBD+qRc(gvXA3c+~0rD_36 zS}qVO%L{VrVhPLM)reTqgpBa6KLQ$?h(s^p+6?vk*Y^fIV&Q%mSUby9=fke~Uyo!F zDZcycQh;18IY@(nocUsR`?$}GcY-WP8QFZ-h?~Ld@ba?~y;_-H1h`GA#w3VKg7U(N zgR|XjZ^a%0`8wYQ`13VwGE5@QNat!oMUF#F!V&?7oVNoyPvhZqn{S&g#!tTHJIG8Q z?nvMKAR9FhWnu6?`w0se{eeQ$iGz-Z+JKU5zZ7lJ#%Q0d&@JzCB<3A zEeR^4#u9c4(g$O-*KR4JF#EEa=Qw$zT1^b`-h7Tqe(8PuiR!G(77EJ5; zt!*wR>2gdmtLYifr{|S5?Bz%;Ix%Jk{_sZX_;oAIy=14jOHlFEX2gJ51q(~GcIGY0 z+pOzMMJ9OUc#s0eS~|wx-sWU~oEXa`3XW}U0KhFg8GffymsAS1-o%FZv|Lr z>}C@h<+sm4*b8^sz?gIB9!#jz>`&_rkY`il0B~a{3#Vc^zWbqiVo}~2SiIMp-wG_k zH*0kyf3&%L5B=0sUpU46dA|4j>N9#I_TmJ!UL86(fn3qZcTS|~*ay%r%7u}I-q#CRlBGHA`%W~P6|z5>c| zE)aUGxiwicSEwQd3K|l`)x(p2u+4b?6uX5G=Vo>;BCA6O^$b107W;zm5rdD+1e!N{ z5X4kcR;&HRd53f+9>q?eh0a}VVA-G*(-J=u?~kLuMKzoOEX!y1@sV_*q)ilg!fW=# z>o|3z#E8(+xJdt0Fd}a#NAVf4+RKIan+2gZ6IPmcFMFoV$SL9o+CWP|AF`aj=PvIspwq^Q+!3qrR;aOR6@S5n=>(W61%rNI5}E%YpZ>1BA>1#JF$6v zF80ULle+NdNY`yIPOzG`OnxJ31`jZ^euPlHY%}vVMmF$covf0B#?%)r^XW-AjqsJx zXJ%iyHtHo%Nw_${BSNI7?oYC8aNvv;HZDYM@sU*`KmKCdX?iqqMHT+NuCK@Mv^F#! zTZ-S(l(00gnPD0(`#OGl$Cto5DbY&qBrDDRErk-bG`z?+u^R&3tPAQ>6Zgl4(VDWZ~2ArjC@_9x6V6XS@bFT0Kpi>jLW3(bp8k+$J@B zRewQnD~+YDFQwr*@5QL7XwG{u9~;9ScZnt(^>oBevxqV zjC$Qc{(RHp6SD${+}gc4ADc|p8c9C(JA7$ywegF<4Y&Jubc-U+Z^xGPBA0|=L~4X= zB4^~o_V2YP+YU-=7R6ETnDypiWl1KFAY-nJ(6wWKRUK4yCikHtg)oOs!3x40_T6mz zkYg#mFfB>zaHB~il;hie;OnIuC5YAW(eikNHRz}D!& zBp^yu1m$|TkNbmdxbpku(7I04k%-U;oU5Zpqy=Z14vH0VxEw=jF?k&|{?qD$c85o7 zFT)90=@tLZN?R~$$b*v*V)G2(qU<$D&-vuUDOvl|*xLuQ7)xZ&&AoCJQ@`yi9!|-X z=1=2b1<@#xlf8W{_1{c5m%{Qj%F-m51E|;i>vaz~F0-%o&P=2_HJAwmf}&k zI<-Cl4kLmU=6P=cBJ(L)TY*-OX;0ZBSm5Ys5Hp%m>zkBF1cTWd4;+bpXS9pq!%EfUqi!t0 z!(3``kd;8&7x2;uDcpjvK6JOe&O0$t7+%g$iLJ3@Whjg&&+&VFYQf?&-sqL1jFr?C z!u_;9en#a`hOnG-x%m?nBE?OONQ#PSE0}I^ju=)WMnc9N!hx6>o%2fL6J!xew9@44 z7m^Hec^ReEcp^6tZ)%e9CKLP0s|@09V;WMk-N@nE1g@O-@CHa=^dRNWo-;elUbUE@ z&zs?wm3@q3P!OgIraXf zO%6Mi@?|mGp@6>14C`?wYOYG`D~i>OpMKlf)AMq5d3)D75q-4CEM!kc&9fGgte7*Y z`l$5NHk53>@IP(p7^7jYwcy%3cD6{aT$6{VWLe`^L;6&Bc@EIl-CO_uz1S-{e?oZ~h_iSZzB{ zp&G=-F)_a*x!!yt;#Z26CzEprxk-MV!~(fYPDdwyqEM}Y?3lokZE4Z|@C)~hZVC)}XFM$sc|XUTyC671-K*IUZ$%`PA8prihg%7a-0 z+bhZt`fESJ=%X3+9zHSs1h%dx&!-8u3d|3)LwrLmEZyHSQ-8p^0^KVwoPL4b)|bA7 zy_EfBuQx+O>FvBzC0+Ok{iNRg6P;*PM5Lfj8+{2UDI!#hYla!Q8Dj5p(z>x}6!iqK z&cuzU)?ZV|g#OTSiM}xCn~^MJU~V1w&4~BM<3#wptKJlEMh2$hr^asI99=>kLFFo3g;PG) zVi^mvZd)?tr5NaX>(_D+veHm0=z<+TCgM-)U*o3&KpM8y2H!6lY?dmWa~DvD7z3?< z7gSVAQMm{p!kBm%wASeRTKu5mIeh>4!@hze-ST>^>|X;#uyAz;+U;Iev;6HasxP#iBfv~mM@&ReC-7a;TcJ7FPLW_ikxqG zx@0jJQ`yo!6Z2ygO}y4PI6o7(g2uuD z5c`z_G4mT{q7a1G?I}hwS_d$VSST!Qp07#*E%2sLp1ntp7VfXUXy&+Hrh1;yvY?x% zUO4>6g@}4p5aG^)qqk_YD^qw=T1iqYy%*XLgqT6eyhvTJ-lipWTZ}PsUyOc$xeL*5 zG5Bh6V11r3@BS{@47_5Qh9>h^A{PZC|UyvNRmL3n$WK zbSecB-lmACMTS zph>A&C2@^1GC-&lIy@{?ZKhgCNfuV-;DI(`+&7L4eIIZS`z6Ui`ELXn1iI!L(7__^ z9tUu1oC*Kqx) zj5L`95Lwa%RJd{YsrMUpqQR4XZ)sU=@c$MYO+c*l&aO#p2R~Q-c7SpXgmN);!*JDK z-iks=uML2id;plcf59`#UgV{mFY9y!O6Ro`Ek2->HwR2_MqC5uK+qB?yHK~%kp8Lc zAaK?4`7vC%*GR|1`I*H8AEvP!sM}o|3RIhlU!Qd0cbIQ^49*4g>8hhJfT8CkP6G_6I-t zn`k0K+-!C-xDx54@48DzXg0|T@g_?d?;sItU&dj;CywP)A_Ye`&8}3=JZ;d}O$J(P z*ZaM~OYpz(fBYHI8p*c?_HR! zkWtW86^J!NTRYx?_O7O4yeT0o42j>}2LjgaEhb6VJYdv!4&!-;56e$@)^o=MW%u$a z)h+7tK#g{vArT?|N_)VB5Ge~tW-U5c>bQ!#>%rwjf*%L&KY0EB{V^&6 zYvSD>|0^OuC=zORVmp&3@c?wxpfcQcu-p9KzxNvU6;%XR-oFgS*{(<)FbGAyrCKbd zKLhCzb0CW(=ua3h^ZlOy0vJq5?3=rNy|jNBMBO6CNWeXbt{kcH3@+y^j-t$G>e#IR ziO>KK2=p1DhA!X762wDpr;WABrIZSiu>2+pU?Skk1#~+2XTKyF{@+XQIKz&Ka&9H? zhl{)jl24~YEdY4<831OlSl1yn`47hOKX1E9ay8lNPJ_RjplQ~nx~>CqktKWy{1^+s zi#D$p|7q|4az_G$!oe7JiVDA;H)$>!ehTbOSa{_JVbXajRZ*?ak1NlFU?MM;jt{vx zJi2yesZz4Q@#;q0xgUF98AnFpU6B1=S=|-!;PRBR0IK94(D`)810c31dxvd+jy@_+Z#vP1^k>Oe|HVdtjUE|JjY#3ID~py% zbvYr}w>dQ8||eE#!*)|JOChI z8c#B5a(p$p&WHHl8&5T!tP@MOh3wq3_0H_wAK&8;*iQM4Rd)h9Y8A$W&{i!M zFy1ZzjD8xi_idTX>(0}vKtyD0J6Z;EL~6}7HrLa%W?dG6%eQDrO!lZ zd^`0T)`B)7*+1y>GavcnRINLwT}CZib%`*i*!)}Qz50eisK*`jzRN(5r3(PywUeQS zFfh&d3&I8F@YHl+lJ)!u55SzwyswU@I;wJ*X8`IMT%0skum&4ESNq9~ngMF^_dkK= zV3#ci9eB=b`l!Z*RAR%FJWi;E%rDcl&lg&#{~45y^kylKs+R6;{ zuyxLPLBH-AGNMU7D$7`Yej5mV%63FxffG&*a^4}T%#dZ(O=w*8677Af{-h8c#s8T? zSx9cCHQLs;!4uCyYnbEr?IS(C)@$Ry`!(qCbZdMM8g}=))*uYiFQUbRc;SKKxaM!& zhl1uQxIe<1dhP!+d*xFShiMBZdUw{#KXdU%3wDxy?xsrbiczmaOJ!|=@qZo=@=ch>J-{(@4jAbO+h-)s zA`ier&NA7j^@ftF8{9$)&!E-%9n{E$V}ZtK`1TkswZlC7vy{Qi`oNeL_7#+a-Qz@{ zfA;5b9lwPYB^PYbd?e3MOH~fAiI!PF%~?F)XC~RUiaRaOkEQ%4tuOUT2$nzWNf7id zfOYZ^41ExT`zxnO@8p3gEd$jlxMo&ASb^&aS_6nIfPH$7bp#P@&@rgJk>g?s6v2m} zIz&7a7918`C`7!(_=SECeFy)tlYX=#TFiWPBt2ra>yGXuDTL}`Da2Uu^ksWM$PoSD zb3HI@iZ&o_b|U*@$kl?-GT&=x=ue*UxD;JAxRv2J4(i?bwM{MiDP;@b zR!8035acuAUi=~2OMP}2NP?C%i)^J3|BLOpDBY?;NfPrx2|WSICZ zPD%iD<`p`5Hv@@9GmgNBm+O?Zx|3fwkl3;h90-$td)(~?665rNW@>!xSkw(k#*g2` z*eFr757g!jAhV#O20Z^9 zr*dx4`4P3{4;nX#FyMGeFLd>cQq^{*Ap{BDXMX_^>fn5n=a=QfRXR)0SPp2T@&ZzV zbyOd1GILDu^I~w!AgImyzQuoys~w93brxG6$aqir4GuvW0mK<=50gF4OF~jW4RV>j z4NfipS-HsCI#Ej%;<%i4^t)^<;b<*UXe?Q2cFvPm9%vJt)&rxVrsNnwPBs+uR;j=r z`DPAWqK?P5%U{Sd6_^>1;1k@`#9EIhx4xBmt|pHy z(yjX;xU=`{B^=QS^#qfffKyqh5O3;VkB0ll)Z*+=0;QuAq>;JrqlO<|cK>pX#tI3i zhdLhfj59;I{=eehI;yHKY8O_d8>EqLkP-x?Lt45Um6DR~@JI<#(%p@MbO|UZC?QBU z0@96ierux`zxTcWedAsS!@)T0v-e(e&9&ByXFfAH3AY=auxjgOco|qnf!0Kn%1}!d zV~`M$6JVyY0`LMt@q``9bdTO=4`v*KnkmOXkr#eDFbqNnLwwDZlt-ZUPc_wysmyP` zaQEH+jLE;MG%88FUmPv;C|ao-D8BY+D=B(Ig{VIj7YG;AA(6WpHQAH1#k=Oae@+jw zcE(aENcPyl-N%j*()5@$0O>6|1h1aYg7zrgjTJ^8$Q*k_bL3a^RzcyL+GxO5I}ai= z&VH`HCi=tuNIE@VM)GAM^o)aqq3tS%QwXc%_)^LAJpv*67^tUnrH=XzD0Xhfj6kKO zn2%>)SCqFWy+Tx{nF?pl>z#*TZ+djF+sq7}h**3rrH(*@E!$l^i(ZwqE1Vd?Q zvj@|!!}JYVwFi1Z7}cP*Rw`Jz^8tQ2L&&uJ=^nRNL4#$+uEU_?318L1Sahf79x}3H z`7e~}$kAg$k^R}W!Vv~a5|^6&B3>J%I@_+Pw|Eni{nZGUwfma;DGb( zsz=D?)XHso%uE+-WFU@mC?!qjo2i0EwD_13VLp4Q^+LSPGiRSD zmv&bAMfVHTNt&IH9RZy97_}@;yr3>(oVb^60JJNF8*}ZK!`)&KSBA@Kz2-2qT(NPe+67fSA6i;x(sQR2D1UUlaXVz%A zs54>#GDsG=R4$J0^gQq1SvP`rhl}tY)7i?8t9y}w1+B~*%l1j9!X>%i|fepA* z8ldP>T@3l0^p^2+h#u~}gG4-)FAO=rfIAYI|Ixa_&byaqq|U8vXi5pf=W{ECrs z7coQ7^-(I=-8U1QJpn2yr&f{(Bb$5Spkmq{EEGS-4LfobJ*i+fb!7nuBaYZY|5G=~ z_Yq;3ruN=XsRcmXYK*aQ&lZRep&pcIsVt)#GR)74v(2QaXF-uIV`0xmjM^sX zcv9!qj5#|#r!wihEagJrX*z8@?`92bya%KXtU7(Bk~5OlvV^oyn`Z~$lwgGdU?|aQ1?>$b59tT(WF$llwIf5 z6LWXu)#{bF&U;U+&aT(k!(68f9F4R>hPi66gF9bUs^ufM(JY6ymw+TUw&7&c}0#$R-qbvw<0sTmrD6iJyiW2k#{ z;8MS*9ar9cu!FOdeU~AGCv*Qw;>3-cWFQ6xD3&NWhP!pezQ*cU6nrB#&n0%XMAoIz zhEGkuqnZDfEw$@r#0$;9``$6Q3egl%b6n(fZ|@mLI?OWOyb{98959o!@~-G3T$8K1 z>qmZq1^kcpt}(;0dvDVnj=0F}%3DkNP|(W)Q~)7I+@~@vEXol``FQFg#RUF7scN($sj6qLwRR{|yF-|40@ldTt9zhm03~gEDFYs+2{r&$C0Hn|Sh-QC2_G!> zH0Nrdfq1*4%Zo>C5(OuA+x1(};&b_FtC2q-tk zUwtN>@}BM;?OHs&o9*qdBN71Ns6}FHuCck7W`*7TZDcP5#7DTrpznopC~E`9wjKU^ zMq#IfE+n|Wf700ifPlG-^zrwm^<5>9rTJ!g%9wLTX?QI~NEptT#>RU(o;7IEm=6I>mR*{TQ%VV9_=;lK{i+Cj-(>=VqxHUr<}|GL(i`BRw4mVFT?o_- z#fQuBHJET1@4%q+L2*E+r|_~yE;L+ydFbdVVk)$QUJQ7uu>gC@_9P>b;g>MFa6VaM z;!iS(fnJY)O~S?1tWZq`Q-6uB4RPg94gakbl(?U|#PJm&KF7a;T#zeH07OzZCtFDN zpWD@52e^KI`xK@B=$0;k*)GSj6gXG-y-Rx12mtK#I4}dsZ`Jiv8x1`wmCEoW4CxzE zvz?B9$UDp+blKoOH?STI4`DF@>>d)`RVdrC;U5dtXx{<36%w;tntSP*DBok}N;I$j zoE*S@L+<)+HI(;AXLD?7&_-s*S)MDD_(i71b;+}(SJba^Id}WLHds!F4nO+5Sr=Z1 z3L`)VH0lU+5dnPnWXWtbAVwVO=K*M;Bxrg*{opBD zhrASndhT#n@&s-2q*A?xG=wk29m^*(h5#u(>9aLnxzpFkgZmsS$Q>)1r!c)xf|3P> z!x;a0yHJTSK(f^Qi6HRp6uy(!%$;?2Rp%gqDg}uEl$~sw8m>W$UOieJK%#{JQnXLec*S-wm4O7y&UieU%iceZ~CUYC3-zTJoFB$FNT=PmFOA+G2QCh5EX3=urQ{2BNQKFxXGeJ=Uy%*)vS{W~K) z?Bx%Y6@aw=yrRYz{^yoINBCU~3goiA{(8Y9{d;wQUnr%3`0V{&tNuqCj)XvkC!7Qa zNdA2kz+rt#Z^x-p_<2_RUWz^XAQI;^w5w)`D1Esb{XMzbKo;aL}GD zc?P(9x=@+TqmTib>sR97?@P;|;)SD$4R14m2KFIHQImv<_rO+Tp~Gd}v$^u&&$D{$ zATymWECdg<+5J5?&@V7P=r2UOhM+kZ@EQKQ8B+J$g_ryyr`{#^`9;N-|4d>8Aa)f{ zbGt45U;V>|y@iJMKfiIo66iERr|$drK)~43B+uYk$vmWcz;fT5)4w1tTQZuO%VkA3K{rd@&A1qpi}|-9I%i- zWk$6F@>`sEumImOaE?z{Bzo1VYGnom4-bj`ako? z2nSSCq6W1$gX1zv@LLVRFA*RVtbJC-7-g0KLV5Y?OQ7X1_oIv?#@LlBeA+ViMAe`OG(g`5cXLtgN z#!=A3(Nu3%ClXMhHY3yES$sc-aR0z<8+G1N4hP9k8jeIzM$$sK^vxHm3dBwMi3C7P z$4pkhto>e9f~xsF1kv8C2JJvcq1L`cH|FdCEMbffka2?XXIKM#$Af*qDs81}Rq77~ zyr$zp@@h;0Q2YTNX6tsWlH)_rShGksR~^B;O+}7%pl}&cuRDb6i#d>!b-vm6m%1y) zr+{DE66`|&4}t${Ex&t;(z$Va^;^@S+vg|VKS*_5r~O}lZImc)<9u#dl9j4CR(~qc zwL=DEZlw!~bgW-W5EQ&%05(uD{awO%;I><)zO-#JM^?5?wrefWfnrM4AXkt+bq2JK znn=y+1_(zVuf3pf9)5t-VpRVac)AcWYEkS_02b_P@j3g=ZXbpv_y|NSofuyir>xvZo92A&ydksXzs{lBMMNg}F@{ZKb8?Vda79wS>^JUE@lCVR7 zQ6UWBMQ;bJ@Iw$hmxIJN6)K>@={*1>g#opOuB+_`8yIUJSpqA&Thb@x13IF4ux0N8 zM$!=wu@*TTp8f@4n*UZO95Vh#48pmQCWNoIt-73>?uxeK!4xe9Le13or(C#W<0z* zP$kLfe46I$z4JBDUmojJ@~EijRatNs2>b7W%GW?4+Ht6YlN3h_`9Xr~czl?s2gsTo zCLY;MLcj*|V-V4%IC@iR_#$=V@T3-0(=hMEl+v`5;q4X+Gj#X5SgCEuhPHQ!`x6HA z)Y?P=B2@+=jwNTD^)?I*MGje@R#L|x97)`GZ~T?(N9ijtPX7?FuHJu}c_A``hK>-z z8WIdLZk5Iyf}peK^gNkKv@&}45Qw{5fDnG*$DVhfg0u@ToL4F#ZV*#B1m3WJPUiOg z<+On7u=Mr?%b-2z^sIy&h|V8)_K4fR*4VSQCJ4prRa`cx&P_!Nz~H;S{QQAGRltMi z{e2#b1yI!v*Ssw!)&&VXM5gsX({pipdJKf<*T5!R0f?*w?=#QAw!pgmz`XOjJ3|j0 ze7wEr3^HO2aX&q03y3=tBIZA)ZoNWw3d+G+;vb6oyigEIIbcNJvA^ltu*0$_-lG`g z^+K!orjutc*c;8EoI!kZ z^8#b#7nx6_UTQ~1JMx^l*yu^hQe*?t-|<74GV|hZiJ!-a9#;+-@b5^?y0sfF4Ml2& zMASjeFT+3<)~NrX`7KHXIOHSGPOK~+sWdKDFkL1qyJSp#`I9$PmB%ftb;fn>AFR;Y zwXoS?pQ<$i6>8>b`g&xo=(CAN-hL1oc|gc>a)@!g3QFHDWd@OkeFmT#rO`cr{qeUK z+I|TE6NAD9XoY-Oh{04{Z%V*bv{GhXp}7fo>Ez>TqzWs#LX$sWg!h^?$z0zTi4i)M zmYQi5w}%Q~s6h3{h1;}VLOsvEnNK0&ke0R*U z-G+{Nhw#}}_?tCQ0VgwV#Im-}?l!zJ`KFhr0SBf)-xCC?bmmznfRm4yaK-EPxyPta z+fuzNQhheI*OiTa1|e6yQOR>Z`^hcLF^D7m15)Aj?ur{i^)B?Q`;1AwoIK!wz6XRA;V@=d>)2mCh0N81iFzeO^9z+zX?Xm~K+sw=F?9wqtm)BC1qsv_#iyJvR- za47V)e|tHwb*iWPfsD~Ry<9trpW0Uu#gJuMeu%c3l--`iX{LF0@Xg-!#0&ChyF(rJE#6@z^nO2Z%AQZh5{!S4 z)@Ufmi}Ws4XeK?3JT#YiETlc2_dBKl#PV$ {^9Agy4$~ztcxQytxxHf}1=f3hk zFyIfsZGP+I*NA}IGTe-G9=C4z(fUFt?JfWummnoN0^kW`A!viQnB#U17xMt&*2(az zG*Expl<1)?MnwK25sKd1l81HMD-3T*Q4shBRc!msOJq=dK4HJ5iKQ0btHO5QO!N~= z1fVNh;a4$a*^LcE&Af&kg7B7Ax$H|?@LY!>r}J~-P!#G-Q%-j!AOp&Rk(c)2xHL%P zex3D&^yBdYdIInzIxlmpRoQukw+dotUrf4g6XDB{zFw;PhK=!&Av&{c6(t1gJ*IkvG5SN8A=7?}fNaH32V8Ck|cN9?T{GkI< z<;NJNZFR2lX+^1DmS#!Z6eBt38v&MLFJK|WJcHMf{`B<0&k=~;ODj%>jFmS^5*BwmEW02l zONYFa*vO~ILN@j&puHdVp{v`t+uqAgvAo?$O;UP6-nK6@NA2Wus1t-nj3Be0t!e!#nhhq}1Sx?fQPOr3rHWjB)8y5qt_I9#2X-STXU zaFc?sez-wDZ&~_@15tg{wFbBtVtX+)wj8|tVgyMQ5nfnX;ddUE!QI+^12%Ji>n+iz zIQQs@8sn9TCXC>p4u6*w%csr~&o(=6p*X{yPlbGp2X>Fr;)!j(-*n#q=Uj&SEp&Tt za#;vN)50_Q+(xK?J9~+?1uKhP!1yuYQ1t5aKby!Gn#~cV5+tt(Mq@^LZgg=fRIh7q z8CJ8(71NMfU25kEnt0iSBO2@x_6!GSuhj9 z<^eAUfH%kP%) z-jC0=Ij2j!sh(6@PN}^~xQcHhUI7Y+X|2GtgRYotuFf5n)Zs^V@0YI3`wQ&JCcLGuOE z1V3x#r1KYsmq1y8-jT%0|Nu<)k3%5pY6V`1N)}$2FbFCwqW()jF{}o zxA(rf1LEZ?R#XRJvpPm6WAFDlcN1Tc2Tzo*z2B$WRamtlv;;kcbbZ(KJ9 z-1k*CKCK6CX$+~m7S*rhs81=E#fS;kPG@{wlQR*J-EDVoS@6^qO z(B>>`UR{Uz8MV;C)N6ZpJI|qKUuIv=yC*5NQBlIfe{g@$Q{gP0@5!!4{Og*98Q)2q zU9$U89z3iE-m%uYulwu<`YNlOgo~@|B%f(`HK{B1>*Q^`DfVX?R%kG@)PMBoQ4#4F z%dkn~N=fLpmrFkn*?_KdF;zP2wNG6WtU2MY)t4(;)ywi{6mA|QtC%-t=E_^=1o-4Mc-7b%(|$vZZK2g|vd8F&MZlNaU!r=;Qpr@V+fA{3^^3*^ zR}Jk4c5YzWASZm|_pMSzlQCs5tYwpI>@G0)y7AZ|C}>c5SGh>b@y)c@N^+6H$l%Ix z)Lzks(#ZPgQYWXAnTpg@Nxy1k;}0`_g{r!kGPbGRp+>aB_53y2*SrlUGNtz=CZgkt zYD$xyrJmWZb-Ub7IBXm=TREh6sJAM?&#YTMzfydgb&P6k5MSkqpyHf(=EfE%*q*w% zNFXsJ8&{h+giQJL3LG+}k_p_=%zFLE7iWtjgV~no$)UZ32eKf2*D<8?i*QA8kt~?KS%U2e<*^u@jbXSL zojktf8=?%VD{~COhccV!(M{Fr`&la3knn!}M_CJ3icly!wH!At*JHr7`!G**$Z^bw zs#>|qCz9^+D(L!rzbg$V>SBUzoCBmG>$^^5+VOR8c|QkmF^Z=mWeCV>DD^@K`$P1v z+|`hNKLmW88VY_&0u%jdQN1YXw`dYrfA0MGlStSy4y7}tQ6Cs)P zvh=VP7gkqQwK_3|{ka;t_XlbbH1(UEH{kysbfGeM%TUc22eH3j6BGlgm2x;>|07(< zt8nVK4b;D&{QdohIFw2sn>$5&sliCNO-T^`{1GMw?&8l_D&K`{{)#&^K~nV3Bhs9 zagxOUeu{__3U+?sin~3}r4B&Tft_^)#w=-t>Hm9ffV2nbMA8e3S3ZwKGjcdBtM$b{ z#{28Bs6cA?nIHbS%~KKZCU7y{hiSY|3}Ttzpu)cYZ*l|Is6Tp?b^G>hd=-W&(e#IK z7#J9>@sIIm7v5dImB=#uLo~I5sEp>zPcB|fg#6+GyA*frjZ>F5d46j$^w$V-n2nu! z%!0~HWLhbaM>&r>xl#YR%Kt(a^kn7a zL;ohx?8>*gcH(uUlL!H>DQ1^59|rMt0Rq_xNs==(4G``F6xV@p8OuE5T8Xu z?Jb+yp9M`RZnU?$hmUfwK5Rf7i~WhMhxK@QyCmh7iR__^&IfXj-DP}Fi=gA(7Iwl? zQlURKBR0ich9{~j}X)^m_JnLmSLB3M)?f^08;gIe7wr9 zVeMEZ&_rm#5;4@EnwN*{sQBWS?_;6zZ91?E0B{9%v%+qz^ySeL-&t}*>DNnNo@#2^ zUHsZI2|6h{>N8oqxV*Jm`8!R+gcK(Y*MQ}TC(V8qLb_jG2g6&iDHR;=u~v`w&VNbK zm56((`g;|qu-SgOYx<-CKlMWpLm~!Q;`a5-qe^Ne2be4&Xy@iq0Wt;U8pWIoBPg%} z)x1P8-pkGLAzsZKLHSBC8rB`#^bvKPEacT<;1P`po{vk@mqi5(?k1(R($=}zKh&+S`2-~jz=k}Q?IJ7U=lC$|FL z>N>!Yu*g5kRlYn*5n;MdZu3vWmzOm0xV(5i0$glBa)^#JkBuCy2-d^E@UXeOy!>2| z8;S=B=R?2OdbPaKKJu2S&g~<6g`Wxx4_Bm_B%f@~S~E$=%RNuH4e9b0SK)*4YBq$*^<8d* zwq>?RAPNmiDIdtGzpVfM2w6=B+9eiLmRD9-yc(_Bgh5-5wa*M`=r*U8x}V{GYq`iF z0s$L0Xg6C<6V9!Q7So>Uktlfnc7!U>ujJ4WUWFOFiELtbzH*k;NSO&l`a}ELqW`zq zkdJwFzq3x?tFrYKXZ=d?X7Zq&p)d%_kCzjgu5+TJ!aXL;9}&6jx@n~P#0OxRpNy7J z?Mhv0X(WT~{)DMbJ_)Jtq}A5U8*ly%tG%5*(>Oz$!3K}*<@=%7jAH4gfZ#pMvFi(} zTG2os`lSkCi=;ckkSMrGc{<1M(3+N;1s<;qbgcHubs2b%jEnE)Q1}W`Mo_{MG1b~? zJ=_+$)4HvFP>y_GIjVDOB(~nqx-?=K;CE&#bh7cX!|pV%g2FqjmA%F&gz8ppYOOJ&-6+eEi_86is$zM$mr~lmTqZgBkF6Z^ z?<4^Z%LIetOv>GH2h;tjC+je$rr5i1>MTE1qpPbc$XDR=ZSfL&8pkP6%n(#WiXUwF z?@?fk{7MUFaxl;4i_7r))xsjs99Ur7ykWQV?1wP4IiRbw1e$rq$04hgK+C8*r=p^Q zR>A%eZ+dd;r#geKf#^iS9<9em9_hb?^^O;AS$d|Et7U9&XX+E^LHw%nwr9Daa5)+q zM-ufYGnmowqH^aynW-?hz>cM{eqf*!0>wJ-JMfo_N7@mp-{IKK>A7|Hd6Z`aB{ zN3}$7TMmvg2tV{2sj)Lb(ZM(GuX)4;?POw7zXM?yBmL` zY}9-)H$ZhilXLefptxKKN@ednHWEmqxzJU*1^T>M(xysNL-}@kHV*;4)h%5guX9WX zcfVlhrl!3LFkN?%-#e(by|MwlYzp@Hxt*(H&tAqmye6qXi71_oHmqI&PvW z7w75n&&S&Ws*;~yz6*|3oOr06TtK~zP`3+>n_?0V>o(@)EuM7edxO_$sOA@Ysi##1I_^Py*Pd61Sq|^a}YPJ!oq%AV^rNn z52H1+Foxxe8#Or6)v41B4*OGTa{-Afz-g^A%m=(EJwL$W!f5mueqib@FM0(Q30{lX zbD2lKTQ!tIcoH=5kkBHXdRqIj#ufWCN#REkv$m>k?>|G6F=rryAqMvm2 zS|{e+4_5v8wqIU**1mQ{g582FpylzFu)0u-3dvywS17l?3R1Igg03^vnQ?{NrHu8| z%Qv%Yj}+!Lb<>6?XjEM{k-FH_KjVWc$2a>8J|{GuXEHwye>e=efJT3NBWObshBgK;NamXL&vB%Djf7Dl(2M z7eI|wDaaxo+bC^P)KGYh3v2zhrql6YC5^X0hqJvaIoJbVe<<>4Gst67B0wSbNMO2-7y zt(%_j`tbKqI3Y@TK)_y*Sv5b;tUf$0y-0eygD7H6s(mpileM(6Lg&jTnse`jTEMJ<0uN-N zyN`Qm;nc~+CnYZqmh?h+0(Ksk60AfCi0wzGl{B~ks0WqH*fxg;$a9UlKdT4ikG?RYA(Q)oaI_;o{YDrb z$}0mhp{eXx`ia7Yh2>@eLuCe>`c+&+1_VZVR8>@UFg5wiV2Q)xet;}@CA9g0)Z{op z?#@?<7X-CeKU27f3)Ym90|8#x%Ml7N1hi(|J;$+Boh0s|osWcv-Far+95*q@Ean+7 za>ufy&brIM>|`4NU-ccGCqSdD0iiloi~2k++Ijjo+tHy(1kfB;!N)~DNZ z9T0p5ONVDvqbx%9YG_QZ+5AtjFkx#|b*`>ecTIZst-h{qmg&&Z{PfzLPa<@oq6Krl zZb#o?)Z0TRMP-vg1RqEG8FYZ9@R{E%k3Oi8%&wOPo9I8=5&QWGocz zoYmOP_PPf56JI!6uwwikR#%J1D&~n?W?hF!6pn1%1rjn>WnRlIV~aU7y;7ioq%0Bw z1(vucmg71M1y2h+cEH|<=zF@XAzZ0&xZNgv_EKbyGycL1?~i@jYLzHFwL0O_@-4eI z4KG{hDBppcWYxKYwX153EqUjD6qcgy6YE^N)Qelg!5<9eM-gTyrU^mJq$Ba1kF3DP zQ@k)RWLAmTez%i&zJ3ts%_dLwey>eSVRp74cdf}&uZO)$t6*Emb zKU;ru_SgnqBKPBwK>7yfhEeCb#RPhP8u>AsvBpkt@1D6yG%6<~IT+j)NgSh?eH%bTM5&slDoB;h5--eg+ON4;Oi zs`y$a(bH)A2kcsx6HIr!@emI@ObG!5`OVg)E-xx)h(8YUf8;Z(oYcw%3Lw-5#U z!5p(0sU>rVskSD>a=XrB=VExr1mkML&mbi=oS~P=dhBys6^>r}s#=}2U{*o*_<^CM z6;>v6*~DX+_c9&D<#ZR5EV2a@j`?(1*!#rQ(%hV}vC0z{#zp{aOcJh8$R);OO;6te z+`~Y+22u+;XRGCzSU8?NfpiGDtX0UqQL$c6RkTPr(kmWqK87Y4_7>7$;9Meh=V0e= z8(infPZw!8rf&YBivRY?4H6hF+F}*oMDJ^vt4u?lUo(5VGoKi8?Tu242Kpxm$l>YyX(^)1x(1H! zUHHM${$S$?4m?ra+83b%wk^yNEr0CEgzWxS+gcw(nx80>u2Jur+Z>}ircA?xb(QuV zP3c}sd2elpR88hUo?~cJz=hmB*?w(R;)@~eAm`&Vkwl!7J&QCP)GNWy)hJ=>t}Upy z74*%H$WC3qP-S6F^W~9pfi`3Ii$kOUoMAcZk;)Ii9VkD-)mg9iDYnR5Wpo%R`Qgye zWN&h)P~)+`O2tG1b{v3u+a`+>ya?>6y#}kvf}NEi07&1l^y#xOJ|m?frTOpIUZ$}S zJw7x%XMU2yP1^!m*zayoAktMbO5`oNBhWeRiTsQMq=!H{_ppowdU7S4TAzC zz60wp3z)yBB5}yr`E`CVDY@$DIXNGpP%JDMIN>%ECOxmS$`HloqPd*pJlV~PY ztDVKKmTFH4k!fW_e9}S^VzXIPV0Ab3Nf4?0spU~665ccFN^g%bVmq0zykz)VE@^S9 z9EkqaDEg4^C^dGs5iu_3(tOf}!kZ=qto+a%*rRxE`6o1oh(*k!-VwrfGBqgX46z^J zsx#lbtZ797A{^4W)32=SGyK7vC zu+#bJmA0vPVD1rjz9%M(y$QWVksJ9?1TRG@j;nfiV`(I-=7kMR^^Nl4){n(9!??Se zw!O|Nu8rooK2=Y~tY?Mf!UBIqm&hOs)Ulod0rm_5K*CnBd_bgzr<_W>)3lTEAj@i4 z0&W?>X5<#v1nvTZKdHQXigl{BzT=pzA7yu6d7gHf&;#st-fr}JLfpHH)BCMAe#I0m z)WG4Wp4FJ_YzE)oYrBy0Ai{lbWKWUJaI&#e#rvuo)4?J^&7uHFKxPN(ov%Z+u}w$J z=cV=TXo9@`^)AC@S=iYGsfz**wOQ%kZGO+&6}TK!?0i}L?>}M@_Pg<<{I%l=9bb2T4YWrJ_5d)WvN%Ci&MB5Gs)J?B8uVdUB zX^yN?&j*gpFF~Myz&G7@u@BmzfK~DXrr=J`9jw~DQi7Wtpbtj#V7|+w_VMXrVlk## z-P--`?ZrEZQh8P@CsnBI?uHJ+CIQKvLo6Eix^KF52%Rw2n0Hcjm9t9uMof0^jh6HY zG5xYA!KMw~zi_8{p6@jT!u38sp3;2Dc_k$ERJnQxRUlC+?vzh>e`BdK*J}DxAbr=s%8g6g4;)KIR~rTK?+o$rnH&S+Hw zdUF6hRtdlcUjPT#_2a?nMw6BR475S1^94YlP@xH0MQGH-hT+gi4ujfvinpf#&^LT? z_ygQE)fw=7)gVT(-)B`W8fP_4s)6xA$adBE#5;DG^^Xd_fEMFT+D+c4Rsbh#Gj<99 z#$>A=wP;lQt8c`WG_6-UZVru2$`T5JW|BolQy_d6xCVf{XKTqnFeW{^5rt0PSOc$x z0R3U2L&uFL`L`{Jy9QP8ZSnABw+)Js*z#uD%Y zK;kM0QIr933L;Aaj+?R-T_bJm{0gjzK^F)-wyTuByavL7PWo&s<;%?mC0>UBl1vI; zYd719a15FycC3c+ufG9s_)im8_j4dP-rlPf#u)%!aSrF3uttqQ4UE&X51RU@^8-Ww zU~GNV@^$x0Ny)CkO_t>H0$of0h{f#Gk<60Hldihzl&*dx@;7PAZ>d*?hK0r!ysl(s z`D(GTBK-~oTNr_^6ZCnnyk6rbff}DkPUrf|Jm(j<0+91Qjln&Zh zm%7+EFmq336mj=J{_g+snh%5LCa+U!G z9!+cjd&v4As(3aaP@g5dIiwvpM}&U)30!7~CA-2FcijbX4e$yff{?2HfyBy}E9-94 z&Hgr8d3$y#{gJx>zMNQBwNE7{x(p(o6)WRLnZl<_ifbT#fD9KC2BHDHHRNUbnj>!Q zktxn)L*83xEBiZrgBt~TRy0Fp8k&pN-e0|n5P-{GEeLOd#$7uov9J1Wo7%;4ngvy5 zJ^A-S1>3^+>tN?)GG0WhM?83)o9QqLqItdqH+&2%P7mW!kAv_KaQzD*bm$kseZ7^j z>L+9M770h~PBO34XI|K#jDy-3-?0O{{Qm8Mo}O=0nDZ9XM|IDi;x>dT@~aF_0I2I& zbM)ZS4FD(R^5~q%b(v2vL(EWK!4Dpg%{9`H*5)*N%4+*shl1Q8#CrN!dljlDo6=}j zctH0({Icu2I=tSVRJyTmLqmwUUK>j~U+!ec{UN~gDI}2?ko^Hbn+Vmz5@{0c&c$AO zE+_Niv8K*`;SvBiP4&4>OXz~Va$CS{QQ19 z`0Bnno*6R8Z2IZ<1S8{}uhm2i43wN%OyuXzwD|L-e$v#aSL*5SHwK~h7pL>QO(@Ue z&qMaI783hm_o^~`vbu|AgY2Wt(|gNoeyExeCcM@^+q>^np}d!A7N*w`W9iNXf?jVv zE;pTz>6Pp4i5FB0{SNp6X`n#!x1=F_jv)8|Q9?5>4h^A>RiT@2DYn7ft$(67*DS{K ztF;rT4GGr=nFLW=6$Y%vMK{LXS9z!RUg*37v`3Tav#95u{cyXkgiLjG-4wdjGX?ui zo^)Jyn7$XxqU;J0u`b^iuinGBafhR>?xe%Q&7RD8cy}nNE6I$_Em1z+X#GLi-Xom+ zQC6)#d2DHDf|xnY^yY)1dH~T4*E)Sr-_I1C?|`yf1=T$uN~xj^LSEEzve-0)St`rZ zm%>2RXI?gK@Hs!*tDidkLa&%=1=-`iNMkNav$D38v*3<=PD5XZM#+xwG{j9GM$h z;HV-Ix+{vx!xSZsH+nNu4|iV0^JUn=zUXPLId6TOXa1TKXtHp5q=A)!NAqT#U5{zr z_b5b1U#+kp3&lSa#Ly8J)IXi)i=$2y=8>QDqos+3=y)pJnfCie%;101iM$llrXazj zs)q~aWE=LF@F3>;_gV@hs_)O znQ(V8Z+7X2Vy(fhu9QT^ZO}ylO>w2-|)TrKgZE&}V?3Dy2Cj?;- zjpB#0EkJlSw)`{TG_!J;??jm#a?l-(YnuB?(UevDMYkH0xyPxnMuF9;>9m>rjBi(z zM>4$*zHn=5Y`IP>`6A~l@*1_jmLP2Kp`wA{eu0{5iMhoCl@|(T+Kyu>I$r&7IXx;? z%IV4no3C{fj+HxJBA%%ewd0%PWqPuE7REh!zkRG?)NeNM+Pz{idG}@G8mmIrxKzpvG25lLUxG2BbN1qHV_`z~XC(7o`00GRIuSzAc{o3??z><^q7RArF5 zzuelOx1@Xl?YPCoVTNbrdRH{)*)Ite+>g$x(r(NpOO19E#C!^%g zSR4`tiQ0X#N9X9QcykPG6(_l?IKQt;--er*DQpJ5$_$^HOdBys&RPS`A-XkREHlEz zk#e&=UajBxV&HMfA^{z0{PYX1OpdO_eJ06=OKepzYP{s#br4%P!v z#9jE`lLx-9mIpMSuY|?y9{><1(BOw6Ag}({^OaKICS(O)gV5iPYYMS->_Xwt<0JG% z&IiM8+`V<2bP47Ho_jQ2{z16M$6`_>$txgP*wa*flBD~HFs1KezW{L;{UVlve5RPc z;bkC}LjR5n;M*ww*yND*_tF4e>aNVp{rhDiIDoO<%!WMh+cE`f4YmVbyVv1$KGWYL zq`eP3syAq2PXDY%K_H7AJr;}LKdKB_1Hi#seWvusD1p>I90*u&C^+#SwQmM$|Nr(? b4}7%BKOiZi7=&H{|70W-?iGs}J^TLvsqpK+ diff --git a/docs/sqlflow/doc/contribute/mysql-setup.md b/docs/sqlflow/doc/contribute/mysql-setup.md deleted file mode 100644 index 11bc56a..0000000 --- a/docs/sqlflow/doc/contribute/mysql-setup.md +++ /dev/null @@ -1,143 +0,0 @@ ---- -parent: Contribute -nav_order: 2 ---- - -# Run MySQL Server and Client in Docker Containers - -The document explains how to setup MySQL in our development environment. - -## Run MySQL Server in a Docker Container - -```bash -docker run --rm \ - -v /tmp/test1:/var/lib/mysql \ - --name mysql01 \ - -e MYSQL_ROOT_PASSWORD=root \ - -e MYSQL_ROOT_HOST='%' \ - -p 3306:3306 \ - -d \ - mysql/mysql-server:8.0 -``` - -- the `-v` option ensures that the database is saved on the host. The default directory where MySQL saves the database is `/var/lib/mysql`. This directory can be configured in `/etc/mysql/my.cnf`, as explained in [this post](https://www.mkyong.com/mysql/where-does-mysql-stored-the-data-in-my-harddisk/). By overlaying the directory `/tmp/test1` on the host to `/var/lib/mysql`, we "cheat" MySQL to save databases on the host. So, we can kill the container and restart it, and the database is still there. - - Please be aware that the directory on the host must be empty the first time we run the above command; otherwise, MySQL would fail to initialize. I figured out this problem after several failures using `docker logs`. - -- the `-e` option sets the root password of MySQL to "root". Feel free to set it to any password you like. - -- the second `-e` options sets `MYSQL_ROOT_HOST` to a [wildcard](https://github.com/docker-library/mysql/issues/241#issuecomment-263011059) so to allow clients connecting to the server via TCP/IP as the user "root". This trick works with MySQL 5.7 and 8.0, but not the most recent under-development version. - -- the `--name` option names the container to `mysql01`, which can be used to refer to this container. - -- the `-p` option maps the port 3306, on which the MySQL server listens, to the same port on the host, so that clients could connect to the server via TCP/IP. - -## Run MySQL Client in the Server Container - -```bash -docker exec -it mysql01 mysql -uroot -p -``` - -This command executes the command `mysql`, which is the command line tool of MySQL, in the container named `mysql01`. - -- The command line flags of `mysql` include `-u`, which specifies the username of MySQL, and `-p`, which makes MySQL prompts for the password. For this example, we should type the password "root", which was set in the previous command. - -- Please wait for a few seconds after the starting of the MySQL server container before we execute the client; otherwise, the startup of the client might fail. - -- Once we get into the MySQL client, we can type SQL commands, e.g., - - ```sql - show databases; - create database yi; - ``` - -## Run Client in a Different Container on the Same Host - -```bash -docker run --rm -it \ - -v /tmp/test1:/var/lib/mysql \ - mysql/mysql-server:8.0 \ - mysql -uroot -p -``` - -- The `-v` option maps the database directory on the host to the client container. This mapping is necessary because, by default, the client talks to the server via Unix socket `/var/lib/mysql/mysql.sock`, which is `/tmp/test1/mysql.sock` on the host. - -## Run Client in a Container on a Remote Host - -```bash -docker run \ - --rm -it \ - mysql/mysql-server:8.0 \ - mysql -h 192.168.1.3 -P 3306 -uroot -p -``` - -- the `-h` option tells the client where the server is running on. In this example, the given IP is the one of the host where I ran the MySQL server container. - -Please be aware that the above command works only if the server allows remote connections. - -## Run Python Client in a Container - -To write a Python client, we need to install the Python package `mysql-connector-python`. - -```Dockerfile -FROM python:2.7 -RUN pip install mysql-connector-python -``` - -Please be aware that some [documents](https://www.w3schools.com/python/python_mysql_getstarted.asp) says that we need to install `mysql-connector`. I tried; but the `mysql.connector.connect` call failed with the error `mysql.connector.errors.NotSupportedError: Authentication plugin 'caching_sha2_password' is not supported`. - -Build the Docker image: - -```bash -docker build -t sqlflow . -``` - -Run the image: - -```bash -docker run --rm -it sqlflow bash -``` - -and we can start Python and run the following Python code snippet - -``` ->>> import mysql.connector ->>> db = mysql.connector.connect(user='root', passwd='root', host='192.168.1.3') ->>> print(db) - -``` - -## Run a Go Client - -In order to connect to a database, you need to import the database's driver first. - -``` -export GOPATH=$HOME/go -go get -u github.com/go-sql-driver/mysql -``` - -`go run` the following file - -```go -package main - -import ( - "database/sql" - "github.com/go-sql-driver/mysql" - "log" -) - -func main() { - testConfig := &mysql.Config{ - User: "root", - Passwd: "root", - Net: "tcp", - Addr: "localhost:3306", - } - db, e := sql.Open("mysql", testConfig.FormatDSN()) - if e != nil { - log.Fatal(e) - } - defer db.Close() -} -``` diff --git a/docs/sqlflow/doc/contribute/nav.md b/docs/sqlflow/doc/contribute/nav.md deleted file mode 100644 index e9a60b8..0000000 --- a/docs/sqlflow/doc/contribute/nav.md +++ /dev/null @@ -1,7 +0,0 @@ ---- -title: Contribute -has_children: true -permalink: /doc/contribute -nav_order: 3 ---- -Dev, build and run! diff --git a/docs/sqlflow/doc/contribute/submitter.md b/docs/sqlflow/doc/contribute/submitter.md deleted file mode 100644 index a312597..0000000 --- a/docs/sqlflow/doc/contribute/submitter.md +++ /dev/null @@ -1,98 +0,0 @@ ---- -parent: Contribute -nav_order: 4 ---- - -# Submitter - -A submitter is a pluggable module in SQLFlow that is used to submit an ML job to a third party computation service. - -## Workflow - -When a user types in an extended SQL statement, SQLFlow first parses and semantically verifies the statement. Then SQLFlow either runs the ML job locally or submits the ML job to a third party computation service. - -![](doc/figures/sqlflow-arch2.png) - -In the latter case, SQLFlow produces a job description (`TrainDescription` or `PredictDescription`) and hands it over to the submitter. For a training SQL, SQLFlow produces `TrainDescription`; for prediction SQL, SQLFlow produces `PredDescription`. The concrete definition of the description looks like the following - -```go -type ColumnType struct { - Name string // e.g. sepal_length - DatabaseTypeName string // e.g. FLOAT -} - -// SELECT * -// FROM iris.train -// TRAIN DNNClassifier -// WITH -// n_classes = 3, -// hidden_units = [10, 20] -// COLUMN sepal_length, sepal_width, petal_length, petal_width -// LABEL class -// INTO sqlflow_models.my_dnn_model; -{% raw %} -type TrainDescription struct { - StandardSelect string // e.g. SELECT * FROM iris.train - Estimator string // e.g. DNNClassifier - Attrs map[string]string // e.g. {{"n_classes", "3"}, {"hidden_units", "[10, 20]"}} - X []ColumnType // e.g. {{"sepal_length", "FLOAT"}, ...} - Y ColumnType // e.g. {"class", "INT"} - ModelName string // e.g. my_dnn_model -} -{% endraw %} -// SELECT * -// FROM iris.test -// PREDICT iris.predict.class -// USING sqlflow_models.my_dnn_model; -type PredDescription struct { - StandardSelect string // e.g. SELECT * FROM iris.test - TableName string // e.g. iris.predict - ModelName string // e.g. my_dnn_model -} -``` - -## Submitter Interface - -The submitter interface should provide two functions `Train` and `Predict`. The detailed definition can be the following - -```go -type Submitter interface { - // Train executes a ML training job and streams job's response through writer. - // A typical Train function should include - // - Loading the training data - // - Initializing the model - // - model.train - // - Saving the trained model to a persistent storage - Train(desc TrainDescription, writer PipeWriter) error - // Predict executes a ML predicting job and streams job's response through writer - // A typical Predict function should include - // - Loading the model from a persistent storage - // - Loading the prediction data - // - model.predict - // - Writing the prediction result to a table - Predict(desc PredictDescription, writer PipeWriter) error -} -``` - -## Register a submitter - -A new submitter can be added as - -```go -import ( - ".../my_submitter" - ".../sqlflow/sql" -) - -func main() { - // ... - sql.Register(my_submitter.NewSubmitter()) - // ... - for { - sql := recv() - sql.Run(sql) - } -} -``` - -where `sql.Register` will put `my_submitter` instance to package level registry. During `sql.Run`, it will check whether there is a submitter registered. If there is, `sql.Run` will run either `submitter.Train` or `submitter.Predict`. diff --git a/docs/sqlflow/doc/design/database_abstraction_layer.md b/docs/sqlflow/doc/design/database_abstraction_layer.md deleted file mode 100644 index f1a8e2b..0000000 --- a/docs/sqlflow/doc/design/database_abstraction_layer.md +++ /dev/null @@ -1,97 +0,0 @@ ---- -parent: Design -nav_order: 3 ---- - -# Compatibility with Various SQL Engines - -SQLFlow interacts with SQL engines like MySQL and Hive, while different SQL engines use variants of SQL syntax, it is important for SQLFlow to have an abstraction layer that hides such differences. - -SQLFlow calls Go's [standard database API](https://golang.org/pkg/database/sql/). The submitter programs generated by SQLFlow call Python's [database API](https://www.python.org/dev/peps/pep-0249/). Both APIs abstract the interface to various SQL engines; however, they are insufficient for SQLFlow to work. In this document, we examine all interactions between SQLFlow and the SQL engine so to identify what SQLFlow authors have to abstract in addition to calling Go's and Python's database APIs. - -## Data Operations in Go - -### Data Retrieval - -The basic idea of SQLFlow is to extend the SELECT statement of SQL to have the TRAIN and PREDICT clauses. For more discussion, please refer to the [syntax design](/doc/syntax.md). SQLFlow translates such "extended SQL statements" into submitter programs, which forward the part from SELECT to TRAIN or PREDICT, which we call the "standard part", to the SQL engine. SQLFlow also accepts the SELECT statement without TRAIN or PREDICt clauses and would forward such "standard statements" to the engine. It is noticeable that the "standard part" or "standard statements" are not standardized. For example, various engines use different syntax for joining. - -- MySQL: `SELECT pet.name, comment FROM pet, event WHERE pet.name =event.name;` with keyword `WHERE` . -- Hive: `SELECT pet.name, comment FROM pet JOIN event ON (pet.name =event.name)` with keyword `JOIN` and `ON`. -- ODPS and SQLite use either `INNER JOIN` or `OUTER JOIN`. - -Fortunately, as SQLFlow forwards the above parts to the engine, it doesn't have to care much about the differences above. - -### Metadata Retrieval - -To verify the semantics of users' inputs, SQLFlow needs to retrieve the schema of tables. For example, the input might be - -```SQL -SELECT name, age, income FROM employee TRAIN DNNRegressor WITH hidden_layers=[10,50,10] COLUMN name, agee LABEL income; -``` - -In the above example, the user misspelled the field name `age` in the COLUMN clause as "agee". SQLFlow must be able to find that out. - -To do that, SQLFlow needs to query the field names from the SQL engine. However, different engines use various syntax. For example: - -- MySQL: `DESCRIBE/DESC employee;` -- Hive: `DESCRIBE FORMATTED employee;` -- ODPS: `DESC employee;` -- SQLite: `PRAGMA table_info([employee]);` - -The returned data format varies too. Our solution to avoid such differences is not-to-use-them; instead, SQLFlow retrieves the table schema by running a query like `SELECT * FROM employee LIMIT 1;` and inferring field types using the mechanism called [DatabaseTypeName](https://golang.org/pkg/database/sql/#ColumnType.DatabaseTypeName) provided by SQL engines drivers beneath the Go's standard database API. - -### Prepare Prediction Table - -A SQLFlow prediction job writes its prediction results into a table. It prepares the prediction table by - -1. Dropping previous prediction table `DROP TABLE IF EXISTS my_table;` -2. Creating table with schema `CREATE TABLE my_table (name1, type1, name2 type2);` - -Most SQL engines, including MySQL, Hive, ODPS, SQLite, support both statements. - -### Translate Database Column Type to TensorFlow Feature Column Type - -After retrieving database column type name through [DatabaseTypeName](https://golang.org/pkg/database/sql/#ColumnType.DatabaseTypeName), we can derive TensorFlow's [feature column](https://www.tensorflow.org/guide/feature_columns) type via a mapping such as `{"FLOAT", "DOUBLE"} -> tf.numeric_column`. - -### Save Model - -SQLFlow saves trained ML model by dumping the serialized the model directory into a table. It first creates a table by `CREATE TABLE IF NOT EXISTS %s (id INT AUTO_INCREMENT, block BLOB, PRIMARY KEY (id))` and insert blobs by `INSERT INTO %s (block) VALUES(?)`. - -Note that Hive and ODPS doesn't have `BLOB` type, we need to use `BINARY` (docs at [ODPS](https://help.aliyun.com/document_detail/27821.html?spm=a2c4g.11186623.6.577.768231deoru03E), [Hive](https://cwiki.apache.org/confluence/display/Hive/LanguageManual+Types#LanguageManualTypes-MiscTypes)) instead. - -Also, note that Hive and ODPS doesn't support `AUTO_INCREMENT`, we need to implemented auto increment logic in `sqlfs`. - -### Load Model - -SQLFlow loads trained ML model by reading rows in a table and deserializing the blob to a model directory. - -It reads rows by running `SELECT block FROM %s ORDER BY id`, which is supported by most databases. - -## Data Operations in Python - -### Connect to SQL Engines - -Thanks to the Python database API, connecting to different databases follows a similar API. - -```python -conn = mysql.connector.connect(user='scott', password='password', - host='127.0.0.1', - database='employees') -conn = sqlite3.connect('path/to/your/sqlite/file') -conn = pyhive.connect('localhost') - -cursor = conn.cursor() -cursor.execute('select * from my_table;') -``` - -### Insert Prediction Result into Prediction Table - -Python database API provides `execute_many(sql, value)` to insert multiple values at once. So one can prepare the following insertion statement. Please be aware that MySQL and SQLite use `INSERT INTO` to insert rows while Hive and ODPS use `INSERT INTO TABLE`. - -```sql --- MySQL, SQLite -INSERT INTO table_name VALUES (value1, value2, value3, ...); --- Hive, ODPS -INSERT INTO TABLE table_name VALUES (value1, value2, value3, ...); -``` - diff --git a/docs/sqlflow/doc/design/figures/sqlflow-arch.png b/docs/sqlflow/doc/design/figures/sqlflow-arch.png deleted file mode 100644 index 0a5a20236049dee1546fa2f1357d6f3320edaf8b..0000000000000000000000000000000000000000 GIT binary patch literal 0 HcmV?d00001 literal 99876 zcmeFZWmuG3-!}|6bcaYYlpxX|-AFfzbax{SQUgdxry!k*l!SDr!T<_L2na|?iXdI@ z8r^m6=epl7_wgRb`{CIK8{uST&huQe*8l&jHBlOB^0-(OSV%}nxQYt*wUCe?QbLGFdXIyTgNt4q3krpbx>;HaYu%Uo)f|;w{Wv_^{{hsh9X|q%-qG(LyVpt zaiYKe`+c1rcGiEN$=Usn+X6SpiTDdAHwPEzU#|_GDvCHNtYziy;^>KZd2MGq4{;vR zUl0EG;lI!J`&lwBPOffN?(X1e;`jb|%0G|&`|;{-c2?m25T_C6{^RlgJofL$YuLJY zxPY;Cv$Ifi_ONmTC;Q{se_r7K{D?nZOOz8aT>lup-|x+@qu|XE#}eiI>x~x23LmRv zL_(58QoJwq&H<-+Iw!S(TM9Ov-7kM<4QKKH<@b`V=e+Hn}B1my?h% zi%4O_YJy8XT_LE)q%0bX9W7ovXAO5Qb=|j>EA-#g{l{ca2!yJ(5Um%%n*M{66$VzTgQ+~e`zrz8+m63f-$2Kk zvu8cDR?>-i%8X0|{rfAPc(LP{fU?o{W5MOo_RInN+0DB}ftP1={pMUmP?!hwjr}K9 zyU%qS`GxZFcs$2Zi5nU=lZg<8;J?4L5|L7&`ET`a)m~V2Z+N}hu!uWG$uGG%SwShvD3buj^B(dvCeLy79Gdj^I<&6AWF$;* z^p7veB%(YbMunbqE3PSzt;tvy&rsI_0nCKXYm81?9cB(WN&G#)z=;Ncu3TDA0x$B)NKTb zt&7C5i>$njFDq9|-ien4WCy@^2tP`g z(JH?_zk{7D4Y@Fm&1aEX8_oUBXFu_dTUjs5|8!r-8Z-7_(!ix@WB$qawzwI^Zr|Oe z^{mft9Yj#ENL)vR)~Rg1_C`^xNnD@CyC2QiILv+Ri{3izeu_PR5;>J8bt6`%)>?JF z{e0Iwh8^Wl!fhooJZS$tb2g2bpWy1yb8n>-mLxi6hK~aU-BdEpN|8eZm>t)S-!)BrV`edhZ@43A=t>ZbrUq!tvh8yy}h^)%JthsW2V; zqw14G_DtgQbf&KFx`YZ1;-p`U;`0?3X;A{Jx=fW4-a1SPqhgazNX#^R@hjtXU6foe zN>+1=sX$r>r=Q}>N8+9lv}Vn7Km4R4ay;iAKf(cXOx`FGR$J*)j0)aLPkA;f-IYkoAI(uN=l%nh;_kRxdVGXhl*AZL28u) zFMh}iz~`FW@4<%!7Pgy^x|l;YvNP>hZ7m+=NzwU)A`7@p@xJLYK>jl1p4xq)Y; z6jnP69UbXb{FPIo_~NhDu}Rrn=VJj|P+D5+3xd&Amo*>GIS+tzXsS$>+U(UqP+w z<DnN{@ASFoC4ZKuytu}*lYZWQgNlapSf6$?!+1WS3|^Z!L(fmz zc3QJ56(!t9cWM3ir1nV@m>7_W6qHb`@ATiBZ}vRc#0e|1+^y1>Z}H|d|NK=yD6n+5 zUp$S9{T=wWv$(3|(-iO9G+&@8z$#TB^--OoA1zY!`r=@$)<#{g&F}5@LWfCzEah{i zqH6r4zG1c~%?7PvC95Byp`N2u$MF(Z1hqJFF}MT*Uq-S;=G~vvJUMn2z}-qyTX2{x zNA$kC*{`!wJY^~K*nf2h0zBkeoSz5F<2}@Jkc}5vT!x78qkchFU?|JJ>w2%m5IRn zl=o)iW8-5B0sW~;qfhul4@=atS#$mOqsSOuYBhV<;ofMW)5tTut(eL0w9ovD5J!VQrYvf?lc5~9kWVx`oqhK0acp&z}(pN6LoUEhe{ zbpQPJQKn%;)^B4j#N_S9y4Ed&WTDzb%b!M_eP@w0)WXtf2d+ssw zgoh0+J#{V%U%L34>&vH0yvK?Zoj-5*ppd4lKo>vNpPc*gnt&b)BO_3Vqa;%{8zAMj z{Tjyneo~bPrGz;z*H<$sj4Rqq__@w``mL(WSD&J)dO{0(=jnV{B&-Jgx<_$T zLIujMrkM~leF2v*F+o@6Dct-Dp98Ne4pnE*m>=}DeI|YNobZcDa%pWfn=w7X5BC{m zf~0&7n)a#ccqSQiUMo{t<)rXES8ma{0ns^op(_1v*Y)l>I-iS!sm(zv&dUE{eLHgT z9b4bCo5WO0SR}VUz1GI2s4(w)Ce_3B7Xe%u{ZDOYS$&_4`}$0aY+c)!TfMrVK43Q z=Zm&+H4#+`en*z^5>;%qZU(wm`w04*nFD1`U5CY`)0BE^B=n0L>uy+YLw!z+2gK7D zsS^X(6e+rU`u1FD4_wPuTkW$jlib5qVE@OpA}$vdTrQCrLxv7ye@Km;Ed#>rhqjFt30K+s;-|aLka2V7(COT9$$?4ZS-Z?-1?maHkT&7bs z=_7IZGkI}z-^hEO34pIO@qn{;l}0Tq;grq^pif&XP}-dgf+BTyYpUuw0F6Tc3)6vd zPenn+`sSN2{dA?O<0=W8vsI%+C8NM<_e)y~JX=I(sL^ixw(92j_YR$gvdU(q1R4_V zls$}KZW22UxzoXm(}Pc!(VXoGZBLGyjU(`=$J!yX3cGVnM#tM~2}j_>H3uakzeflU zj8M)EpBM@v4QFkajZhjV~McEV|{@$uO#%fmS?>;8!53_XU* z*>b(w2`5lKHzzBNL@verPh16JG(!aY+aoMexZ_THwx{c2yL)j%lXZ2g?xA2Zx}Cdno`K^;h|T)y`ETS&M-V;4 zU(V-?e@)9V-fG-7_4E_1Ibk}WDr}^eg955Y+@r_U)1ol3BvzY{`yUbzJ-P7WgC50E zg=owsJcCz$4~^Dkv6`fx<(5bI(_xe$IRrT{k~5#*3m8`wXp$&;z3O zHbbrD=ep6W6h*|eg5X^{vklnR_V<4^SSb67)m7M5weNv-scv;a>-E*Sb4|L~{&wx? zrWSN6k(rLTlZdsDWjxT_#u0%g0lN_C;vwDZb7Q#s&7FWPGsGVNcwv#3fO%^b%HcMx z=RD#0i61soYZH&?#aj8&s76VieK&P1btl8PeiKC49wt~T;=KJB@kr;oKCNWJ{$R~4 z@mk<~Cjb>Lq<-#=S!n<3M-}?v)6Cf(NuPfPXX6$cq`cNHv)&tsA6J(!%6s@G$bTuk z_|BuNdX4gW!E#DRc}voZ>ONX4lWHsy196iMjnV3dHpjd3V0~&yD>` zAH05e$Md+svJZayV`IN=DdiM=Tcbh8*jIM`4ORaIo!8~gd%%?Vva033lFdT>c3Y^> zl)c_SdN>B8S^@dc5N{-xd8HQz%AaO!w;P#7X18^w+d|3lZP|=iE$LYd8#%i;uex zGR7@v4?Jk2hl!^dO0Bca7ZacL)lj^(|1`GpQav|^j&l2z+pJSVRP{!lNSSE_sRb=q zn&cE*>CYNYg>I5b_w7@fWbiw^pb)Omt1Y3>f0}zryp4AHb6mZz{MtQeme^gU?xj{o zU;qsHAzZwR2b)Sa&SE?0-Md7Fc~{Z*Lda|pT@UX&P~GH}c=H4w((e{KrbNnN$O5Y4R#Yeh0*S!EH`zinggsw#Z0upmzx@bLL&>yLiQ=IG%e&W3!C*s{8G3`!gcSw1wYV z&=TRIn5d{R4z)O#wR;)daz2phf$rtL4`1kjk_2(WY@wbmAjOIppn0OQ^Q05EUnTz} zxD)>wI*ZJn98E~W;rI6QN`$6Zmhxs&XJyWB{b7u_f<;y*jO>a0uGRgMCP;O`e zLZ|H~3eW(G6=n)GXtPzVmq)YYW4kl8FL7u@b^VTa+0&cMy^23deBqkJDtwXPceoif zbMH1QuP&cdU?b#1xCXy!j;P)&Di%Km&Tcn;tqh+n^ZsqWi3)Nq$D*-a{mmY@Teik!AKtcbpR;5wwB%ufUO*kFHIQVZ{-UMA= z;q10c!1sx42I8pO)r5p2R0%9nw8y!Ep4|&7bp4Ted$7T4N1s_Fs3Y;pi!UfWo86!u zf|ZmMSTcHBwP}4cH~y`HS|YtHRjs>5xSJTjE}J$u>%xF{x?jGzS(9<+YiED-&1o3- zGw!#F7EU2#GHr1#9elh>2t}yWh=P*y0X>_mz3vDMH$h4lMUPTSP# zE0g=YJm+NRx_sx!9d*+KqO$|$2N3Jmb~2p{m-PA0?Mew@q>Qq>0%8JoO3se`ugp=b zD7Z`|!`&*3+w@bRdzoBjAzL=MoApk!k5Xxt2jX=?U-v(XwWxRY!Y!oFnZfPnt_`B9 zH5}xw4drW#7e9NadiNvq|Kf@shOFRpY%~ceLkFv?XR!xYS{M?B3jihAjG&GvH${t_ zA)30QSrB$czWiDZEYBt_Y%0?w38e&qRGxt;@$fu40YC z%;>(9ROHy$#HgrwrGrB@7)DJ{t7@;6dRH6k#yR`@$&$pH*~Z6(pHH}yt!q@x57vhD z7c93qjqR})&?MVp_O3RupRO|!h@mJR8^+by8v}`>`NyYzY3u!|hl$cPzq1cOgZJAq4N> z2G>apDD}N1g7M?M0Ji%m8{}*{evdNW{)hJbrw*+k05E2eL3C~VzcgLMuRzIv1yLLi zw|V>D7{@$oj?AihUpgLSkV zB2>W(q|L3hc;(fZfv+JM89re^NT;z=c*IZ*0XS|m!^xXiIjIT3Xe2KXoj+)8BaPu) zxB0El#nzBGY0w&++Y=L9MYGdD(tqJSKuVGVQaP6?Hwz|&VGpQ#%Gtu6MW$ULZhvwB zxtV&WOg`HWlHzCI7<|7*F&6>wz~lGR_E5eRs7dysjmO^?Oz>#M%L5zc+@cLVKFYdn zPjd@YARz5{gV`d?6~Lk?E=qpK1RxnG{rK8fED1x=YVmT)5PV#4HWOdfex{^fXGj0U z$E!>vE{&>;#cVo*k>fN~zPb3zgyT`a?vS{?aK8kg+nHIX(r~lh=ZCo`?;ycTjeKdf zD&w}F-j`>G)P)zWXGc3*q8EETc>B^bjwe<_X+i+1KJZ?#Bf0$o0oVBgy{xoHSps_t z*b8)9Ba@}u;8MUQh{T0ORt;~{@6+T^XLRnC__p~~VH^Iy@Sho=54J9N6aRE1Hh|6J zQ9GU3++QWc+DG+Ytt5b1OGTKJ7T1G@#gh zZW8!-f93shdIk+13HhTuxMU*`?1T|a=D@5cyi|aMS;IDc=Jtq1yjY1d_AQPHuJ0Y! z?X@lo;$q+Dnrc-?X6o#c(=2);2{$KSX#%0BSZO3T=xQUJoX0Y<-=N*UX`!YWxzL9w zx<1nU2fZ5BArHdNf$~G>+xcsw-#zj1vP5agTsz!s<8bX`*t4RIMtTv|Elw&da8&(4 z?=<|UD*~{jmzLlnFKmUa(?EkWjp|VGY2dwrK6|F&MZ-0>0)ne_tGl(?>YV34<65K* zy}Aa{{d)6~B^dq9z~g!EG7;-kK9?_sqv&{4-_n_Q>d4SK%&n2(}<$R#@n!4-SERDNFiGKbZ= z=^2!!S9E$JFfv1v9x#!_+o~^KsS?(v$eQB%UtOGX#C z2p_VDv|sEi+s!prvr6ZE;8HeQ0gO~bHfp#Czy=%Msu_( z4TW$gTfAL>F0Pu^1BX<dECZ17k(` zda6W`H()_)%0J-55S830KQS&;!i~VaE_Cv-exfv>%m{Z}9%Z)l_Ui6<^>j+!`KlLM zGP(Tpma)e1^$_2Q}OMeb*1C zOuWUl&-Hy}Ie)xuzyK&a+VxpQp?|>} zp(fGG{#CI#&a}L^-(^St*!2b|ILg~WOEUw~>-}Jb-d#wb0s6Nu3h$>ny2U5idviA+ zuC>wii(LO0 zc}T{mPd}VkD@JdLyR9gAKG`gNX(*7zreF6?XgP*=0^iPDAOjFc;dCjKF193_9N&Hf zJPC)5^-beEPUA3oeKnJ+Emn9!_`fSNfEVQ~Ci?Z(P&BHUXNZr-@gf%z1d}paH@$GO zu>uYaip5Lo^Vx(080rz8bV-)apx*j0@Q>aniq8s$c-KgvHoDrzXBS{*++R zt`*bgM+uQrICq<(Y|9%JPC?|Vs~>pA2P@c1pX&GaXhT=F?z3qM6Ct3A{`~(0RbpCt zifmeHC;uB7@j{@HmfJQ@{~H<+L*Rg%oe};2f&(N!69P1{^CIi!f5AYopNxQ^_|=Oz z{08~|8M3jH2LPhZ!R=A{e~~fAk^x@XnAt7A_kRE&V8~M=vr3IszSlGRFCw#oIs$h6 ze|V!nsB937lOe#o9n3FjfLP-SP%0GWEZ6SGe;Gt|vj4C4&<`c$8X7uc!HuHil_`@T zTp`gL{o$P;TT!f%!Bef6RSIyQPnskU0FOYoKUwP&YX!*a39(7ix932@I8$Lz1d=e3 zeGD-kGUsu%t6-6ZE*O#OPvfm4_Z%iufL~i8^t4xp7FZ_;As|6g5VF5MdQJR&X^9|O zRkuzGhJt!UnK?n6$$GA9emXlhGr)!!IPfg^`MGX+x-XLvt5oY)Qn~y)2uH_-TZVXm z-@7GpE`G?Dz;majug~t9CksAc0_chjX4K-PnhgXy@zFh7F_6HX6ShJ@C!W$sDqs3@7NNq1Q?B@nqy}b9(6h|IqiIBxS?XIIfaH0^Se&bff zhsFp2`-yMyK$LvQW9O$y*C$R|F+yY}ylp`kVo{fXStLb^ogpiXwxW{Y;dsbb#e3xu zPsS|`0u@ZU*a}gc$XpWV+qG|TTVEkXajLG#q1Y*!=AIq|7}Q!1!c^G-iZF;Dq9$A0Gxd<6jQP+yoT z`D6G4yKW350?6;#r{v-OwiW4M0zkSFjy&GO*pMZ|J>C=c+@xaFulqKCJ4%2-<+gSK zFlzFOfv1clHe91pE1qUuF|;X{jFi@+A*!?(EAYh?5z<-jr*!eK9)7Xv_@#A5D3Gd! z^L7=P2nAl;LxMMu2X9a<-k`ZxMK>Wo#x>&m-mMZ1sbT)m#}+%|l=Aw?gH@Uww&RY| zgLM(PsFhaWCC`w)Kc>O!H+6bJKhHOPXKFypGYcdqF2uvPZcedRnkd;qd}xQRkY9|5 z8`jda9zK4Gir+_YM;VDy{br#3HU2TM2vzA;gJlmj%(8% za~CRit4B)&CoQo)LAdA?Jz!ZehDZh_Z#-BNw2v!;|7aItDOK;pBRHW_x~#9eXX~K$ z=B*ix`=!#}O%sPPJ<{$Z0t5HmfAU2*8MrGU zm_!QV1>XZK`ais&BDCTC#lqS*%Q)&aTgN!sVJSs1l%sf_fz#3{ulin_^yI!Z!S_@y znald_AM86|l5{Xzg2iEOL0ML*5MJ|0K?Ftpb;C*t7+u^Hpa2zMuJEGzj8av=t7oWU z;niJKDHHUdJJ`R~SsTl-r%FpNED#dslNWKn_$?fyV5~gT5>^TSM{|I_1w;Qx-re6G z?*jRJnrihoFTzJQbZOjmy)?UxQ7O+Vi37pHypL!fAB*1C{o{Rmi%G@T=VB381=k#= z;u_Q)`}NjM6LorqZ{bc+S9aVThHWJx1RX0BLp)+CmghYXn_hy)jGe1ozXlu}q-$J> zdelo&?EOB8sY=Y_o5owY0zM0a|U{vhj;`ovro!^^wrYp}96d1QOpHeeQi82OM1|Fhzck6V=?`{E6odiA=RBv2&>S}a8ErjX%S;jx@R}%QWI&E1PO8?fN!tJg30RxA5W02BodPzzTmD zTqNsWo3upKNn6|GzOGbl-uDxK7Q}|w4DAUyWW$+ZK^Ti?U7iqBjLN%)CJ%!zdLnER z2|M@apJLx4<+6CeeH(RJpwLnV;2+1%4;jB^IMfN_XE-i_3PUlNLnq8g=Yk*_10&%u z)HMz`)ZUn?%DoQu04*0o326$ylEF~DvC=g_Qvpf=84*2DVXN+b*!Jb4!f2VD2IkjK zX;y8&C%!bjsTZuN#Gt|1--3vU9Zksw<(9Eee(U*eOA?4VAku6rFmxIP8>(^<%=`C| zrf@514dokiT#dj>zkrxdQ;W>W@~T!9%Q8hy>8Bhb(u&|3qC0_`0M?7;WJ=|B?D5&WJ0G$Po%(u(x~Rz*6GTQ zxG@rXVPnjn<#W{6-+)FK{hs+7pgDR%iEMrahnqp3W@;0%5BE5&gU)7d&L9RoNlhfPn=vg-O4}`mbs>*#Keo;f{l#)CP!@fo!bg3jr67=SE%`z$%qL2->P> zSh>FXeqAK}6svg|sYGAKZ29g?@g+#w*->;{ot61t{21_SxBW;Tdj`B;B_M1S0htF8 z2|}zo@ILjck`CtoA=r6FY3zJ{>-oBYGMM>q-jQ%0K{pEeO*MZHQc`^9+Hj^ea!4^3 zi1MWXnb3-51;A`8u-IyT3H}R#=M~Vnn`~uGTa_&krI3Fa>7SuIwjZDyvkg_IVw~N; zCO4>T*%kI(juqg8gJdJ8`}#=gFQZ(XgcZ03;}`kR42huj^lax9pj8#V(v1@!Tsk+&1tW;ofd4+0jhz?Sv_E*aPd97o!T{)vAo4U;C*tu9ioiM`$@%W` zc%kaCMw+_Bvj#%*;l zj{EED$9rb=k08t$W_s7wv1t$z3 zJ$&`*L+l^(8wdG-*~CV)b*~#6$W92u5Hqd+K7G0#j2!w1gj?wl(k_Qy7fw7#tAAsA zhVY54M{_FBy&Sq6@!|9^Gi?m7VqXxQWdb1;{yUfbczi(4^_=HuKQYgCLW$5A_hPsX z04*>VV=#)Gu3jab{6V$3Y_?qpIb-?^Pr#W+_NR~;X4lXq-32|CjPcc9vM*T7;7fr7 znLmwD);UG&^=ykv5e$kD;xho?A!(oPvmjA8+NtsA7Fll5I*y9g~KnuX`OkncV zrLRIaoo_d6Hhf2)JKCM!+G6auzIqbAaDvc%|7Rc3Ulrg-hH%6_1<<_lLZook1UF6918chY-Pjy%mp_j0Zm)IKEcEq%e0Z z!Ptb&+ySEgEtbaskuU*0606z+NAB-^7e5f;_zPX$mwe1zmhT8Ii4TBqb|)p9sJaKK^h(hmRcaDsWrsQT?mA&ce~vd zJCP9)rhD5m_+pGM?WZNJK-L|5>!8!|W9bUus6k)>vw*C5;UYyF_=|RRh?7@!1qTUR zqlxLQbcCrI1Ji`H;Ttw2D!sB}z-RG}sn!8Gujo17_3RKPsdU^>zpf*6uiP$Y!RfeL zwUS&ts3J2eFZdJ@O4K1}z1e(@(_CX;d5oVCVH^rYI6dc;c+PVzz7$3bk+(jEcRa?O z@ejBJ!aIY;v{~sSP*M|#?mLMIszsJJaeaZ+Oy+hP6P}@y%;d+^=FJ`Ls9%j#TM33O zvjz2T$f3Mf1)UpQ?6y4SCX;N}croX_5U?5e^PA==2%13naZMK%dx_OkK&q)f zwucyyoNL0)ys2E%C;E;zu_?&_%_gZ0edg?R4nk(H_isE4129qtp$|~spXJ+&CM$#y ztLvjppLODp%-wZv-ngwj3B3fp%PxCVc#CZqPuawD@utCw^xnW*#pK(5>RRZ5MYu4H zT=zyf*EWb>xY?Es3;BW zjlJ5L$?i+unJSKfg)uM<^e*6eYRlw3QF<5Up{L0OX8sKeb%E1swP1c9y*!U)=Nx{z zi3s?EVp<@BWZ33GtPYAoi`v{77znRx#szo=mc-D7zwmd(q~{hPTaT>35bl{^F#>Na{(7)?m^{jZ9L#T?98w^T^=GWaNfjU#v^Go$Q%))o1^?PX#!3Y^8d2wuz6TN1OKyo-c>(Wt{dy zB}^zp0n4B#IPuhWOHD1NLslnf1YL)B8uLwR_t~{BX3U4KS9Fk2>1hZokxK!z`NK@w zKUy&XE=rWoNd6pRqkuQl83-PFDsO?k3fh0Thp}>>z9z}hpZHp31$%KCHoB&Hbx$*= zeZ3i2ZP{C=2+ctFu__f^HASc}0RNup6=oGzf?Z1^UrE^`vPHa`<#3%|a8+Vilyb!( z2sWbqz-$Q>dfoA?A~YA^V<3N$!DS|Gyji^w?I& z=c!4Z4;M9&f*0U~j8Cwu{c{+&OgoWEO*#V8{LE(SbrnIT*+W?^Q``~O84?K1gz&Q= zrY8bK0R<9qhyEZ_I=-Xy7*FKh&AX30yr07iUMu{*ujeWN1`du*O!u*-DG^ld4WY#0 zF)nI?bNSSn>9G^-F$L?-S2>R{?!`L_>2Jx>Cg807D%-u_2TXWYXFa_yB_7w@le0rVjwcv;Zi zR&WJ_4wA)Zmjpt7iV6JxXj2N%GXM~y#2=I>pde$e46a_+42_7HnrDb!7B7vmKj6Fi zeR3%572&(xrsV9FVZn5XZd5GdccW{GSPL_s7(|X;H&-}ugPfO>6*+d}8q4o#Vc1 zo#$_nc83@tRl~@Lxj0vT%W%TT?&@D1F31+r3$t1m`o#h-8V^A269i4g6Yl3SKt}9^ z2E0NlVGo5K6OcHI<8FPmZvEu`I;fOh#=fhmICh8_?Fi~E!^P%|iqd;o+-lok(ZHL4 zoR$rGj@%{XJ^Pgub_3FY0q*|fY>*}W5PJ+lXSPCEq-@^+ZJo{#^{v^mv5ptSrQmb* z+MZql)?eQp?bj)_0q0&*xqpMo4cTCekDUy%*HuL~2J$)o=;J;)k_+tteJo|-hXmOWIFXO`>3~v~-+@L`Xr1zK+y&QxQu}icvy%omZn92A8 zxpx&X4BX5Qz?EM4_OXx$v`hq_=pRrcjThaIyZ}*LbL&Kf>#IvE5X=XHGs*l{cfg`} z%zH5q+l;22>LPH+=@D^ML|Vx->(d|*6jni!o_IR}A?5vsgy8$0dax3~wc$6V-=MG` z!X8Mypg-97> z1)goZj;Ti8nkZ%9D!?j^NR6QJi^L+p4*eszg;+-kxM1$rAj{bNt_eSxg2~AeKyR=L zb{$h>#*9-a0@!bVEz<=nV0mzBp;!C&VW*i2WwnVj;ym?Gf2VAzYI=7$d(Z%da{B?&1oliC=TE zjTE74kUeJmv!5F^Af%jdAPf*zJ9~fTiVtSVtSR3S*F$qaQ!7lmEcd4Y@`1}eIj4fpKqCaqv? zMTO&B<45iEiAo~_kZCTPI-mDmsAw6H%s>`e&vO4c%lg1n6oFIB>xWa7IhyX%1fH02 z8ypM*cTE(^eR#b5t#WJVn2dBV1Lr1LTrBlWe)g5IrKV6DqgKW*qn|Xk`W{Qa zD!PbJMh*P-7%ib02t-Mx;x?7j<3tJzfzq}ts6sI@!9kI3M^n1{CnHfAS zJR zaMsil9sr|T^%@ik@1E%}u#>rX0~&B3Kg8{lkPY@b4*6|)vzWnY?+#p)x-Vxme9->_ zN}@2Z9K7az_wkYI9Kwz%<&<$?K@_8-GeA$Fu7&r_bOd$KF2ty<;ns)ud1hrV3iH>? z{tPs=_O`sm7A^91iN*~W=xl)=pxqO34b@Tcp1<$9`~tu429IF~h|Ac?;QiSlIhugL zIWOee{ZFYZEaXRTaTvbp!}*ydgH(v)5u<6o4ieh#a2(bFHi_}|xD27x{1>qOW86p& zIGjUufQ-cpPNY6?{FxU^K49&$;FF{6k2dw({M(?acMC^c3_q40>Br39$DXK!_3MER z3dQ4K)9yNGU=!8pnvYtRGhF-iVEw-JwUDjC#z`*g6VSFCN!Fv7b3SbK!S8#_`$(fK z5lMC^!4sY}`8}0v42WA2Q_BN1TI7OxFFd4&|=|8V(XL1Pt={uUt3&Kg3qJ`s=W40h>t~o%*R$~%+TO=B?6qt z#CScCilDE2bb*$1NF&Bg$R&zmHFb2<3earKMe?WTWH}ND8Ot5)7HAwLuqe{`2ze4- zZrC)&&0F~MH1Za*;5DFoKS!T>K4wXooBc1%8yhwSgUX=RwId)GWX~E1~wWG|NTETL-UYHWm&p@z5mP%Y2K0 zxX-jPu}n_O8UuP={k{rp)AvHSGE$*nt6^grdcfE;OojmHpZAfDg+M9n#;XK ze(znwDcl>17D1A-oJAC*O;D`~h znOhnw^xY%OkbNLJtPXO-F^oYLy5U1eSaX<1%8ch0;g9VlV~48@nAAUaC}{bHHiuJ zkHnJ>u}bA3aVvlLsm^niCC5s|yyBQe5M%OcXVWF?=~2bQ(j*IY zJ9-f=a1`L5{21)ltJ>8?LxvGSMjJr{N_QR|pQd~_G0-(@uY;+z`s*$G7Hb-td)y(tKZSy0{;FWOw=DRI+Q05lU-K`V_GbWM;SC{EKsBMm619R|Zq1W?ZB5 z)mSA^x)2*nDEi?O1ci?8YeG@$SHseg^TLSM5Zh|@02NtTjRC#*UqE-2m_?&HG(FvssXk^zowdDegbsZ z+Qk!qB#HohO%n0mWearel%YNUJetAF9NnQ^*J-gnlC5`}+NxYnSH*#17vx;S>{44m4^2H?1+tnIitI9YZRJ7tIXSp!9m-`ZAYQCNH57 zfIhI9qyO^KJQ@7~6ReINi3#1>4B;H$Q3)G&B2JXyHq(ny9 zWH(-7UV^ptD7PP4P8Rxxe7|Ao3|WbrNq;xoppBOKgq4Vd@wi2 z!&{*~N2kv&oi3NqpGHAeEVT_{NPA-Gj5ES(ma^w{o+~v#lzlOsK+o2sVdC5ZR*&gNDiy)a5dM z>hhQT`1bR-X#ESq-n->yZwggjnGxXmo;xQxOj%kqAUz90v88p_(25AY4llZkS#$^4 zfq3O*Zs4FKQcIFvCe+6m^~ng7XIrm}BQisKDcuSpmFpg9fm_(EHUOv&bs^42qX{^> ztoMDHbd7_7M~9J94K5}&Us`=ffSWeq+H~5p;hgLM!v97XjwWX4gtcLO(MCzvPcbgu zKkmMEY`8D1hVj${Oe&Wwt8=f}dM8h~b(*-HNMWvdw7UZ_B;fW3!MiDsjadRFK0~Sv zR=T0A+XH6XR`V#w{AWqQK`7sPK+r(47jkjT)?C64lf<`Kptdmf{~J`{!hAO9QMmE&nQy<9Lx))hWJ1UkVmyl{@|i3 z6j6h=M2Huzh4`cx9MXLm%w#>x5S~00YsL7`YzG!S%!_%Bwdf>zNDV)|kpeUpb6UdZ zDh1ju(jT^u(}l+g{L;zwgbcdUhd0F536PWsSk&@dyK4LA^I{Sto6PGwsVK-wc-jorN@bV?um+pxe zqI_Fe%d?b$E}M@!q7*mro*BMiFgI6JBJ5upgX_s&2RDR_xC%BdBFGC|1m(T3_j7sawJO#;Ts5fZR;vcG8tjV z9!a<4Uw(+QRj<4IOXKLKbCni^jq{aTFv@E$G`Qr(P>Vc5csqzsl>jo~ATTi8Q*JYW z9Hk{=!d)GXJ4VfQ6XkjjfS1Go_6?-b%RaNjA$e&MG&-?#c~7miFhMo$`++9Qjbtbc z2O{Y8E;v*|0lEoEvYR6~9OUk*V@Alh90>?>*Ch*M<7j{mp#y z{{PNA4|5M5PF?%jd#ztB2+!%HP3A$qfVemClK1>)Ai2U2?y2)oxFkXtk7hU_4e7qh zOk${BJDyFJo{7IMfLQF>xhnBs9iP3&-cBUa+TL+4t`9k#&3;7F3Mhf)fvep5V-o1w z-te)VhruSHL`->Vv-WCmi2jVLN7HLckIQJQTgyo4m(sCvQWGRe;gL zxQIBU{dXp)p1pglJc4z!RUDX)I%m@8_sK~_Hz*y0Ug*vQVH6sb#*9vd)DoroHXq=s zze3TKjyfk7_4ooF)>#ek!#)ATwsMVM&#-kqM5FQL0`>HOJVSNl?R$y{R z89)6-D;m6Z#T>Ae%tIiTUCU3mdL)0Y)Fiz_yf66-%j)K7ak(IiyKe-J_ZHHTL=dXTVRg=~c& zPqZ((M{->wh>16b4jp_2HF+M0H^NU;f@rX?(}X5IJ_nkub66mb%&EJ4mGqJ8@pp9w zSr8|`F8v6K9{}4t{>{*FXtyo_wz_nppY0n~Gvjj5Ho7j|{|g7b(3=N>KhA=10q$;` zccnx;6&=S?Ii{e+)5EOC_X{*2?Zo$;X&efPL${#{dw@V+39QPbCeQgVVAh06eZpK0 z_Pk&1GEUWfYnSl-oj`=l_f#SOfPgUMD)kG;9W$*jPo*2kRL(n8=KMlq|@V-SE*dZ=rAv z1(Js}Alw+l0OjOa(M61~hQ?w} zmeXG}zXPq|)o0y9-0@TSf@;mX*aywM_?|8&tk-b7)_(bV5zyT^MlwkRfFzJ5VkJ8p zJVG)zkgwfyM+T3K|1{3OAXN<%=%A7Sdo}jjAFH%~kss{6Sb8NQ;iZb z4&u~lKHq3O41m!NZ&4u@?vL3oXIK<=dw$%Nj!~JgHjM=EEOq23eshkf^djh_K zIO1%3GsEh)KiSi4D12i?@*jXa?F3MW3u-yCCF%_@rPW?dyJQt{^w^7m)IC9I25VE^ zr5u$ujo&y;I619$yzu`pYp+$H8k#7(y=7+tIw@rh&*wz)O}t{H*uO<{4J69H}w?$P9YlqAr>dlfMHqO zh1@v>a5QZp@^`!XGQe3x8xge`zeN4ZmTtO4%s}tcjQGj$LG()!Pg>m7NdU1JZlVm_ zcn`qlWX&%6_8OolnmE<6hq(NdxU*EN8zEFRKk;iqc!fiSj;}Lh!6?ODCyUYCdS8U0xFb5Ps9> zJ8(k73ruscrXE27I)AOnlQKe;5dSg2Ki3ZH(jF(?uAXJeGL!Z`yEItuX!s{xUy||krcwIE~ z^)o-XQhVXKPz_xQJo1xB*FcOomO#IR&r;7NJ7=$fEgle>pkEbN8 zjW?HEa}Q5ps^GP$Hn@Cl)x5EYYZX>(sz_)Fu;gC!C{r?PNa0cm5%187Egu&joF!bg6O2Sjqr68eQ?~4y^_;ve1@Y>C7wfj>V8wWd0zlij&jfk`i zwcSs$WMsQzehR0)<*=-qN10#kn>t(uGq?gUj?Y+Bkq*I%tf0Mk^swhYY>U%^zA~_HNa{aD{hnB_{ zZgpNL5lgzk=UFm?vukwgy?`4y7rKGyvw~IP25=}|BFiY=TVm?0-ThmO29ox7FVwjw z3^L4A0=KtVHiHD)lv#K_WJT`p!p*Q`${~Pr;ep7l zxxQ_O9k{oic>+hvto`-QV~f|jjU}KIj-f@jm3PF@abA1*Ty7A8ah$W}BB^)eBc29b zjB2ycqaV?3A-MMrRYc9ps-NKV1Vgg%@M9TD-31$NLH|#CKaL%j=F&lzrOl-W)l>MG zaI!2%_6VGHONNGsNlFUo-*EEDMNwxcTbJCy5v$rI6IUsz$OS&B45ZdXcZDu7h`PIk z&O{Tpnx&rzwQNMaSc!vRrcU?>#o-_Lh$wwz<`K)uPk=TCwRRk`Z1k=8(}9c&2`9u!SMj=fE}85&M@>!mv>P(%{ilz5w9gqLN$J&UxPK zbWoui8?zWrLio@}ubOK(Jnerr{o0qR9=R8KBMU?$RF_M{zH(6$m_8DwwruyTzK#>h ztp<2B4-zRELow;QD&gPxU}CpTs6j{cB;8WSNtsP*12aP_uN4^yQ(VfrP1e1unwMh8 z(uIErBoY{pu;R`rCq5D*3c-aPg((ejQ>axy`e)wa)l>AsOnl~h(B01`d5$Cu(Q>tU zJ(KizAMIrD8qWwlvosV^xOk~)y*-q2o0$(I7>7aG7?SK#q(fs4he{`@oIOfBo)$Nw zX)9KN^z+Iq1#`SUlH#=62xRwWa9A~ZXs&=e^l)1>_WL?>5T+AsdRxsRzqhvLI;^$3 zbb=zL?&N{ZrIwYTnY{###p=FTiS1;`{fm^%BN)7wN}|sHD5otKUps&1qkEJP zi`yxP#PMWDFA?f4p7OQO$VGFMyM0ySFsOPNBSEqPUD;5DnTmf$9J8Ifz%3dF(fosG zsU?l`W_4y!884%&K$(=0KO(u8%95-X>6zIlQB9$=W5&e0!hIH$Chnf`MRJ| z(^SmtYO?_bA?g4Rhbg%#f=?H0?CBLHpGzgEmn6~3X>96mJ4Iom^`g9RtL5G43&kyC zOo59U@l;iPj^K-TGP3SJwH+-a;~j~Gt&}-sgT-Ypy4M8idVNR27>*3GAkAi%;;%j0 zAE`H17sGdkqpm0X)En9PX<0 zet~1GxcR0Lc#6nAuqUG4+!`lQ@}UmLYGsCQ*i@h9m)S{TpQ7n9 z2w+z@pt;y~+mtYAjpU9iKlNjLc}QUF=5~3FN%_6R@l(b9TQToqersvsu}a1VY58{a zh!65Nh0d%|4LwsOxbwRE6txIQAFa1o-y=#7-rH49$sQ>&;r%UVl<$42a^%url0Kbi zQ%Bj!uZD4INk8=Q+s6;RMm^8pKCW-0fJG;@Xf9Qf>>(r?!EJG)!p)@}D$wP+*0v^F zOTCF?r{vvi<-4#{NiP*JPUfNP&dN94lB!rA=3p1~!+b*6Vf5H-7=)D&iIFsL1v7-| zLUH{DzIDhR|EYF7Nxn?KF4rrG!Iwm!3yM2V+EX!8e?ZKiBjD{BEQ1*Bs%>7AH!S6n zU0G~&0P!q~c2mG%GhXg1Ucwq8po5RSaEpFJH>~pu6VXb!{PL97wl039Rjw=Xq#u(7 z6kYxd1a=x_Qe3;DiA(>{w(v;^br?>hGpX6s_Zb86Z%M>h<|3ZEpxnnA8!)eayo7aG z8a(q^b#111sk(mQ0&F#w7N;_E=bb?G_ZsF~lMw$R!irqB5HQ7jJ26#qvb~hP1B;hO z=8h6ZQkNMPWxpPzK+)Ophj(LQvdg3=hcY*TrcNLgF00A+u=4`Ry{ad;uL}Rq$-M2j*Lq_`g>8k{E?dfUCSMx((UKY~^ zye3Tsk(CB!LeLRLPCppzzXDP;T%~|Moy_`2mSYL7IJB@@2c`NuM9(Q=eiMzd1^r2! zuxD^y+HEB-qhM96KqO?jGJW|YuUIv>Af+XDr$d+O4{JgDyRRPu={+Bc11lK7-y-w= zl+s;G5^Cu%B9c$aiiY-2Al?-zqF+dVdh;+uXtP$bzX*2z8e4>DCMa=MI=3L>uprsfU8{?F&@_Av&aGYX05LhoE=|5 z>yreH<}Oa}4epj!%X_8)rc8G)H9>*=xWW8r9|-XWP;0e|Gu|Xg#uz$)UGB0A(;rM0 zH?0Y=@2uERs`${)y*eD!Gnp^<0w;QzWhbeG3G|tylMwQ@3ArkWcqw16IPJrwLy-NH z^|T{x6Ho7Y8&W$$ma%6zA0&q7<@Q5PN4{oOxckzGCJ6JNG7|=1(N0<7InbBb<8IV} z`U1%y+@LyV=L9-HBC8nEFT%|7yVT%XV!}W`-{$)d^?)h!@QSj2=6?;k{$xi%2qk(h zzR|%VYH=Lv0KNqpA*Uzf-ZRm(c}fpCwd%cHC5jbTh2i zDIZL1^qGE~g`*p4TCgw2Lww?*59D0x<_CQje~afmQQ6Rxmw~O1 zVQRMf(Z}0uv#B$8l1+)kG-JqG<=)}~-pWi0TMKhY#7(~O0Ytj3d%jtLb35bX1&vFl zf+r`B5JQ_e<1Uz0nhwB~l!P>w2nIU;j-;OAHdueQVh=&j?GfGywx1UfIN4*6Q*#{8 z3G(+n#~-2~Nt#70!~Jhb8D_CVZIfw2E@i%?CYSC+A)fBp+Qp&=aCa}9R-#U2Y?P0t z&AEJyPwXo!pCQbr#Zn*tNKJ#V%1E0dLjkLsD?cc#o=gC#(Rt-`jkHeh)2@hGf>8a5 z3?J8ta!J#2!jFNtOBYSQVZUXkX1wk*9efgrJM4W1UX#I6!nau7MbPmpBAK?{ohX?C z5_YsHE?v29hD{4$GkkOG<}bmfAyj~Y2q9wv`NV?$s#8Un!G|J}@hG$=`D{cED_^4k zcyTUp_`_ysM?x5i`vOCniBOmcT-bxF?-=i$l(Bsn&H58u&Wo5iZy{wfqx(82wie}D zE)Y&tn3+|zrAQ7{l6+@rr;jj zX%R;+RRpxYEQ`PMc!`eY;~^yD&PS1md5+Y2yl`TPeZ+GOpPW7YGJ47RO^d^6cENPn zmJy6{)e_4#{8rPUj+KeJuZSSe<*m>)%rdW4V;ZRd@#@NIjh}z$Ru&W_y|Va;L`Jc4 zUT2HU&(Yx7jd&UmVa3^~1umG*RH0(9Tec04N>q32r1@enhnYkuJp;r zyu(>@GQ_LbRltYkl@!J2+1z&{P+2GAO$n0H5`6ZJo?kmLbzSo}%C{ViA*{kAQFN^| zQ_bP`jTQ*#%#$&X2OA)8>=Q)qoriUcgz*F(q4`?)x$?eD(0-*{ntZa0kHe8XuUTfP zXkxun55~0H#91#qdp_SzAyZ)iS#+AHnPXuwI3qWz=B+<49(nI$8JrU=Mw|JaG)Nu{ zsWngG%I*5HtOZa~{JFm;wgM9tiq9}2hp(g#bBzAl6(6U4$p?<&eltFefxiMOiM}ny z+?u+PdX%*OJm6i_WUTBKn=}Pr4~Y&G%^D9EUIo1F#A;Q+CzAzY^a034U;FcPX8b{e*Pje4mFeFLZ4I|s zn3Ej8v92|2Uu9dMDB>daO&ilFQ|MNGB4}P*4>!k-z^7K&p?GH#mZAIRHMAT`raFUq ze3NME3-5`n{cgeDuy~1KL^POL&%S3u$T96tkO#U37#h4s*9luK*dD=oT|h*V>IZJJ zEyt5l>8*?gqb<2_5rCv^;I-SpX^f3vJA%UWd!m zv&)rm^CachgH_J1kij|Bl;TWCpM?4R|nN*IXi7U*ZR$|mse+p{_CN``Z`txj?tx@)Q8G9ezqDeKxr#Z82gOCz1HbkRTKK=}6Nb7EnzwB(qT3{fdNu6g2s=)!<)%Gn`)46=|~oP@5Y zg&NuX1qsop43MsJXh662Q-PP^63DA5*~pLd!W_mJA9W_F1@0{4Q(I zjV_HK)7xB8NmhB5=uh6|sH=E4vn<*8YH&gwMlyNC@K{wmU1#QFgp@v4*@|~RQuDdG zuaEbJCvEv?XY1|0eL9XJPg4ny-@1NDZwat6{iUKdA4jqBf=^g0OwQAAXK6G`tku*n z*ox-}ORTneMIHcWRE+dlnyKAx-V~>GKV#YQabp#(Ps3M*qef04ByJ}#Y`x9gt~qbCq$WeQM)*?hZVmnWMxQ;`?0Utmia;4^?_*%ryV~(a`8j3`mY2d` zxDkK5lfk`Zf`3c1I~9Xw?#wz+B}IUNrw(CTEyf@fUdz<7^NenRPCOzvLyWKxTIZ>v z)CZ!HSKyejNDw6ZT?IiQbdNijl3TxK1}BOU7*J+JnP0cE7r4EGwNhaFYi752HYvg; z?wgiugk1b{+^bHRZmf?lsT(~eV;#7f8)@H|UCWuH6)xK)A8iDwMiB-2lNf^RVLJ%Pi!v}b%LFk$o^{f|| z_c<|3LxIW@ed-GD?;9e2a{~)EB=Kr-;;O{1O59uBNnIZA&YGcG7PzQ^m&bamj(KLAMqG1#JuoBrji^kI|ttK(R+E_%(-DJ zY5sCSOmd7SR2bEEb=k#!HWk^XmW1q zH!q&(QSYd@&@KaiaYW99yODQwGEKM0a1@LrRur5%E_xdcWJsouk>7gT5OhvnZAZ_u zVj@xne}Hm+6mUyt=viZzp%Pt*on?FG*C!@BJ_F9!S-B^3vGbsB|G4}MI4469u&z|V zvg!6HgO;|x!*}p=E#P%3_R4dyUuR;(c*jv#a9U75m8TB}V7aFck?5p$$ zFFtI`$-rO!{mf!E%T5`$$U{ zxin1XUrm4O-Nhl8l(W+i_{lbJl*+_U^PKu-q}QZRdyMcGTro(JJIivCqF%4ruzd~$ zt@|I=?W1wcv&r0^5uN8P<-Cz{qy~01P$S6ZlI_UwvCP^bN<3!j7AS+RIFA#P!r{F+s02?G!M`H0wae$ z7SiReQaJX9C5&Wa+{;}LfK$uqQ*PrU#CbaGbsbt@Fi>?E-FXGL$V0P%#Sw=Y~q%d@a_`NA7}l z3~_VWVeSth7Uj#CK@MHh*?9RguJ!>ooP&@|U|+XX7CFYaA1mj@u8_$*uyVN5cypO!4gfC9_zxqVZvWxAkmx>vVZ)iBC@m(V#cLe4>pNghX~UKp z59j_hSFck<>CSqn9u$CjhJ{6fX8MJS!s-m9ll@={#bedrHG@&XF^8G(0=tx`@U+{C z{iNg_*I_D;Kd5ndVnrEi6J}TsKfmN4gLnF>x-;PDr?rqMTAQw-@eYc^xUU0)L67AM z{}ov&{R(Xc1`osf?B=_aE9Wr2+W%R~5fQX? z?VHP<;mW^Upab!dR4@B7y}R~7fkP#RFZ+|iQhP&glf8uEaEAL`fq06w+Lhxww>qfN zByoTWkO`r2^V~g#Rk)Js{hEpV*m5i=fe4vH3|@>(e|?87=z=}_owxKm34-SK6|={H@s@q23evX@0|-=8Y{m60A%M4VqfWAsY%vA(bD z>zfwZoBAZQT?K>lifsfP%F<_@ua~(MFbK4^re4vhwA1y%z%H$iKa-M@sb=NSr@*MH zxzRTZ97{h$;`Wcteo|#YS9nHA3Z={#8EWm%muJM3eP}0m_uS8DzW*Vb*Jr0zTs!!9wYuZC~xU9GVrM|9gdpMYVnXnl|9(;2ZnjbQIW zip=S->NANeXL1SH^sQB-cL8Osi$A*{#6Dlg-iu9cWcm1_MR+%v|HU>(+@DgLWLJ(b z)A!)!r3;q&)wT7WBWzW6K)TmREOB5|L+iWDi-L9@mx4B44=Ng?&f%`r+KnuVPYu~p zb{;3GxMjbwdiYhi6bupd>Yu7DEO2^Qklm)~LFj+c-ZCV>eb5HjBd6ULUW8|~$!Z{J z)o2V6xD;k&j-BC}b^)a3)rq-+s-aeDFhkwFS# zMY;7hwTX}T_hhn!6V0h)2R|{@UT%h*m$QpqbHtZxS&QTXZE^xtWBcaLPU!2kZ+vL+ z-R$~&`uOWIL8U}wo~FgWv$*qWk4R~OO_pu7vT>E7S);^I+XTQu>dsu-hkHC)O|O+} z7lBNuwaJV>tDmJ3OzGG^r?DkL&bPittS_uv?KvFs+}Kc69z40*q1A`>g2+` zrB88(EWF3QmM;6%PFjb^XRuIb>tt2Y^z+c{vVr`*K}Kw2{^hf{im!wE9sAuE3O#!Y z_Vx2*f^nUNT{kbf530US_83&H>Q$8HG%_CGaijNwA_Abvl->-A)u`yCjCRv&XhvLgqKaPrst@ZAq+72ZL zmaPTw2kI0uH5u^vZI+2tZ|OF*ri$jTo`{~)kdleaej%0KrnDuI%~ZQbzQ=OF_-pqr zL4rMg&K`DcbZvlq!Ly73D58bBX1chF-HRJ`*y93VHb_k*2Y$qJ!-?~H-}B;Q_awD{ zNe6GiY@eYmsroj9iD3NMXAXfTJ?aXFp&H}AA_rnM_MVI$(I+0g^JYV_67*21I_Yc{ z7ti{PrZQuy8fo+mdcL1P#wbqA5fzhV`R`@z^vW}OV3I_0!bvJ0>2x~_-yK`7y)Pho zT%)fd92rGPDIVF7SuEA%v1Rh~Y$0Ywb9=VpgYk<3pU{f^!tR8t1qovvQXB*w<8_y= zji**z3Uu{yy&n)Onn=6;fZZ1V;Sl4xB-=a|+P!|nXXbvFDcy0M9)P|1G;{s|z9IX( zd03+b635-RpG75nINWq(1;0%$Itrc~E47fXPP>a@Qd*E_CkX^8VJ${&)%?BvyiAG%;0%l$ zv~6sDfX#6=oW;H(b|{{sB60o~SLedBZcnzow3n=YuXtfGG`_@U)9pb$KJb6;(Zb)S zlchl6uAaz$R*QvHfVQV@@p`5ot7AXAtpBgME2@-N7=j-$zi9~v;M~u{eH@hdJO5fcd2_{?ji0w3FzN!pLT1jOIYdS`2s0m+YtM%6B~*&b9n}j zL`O)KUbz>6c}Lm=s8#T$vPr{#Oj3eviPZ#21=S6i{J}-b{7W3OGfhOgYz| z7mH*R2trO)9WX(LE;EqZBFG!G4y4K`G_BSE(p3WeIYxdCvD_{>2kxhe?Lwhwl(&iO z7)Fp`k~<75v1x4k*{_4V4!rTeUx!sNi0TDI0zLeOK%yLO;>O4uhR$RfY(JH7vkMjC z4@0Oc6m54aCS|^5qBhAQZm;3Pwd<(DsP1@Q<$}r31o2$kAWtIiq1peX0}rg}9~B^Wqy^#r5Xx^tYl?`$|CRJqj|>nkaB9qn@n z$^t5>DMLsk0xhr9flI1@eUg#qsMQb*%d`iP`%@sWszBlK=sT$Wsz5zd`@PE-s?OT) zNVeHAz5(91l%iexh_IuPNeFuz1N+y}FoS`%Z6Z*Pi0-5uq(3)6>*@e}V-*Yn9ZBRB zwY6)9k!K4Z^a7Wpph=CNppDP0^@DoI6R4?_g(cO&dzScX4`4uNO+15LEQ~knz4_sZ z?WbH09^`mOTu+Ew$^c0AiHrO{iBrmU%0F+<*UZF}0|EGH^Rbo|TDBz) zWUW;y?OQP)e`6j;*!-DLTCBQHl|w8;Ikcj$Osj`^duTXEidOF+)hG5Sg~Kp80k188 zW~0oc&v%AL>W4VkGjKcH6f)P*W~&SYvF0``msj6moa^LzN zbe*0)kZLuTE6Uzfv|)cIC|u(RG2)5Ee4Dka2o9kDH-~FWIP|Zz`KJqVQ9Pn)C|dlI zxrJyH&+Q_qCxaZlr5?Ih=0Pg!#XZ;e9gajljrK*}z@L#~0oRW-nH2gB5d`a$#Acjh<>0j#bp@4`-?VW&xNuL-Kt!Kx414FAqNM(LZHe zf?ia3&XOKOP&`Y z8&?i0*X%|TLR64xWc~;kQ5!G!;dxdb(<9LUB_!bDDbVb${0rVrBPA`4dTSR%m*JKw zni%xyd(KpN5z$uGU9h9e>7yTRi2sAN@uZ_$?%7#*n=VSym$xTQYiU z0$NwJwKwm+pu*5R6P>BCkt;FxN>msgQVh_YvzM;IM>5t!%)%EJ7CK%U)eSw&;X*4- zGLGJ#%bgBX;^rNVfF*Ihn8haPz0~+Y(3qjK@CA&t{BgFp2iVq7OkG*MkMVEH?$A{jvEVjj zub$_&DLBpA#`<+<ro1k`CYWSf6g$TyuWd*^Jp7IK-Oey2Mis4~ z=^_^^t+q)PGG04Cgo$;)35N-NWt8YCrP5#w|5uB##yK#)txn!q0(Zx}TM9qLrsla9 zq2mr^_vXgjscFl65F_xQGngwT>&Jj7V)cz2(=!yQCC%Z|QL0sf%CK2o_G6(=QjFn=v0-FIJ61+ zmAT%tRw7QBWg|!v#S~|Nx$0{|MVI+t1ta$vs9s*ZFZCyI)Cn@u)~)wwswk{kbzfFa zB=_#UryFfNU?bNs#Qr8htITb{t0#;+Iy_1hKgL^;MV+GO)a}e?B%74E0npLPBQU-l z#rrJXzSi7c#XxHNLK2T_^#<>3)g0Zmtfg6a_!H7A6!jkP>t2nFvz08CJ3IHJVq3xX z!Ek5q@@56GqC=l9gL8qZd;57olQ6yPfFjWD(EUcM2ej7frhh2QWtOz2tho|Li@DS|fm^f8jnHoe7AaZ}G-i%B|* zyCIUj^bqH(E{3r*mNVA)jkWR= z2^to7*o-5VNqoj$uRS;GOc{H5`V6h-%PrANF|RVsASJ1V5j+07A#}!l_nulZ=)Qk3 zr0&4E=xPz$ZjThV!Q=Q#AT%#RPk3DQI{hPIWuRW;k1eAH#Jxf30e(ffQ-l;L($u3HjoK5rJsMSO2>Dh+q(=gw1&PWZ7a_+(YG zQp^tAYW`53Poc2jphcx@q+`o^gc<*^65zh7cinyXRN;I^cBOO$ZULpnpm`oLiO!(G z=tYk2i^e>+F8x6xMFK0%^h~pICH?EFULn1VV<}fMvR^3ndNI4wY46HQlRPb^ywTrk zTW~k`w#}W_H>{sh%tz(8Ja4wvL|ePqYKQiEj4b$duSME%DP*jAPTW_)#O97G5)gvorG9{5EJ547mF5OUWy-raAB4 z_VpOX{kgBbhGs_oeOb9n>?4_`(cvFBeSX%rkIX=D-tTUPlwj`E44~&9`8Q?|6;idN zp4tS1ob)MH3A=bLqqrXPMWtGo%*Jq`0-bgao~waUi#*)5b*YWR6yCi?rUU9d_N>vO zy=@O+@NAD!dhf*NOl{7uAd48=Qg?Tl_-3T2eM>%y`m@t8oxgb6@z>21M-=>}iK1Ln zAJX%?J9=hs-L!a-qZYIhqh;8)?%sV})#sjZ$0gAwQGVJF1@R|p_a5W}byC3a;#DG%cw<=Vi#P|K(3?y;%S2l{B93BmVoeO%uCdatg&4Ty-`D~BDa zqf=}90uP9iCoFOfeI@JiUiZFc*U*}hrqbM$c!9nn^y%`DqSuXHTC%n+Zu1{`8~Ow< z3Ji_J*D0?edbn|S{Z$XncEv<_`?P#x{HR@yuN(WV>Ra10uM#(ZolX7z?mJUzExp(n zYV+y?f__Ey-%at4p9FQY3vr>m0>Aul+3`9NU?$c4KU1&-Q zK}0#qdyn+@H=rN`y&3KG8}F2)m?LZGX@lbkNU-pJo@9y|bOyC+1GFG!FKdezUJ7a$ z#Qpx0wFel7Sb-Hd$d41&!Dtd_`LRlM^R=~td5mgDApSk!o0;`CHaMvqz}fjeKsxdF z2SU?AliJtZWqxfB@5Y<2J-Bxa;gfYpVVJ!nc)eW!6l(P4Z#38H67xOyPR%>YQ#L1?5Qbl(&atG{nuynt)7 z86Bbysg9Q_G5=sDQusi?s=t6Av0^!A!7LoYWM=QWKzEuNMt>v1f23rSxV;Hk{d+f{ z!OYc6t<(tB%9*I$V2< zad>A*2p4|PX`@iMZwPjGcpG+ay&q2EZh9t^r(1OtWDJrn{Z-~$M~DK~8b(-FgYs?+ zc3AQmx^9M8~QS$Aq zD(~<3t*FP7hVr1NTAH|3q?mmjXYvBBIB?$@b-u@)I*gmB8`NmyCM`PO>0cqV{NWvi-Dgk7IIFh5Q-A8iV{P1I4F9`90ybw}$Vf3?-hH(s zWce*h1RoWM-$QlSf428BrJQPZtKdJ@KkgH_GnxfBD1Kk>zZev9&dS2=M4Dd$_1=Hp zwD1TN8=B;ti>gxC|NI`fp?G*USxyIpF8{B83~t7QOK^I9mtzf1V5zixi$s@&A3C|Ld!R4wI!Dg%bHbHb5fD-txrNPLuT_cXJ*Z z4!{%%zS=hrj}N(fP0iUi{!gZfOKQn7{pdF`K6>CfVR2~HD|pwpM&@m)?Gr?2pon6{v}z~|fd z3b8M=rBX_D>S%RH`2kBEqHr#u7>jn+-RN8T6G*Ru&b?;}8G1%zJg;UGYLuY6c*TQp3GmxR zuy%|eIXJNE6+L+m8j673&nJZnSZbpnIeer-N)#hA{oaRi@Xo-QCvFdx?pbb-u9S^= zLHzXyK#q<8*BM(fs{j4i0?5Y}i6=ll_Ct(*GdEUMkqlCAApDmg{lcpy7Nmx<3q>3p zFk7Yy^a|GjSt|s@WgD*$Z=Sb=<^7euJyozx#a3JAVg(=i4S^?yUK%-CPDZ4S`(&T%9(x` z&w-3DW^jFTTF_-0jbLmH?mg1+VG2^a(d#!0&05+YJQ`y z&;wK)3mgzL{8wY&kmE+BUq9D3GQMHhll9eBv*#gtC5e>})K4kvM$a z@y!vQHk?r*+_YCopH`$7pcqQWvVOq0j&}_@#-x94=|p2FV&<2D9)l&v0ZZ<2!mKc^ z!)ald#%h=@(}IW1kl7D;8rEVAS6+q!$!KZnhLA+sDB0~_*nVgx= z2NL>9xEp+Lj1MG3lxL0Y<@o48al}QF+Lg%^l^_>~OzOas_kB}zdZV2=Lj0u(cz(9! zD#%AW3H}@_3OC>#b?I0lDyFjY&8^RGWPNQmBkOUrrtqGFx!?0VT&0TIde@4LpRRr2 zhAuxJ5NIa`;)ZK+vIHkVWR@U$rUGz9xvKg5a?flePDda0gcRO>2{Cg9 z^0F+NnGa$z<^^dzG6xsQSg|z{&4&pU6CXML+3Mo1!`m>YszP|0SCppb6pGq7c;isC zju&ADfc;2GhC!?h%LVg(HL+z&co9#647-a1bBwAU>so;7*5^5`>on{FZhLmDo(&2N z_m+ghs!v^GZ(rgQ#k7f%3>zW z{n*%8zxQVk1Oii?ee@nZvhs_N-Bc;X4aEhsoK`k6`9aRt!5%Y$Uv1m}wy-|~Tghw8 zQD`SC{T_-EEc>?{Ja_cHLVKTD0z~{R_pSG1MAIYZf()b`)0x$bd~8>Eiou|Tw2&-; z{i#bU9#3<|MkUFM1i7?1VnIazR+k#JAOyL_f5mrjxJ$sEG4B&khwK?uuxFHUzgs`S zn|_)JUkRKiLlrA`1I!j@&JlAX=-ZOP*RbPp12Pu zBSkZ)z3ri~EC!*Gt5feF7~em^t>cjQ#7${dLj2DO6+8gILxm=o za?Oe1g=x4~_vi?|y<<#~QC)4A3WY`wE+Om02ZYP1{37Sh9xG>MS) z-<|FjVnlhqZ(Ue~&g+sE^s>?-3%GRSo+K6;FrlI=1FW(kNzXzo=&6$&7f+ z>}fok_pvJ30`@~S^W44FJ*r=8^x1X2qVu7C<{a(+@B^NH1gNd=xzjdS_xqABS}cL; zts0?p>OQoUunEGv~I19ROzeJ`{$MBv7iLGOaVR8_PhV7jiH`hdHv6R4)-2Zph5iS z2LDB;{w*1TXW_~v?3Ymeqf|kig;IryNKxSLL;v*%rdRN6vyw|!)6=os*|LzDAHtal z7V1TA{Mm3i8jPL=>Kbps@ts@3x^;IbH=5&TnQq0Nv5@y>JRz*9-ueL%Cs`JAVOo zsw!A2<3O$o`2tk{Rgz~FOjt<;EuL>tr36Chz$_Fmdp|2%yM@q90$|b%Y$rNR9>tMe2zbAb_+TRazgmjP}(-HuXGlZ+xIygBX7>j~K zs|aBtoeE>1P5ub8&K==~5tJx~%{jHT3=m-hL>2f0F4Y(W<}DRgPnH48H-zk7C^r1` zb4n@rSr`hyOi!%9_SO$$O%~FcfQ9vJ#ymL!$v$SGI7DSEF~Y4#{t=Y}%x=)kd|mpt z5prXP@9ORdImj$CC{0CGo~J(R`mQt$+Vm1&-NuL|HZ@YA>!d5q;F6^MdeLcISjb$4 zMFb|>eW{v%l@LOu1XAZJ*rlgZ1RWQQ+mFg!K3}1_@Cx3yh1aX$M}}_qVT{}d?%IN_ zD~X7fISwyw6{-aW2@NTBK=`Al(r72SFG z?jy$(`045Lg<1rymL5pBVp5H?&(btDdYWf~kc+*%#eEuqUV z%Rjh9X51Fzw__2sT~KjV)HWj)TjKL-e*^_3+O{V>*x)*?`PXp8t)% z!&^Otpd>+R@wiMR`j|slc#i$r>xTI5fZk8s{|d8A`6sq=VAk69?o4dmmar2r#r;NR zi52Q_sGco=!k=@U{1LaoO_27NLU1y_?Ie)&eMs1`*j#$(ID;KX6{zaK(Kx>BQ#pzI zS5C6Q1`9E8J246r6=Z-O+fxiA~&<5^%VsXHe^&KUi!s6@9?o zy^b)CUc)rpMJaZDSDx3>sX{1{mcMeJUx6}nHYh<}VVmD_E!tx!VO`%Bpc;A1!2(*K z#~evnOlx_7Z%A?D|D-v*vpd8EYLTVK-^->*3Q8Bfv?frg0VWhm;R54^)}`(!@0#3 zmU=D;%iW!wnz(8zxMnjrD2#VaPN4APpLucf?iLz@Uk}0fK^=$xO#j@_)Wi z*stMJ!hU@VurGvS+q+qRBKQe`1CaYe5Gh&XR;jxG2VZ6)F@`evv!kyaG_do45Gzha zc`?ON1q0gI_Au4P!0`OP4*qFu(k(nvM*?vy3Z(l{B7qNxBB?vQ7KsK6|4S$s;{OkM zZyru{+rE#oTFOvarYOmrA#>(&8AIkNGldG7hmy!L7nw3-iX;*WWlS;^8jPicS{0>~ zS)s76+tB+wZ@+zfzklpM_CEG;^g28**7~f^=f1D|y3X^w&a1D-$Bpxks%(yKVLMtX zMF>no-8&;TxnKh2Ki6~UP^W5F<)w8P7eaKmKJg%09{Kwl?F`lmeLHfsJ5&3XhZuR< zv^EcGTe7@gTYI`@&l=IAa(h81H?*rWs!y4Oz=Z)VRXG4C_xGM%m#L%0C=Va)DmETP z%ptvsXO1#fWlO}vkVaa!<7?_i zcq=x0_$)Yd9UZU`xqcq?Ape8YLILp>NhGW)T~YvN=EF*K1v6A?S|JDmc&L8xKV=4k zD%7>`iwwcP=w)`(!s*K1j*YsOlRy%o3)QV3=`M}uvZ7=Q7WS}a0R)ypCU{7GjQTVN z%Z`S*y{%8r+0N8%R8Wypj$s$uz-0``(djeDnwMihzDgy12pfuaE3uMKLfg=OsK(15 zVJcZNKBvItUuW9J@$jKs0I5}#$lzIB} z>9;qCa>G4DGQTr)RpEI25q~7F0`j36I(Wl#AFtdz2FQ4J0C?{~ymjdZ^!_A%on zTmR#PP<)V3kyYj68;&5zep~YCKc15)`svNNC4YJW(wB6SBtP#I_?mk2=;? zS~!3Xp-PGk?N^!X>Vp+}4R9u%Tq>Sg}!I zJ_Z%mrw?dlT*$Uy?9O@z#gNm3y_D-=l8ptGrzf3+h2b6#cspX7h2XP&%L&2tp9eGA z0;`I|nFG21GdDxG5hOPD2;A{j*lia$k8b(v2UtUh1vsk|**_a}Z7HHhgfMz(+RdH5 z_Sn{+-6jR}|H-#8I7>iP46F@~ZA_qdOcAzgXxERRWI&Vl{6R%Fm?7cH5pekI40I8` zh9q)Nhd&yH@lnX~9&g1Zo{if>w1Y>r;>O&$9R#+V2JG^tHwwrd?d+jCDqt0ct<&*9 z5k#b=sCu0R81oo3cI6L)!%5Y#N0fmO>>hm6VnqfUu3$Q+n=Zzkr0-O;ee;ZgqAoZT z6?Kqk6A?uh%xtX6N}dJXny8*=Pe&ls*O$d7Ad#%h30~^BWmdXppp9DCl=$7j2H3x0l_GG-1=_-RCCz?gJ()+)ReKbnGH!RDOBVQPg7)&uk5%9|5kh?*41R2K zRmT(1W%NU<*nRNox4sHntJa&Iq-TjcqmC{xOY`WT6Tbvdsdfmfu*jp)Mb%6fot?n$ zJOA5@8G6Bg)q?XW2dCUl*)cXKFfy`zd$kJ4^xi>pQ@fUPX#&y@19*BS6AV$UJS#zU zI}rd~)*Oq-v@sBLl;0uuz6Zcbvm0pEJ`Z2V9DqkuWCIST?sIKQE7R1BQ{78ksCf_a zdG$D-7lwxz;QIFexoMBlP0LE&f`Gh7lx6!h%f-Lp*?cgR=UNM3Qux3YYI&mMfrM)` zdy(@HnRLVQ534{xv%iDEXNn{`oj-Rc7Tq0bzHR6{zsLaRxxfxN$c#cJ#(JUY@^mBgi&;Q} znj2`dFVS~{FJ84qJ2Xjs2Vr0tjIFzWzu_dkrWPpu{P~mAjVVR_l$)CdWH|3AL~t>^C!}zi5styLr$5}X6YG7z6IcUF(83L4iZUL<5aiq%y6Byrt>1OZ zl&eQg8yT>VEJtJ2#`JDtIAJMnw{W?~Cer~cH5`<`Y{N1lO%g7VID>5n0lCMD)Q zTemaWgAtPFX@U%Rlg7L-IUODVz0%)ML#ETyREB{gLorlBHH~u(2qzfcaU7Z%Cptyy zN+&>C;SMZ}gO`Vs+(yV=(1ef>zbDKXGMiA~!;QM1w{`r|#gBWb-RXFrZI|cvwqdM| zPEWe&feA@IXgz!G>-0I9-L`f{>PaM#qIf#}mY8w`&iLv^7vRuQ?YWLOfZ@q~*zRr+ z4utNfy$bZ0bEpXFS^UFaa8r#1O7xs|&2zgDr=i5YPMZ?uuos2)C%g@8NeXu)E&FgHTk(q z`HSl#xfTx8ZE@J=^c4h`vGHImar0neTLJbv8SXw{q5LW~?f$U;V0!RoYHMH)3xKlt zfwUFqnyUc;{ynUiWuu9OwM2UQ=M{vHphEdw>zQ^Zk{-6FJoz9Vvdheg{=D~TT$`m%BecU3wWbgNo@@hW>%a~f48~~}EGfd5s6V-F zE3BGmhW8Tia1G9q7PKV4jXRBY`{v=he$bM_sm+)B_=5?RF|oXDpHFp5AkcV07Niez za)f)N`PSc4jEnel^60}HyT}Ye3kqpal!V_bFLDXS`pVf(@+h51s5hA$n$z4Y1%@aW zmIGf;eohJYx4EP@KMDzf7IsLorrK7ln6Sw4XH1Azy2K(U={*H? z(J&Y-3m040R`|h$MVNGk8z?k=k6gmgKzSC}n*Qy8SAk6nF-{a33FM%^yc8-Evm4qB zOU=tUW^5I-{J(MR4p>hLp^8EH($D=wxB|9D*i4wA8k6Ta80Z)%MbKLn#Pv7O z-9D)|16uoP;CNnreNp59RvqOpi;oNfy68d2qv!hzkz{L}S{0O)QTTgv}VCT;%10Ds(d9rQ)zX6u{0UN#UGlC?kVtMtUX0yr72c@cz=M4084Itk4?I3j7S5HD7;+@ly z6vq-248(y)v#g`5gV5@ok&0o{JAxP9-5FRSIrt|y;YW312bTq%+0!tvfm07fK3(&TYr0K~NI=f-*vRy4RT@Y_+? z2oIO1pLW1-UT))V2Y}GZp{e1}Mc09f8wxOK;0JKZ_uW$XO)J(2KRiczmHZc&v%y7@ z_Z@Y&pgx|n88hWJn++l)gnD2>+nE)}P%40yXh~C%`wEtxq{!xe6_txTtdD-^EIlm- zM$g{yfJbUIU;YJsKtn%Z4hjHEtp(ZsZ;ILrW=2=DMpM6*$LtMR@9Tj&$p?}#8JDLz?C7Can1gs#@NykDpuX8^ls)9D6xz)~QyP3XJK zrwH4tW3UmmcsUZDPtKg$ zcVClxy>HM00Jqs868iZ6imt$sqPxU=Li=x6O&)UY4xcRs^8XIV@CqXu5YyeW_&}fX zU$E1R0Mu$-JMh#0is9%@6PSUduJtOQ|L;G+wcmx+DlbLy=x=b#9kHd4(~bV?T_7-a zL13KufBXYII)nj~tQyk9Ut_V&0sJmoD#>1z!;! z%BWKyKhv5mJNOF&lb|bK!sv`&WQH4vhBGX8%`FFFEWYqK(xj1HnIZLqA>r$7|w!b5wmxUe!wMF+x%Nlsj^}f7Ob)|7;+HL)(&?VH5~|E z!$X7v3hb!=*tMvPPnk*H2iRtzzpU=xHF)^%s;$f6Yf`17xSP00ZKFfjV`MJuiWBHHW}%q2E!?5aEK_7+(a}EAwhjP>J#weD0l@n)SoA16-rb{6Zu+Fx!P1sIXU#nk zsw7--7k8{JNJYp?5|+0Dv=ro^4dLPp>FUZ2#`Cuq-&BnLoc(m~Nk3~hOsmrxS8(ji zX@qM&1><5Q0TO`_+I0#nBA`dH;CHq`h_;YjL;V-CFiU4Us^`vE{Td1y(J9!w{LuPM z!2)OW$*i`q?DoLrF@PJBcgX`rN?} z3Z$({Y0Ht%vP2Il6ZJt3eBP4l?BnK}7fjZ>G5(cI5fD*Pv;T(5-+L2D$v7CQ{)Rt@ zumQ2RwTf_m#Q-jZwQ5GV{q&vxT6_7q;#nbYy6_j8B z&pds8zbAK1fo0S^P)WQ2NfSdr7j&(<0F4k|-#L1@ zZC~KCm*dci`P`{IHOAZFDHF(;B!CJooX;0QhYx^Bkb|9LZVQe&(SVH^kAeS+6l+t1 z9EYLV*wkguR^%5p0D34CTTwc8WE5rA2q^~TUq7^x{(=lqS2;1M7brtVB%VUev@mGa z<_DzM8NoA&Cc=|EX`_KoUq!9>sb_`^9o~kyfC0b$W>d zQ0i;XL}|$5p{#1|>%VsE3&7?Mh>iY*+U5k!%Nq4$M(KY+y6Fbx@wJNFijA9QU;;dh zA+;^qH%DPd(iZgZ46vaF@|33xW}(~%VtY+005P9~|KsCJK7(E0<%`-cha13d@Wl`4 zUgcP%H%A<+=}m(sz^I)7=(A2-2HYewQxbrtL37AO*V<%3QXrTF#;={_cROXDqGAz+ zVK*3KP7zNRpem}5oWP`>7eexG27uW(w221+_M`3%O!q9p?IAP#K16=ydYtSN(gBG6(4z4^xq_2O#*Z*TINNfVS;V|Pv~ols=}r|ihztV4z~O4N*4q6; ztN{cJk|2ySKB3hV1|?gwZYCOPZk>wI2a^|Q<1q};Sqnn;4Wu!jQnf@JXrn1$4^b%GP-(EY1HKP->|+c14D z>X{}}O+GYu@6&}O4A1If*ZC6I-zd$s1RU!dINFw@F_Z4+wN{b5P)KvGoipAXXt?3% zOw>CKUG$#|*q!oXlR62@>#G1H>%o}i@QUEbGfGgLN0PXP(rd_hr8-mAL!DG$$fAu- zV1NYPASXB}ZRMCuI78oEs_#C_y65G!dkBmL>oU~Zbr+eJf&KjWTPfeM?)D5ep^vHO z#h*0YKbD_M2;H{FG_QvV&7ZbOP|q%U_d*!wXWfi@=1pkM34Ddk>id{CHuk zUpxEam&DgcvqrcR0J2Gt1sZ9gnM{JV^{*!yez}9eR^O0Sr{b?us2W?VDa%Wo^1b(s ztY@-prjSm|(HSk5B!JQ;6mhO++_W!Mk;{W~!3579RZg+Fs5-FZK=RtmAJYJ%-cQ_a zIc!Z;fOK>{9x;Kpzi>dishq@;vKCbtn&IP)jV^eI)EdSjD#{7)Y0??7TK;<<@L2SY zQRznXd>w<|U-a9Sg*24jNI8&pXnq=XfwDz)D)$8dK<$#T(->yrSlvB-$@%4PKnf3J z7DojdH7nh^{PsN9R;znYKmmo+xpmHp2>nWb8ol?*2H@_94+(CTzn+sopr(z2Ko*La z)P(1cwbHmm&P>@e5~pUqLo2U)a^z3Tc*_=*>wMyuR9%C|{A-ZV?mIl+o`DZn51gUN z_C0LNU23udb$Z=rptGN4po%hz9oz zy*znP$ogKlua|NOkJV`5CCQ^_5t*Lh3N*={+gIvJo1Mq<)KX*4D_~vpi?@IR5(={M z6VX4R;C&AfpKqw@hv!UPGHG)_3TD|2fHP@J%E0+1D z*JZ_7_P@Lb7OjjF6K5v-UT~DgTy0X`^K4TY*7VwPrA|SUYbq< ziD8oo*+JiGM(7ZI-~L zFi@wD6khu79)2R3k}vmQtn+~7s%rHBxGdjxX75ER`C-MX{CN!{zLuxgADR0xriSg5 zZAYFBmLru=T~CjZTIN5Nlog_U=8L zZVRN)`ow0yk3hbQm!!zl@z<5^Nw$O3k+qubRzVDJqljz2z|ZT&${k77lF$# zfR|q_c1|bjE@1d9lx!K-`$#hqUtVPEjV(Z2WzuImxTl{bqt(si!N4-AVP33%W=FE8tS)NFY4 z6LQK-eN(Ny%VN_D01m&w0bq*0_EYld?tYZQU?bGELH$;(J(+=(Sg&&+m3V-!9b&Zb z^rnCvQdfV|Poqm{{n?{QSJ$VA5i& z13+|(s;aIJH+tb-;10$uP98d-tHfRvhgCkNe#S-#EhaUvgZ_*#AHMCV+;g_tbV=+yF3+LWPVg=r6bp zy087D?upFn-~TKBB@YMPf@U;x&aD$i@TXmC$MRyV+0ite`;Wc>mY6ZDVk})ZfjJ1> z9(~%=>3Ya^VewjA`fbLx^>mf350Y&$!6(xwYkUD$o&QNq2-~}>ThLp|eQ4}$q1D~9 zO3tjjEgs!AUiV%PfKxOB9F0Xf0hQ!9;-MHQ&JU;_N!E zz6P$Ze<6q=sGE+}W9O|94L9)Jc#P2HKF^mh3u_QUHF&cAu*!deykt=$L={@3kWn3n zxt^8p&aWy|C=DaY`aPJYX$jOuDR5N%0C+(NvrKG3BLNW@Z4LxaDUiSYh5D?>(U3C$ z#!&8+gK^*BdRz^%@~XFIU0q+^N&&wWAE7SDqun6at`7``48d|P&!gk(ELiH9BjR*s zq{APWJ}Xesl>pa*#0^&E$0C?b^OlhU0uC7ZXr5AMa1uQ4Mw zR2WX>1vi}?huJ;BVdf2p|Bnxzx_c=bFcGu+{%*{W4g}tA;bSU?)N=rU#_v#uR{>4h zg60SQ8hp26cCq&vYISE!KrJsNKtuY&c<*kH82<_K7w4+@5}W*48LUyz~V4Hu;0La}ex} zTmh=g`VQYn!c4#W#84`0Jerj!G-d>fmOdXc?+v~Mu%cjMnQV%mw++0!lmxnq4M7^*-}taaoRGwssJ{mO%00k5ZtxLOw~zBCj#q zTubEGQz~lTNFw#5`DuQUD9k7mdGPM=p6~C$T2q(va58k0eBbA=isnHcF6lz7$lHS4 z7)FwwNOXLm=8Qjm59DF6_*N^ueIJ4W36tXD3mO6u#D)}lRz{b%Qvhez%2EWXB=tD3 zRFTUF79Vv4+I3Afngh6g{Q{uy$D~pgES)4FQzetnJ2(mX)Wj%!t`8d5#lbJ@Y<1vm zdCRVr`*fE=*?Nl3Ev`O4Ydp2zu$Mg9WZZbDoQ`F)No9ROm|dFq3CWR5%K>2F3Q3f4 z=c}krq#Rn4-ljA_%QDA_YkpGTy}G5^{FZ-~*W=xvY!;(CZ_==Q*`M;o?^fQsaT~Wc z!2b+?f)*no^!8+?kQ+(4a0{tYFxT%q>^yF4?Rg{o*ua`kNeEsU3Dn=5;+_v2PKel^|*=u=$$6u37>PCljgFmX0P=WiHD*{OEnhZMX&9Z^=C# ztE#s=?bw^Ig=+TGgKuFzP@#2!yIXO-T+f(Uop$=I_z7}OMnQr5$cp9E%0}g&2~l7B z9a_}+s}wKD2+a&2HGf~la^fmCsbHwA!{hCYFPJ1;lC4}-2jgTuG06d{YEyL9y4pS4 z`5}fDQe<(zi=g+t72ao5?4naZI0 z$|cA1V;wXzI0={i;g|%otJ|_t=?boKshdr0=gISz1NX^b#RO6Ik@tU}eYmdEBL591 zVUg|}zbDD%zd$R%(hv}jECV-kivRp2DUc$x36c4G{_}rre3yXw_q|zv{LNG2QC5@B zUbOWu;Eap$X2LnCK$@vS{C@YKPhA-)=gjJyGJ9cYq;{vyMz*X1*)rTBa-|<8yMeo+ zht9Q&29AHhW`t%WdwBi5ZQyUVoV>9G$1#7zBQ zy0H7m?J8&1%~piw1Av2}&JgI-34#UvC3AWwc7~(-j@`o=Z#uV_r>$Ez9tC+lK2K3O z&y5h38ww8_jTU7{lA^Iln%PNKik)P?$^&EPBcUJE)cm5Gu-6(^R{(HdkB-bXhcACL*6sUR&6A(~{qm8`Ch>1Jh1- zJ&o3XYe5AhLBbh?%kX-vz*tUuee`r8=b{}0$Tsi!f9w!qgE6ca^jS9g%61So`JEq- z1+V!@K%G?q$ZtgMr-3_pyPT01C7{mF>ioaEAS@DOeU1<9fi_yZC90@2z@`&L=ICg1 z%iv*D2%(-=o&pjjs%KY3Ti`%D4xPz&fhHdSD8LO`*UJFVpS?`F#o~!cm6#cu+NX^h zO{om|q5m&ch}`eblft<;g=M0kt(Ux^!ABVj{(8`fz7r`qzC+6-Uqj5PFXTXcbpk9o z9*o*#0-+@w@oH86-z}NNPa4QZae$NoMLQV=pu2*ADHxC{z+!RHNagF%oQfkKd(sJA znr=W^0#KWcd-0mp^b|n32R#O>N+Ik&2iER;Jl#iu^~W2HYORRGmcltY?`T~^R91Zv zVR!#sW@h6P2)YeRt1^31(ggW)+y?S8=vhr7p%^QrONOH3<>Z|5G2E| zr4%{Pz)AWBpmiz1>+L&jM|5)GZ9#&sU(M#$NqK{9eZd$3$eVyMO1%pRy_iVUAvV-U z2rk#;4)dS8ay%fzr*e21WcnF~q2HIAS@@bSy&L@@p&J1HskTF~FgCqmAS+$tYD3(^ zie(D(e^cJv@vlp4Vued=V*TD|!G=?V#jbqH^lP$-JwVSB?%Y}j%MXrYGG^}s30ab+ zikg|=#K`A_PsUREZ4}ho3RZEk_}38O!k6iA%}XAjeDysnX%tNM;(z9J|5V89HybV= z5olgq>kcXZlq3C7Q1fi;r~mvg8gYAvnm2d)e}3?XI{ts-+nA6ayK$2aQ%i3SD43v4 z-3usbtz3KaFTj|O+SebXJl>u{qEX~4ibh=&8xU4Q9~~>-f;J!tfdsXz;5m)RK!uSE z%n9aYmdXA!3>s{L#LZBZo&pfDGF8yZ5*Ye51_O2pgG>t2^M{abLG?yIOv@YuBWmmJ z5m4^>lO~Tnqm26o2y9O#0Hyk{LFiV-wP4;rkPm0OL&!kE`$h2{-I zPXlitRicD8SDw0sw)Vm`T^h;(ev2fxm7hM|g>tjm%W^)S-==p3JZN~AUi-7QYbN;W zuV)88U;Uhw$aw>b`wHm6Ls_6au+vF`R+0grsGf(}+y0u+-!lq$e-d;?YlElH2)M9c z2C-!aj3BuIpSt4A^syq_2XzHen9z3fp3-Z7+=Av2=z& zf(9}o$5PYK?Jy3O0l7-1AYc8bC9v0TX*z?E?cDe)eWbirVC={Q)%8;t)ySx2RKULZ zre%pFuP7v&j~jBK3HRKj?ok(LnJ!-@AqP8yL&^(Qmh=bAbdWl$Ns`gr$DVZx#@3^r zx+%kuEA&ShdT-mvSpXto8F{6RZh8rD49%Yt!#5l|l9FJG5t3XzyQ(l~^#b~GueFu9 z)qz(rnzuRArL?+`ABKtAoYlq!BbQ!LSo2`_?UvVuc}iHA^(lRg(R|x0nvj><&UvF5 znTQS>rgY$^w*QX$PW{f-!CY`o}TN|r`lV^AZ%;Bm3p5*6g6uqY^tx^fK-+@sIZ6n zH=geRaAy8B&^)8bVKZf%e0pg0e6lsN107Bs2EA8aWY2#EgNd zl(l)xA#0;XZ_{D$@7dACd7`k8Gf>H%cUvO_nd_v|HLZ^Dsd0 zG|GauYU8fVa_pT1K;q8e*<38)A)R^D=7_iuEQXFKOBfGtv_8Vv2}mW)0AjGy_=nmunEWo&ZEva4`zz$FyYVz`UB%cZ_?12LCFPuba$vY61?z zgN`2V*`h(`AiLwaY2e?;*bc*2u3yBRN?U}-_);pm)@{G)ae3HJ>;H1GHF`w++GF|* z(-m85gIY3VuHd0V`{mEARSy@Qg1UDer!u;W-$WB~n*Gqu)OS<#xaGp95aV8O?YJq1 zH9UrzTW5WT?4BsxMaJTt1~C5ifMyu3UI}I!M~Z(Uz=t1C z-E}HLnZ>>~Ny60(nIzvh*^`GErw?yeUxvA+PPY4@7e57pB~`NYC&{bO!*0m=2(K0D zCN@ngH<@x~E-?SQ;&-aZS=ao=q#U8I0g8*J$9EWA3dx>5daFF58)b?f(hHay?`0K2~R*08KXtag0Y5M=qP&+|7LKIKy4BV|t|lF5<^q?N328lU_JEGFvT74;|4`T>PL=uj*{mz)i-s zj;u0tJqtHJWr^B!`YZN83%#p3^j$WLsx_^C*oVN-?FGVZ<}Q7zc>4fLzm{z z&^O1ojC;4yoTj6bO+fcvj=NqrXly(nbu(;0|pn*<9u zAskcL?(l4$WNRC2BOrQ`k!nPFsY{h7ap4pQ(7ohLZX>M{k+uX!uv!kqoj<>;sfDrT z$#<--K<6I~E|*&qn{b6IhyrADsx3WlLH()B&XzCG+=C=>Co|0T;GCDsm;~05=H=)% z>V3&DZ}l*wQplGNjrMkJ%>XLlDg5vHlLNi2Ad9H(p9kjW(j098=f;)3LJzq@XhfD> zOiu@f2ij9Zfc*CUnUS?N5po4`YA$G%ht`5P$CG3#&%sgSSH{*aANXXvM-q%~XmlHyU zWq=F{ZdL>WG(VrwbN$fa$+r(4ysGkeXAA3Xzw|2DynoI4$@LCWm-T!F!d8JEVo<$yPiBV7AXFvY-d4Tq77Lo*NM)o zBw;&6XQ;T27>2rupu`BSa)}rJEQa;$7<=VT85SvP00KGTx06s_)%_R}o~M4VYm0)V5HpMaM=Yf-auzp3S+-EnzMaNy3rL} zJ$#>T7ZAzK9eexW{$6QMJUm@aH2RRu<5W;n6~n0$1n(^~(4F=Kx~*PNlI4K2mV-0o z@*`b2c2Gj|8{|F9QL^j98-_0JRXA$`<13;@c5^{1uNY9jlhfd>WDBaFBC8X*WkoQ~ zyMFpHGb|_u5ZuS#JK3YMoCPs;7I>gi+d+^LrOO9Ypy9e`W;UTK$=>NK)ko%y{iS&n zwvA_YLUT)|E+-+G6VI*r#E*|TnH)1!BnmnSG&Tuko-rW|?E@!rN-0UqJ+D;weAYAF z-YCPDOo`*A${jNYG&lGvfNU>j*iJ71?ZACMp%^h8 zraG`AeWaHUHcI)*Oux!z?~ItSrW+k zia9Vn$4W2}k8Y&d@&lV_Mz*zc8Nix;F2zaKz;B|J(xTMIVq#702M8y{%cMog+H?>R zG++9#QVX`~bFOn6qe%((xElRJ$zM@8%gha&r1fMWtOgI)-2Pf3Lh<@A3Hym_z9ACi zD^OA02tN4Rm6VL)+{wo3ZF|$3_c#`pKgY%$(k0B3|8`%XIv0+m<&|NZ;y%V;9^J_I z1=ax@)KxFP%S4;g@tOz<$Y{heu}_iF7B@a2n}K>`!IAXDt<9;dnkV2kMa!OTV}cNy zJ&CQ##%Afmjr#nP#&rybF;D}KHyzva%YJ?$Z(vA*@*Mqn3n+~F*^z#%7Wi9J3k)|| z^S6vA6w>Sw4#vD%It1?zP$ozzE!A^8*-p($sI9xfBbynzrrg?vSsZHf$xiTZ$=R`=^axF^gUJy;#sitB zFd}~0fIgXVcAGI}g*0Qtz_a8FRJpSFN0RjI;Nrvinlj9@7Mxj+6yBODBna|E?{kQ} z^RhK~6-KV|B7?xCaeVkLHP_uT-uXyFjMMMXjr?BUVgw>S=o;sW#~J-H7mP$8zjD4S0spy^bKmcc;zdWyvk*61*a$9&Vakm z7s*s+_ZpB;JHoojC{52Adf1q^#mb(Iyq3x-hHvU+IpAwMT3q`{Kj&H#j7`m> z&6k6&yh^`w+~U)T@#3OfWoBH**3mqnvdCJHTaCg=*hT!)dnFvcq6*{Q|UmnG*DD(YVV8Pv|wJX4r?;Wh!dSeC}cuNI&ZwRZ0FN;vr62{!$^E zqtM=5_Or2mq)fWIA-nRQmwR*mEPMN+L`GA>&^^Ph*^W}7!}{tN|IVnQcHygT7304` z$HS){q(1xAeA^3m6=R%$pTpMd+VMPAQCVg7y#;M_I!RVEM@CE$TSy8S{nj%!FuvCQ zKnI|SqYKj?F%iu`$cYb?O31tAHj5Uq=NM4pthe3J1|SVi5Pq*CA#CSCl))c6pt$$? zpR-s^1J2@hX;B6g{#r9ha!ew5_WUQY%*q{5^8a?loiAN6sE%LVewZ}o5-9B2AsvyR z)Rmn}&i;8ZdpXX3ck|*)g)r%_0dEy_9dw6PAZhFt)JZbcmwPB2TA4T=e!hciw!yPm znp{%xOcu782BnSc|D85w&E1G}6NDaUlj2J-vUI#Z-S;cI+$4kCuWUQ{+su#PuHANJ zKk)4gx@&P!i6fM67K3PR7g_UgdBoISeym*M)jZ2Z)DLkrJkQaP^M&bU+BT)ckWwsB z7<+zm{X)($#RD-(i6(eIDteyL+YJ|qP%o6{Vziy@lvoPtoLtgdMzu$1035Vt$)64LI@SP0mX+dM=y&~ z<2!X@Tjt0jHhw=#Eo2;fVs7gziq5fZ{h?T+M3VD_{9|`2blE9TR^dM}v|DV4{{E|t zfOBzyo$LkSXzq=Rz|%ur8;vo;T=vdRz}Gy>!1iXrjc`^Z=`CNpaX2j%y3TJn^s&4= zK1aF~x(TT0alkV&5`;H)m zO@aiW1BlkX^2NB5vylE&Qg<_Mt_M$AIPjC+F~cHyNNv7_R?`XYe7^T&g&Qe_Hm1@ zP8uGDWH^EsTJJq)<1cS2iv`Tf@E~z|zc5t`O-yvZYn5?tcj`3lie;8?;i6Vt$PH2W z^B8s5GW?pMmr@vdtI~k~-v2;L&ypu(TTO^Q+k07d9K${+J~YJ%2Wf5alp`L`x~|^L z0JX!3b`ZJb)kR*oe256~MGRzG7T{?08(@rz)04I`@ZU51$tv7AkZXq;giYO}KIHTI zM0Saw$fFp>YTc#?nMPdaor&E~VaHB5?{WRQ3u7%OU0n#D&ib9Sb2jmvw%n^7D^9Zq zDI8N#$w%Q0Q%|WoH6BQJCUM80yibq1VTxOXj6zMB#O?`>{|!gfR`uf@N!+8*c$H|E zd?ESwhYoE0i1nU%`J615iWesP8!N-)0vY?`z;evAxUKkhoi;RXa>0$)b#chLrzd@O z?xtlZlw?L1>G|?{7N@@YmIn?KefO2*{37=h+*8*DX`Xn_mCg?xfK=Xhs7qx`gpioi z2x#9I7MhC zuOBMY_*BQ9H{rxV>nRA4g0|Za(6Izik?kjyId~ZMu{k+0xHU*?4Us9NCBA7YQlP7` zMU@6b*_)=HQ2KS5M?V!Z4JFSH;!2c4#4UPd$H=6nd@4MLi6d{14r{Wl2a=ZUGL7sn z!5;YFUW6PIMsP@CB@SN3G7)!8P!x#{A$hfv2Jfv&tCltnhJ;HrVSV}w3{Mj|qf}i! zKXVD8mL@R%zC&P@mNl($_jKthv*p6E!ft4dx~?yz5cne*R7xh`#yIAa#R{?H8_yJ| z&%Z23|4Lfh#d_TE2#fxm^DIuG8~#c$(lg!xct zB$}X=RC!{h_!U&bFFHQMw!Q%*t{6x|mNJ*9_0=YcIaEW9)O$fD$4~r9|Gut?Poe2- z0{$3<0Yws9gXeGP7UOm9$%}BgPiM0IF661?DkQ1qKsP3X5wtl=IhoeumbJHHLT^=Y zK+FVdsE6=3@E=Kq(PB@Zr60Uu-xRvLcN%{NL);onB)>|eZ0Oi^n>lHCi$D(c8)qY~ z_PxakeqWAz!^^*ZrokU;1B-Fv*rSU}H#(+j-!~mpx5wyF1~Jw|cF0`x^gk22<;3x+ zjxa&zvi9b9%KH&VJHG9|`ejB;qU%JR?{PdP7+d2%wIyjJzAcXBGI$J){nRF4$VNaX zOFMgCy2F&pyxu}en%%D=HyFHMXw;51$u0tTodS?;TNS^Y^RT_ubqj+xI5>8Fj~G+X z!d^Rm*6SP;^$7V!5v5e+$Hedfpd=~n#Ec=5tLYFyGeaMKMJ=36nyGbV$9Mb z_)~It`qj_by|{i(KiCNgvOyiG(w9Rn_wVTBYPGCQN833tUOS=toM&px$EnyW%^H|O z9*ONb%!i!}dvHWXeervH6H{mFIqL{1Y@{b>8n+x!6K(3hgjf)eba8+Uq`w*WmKTrR z#<2@qi;mp6`Ww1$z&x}1=Bx=XFN&DuYp|1Cr^vO%DS>|+f%puX#;~mmw?Z(e(yx2o zyMq!Z_X_lGhuQI#^?ZLGlb=;MQ`!x@IT-uqAUgy>IAUx(E_Q|C~S)qGQON6yOasl!QO^%yl`u{xEU(w*r=Ml7UJsH;4p|7v*Ox8)()HV}$OhnU$ z1>?`JVb*MK(LRwf{B-EwZw!@f8q_=12pS2PeJ*ta8YOSLs`8sK4K=YYhpgZquK)Q} zEECIqd0mYioX_!BVVqBJS?iDkOrKMzXMDK z#I-M2WoGb?W2Lp!efOz-A((1Hzt!*??h+}A*LoqYJ>BijXhn7X^JeCY5T-x26=SeU zd!T2B=t8E+??=}>ny1t)Y+!ILXYJU-QDg35DfAQd$y*K{qhzMYe?qs9IdQ<2zH+jP zv+EQg>+`|GB9Y@@4Fr>taWo&m>+&+V={*6!*P;DF+Ge^jLjkOkI^^KOD7XT&X$wLP zJ^(NDwI4zLd?`DX?Vrm8pG5&80o5m|f(HKV)GALNK zOg4kW31;qjZM6aS_rnmG*+Z-NZo4rlxu?f1veGzNwk+&x`MJB{)zq?CfzIxxpDd38 zuqP>K8}wtnCMj|*@}b2-6eo3gXCp-~-8J~y$>DLV;fW|{=}JN#t>$7e9E4SJm}7VZ zx_%CTU93S{CUoX>pXXWZOvDno7n(zcPttX9(1R}BL zDDb3Ftlbb!`dD9rVtH1#Q9q6wwX&F6vWHKACKN{W?vE=ya6kjYRt6XwA}uX5(fm$H zUh9}h4P9=)$EBY>@;alTqYPLjd!z6^s_m!iN98?Km72DZS2d!jk`p^%z!k$r*p5|P z9HDcD885$qjM?LcH9W*S!9l`P6#dBuhDm-oXLQF_;j~0ZrLG(X5W56$#q2-=xXgIm zZFN=ZgG*S7=TRsAq_ZFJB#A6jFLo6)(bH;fB>-t?4HC+0lsIXDd<@s_2MLQ}Jsx99 zycOL!V@D_{q|+PGK?Cl_8X^yEqU6DXMMUT(D9T#(vy zc%~tctmOML4@jS&HGdfVK-_)E913IM+R;=i$L4Sf?>8nx=?uq9n`!Z@*I0YtD(8fk z(p4~e+#!=n=ctXc5XQZp2&1D2kKxm`haPOmB;O~yb25LH$AIemfi-kx&A_!E2KkI5 z07!4kYC4MrUFmm_hCfw5pJ&(eaO}jK6t@pHa{KPQpY0Hlg2Ys~xYbNBD8u=cMfHP} z-63(`PPc3U1X+VQt`(SuiMI$SnJv?&XVuP8_50MkNgy5uOMs*tw;BTcm7h znQq2)Vq9>nb9t!!Pi*N9av&qj$P4sXnc&v^Xm&fwTP(>lG)R+ z=p@9w0Ad({BWhlT_3M_I=6Stcyj^8B+=(kr_76sUcj`@CKlAF6HhB9v1+U+M^{-mo zWkJ9Emz85ZpGEJR0h0&v=37Qa@gz$O%Q3+k=T&J!R|;FBNP=rlPYhWAp?qVfvt7U_ z1OL5`VX3uTiBN}Vg?Huk#o$F|RW7SQtoG9%hw`GJAWVdH`wyVWuI7F&c?C!YLUfb1 zJHygCR|F2KU^yN3;@2CH;Imck7>d&YBZvNllkJc#OEyX8wP%kf_L zj=iMj5ko)~pyUMl&EjS0C6~W6G9E?<{fE{IiV?_k#)?l0JgyQN+7*{3g%ixu7faKq8)hIxdp^e4_<@O3Hx9BqcIn( zuBiqAXol>q5#=C&c7s7$T~Phivxe%1CD=&afRpi<-&wl1BAt<$= z_5;KwY(_vUTN~p;zAI&hLJgjPBIQSw7t(zSMwA1Xx~O`Ow6lMIQ+ZJM{eg{}Sd&Wf zV@L|PgK9OzsNhn_m12k3ta%)7r*(IYdHg>4DE>%n7!-@zE58>Hrg>%9?x0G59#&mw ziYwT~2EvLNcET)7EurD%Edei{htUs}Xm|y>sm(ngJ2ENI%jQROHf`--f|02Jo)9rF z@T67tCqR^G;7G0m22eK!t_$#*mj#3NmKBMRumQr1s!wp)l zUe{AG@N)sJQx2vx)Ru?j<->D_KxULyV80v>)?Lh2c2gs65nS9 zG%l!gi$fQ1RzUkN%iGp6`XCjk&v`DzdTxqf>kyuLka}7|58K`4es!+5Y{Khnz;U{Z zBokDW6kJIm)10ji<}zPTGuuuN=r$u-^Xd1GY>M>9x*qL~f>l!c_VAVaJ3imtEhe7T z^K}tYhM~_{*F>h6QfbaUE6eB>UpU3PyZ*~FiJQ|RdhrjyR6Ap5o2~_b@t>jh;|CMv z*7))V9afAER&nXAvvAYLy*RBaeqK4Mkk&JW9?}*P5 zG2qLXhOTsOpa%+Hlv*9JPsvIbd{NfUYgJW}MlT=*O`x~TSfr!B52BgLmTu^Tnu^jC z1b83?_|aE^EF}WCYmZU4kN+h5oLp5O+c6whZ~cSaOdj^7m||!1%ZRrf50h=;$A-$H z7TTqTBX?{=1tp;Q$Arq&{BfgE#D`Oup`9PUtsMQX$#xWLe{2C_$k@uWaE9jCjRTCW z3A>Y3u|#oe8Ci4Lxnwx#%qi)46NpB}N`X>$=(<;#w!Q8Q@gilVpW=C+T571rWb+y$ zc6RP$koi_6??|TzezjwZg!o^*c4a#rn|RtYKLb4oJ4E82`UOvDX4!swU};Y|mtXWd zpq(FJb?053U$Hw+R2TK{MJjq0NRkZf_{dj(f%fC89n{0VDw!C1?S@rw`4j`u)X6+J`cWvC;eZyP<*did=$1E`v*bcB$Y3J@II3H z#c7hOy&sA^-Zw`Kzoz%duGO2$Yk$EUyCpUXu`zAi9#dTX)N+BFTb?w$IfB?AR+SuG zKC0F-q9PO7-P(k8)jK~>?UA#}ZhCg)dFGLL)eBxzMCJJ}A9A|uj!X9aDiqDG%xJqL z$ot9ag1E;d*<)toqF%`J5a?T4IH!=zk^pW{>Z?TI5g2Wg1O?&~X!^butU^2;|G^s5 zbg0Z{{8>g~NHGQRuyn#@A?Zun$;7AIY->v~g~pz#Pqf}>M@+6iqahU%YUUd#v$ z^oH2+>CohXEAdRC&4JGWC36JviZ?jRy6ic1NqH6j{-}E&F;J*ocDE$KsYC2wd$gf?YV?OD|S`)q%^x} z%pP-DBl5{gQz{zV)0`AQz>IacoJywY16v);Gs@x>8xk z@J1Pw*$?)@C=iLWI>%qNC62hcU(Nu68fkxs4%naU=kl2Jj?NR_LvAvfe&T8(5U(*b zCU^$UGLl+of*u88ED>0CeSmfHzl&x)S)sbL%5s>f)7Br$05DcGGpYPEzFl~hnkv+w zAlB=^mE8c~pri~^ezxRO(8xc3K{3nz!#1HGpCKbj_&Sb_rH-YGWi;ou|82k5FqE#$ zT~(Q%m-$*k20;_B`!u6V50eK`k!>l_OK=Ap*18?rI<=O^^U|4DLs{AaX02=L6RY4?_gF512UeeF zqx9CpUr^D;q70&AQ#tW?6X zae$(a)#!|QJx*;$cIFsJl=b}lDGD#q(~=(|`H1Z_1T}K*$88PDXaf51SKP#suZ)NL zX^I!Nr(RYp4U!x_3YQ&_aHl zl7=A2O+Fak4se5l<|$oq_!H(Yu?F?GxbE6+8QqaAw-Dw>^*iy(F3#h!r}~=u zz3)+dR}*J8w=(|2vT6RKXtl@TTPd$(eQaw7IOU}sNmBwoA4{~uz_9U;k7eWRABh0I zsVRJ-js$Bze{^Y@-1E(6Tr&1Oc13zBjr@Ydq(MU@%vsgvspjP2D|X`nQwMzkjFN;eYym~ z47smBhWv2qu8&(kZu?07k>cY;DAp5fv7MTCK0SV6+2S((P~@=aXk!dIhAuWbPmX5U z4fl;_eRdl=FvR#Zk0md4*Z~;p^hdVzdJOw>(Mx;rt4VpLdu4%TMX|54# zGv}|2cLS)9To0ge;;A#6+j!#-QPp0CX-eIpcDeC#ASFd>49R1<#|)1dAMbw548;!H zJVL;hf}ZXAPoW+%P9DpuWZRm&!cv9*Q-hbgTEqa?15O}*n{DxAJrTtQ%K&qc}yMfE@qqKtTV1FbW0SPl-rHq@?MiT(%a&?tn7Ch3Dt z3qg0Nx_iNQ{Qkw`GR~SLAE`cW5A}+g{_HQ~^&vbg#hi}d%n$ZTl zep#S7-s(HS^?$MV-r-pH?;kK1viD5($S5LvZ;2E_ifq{-Xz^LWaUw6G> zYms@3bfE%g=E-Lfm(JzvGnGZqUrtv4{*M+R|Jk5hD96IZ^1Ab<QE~r99E;rV>dk-wD_-RqZ_s)FV(1}^%NyTdFci+lXF+3b<|<;}01A>U z4D-iC;YZ5x_Zi8t1LJW@!B|?Sjsw~|>Lc&C#$s?ENL?Y!3hJn$AvEucpxeHNV;mg3D-vZ34-yJ+f z(b3@{=}S-teU$z16T%@LlALn0s9-PfnKsrhgPqMsQS(3_z*}Qf_!;z1PyGO7vq)u z!W-OH(XxZ%RaU+q=h^^+5PO%Wy_@}vsy(Xplgo!?;X%npp|iLqa?_vnygvi_w#&jL`GX;Ij6hehLgm2xi%Edt`T3r$_Cm601dx@<06 zqo4YRT(M<>sr--saIG{JwewM%e9@(v)qDH=+e35RzF`UJTv$ z7**aX(7-GCCa2E!M)P350%?iT#JE=EK{F8J8qBC3F&Nj_v4ar`ecPnkpg;T^<7A&u zc+!`*f4%%l^}V^^pF!X~0tRDcWE^jKq(eO)9>G9PeK|3Uvm9tBo}+{eajtf8cESwL z5iBD26~oSV6%(d>(lv4=Q=~4!&-C%z4yT= zy7(T;uO>!1N=qX}v;zJ^9EFGQsBTLcz{uk`db=i1mGE}cU}z`)$U3-2B!O3yX4C%C zIXP!1i_IHFjuvZ5rL*kjmPR$W)>g6v`d`^K>TG)H8sCEu1 zmkRhgQHm%Ic&Q&g2K`Z)@B$@4=yvFh+^(O(9$!zA!mp1T& zE?2RT-jXkWeJ2R%=*Nn~KI$HCrO8XavK{7&ZR3TW;+HkSk+F)nFD&@(Q z6K7J56t*W8o$1Wh5YFT)xnhr|l}N%*4;7xk>+ca-9CrWE)Ws3zUG(s$`fqw!B^6&+ zk~^!eEZcd<;rOmMb7R7n-on6tiJkN?>Yyz@^dG*&CI?oOs_IpdBhs{SsWs^`6xdUEwFq(YL z!pIOvL9_TO0#EyipWLXj>}W^N@Pq}m%a&5!qwqeT8IUz=J<4;IW01{K3krnR_9=Cq`1E3m7t7FITd zy?(qOE?U_sE}-?~y~!aye^ghd8ZzeI#&?`Yp;ApqU4|GHSf~qT-(RI5A~%aj<)yq% zI@^|O9k3TZ{z{xOQ4H5iyo#u4q|y`5*qFqT=F84U5RnP@(*-XQ+e8`&MRI+ zqmz%O<~6}M*(g8o?c=VbK{hEtaYdLprb3L(VE`Ou$~0Y;nJr z1zZgHRVnW8!jd$`HO>hT_u*P`60?NSBjd)1#c?`6!kWKv8Qv5$8K z{YF$xez9IM(WV{;tiROu?uw9{Wy(?xat`0DNSVE&suGqpaNTUih<00~Ey7JE`swwJ zUGsZKx}R0p(6+y>L2htg{#4t&PEvJ7Tio!d&SC50`v@-k*IOHYcA$}n<+tH7CXTa2 zyE*#2Tb1lYd%a*Ff91m=(Bj?fc|qc*iqobLUhGEyBkCW)ZTe1I>I*bJonDRvOd%X2 zg0U98HJfbLnOEC{{e@a4M=CE5X^8F#VqS;&Q^eF;-hUHY;G+^bM;ad`a3M@jb4%GT zVlMbYXPI%`sMV#$ji9j7ZL{_CDT}@cVb@9rUP;+;^7c+N`0Z5hj*%qveL2i6;ld@qKStU@4k;)szDb$ z$v~Q0n|8${rz{_ow_Xv$_tPZP}p(o#&QOf2E+pZq8W$Ge-*k)0Di_p9%W`VXzhD6IQreGyU z{^PH*F{RCb%9?sQ6susdagBUfpE#aXaAO>()oj?EYl7V7SWC`N%I05<+g#1INj(;` z@nASit5!{+UYV+^z$}oR_XQ0P<$Pqo!ABwG=Ou)9VZHQh=1nK0RzXuq(5W_5P~A(4 zXbs`s);T$qTj@)S<^aQKv+okuXwc?K^l(T2Zz=x%tt{+ z9hJD0N4N2sE}_&StBMGBSRBod@Den7>rA7ZMfXJTjNe$57;`xXol>cqMXPNbhp#IU z5!Qz)ifT#}X;Y(wCW7~Kgd^Vh3lWO$sJ-B^zYqqRb2LSU9B>GxYr?3{i_He#W}bU- z&QIkxmDPjaMO8~NyG1tPVlIsmP3TWS?bqTS!nk?t&>D5eQ8ifj{QcuLvAypuEaPxi zx$tB@_l%e)+tNtFBma_w!M(m4x~U9xG={nRnH#scWW^T4Bg}j{kBB=yc51?{a+BN5 zG8>NvOOOKXM4Fws%k#z8x=1J1rqc~MyA+1COMP!(Qhq{eM!CR6egQk$AfGKT0C#3} zGMbgFOhon?K6UgnKqX47_(0jxIkADq0a}KS#ey64%rDto+L&}ms8wRZ?|=UkEq1|_ z9v`)0tRmUm8Ca-&+cmomTQpOx^&u(QW;AQ7c+z{0Ohv5N_I{ovG!apF)adit>%~?a zYanLeSUs=cB6(GeMV62w9MeiT+T5v9v2ZI|%#{Y)$X)`F4}W}yrYlY*=YS3(?EZ({ z^}vlOn-bxc7!hnlb)njrU8Dx5Ms7f9^8_2`%*u{p+m#6I!dI!*ln~olsUC~*(oKEB ziOW+P4amt;g72T-+$?{y#27;n{U{*;IRRAWAb=Gb=V^)Wp5>}XxR6s#kC_Gnoewn`5ZkU7$_l}vrH|8n^V z)3Mhn*@bR~aW?LJjxq`?V(3H0nW4okBcezkzY#E*w-LP;A|%UM1Py#8D1##K!epi& zWY@jPxfVqs870nP6FK`#!=Eju=`W{p(x@m2>^AzRH4dLNr~OZTf7~((X#aLiXxh`t zkLw99%JqOaioz%iOpNrSH>xf7{IvAl`o;orqdP)&6f$!;uu!#NM1Qn(RW|@QY7YZ+ z{&08xuPe!umovj@4W`)#DrXN@tLml+>{CP1D>bMhVzB%+bn**PM0;<}(HXY|>{kRW1|hFxbgvCEyDE6h3idRb-Di3^G*VoA~6E z6a{0exw!$WPydGhH86y@;FBN#)*v5?8HnixM_lu+7t^mPq982I^-VoFprJ*(ct6Pj zvkh=ZM&p@H=Gq!;94H|d8Rt1V-G}En{zc~vRHafJpZ}Y-V*^-YkWgOI0wst9yuX)m z&&y+u0}fHU_&=$UPZB{sDca9m02#Rb>kq~g7jZfCo(-iJCF%q_&u2{(CSwGKskl_C zSWQZatd_r);&{XoJ57shY0U`0;O7E?Z-*gH4 zGkv{+YfkEJvn<1^T~*Zh>xc|_{Q}bw|`4ioFkYy z!T1Ip;*GM^`y)h~ZbXLjYrNX}Ki4007(mbI1*5rVCnZ48$+b|t`{%Ho&LbDlY^EA- zm;5Iu9E;)ug3S0#%ew|IC_nX3?-H@Xw4SFK6izpq4u8rZooJ@CC;!hyAr&eOdK9_= z0jx`O@`8YPX!36!&ejc?^2MWcngd~4;A;*_$(|o*jKB?|(r9NII27f*pG}~QiFK<7 zYH3&>X=_gSDK|x&+xypcD?=WJE5ouG1Sn*YfXS2=h)N5=Qdh6aQjJe17mL%63*|+$ zbh{Dl=KKjbt9q55RT6|fJ@8=vkw+k88$=rimS?<(OSRlh5G`U8xRQ}$g@%6&!T_b{ z!s)L6>2rk{L|+XZT%7USdAacCTfRM>%HP{Qmb(!|A$8a4q(4w_5qWjHsl9311Kc?| zvYTW&OpL1B8cb~Ji5-Z(A4EE@AVdsCg?geXmj&SB5uiSx=|Q%33Y@M>2xS9e#npD! zzg~iEm(Y#-M!ezzu4MR?fyaC|qcAWxDKu;Ldx3RU($(CAmW_>eP%3lR-XTU##@K31 zG6aPvwUQyp==%Cg@E*GbR+Ub+H*Y5g6SliEN!O=cEo1C|-8L_gPm_<$gEE28=Xb ze4GbC!3%vBNQ@CMh_oSF_T@+LzlHe~Eb}w$5VNdtmnGipJ<=cU&Pe%cyb#SPk8?R8IGo zSq(zW&qZfZ^)qFc@A8ciMzY37GKG)R;ouRJDk0QgP>lVu*U3Hn$b+p=KpJS*$qsu_ zINr0ikR|H6>J2PXRyNm34fa4gf{~J&S7=!`{t8sM**Az*W7IT98GRUgc`MFz9A|!Z zh=96wfSG^Q%S1>z0(-Qp2FfpszR3lW7rSGtt!-SwO0)Qu|Q z7A{m$WNv1&C#}9=hork%eD$nuW^*Ti42aG@sZmCj1JRuNtWa`_pW9LA25NeWVcWKt zI(Id_GOfTh2u^{3Ls(aX%@CE@X2spL_F35+SR)U9ujsz{Q9HE&bMIsrLi7Sd?1dlk z2D1ocT^ERPe?$&oSYs3`zToN|{I>-&sE zxg7)(d&&Wg1fh6nZ^Z#3CUB>QVoyNQ`A$~SlB#t0^EG&jg5=sR$Inyj(*G1O3da0p z4!qpo$3WqG?WPkkBQIcBnkpA#!xzghuzhq(^|Q`*6n@G!wE0+S z8=L@qvhxvP;sfx>d$bkoMNbq5p!ejJRgj@wr>!9j@{I*2O$Me;+*zZcqr|aqh*pN$I8c@hXtn z&JPZs+x9rIvrK>`VNBtf{IS^ok?IPTLIW9#EJG!htZx&Hxkq@g^PO}vwLm|zyrW$p z7F-C^a6IbI&+E-*MG*UpCsL^H*Die~Uxw!#B-MiwW)rjohY08io`oJaP+SWGc;aSw z9HLw^0T*vQ;J~bomSteLXiyhOz6&>k?6v^tp5Kh;Ab8*-Ud(x+6J{VwA*f7@kY1CI z_qH78C~zX#XT zl1B+uH$Ku7tbxXMUs%8tCfgZ#S%I-O8tpi+hhF8xNSTDVqd$h3jR)l|pn)dWR@C;_ zDeO6|MOHxzR0K8W0Y!ZC+Ilq*_TN@ty6d^TRrl`_BkLYwH32WX5d9Sv8P4(hiPwYb zHxb0Ct{c71sx|#60Z}z4Tn}*HYu5A{n7zvSSm^??XY{~ioawi0Q({#%;wVIO$yC-( zI@|ROXlpWu$g4(B$ov14T0#t*=R zcZFN%brw(&_+7vNy&(BQM7SyFpA8}u=SjcP+x>ye;JI|@4)Q8W#8`}KOMEHN6sS43 zp7^6wFl(>6jEh15dtf%R?HRL(*K_iFcqo5sBjOsT3B552ud%4b4n$-(UN5!y0<=rt zR0a_R9jpQ#{l|c&*%Y%}d;yJc#;hbcREDb8LHHA2{Ys4cqNq7#ptBGn2@fk*L6>r< zJ+_lG|6$PsEVgx!3Nz3E7{C8Ky9v?*M!umGQfJYgFK68z0=Va}ABd~tlnLK%2vxr% zbOIVcEuLABg{vUWDqEF^j`1bdCX47aqTfe=ryK<*&rNz>>&kw-9=e;@L84%}_o<+m zxIFlQEr5y-NygwRtJn2|bCYJAl2Xv6$B?qTvg~nS8-r`*n{%CH+>pu@=g&pRlD`i2 zQJJ%mm-AaNK{SperRO#%Lgxfgqpt(E-@UyPKmEGnpPlH`E@3=7_Amw31qvB#zqcI& zf_b-qC|l96O#T?*E7g{!CMfv(ZIw0MDZ`oZp*7Z=(I6Z>OR;yx8?0;%swzP{&Sz3z zIi{&foq0WYw>bRl+ih9MT>&c`rZMhQLjq&H86O(W0H~n}cb4bY;rFG0v&AFk@UWNiwIystk{1Oi z&^g|t@V1Rtol#P8RdggT$sxx+?tv2~$U8s~8!2+8d0Kf@R{ zv<$yC-8^5Fi_-3ih36*dNnQp$1M}K%KbQw}EzU#*IMYuR&9!R;P2WeAoxg`5wV&);EA5w3{!T!ROXh~T&KeF zk7y^U>C2ZR$0%d#qPhP4B|6^2Umh5FrQ~kz>8Oq_MDuijJrF;0?e~k9GLG>XdoD?S-+P!jP>4|~b){qmEI$@RQX2L1Q|PCO<@;=}27%e;aRGSL)dl5w1c3^Ss!X|mM$-e{nQhb};qKGdzh||1 zPo-Sy%D5DV_MbyR+#FCT@omVu<=IU+X^Lb`&wum%P zU?osxz9^q&ybn%hj~wv7?Gz~fy3#^uDan6)=_ zq^S60HYJSYkIJ)*5~B~(3|aHa+W131YN0RDnk=3| zV{DU-Fhl>w*4ODMeHXr|E2!s$<`vwv7`X0XW};UaXmzAp>Uu${iRY_?hzSE-aV1^C zIg`e>R>fW18@at?LwmF41(?{g<9Gr+&be$_FE2F~A@JWH!|1ZPu=HO(#s~^7ASj3zIeP0J5wQ7)aQyn0 z@`wq0U}~0N!d?FEl5?!YqfQ9Bl$>s*oZClJX{2n@-MDvuAWD-$VwYu&g&I4grGrIL z^8sP*KN?+vD*;s&+7^4Yi_Z5lfg(=xX5soMMBbv1OU zn!`GWs3cO$7u7al#GDF!N|aexUnKCKYzSSaz0}8wdK>#UJ2&s~V8=?x*yCTiaCfP0 z2KVC{31|a;t^<3Wpp_}j9}&!gTaQ+=-!aurQX};3-xx%@wQtTd|zU*39fcyfw`&{Gh+4KD@0~Z&~P^*`x$G@IZnMZu)io^98 z0w%pNQmu#uUik$*UpLU;km1uW$lmGr;@=TT&9eX+i>vpgqW-uT?ZPwi^F-d7_}_!4 zk5MoWB3oGp1wao%ZMvC=Al9%YbHQTw6ww>P$N3Iy{YxJXlPg!Uztj~f>Haf5q98;H zuMj<|kKkx}9d-zAL;?bYoQIA=ZgZ-LMIcfsn};0(3*tR+t>)$8B!K0dIX@! z70F6p$qe&Pn@)l0wg6JY-}}8F{k0yC1p9c0vs{sG1rBzOcD zufvkU4ce`E5K~@(F7)MCt2ghjAk^aPG1Ek-&eTjnVB&f2jYZoTg8&~0NDk_dFMqa2 z6_D20HLt6c4SoE|RVS6nCf~Q$vZM_lDWK$74-{ixiJ$F-9Q8Nz3R>yGWbPc?qdM#d znNc&1%(l6Gf?U|Kbx+Xf%LBNuP?97~yh%rH`OhX49+BOTTptYiG6=M;#xRN@MXYII z8{qalBDMqiYFD6LT?9+EL|7b%F5*6>6L>ME@@5k*F~Qt4Oy>Kh=1z~-{a;{#J)Ae@ z!e#5djag)Bce~O&5N2E|aWNpmMK-#?Uu`ra#6)zGo`twPt`L(&bkNO@P*ObBr|*_2 zU4DjJwbZ=}cV&PB9{=0Nl-kacE>~2d>_xN!<^`&8y?sUD*z@x`YDzSbg5qU0+Gk)yzXYQ*hmIjow^yKWmp7~)K_j9E>H!~pVOqBF`#2ZKp&pe* zGXOJxSKN2*W zn8pRt8W#mMl(oVZ?%&li>zJp+_-+fw)+K#6>=J)G-(PjN?J;)*+!#XCxi}{ad+b1z zCGd<7nn03#;BXTBx(i&61Z2Fve^6lBt?QEnRV7~DUsa18NuV)%^ptOW&awyQ7cZtR zNRUQI9h%|;T5XG(q{vLwFjWW>V*LwPVydP<@Mr_%jU{k#&Z+THff}bnp3;S(o*ZdR3P^w(j?e?em(1`F=z`J+2Kv{oBYZ zy3;T*eOZ!+sfWys-*SGKJBYyn(q_U$Cyq@S8+Vl1FwhhKq9mAFewQa?r0iPd2^wNgY@iu?vA_sv!fctZ3}(1?!O|!hQ3qAXL70j)wC)a+ zmiON&F)rMNHV*M-ElY`d?X*)y&gjP$m zSI!u6bmOj2b9FOQl%k!~0z=PaCUQI=_G3_)S;Kj>f33w<1Kd>f1LI}>y8q<>ne`BJ z#4B##Fe=1u-g%+VCl@BWp|ab}lamCamFs2rGm^AjfFb^xz% zQR;HvpbzOaETV95D(HDc74JcD_KczU#@C=9T1AR)EOv%K3!Se?{c?`*m^kVQWboX| zdXYYs#|rk#mvrv8c$ZH|=-gN8C2!{3*t8_-?1lcUah|BF%qC-X?_Z7ez-*P9dym2Y zeMh+x;i>H|@Nf7IYx~X4Smu{G=AW>apH+SiI6cG5JcEPFZZJEHLw0Sm?d)UXz3qT% zJ>Ea=Rqswfb*CQ{2s$gfR1+hrMrhZ*apr^8?(ChYGLgM_(hE3yPXb=y=6P2YqDFDX zvjbUwjQe40oeAmv0^{E>A2gdJwv%^YLMy*eVmQYfNHc%w24OBm@M++L($vb_#+!|2 zuly`DsD!OoX|818fH{pi8`Jsyf`3=(JwH%DWb7OICZH8INMXRbxG$hV-{GN4%l$qk-5rk-2S3M(bgbn6iG%h$q<$)987YTq5 zUG_K*?D*aTwbtM)4mZigv$+=oJqde4Q!B7V7F(X+=6(*S9)?Qdu%NB>{7s72__J}x z9lv{n%Cz=2@d9lKE~=UnqSP2JhF5ME>EtOt^aImb#vUUDqZ11s_G_2h@6>V=eg(F6 z2pepBxS2FE9o{k>(j&9|vuJlTexcE}83lNdcO*dU;jnisX~bGd_G{*A3XZ%UsN0d?ErVQoqUN$@|G9>cG1>U>*mstyk;Te zgT_KVmsbMuAfn}>(j?QuDa6YU51~5B-TtfLEWdB1`#rnQ=F8r)vVn+jkD_)o3DMde zSEFTF42)8{j?MD=g0{(Px!AuEmx@oL=SgD(5%MyCLjkAUjCg0r47#GvJ{xNZdJZLs zgAy+O8-N88fKjzN%EPN%+EYf7wuwkYlwno4K`>f>ug6=aTr%x81B%JT0&CO51S*UU zS^($y;7Kx>bStfWL}1<<k(o3 zQ;xv$w(W5uyssRU<1Jo{s(Gsu)z1&O0~7W4$Pi1zR6pgu%l)?&wEh(B9_KOMGFgwl zp47TGP7KOsX)U0&mZxo}v{9#FCqfFFVief(bkS{in<~>vA9=I3Y24dvd}96i{sX@luP5g(`Hp)}4^*Ygk_-q; z3>-?B(nz{=MJ+V%t6n}c=b59fXXs;`vj16f`tqPq0u#sCb3?Ih8o#0o?t2x@qN@|V zI$v>|6)}mrw{TBRS&QNVX%i22ICNSJM>gu_ftF!k`>)A^!T~8$*<_KT|E#du5dmDl zIBs{M8&DJgYVIl?=-=ie`8M{yX<6VF{PV{eDRk1*SF%Ye@m*o1*EL`m(gL~--6yeY z*Z!zTDEwhKFh7{|$B*UDj}@@O-T7ENNJ)x&UE~?1jh-o872=tG+Ng5_!^#j8x~UDzpRUV zqjBG-hr_8csl%iEKssd0>p}f<#}X!&)yb`HdxdsHAH@bOm9I+vdm%`9UG z7~Mt0^9ZcJ?N>%%AlH>K20H9Qm`dusd0<28i#YM;Xovu^*MxY`0o?F>73=|?uY#N- zPge(0mdcjFmO|iyVd;lmWC*$xXEY01^+b~g-Q0Klpw_d2T&(F=F7>l#5MOIQvRkp< zl(OnyVaIsTL3`^x*cP8h(mRi!7+A6=-8?|*dEgfb)_r~uH9GO=V#MD`ZG@8tuMp7eH96u&H2?QTUr5*ox?R( zVwvJWw0|{Q)*JEhzDLP!_7zC_{a=|bu~Tnx5f-pR5%n{xEx5lUvu=1py*96#$wgf8 z9;GR~XQnhzXJHe<42O;GkW3x|hkk+7Py@2h7l6^82rp}i-U1YANw6jADX3rvshv%t zg*fu3U=5=M&Xs%!1%UOv|IubRVn+Rx5OGL=F;y=p-xfOP&5}T@s|hW|b{CMuGZsNt z`f2w)`6wQ*2(WFC;G-8dm<*RDJ^*EhSiuZ^8Nt%2;Iw{dN3rfR#DWv)H@ab85(I%XHShWI#$TGuJWH>kJo_{+lyT)$ z>BHC)kUaDvXcvP!NIHhBW>04Ev-jx*1qKrprKx6^{KEVs-c+ssxwK!#GX^R|x-zy7 zd7nDG-%Qqg*ImmM1TR!%E~7>ERGmu-YH1EM64gzt{531{;^NF-{ULfTU^C*q-qSmmMN+Glv1Y__60py ze9i)5>;pB?0gf^G!7;HR5WQjIWx$PZS^*lBcfX*lOQsXJQ~U~FVSxRs?`{^qR`$bh zdaw!w22{kHl09fXv7O> zRk6PUb7w9uDH`%8KzcAT>-X2uoN<9|`QiI-oUPYzqDYU5LxZ($f?tYD)2CZTlp$2S zrqDq?o6<*2%?;YYMwe}b`=Ow|6#DW*FziMEZjl8XwE#ZEYyGdcVcr)F%u)o5i>~E1 zK2`^*2oFIdp@MMbv{0F12^(WppAYBAchh7 zrW~xxSa5SrzdWGFidiFfoA&ahZ+L;tuB78Qg%?w7(6`HuiLAc8nYL>K4WrW;9R$=~xPFVr0veaxr&7} zw&aPJG)K@Lc@UH5w-&s@oRg@1k%|hX$h`Tf!p!VQ;2{{TpGSNU(nb#N3md#JfyEm} z`S5&kKfbu-DFZug!g%ob5>|K`K`;EDV;bhE>&Ow~FCB;e4kLPoE>^_Tk6p6U7uv0x?~NNFx!Vu1l#gaYh6o3`oEOj0g9!4D6TyEl`Yd~>3cVWJr< z-Rpr9_~AO|=MO0QG0Lf18wIRQ9F+q3#&!1*+m0AAoI-ru3J`DJ>ZLd$Q1^%jv(;XJ zA-}~;&oz^FK5lQo!_&P}8FR*v8}aAo6#^yQ$Y8|zU^~udSK>zBvlCPh=TI6(7BDS@ z`)tfJQ(vl`WV7r>ygE%P#F{}6IxGLs4qnM-hA*b?%l97c` zn)#G|spV^;%UhT#7$v8({=rRypkb$QHSu}8RU0KU7;3PLcDH!`roM;CMewJrp9PF-F80XJW(4OC|d=c18 zD>($Fy?bX`M$-Wj+Baj_v~}tT!sy}cg9zGC8G(>i5ROlFT4Kda6VBEFv+3)b zWL`#FlG9A1?0lxov2`@-=HizwVt-A2VUhUU2GypmyrNSbuhT+)W3MxpD+lZR3Ho=* zU4cLm{gBB}?ruDKlUzC#^f#tUEj#p16@0+r$@ENq_fS`6oml4ZxF$bx%dzo{{oEFz z%;1!iA@x_2-KXZu$LGkUGwbG+h9u%NIu#p#4r}!MD!cnY!L4zkOCh0(@`q$J*NwwF z3{70IbsN!CSsV$%Z9vR3t=y%*vC@FJufa-Q*@{W#K6Tc@PU!R?Afgo+vAuS7gX3Rx zLsnE~ov-VLN)YV0+zdk1yX=_FrQl#z6U~4j3#6eE|4-@v?;hZZ;{Wm;`p@^U2RX_)a35%b@SOr-UYhI^&j6;<*d!svhbHvC>qiWD!Vm)id{1b+;c{+^6eFvIee zKi@I(_kaJ_?|~aEH@LxS$;5s_oMZpjpNLpEs-aCwhyUYyY}hGC#77}>m&4m|4Ad(H z0RASx5UD#|?4>@SrZ6R57!H7`AmXNkxYr?Mma{_U0ZcMp*M^UFrd$x4Ngz5_$?h*C zLhYaO9wB~pAsIf2gHTNV-h)l+H@ap&7SfA(6#S6m3-sWEj*|J2oVW85T8rh-<6_#p z0P~|P=N%q65ySyEP?EYHRBel(hlVkD!yolmPYh($+`Wd^3EM{j)cmi&acLF8a`>;B zqG$0C3#*#*kjtI|mWl6SjTS&WM`sWfVp-R{zSP;9Q-sv#-vc2Eam}RCYJyCPjqW(a zFj9k!8F|dROmCQGNKGen$un!gVRf0>W*ie{&iV)Bh)`@WfNY`$e?W4cByj0zf)5Iy zyA-cqEPceDAWTuwpr&VNrx~BKh$immH}z#~NP0i0 zepx)f%$>_y-L&lkpVYMSR%GqE@^W&0b$s;Mnhvls3--VfPznl|wV;Z?N6_gKn=T|5 z2#GO40yMqZZu7W|-cDv|pt3#LPa(32Ulx1@Cqpq}aS&>y##~7TSvPQv&FvtZ@7L#7)R#{z<~U41|^InUZ-(!?xz$ zq!@^L@WpeCu&EEpdV0-;D9($J7>K6Mgo+#NM}OW@d*rMcy&h6!z*(#A*&B&=88doG zXJiIKkcG8kD-IWh0%~JR7e;>uUI4FMfu5Xuyfz3tzd@28A-pla3C}PhrMQjI!32RET#*Ls;Nu67 zC-@?$09xBuk9P0>39`uhc3SrrkGUR=El!si8MagiCO?QHAEmB~JyfjQ5_OLLnVljY zt>uSQyTRACa};XYsx%+euxaJvBT$DwjK{0xs)vW=MC9ar zj}EboQMIj*f*jw90iPcu$x< z_iL;-PI&3VcH{SqTuUA7)0Wd1{lRLiV7BZMjR?gP6yQ;p9X&IBeswR=zDW-rSV3|{ zkt9XyVjR{6@rZ$?51~4sI=_g|?D53JYr^K{pw%rs(Yy_-X-nbLYn_ocQd6%o89iZAFr164xX7>s;}u)Z`L)bxC>FYV~4d7 z?D6F?$CLRP5hqJ8I=^EaTlLB7(t7HWw>T?}oZfdhJqwpS^wT(W)U8Kh@5I}U$BiR? zJ-vhR&h{-?zjpFZIOl&ta}@(nBtf_l{bjA}xsFaNqKdV@i0sZS+*vPZES&#Iz(#Q% z`{0Lbjk?}}^IQy?ST6wnPVJjsUHzJgm*5$u^u+qL#OpFjJAh0EB+3J5qA?(yD9F#} zNX>AghuF#*tYa_9^hliYN+O2m9ADg9! zxz?#0fp{H_8~Ch>w0z#1)~!&=UnHOFt|+hgdbeq5LIf$Rp~3DM^s(n{D1lN+?)d6OlCy8QB~hu^LyR9HzlrYWU0~ofBO6 zfDEkT+6*20L7p-=Bh7O+?!Te_`;6@2jP53FIWR1?-t@vHqAe~hMc_RxpdQ!&r|Kj0 z0m{{L4YIf%J27PPcRfEDatqy@j_+Em!&3fSQi9S9|O zx62+6n3+i!8W`Bn<$v-t`Yp_r721ZSF;rt6pvZx@Qm6Ix`|qjOo$bbDY9%`nb$4#h zm5t`>KWaBFv3{1~fV855f^WMB}l*Dt!M=@>b zYx@_!1rUIf*pJG!a4X|GU0?K$Exj7qc`PsFGUwbUd9gUwZt7>Tkn+sk;fzwIv+~2v z*E!G0$?bd93XFbFbx0(o$Y@@bB;fK-%H==*M5Ms-w&`%gA^1P<-QQo4!P_T^fmpr&`9pKu7&eW@$;8P&d8PmL zpMYTaG%_Ku-ufRu6o@7iBA}|+DG>i(ulWochAjezBpeCu{cpdC%@hux{iR$9>HkW; zmLw1&nEU@v=l|vId=N7IlqY_bX8iu!A5;cEzJ2?q`f1E`f?F$jyxE2o8<&G$v6*=& z>_YCi#Y$ad9g0yfKP7~HI$=jjz<&GnsU3X%ebM6UTPzW!M&o#@|9)5CCUIPZ`yX;H zF8_Z;g$MkFedhn`KXW+@1)hZUz!EltG5_mVBlkgfvUIZkZa2GP*9Ec=U~nMSv~=|z zJgJZ?yi(WpfvB4gXEQmomP|!EvQDT5Wup-U&-qn5n{O+;A0X6m0+>RG6Gp7Prf+^j z!{n|p5`ZM`G(PjF$vOgFrzGP5zqg+So|GO^yL3COKNw3!U-*mI^|XLM^qQtx2@h%l zmrU-r4#J9ml{LqJJnxDGu#~Bc^9l=6ZPLwml3IL4oNMD@lHUuIg+- zxR}up=)8>N{H)eaKXgP4jFEsRx=W@`T2=PDu`)lOfab9q=Gua9zrTvggy^PMkWEqJ z0Ub(0FqyPQ?YJee`>EOSw_8&9?g%_N0o_XBt5vJ9Ky_^y=B5)n0hwRkfKQmD^Q0B{Udm_Znf^1zu?bdKKY}}i@F~(=buqa|BUAi*|C)86do1li)HNGe3|~G3{fV%@&!ac$ z_>{ln_K{SLNSyosN%DGuvq}P=$?)eJ}MhM7g*r(0 zmR9sN4~v@Z4j?2_;c#Mbp{k2TWbxfbEAb*|SQ25@(hUK+3rOTDaD7bSPTHQ4YdZR+ zg&-T+w}GW)gUCo$q&d@!YduKMiHq9oLoVIN1jN$`*m#byWhQSv77SrO0b$#f>r#s} zx^YN$>z^Ig@uT3nn|)dZ=7p(|vz>BOr}! zv3hv`n*ec|_#L+g?6sRn!{QCmc>Nb+0aBUX{CR)uGwh?kfP$RSh(xIU(Q*MuMRVL{ z+B{hM1&|Fr0|Rat{#UG679$|t*_m;#DQE2m$BTPFEJ=NE@++m8<`^$8m_~XdPvpAC zj0!%#$6Me6^bm@gozd*39-huutqfh$G8i!Fr!;e4p>qyWa^gXzea| z3!aMFP3b2f<3oV7UJ`jb`JEq7NIiuSX6lxJacxbG=2YGD;R!xdiL91iKp=Y$$_K8~ zy)K158|mIqQf%CW3p1nmwux?U!~EWsGam1pr>+|>xxh2$MY?HtD+bLutsgL0=;<#m zsmB7#UeEl%T^yAU`br_%6BTqVI}e}HGQXtUr+m-5;@~! z^&sugj+-(|3CZ8wvWxWmaG>_ATF0X^QfwwaSozKlAcN0)P5}i>n_vOtP&AmukOhqG zB5zBM9!#&k9P>yUn_5GV(2TQ}%^(x(-sPq%Sb?~N_1i5Iyr0YSzLvjKZ+_R^F5va8 zCo-)pb*&oj@3Y-AxwE?*#JNa4+K`!X(SrD(C*MS46hn4`fD$rA0el1Z}}xnckh zfyANBy)%7=#?CgetJk6wj#qEoGwZFG0#ha1rh6w{Ck#>L+H*Er>4GAc?y`0?t*$+u zfI4~tm3Exr@WEx+CFR@mAoRSua*5>y<0dGolLZ7N=^7Oe%&c3!=E!j9=wbV+a=Q;> zM?<`z4V~1Hc4A z4t)NQ#wnelDv|ZY;;Tnkg#6`a)iMoUKOC;FM$;cR>t12@Oj_Y6Y{0Cthcj`EV%k88 zdei=g_iA*!_tsLHN<(UCO21Y^-(|<9hBJ}A!(BnpOTf|Z_zflGew`LIzt-%m7UFLf zDIw3ZN3Z=;iqfS$<5AKR4g&0xI!B70&(EAt2Z3&Yq?R16apWIJYAlsT>IbDW*ExE) zD%m=(1cb}C=AG7l^?OuZSuE&~?mX=1p-pjUC3%2<3QlS1n&tK$!yU#yGPd7kkEJn7 z8|B-4uC!3$pFj*8Rl-Pwi7W2lAO2poI-~s)oVBaN-op#&bP=e$_6aM-4@!Tn-Rhr@ z;EWa8e)a1;Q$b~{t>5eCj*~SdQpY)`M?6oj;k$W!p%I4~4x^1JBOma3HzoFa%Jb7o zokUuF$Mw5X@;u|G;c;@YuWsM&`o5ZQBo&Bu}&=Q{Sipro5Qk5Wt&zy3e% zeR(w1ZQC}YNSO*HnL@}=h75&mo+3latYkdG6awSU)jp678M$8pxD&6~&q*RAFGd+6m} zx7J59ki*v9-;D#KL7XySDj*$M#(7pv11+p&K+AQp`&N=!dPw?Io~zq?J#SHE8yh^y z%bv9m^Q{m0lV^OkJVb6dgQ6yhEMukg)qp)V*|WPStNXP%j0qg4e+MPZ*^oC=Bsqlo z#646JwatR7`q^eM{yTG=>342;{mbCF>0`H7q`dP&1>?K`yGpt;o+T)FRfQGT(Z@q5 zWBmP5AvLMzq3x;;i|K_hI5}tO zTIil2ZY}!#D>kwtq!hYIcgX}#yW+K%d+5D66~(>ze}W$KjN^rl`GTFKYu?@09eSbn zG~9x3Cl1m{SDxe=k1azuhUdq}k`arkAiGLX(OZeFkUF6tew@5xes#5?XNSE*yy}^I zzbzVfCxU9IJ(3HxdvJciJ~!Gm_sOj+GJ1b}vRa=2p}IRmoZ|JCN>PX5NdosVZ~DS& zzB>RVT4R)LZ1D$>iY2Ru;uTOL3}1@|<}deLIh9QH3!mqQPRzuL)w!Vo{^8v8oeHZg zY~k&5HWwRYNXYB-;6W%OP*3Oay){*+F+FxFE@`z}D!R*OK){{<n(ezi=S#J>q{s*=g@U)5#%ugqqZ$Wc zGOJ2S!pqi#_A9khA2RyGq>hZdp?3)_?0peFfZunkC`pF8!eK~CgnZOBw3=)DOJeVr zDGW+yQpBG2sC)qt|6o32lvw0vxq_0`_I9Ox1^!3kHO{B@G`OMmwn%N%^lC+lNWR)j zD*@LpWY)ib8LqkPr1DRiS6zL0LR0g`*xbbA$KUom1bf2ly%WtEtaB?+>CrlFK$+(y z<=!g4gSP|$fs=O-V(GsL&*HRzlzG;pA>@J8?XYV!WU7|;Yz;g+IdBS)3 zLf*m{ljDs3@)p?e{eFqr_N#%gTDdAg2aXV#OUUKE-ZIT zQom@(y0wIvFJ{yN0~j}LysGb@wUhr)^G=@?V}mHTCun4_#-iFYfZ_Q!Dx>p#>nDbz zhPLH;Q#>e*o^5g|`88aN>g??L5?w2^IdcLMR8ziLIHq^W?q~|tD#MZK*iu07MD4;* z8IZHm1MrP~zG*cFTt0b>Hl*@=iUcjYph)@ABa-u#%Tvq~yZ3=P4babfiQDm-Bz4CkX?CHXv+L?n zMByJfb4qjU`p8ggxer_rtU5E)T)zA7QwOUDONgLR-*&dwmymZ{%sdbvPTQt-4mMb!u`euvUsT?GYCjdOVsW_y76=x z^*tnkK8+k6#(x2F{?l&{RfIsbHG!2EI zF%QzP80eiOiL!dHM3&cNNy2dDvHX%&I)P~r8vTT9_H3^=gp4KdJ^lcCSM<}$5RG*f51($C+_#XX!h=oM3SgqS6e zS2^77kkIq{djW~lxcw6R#rTchsFn@7e17v|&C#{~Ar%{V;5IV%5Ql)I90P9Pk|OEy z)KeEi0X1v}3;Qvcj1k$~tiJJ6$d^~@`p1en@tTcLe3~H=OdqX#3qnHNflz1(nW9=Q zdL7M#LF|Y=>^3BO^9lY(sWSe3E?5H5zaZ2Z@gslrNNv zHu8KoGb2r)7*z@uC6^$_r=K{P;$hKWb4W{qT}*Yv>FLvsYX>cFFxi>X9{*K*HhJX5 z9<~ZJiNShrkj%D4%^cJ$XD3!Gp!Vy@H+KPId-h(;N=ED1rook0vA}@G9q_|TCJUs6 zkKm=_`3)`h>aUz(8gd;EE}g6mqWW$y0Hxg3r6V8%cWZ39l~Y)#^-j<&>Nq;JT42Ym z49GoV1EiPM@t)v8;k|F;#UbbLxXnRNVn_y+p&g4=lb3LVoaw&0>dC7C4iz+crIY>a zJ!P(wx+Qj*C1DT^I}SX?yP-kIlF!QhrXf@1I(RH==jHSdF|f@jdmRb-szwmbobugX zPDKDBw>KC<=uQ+Qw-^c$m`1=qI8jVJU)1Xf$`Q!(h<=~_?H=(;4KAt9m=Hqt%Y?~N z(Mjke@wSn3)QAKL*Wq^my_;(|@ox!^vqe3)vZ9cPek}0+nxM(VMqtlNOZ)JQ0Bo9B zHA}GMceA$oy4+J&krgOdksDr9302m3WE?#3vI4A7FhfDZkEux>>!Gfmt2160SF1j; z7&q2)(R%18shagWy=h@e=ut=(kXA4RIuLs3?Y@*M8HJ_6sTOB!C>)Yml-%7rJjL;n;57(TB@^TX zfTofwvicqcZ^3fKoy+7NVdQ(Lv~ux*@Y@; zMaxzS|Kr`;v62q%WpEi9rS|=G-}C!63>>B!Sm^_M5-J%6i$jo zO5%lx^vXe@ie#^;obe9EC?87i0sBeVea0X<0~)tW z0P!S}nY%qlV2q)V8($uGS({~(*O`5QfBO58xvUyGD82aRerWVw7SSSiSgLK;<2H(? zw0eWfcyv%Dp2~THx8CBg>?d@!x#oGUII z#h9@5zT#{5@g(_#@gtA~5$ccP8VJLSqkdZWB)k1XuC#`-K52>T2Jl0Nn5^E>m0)@}VWFIZaJa+N-2x%c_1J&xE6_467KQ%Efm%m9kcQ;r;CS;t! zS~Fr$=9$uSOA;xA4j`RW?!~-0?Uh_@%6*$iv2Dod<3Sf#K{S{U2Ys~4N&Ya%gf1Y% zT{UKLZ-pm?loqTDH-G`#zaGYTn*<&_RX+$jButWu`kce^T}LQ*_AkcyDQ7+Ag@ag4 zr}#d`i9|NEmEfaroD^AbGrKghTYpC)98tS))Np_OBkUZ(M^;!7JvAJ#XJX>0RHTEk zPHz-+)#WO=GBb0sxoTbQH9lmZK5U-N+8^3{2XxYoK@cxA5un*m!uY=E|APtx;a(BL3crwPh_| zli18RXWsJuc0a(>ah?A0Bz6|N&>48{_cJ<2cW}knq<7R(HydrAGzx3D_`GXzoO2vy zuTvS=NeR1vv9%HYf5YbM6#e8P5=1$>|BmMq`40l1|KGg^`>`&AJ=!GHMLPV7;B9Xz zZSkH2-Yb+&=A<+{mm%#kkSlD1wCl4<$eU05B?>h4ZdG5aN@>fyW`7jg=>SfZyGjJ^ zrdR*`+bfibp4C)S`g`94!-g6k|BqjD<497XHZ7?IpSGbPufD#%+f6Je^fJR>5SFR7USwJ3z;VeV*!A%*+IcJl}nOco6a97F6O|NNg>z_)%jfb zQjF>X@&^ijKIfQcBVu%E9Mx zV)R-136Yo)EQM_E@g8_uXI87h@2Pg^E)O3*#4p}pDXKpHLj(ngG$DOsV~%I1iNSzi zI0I;5Vc0LX6CdrK`fSX*oz1|Em;etAb&bIElsaRHet+G=C!uZN>|+v&VIwJ5?z){! znc*}@isHocFic^!WN*3DQD-7zYLapbzFtXc!)Q`}sog*Ce&1ghI!L1`iW60wQSQClQJ< zAO&`ltSv|vb0sQ@8g~1~q#t;p1JmjrY?l*+eOq=*Vp^&4VTmbEtT<|)NJj4We@%({ z6Jh6<8ZSw@nuq^+v3hW^&t~a*b!i=)?faa1LOdvK-f8nWl5COv#06g>e;{95JzC~1 zd-!2G&2FE~jdaGl)VoGYo{H`JeLv~?nN1l>O$PEV+TwunY(9cn4vfu+*8i8LraQ^*s_Hf1v9Zxu$)V@}WGO zk<9R)kK0P_@%q}_k^WFFPg z7Whyn{oA9i=j8s7naWd1;n_BT`~Nd7L@)|;5zUZKWL8up`pO)6nVFbdz%-2sN-RN- z$;18aj`3^jhkmLBK%)^*6nrQF21+36{t{U2H-m6D zPr6;x=ps{CuR|rdr4nk4HC~SS5ooV@ zV`F2@1W%xb=~e5<4#$10)Jh23ZE!YB%S zNgT4rg7_Lp91+EP&L;o9mel6>R9+E-J>?&8o}bTv=F2w|(LIib95X{tVh4l8#o1do zc_ZFaM#)BbFWl>0xYs3d zXT(_ei?VTUvnr~q#)?sOSZA+XRPT6j(IN~-LPVtCJ3QjGpxm$@Zq%D|aq@+nxdtM; zL_g2}$k}jOJ^BY-O}Cpya1Cb7ZdqXMeL@MS-7 zmn+lvHiK%UjoI_jhpTifq2ik8U^c4Tz9B+org~%QMAH8Nlu=um!J5?Am@?~2LJZ0U# zdEY@K|Me#~b!9*WiaZ-L3z0oo6Y^Ndx&BA=JbV&UWBQEs%EU=b<?Lj$3Z@gYDy4Tr7a2P^@6V2pr~4V~1uO|mnAS@@l z^Ws57FA!T@QDL;g^2f2ZP2TRf`9oE3m^4V-Q7dB!QkCU~NR!bM5cBT!MZwS}N&%@03lxXp?u#s@sbHIfZ zi#vE+T*;_peqFVN7Bw&$710nJfOxV8?=9nO>2q^)g&W#nNb<6N4G8fjaNDql^lU*> zrzE0a2}N7-Hf)^ur*K5e}guz2{wrRyD?qnf?v6poGjj2|YL6tso{hfZpV?y%fjkSk$mgs9#!R;` zR9AjOEG<SVg4L16d!Nqr3O6T<~P%?y>A14mq`W4Vymu>4m))7O7l(MU0%^ipgp0S(0Af8 zhiQo7{ELv1esQ=6+ue(ZVuNz#!DXt8$Ahs8Zf#O9=5L79hoB$fYpGAUO) z8w@CMo#Hd4%m6lSXYsY!@Z>DXm39TIsjeI(DMU7_IAe2HhM<8mFl)@_>QFIq-GsQ; z($5AMc_t~&ZYa$;OPD_1^jW@o6)gh_1?$>IOuK>vFa`EY>D^z2*(_98toEbN*1s>;$$h&Sr zjGD>DZG)hrCGRdwg^Q4yh`Go|3@*@q7|z`NUx@Kl-YBvG4J0kE?$xv4*}X1B!NtU_vfs z;p^Rm@i2Udrf-j585e@X#Ic_KAtjCD!vSY2S4gRYwTyK-6ZYe;3s?9w00$Rmd=+JI zZxW}DT|-b|IS`6}0?d5z+{6kF;d6)spg<*AS9wOcm${?CVFf3O!uuC1=3Q8R0Y0l) z;$Y5d3L3|1ov6rt#G|-9bZQ(zaHfnUMeKi|Q)`s^(XKC_?2;Q1+OM|plRtzBM?QmQ z?`Np^@aduwO^vIY^%e68XRf#8-C-0o;eiBU+&KySbl&Gnb^F^;`q6>|c#3MS#klYF*;9sOos9I2iJXi3>{1k83Np#{nzlF4 zer*+X-^Ke`2 z7V(4$ORtYiqk)^CwM!f6#`3iOQ=fsn7}?3nIzMySY_qwa7YiyUznp~#Z*tI zJwi`|F2wHWeqfu3FJ5L@&lryN(75^eWYO?>s_4qI+4byn3BWz^yx$T)MmYvgZEUKt zy}~8=#{noBUj=4xnO(8@O-OIMK8!gsvJKBL%`_N7(dq19I>q2cY&xF~HBcw$<2!Qf z{|E>ps0DFnv#!9k&UnyBba0h@_`{G)Osgzxw5<}(lcY-w?N|DyWdEpMZb0}@NlzZt zRUZ`)6gV2{MdR3U1}dS0vS8bZ4hZCC0w&$>(!npj$bfTC?;1h%@p*eWHF^GjCNiaH zYyd#s2(=-7Mem|D%uxB+0xbj7!zqVtb$imO)e#;>N1dh5T(A#1N0>?G;1nBmP~1;C z)~-Mu8!kg2r`WB_g3z61ys*%{Fe<=(@Mz>s8b?}Np2tUx0DOH=X3JxcKnkN2cu^z+ z#2DFC+pIxN$yf7x*3k-anm}-^W@QDFoqxt>eci|GHn+N9~9dn3)#_7||QimZUfRA`!?y%gHk!i;! zuq9$a7%zvyD{^m=#G(XU#QKS$pZ+rN=MAQ$WIva~#Jfv|UlESlEC3sGd-I{sKn)KU z;NLrGP0}YJpS>MUr#nV>m0MNtcq{utbS==bEuppX8m8?bIIXW&E)ZlbMq$!x2D_qU zX^Z9AMc~@yoCk>gz*7XzSd7PkhL?2J#!mLck!$%|**6 z6aliH(~*B{f)E|_@k^2-GMEU_3v)NDr?D?Y#9Xrisa*KYDh-eu zr^u-v{`*)?Poi)SofYq0F3)|o@vAsrncpjVND^WgIEK);z$8&yi@nOO-{rADVCZjI zdkYY=)!y1rLm)c3d3U8n*R-bW&TKfwP_M=S4BSnIbFg6!jNU7Jf{%LX-U zVHyiaU{<~fu{-1%97fcgfCHAO01+Ne^AXxBmo^*_Q*(=?0-qyBl=iS)MT5l)&)%?v zZ=p(l+EISn2HHo;e5NqeU6Ju*w;Z+Imj)YG_RE`D;R6PH*t|^kiFvKyud(>P>YT(h z!QER!-y%UPeygv|$9m4L22HNBvo6Hv$FvlDb+I0ASVzlh0L%L6e9F^}O)mb~Ls%CY zC@dq+Xu`tE)f;!sLp{=E9I+IyiD7;C%5imi!puU2Rjp$df_;g3T>lO3P zKG!GnX6pKB$U#a$yQ1b=MPv8nt`8U`341JDoz|YpqYASecr0L3c3yI~btC8JW86}q zckM<9XN*5uizm1LwidFNNV>bb!+yT3(aGf2)&%e}mvmrlET)6W%ANh1 z-rdAuz@^Ek#?DX6RZRdJ;ndF7gsh%z{_$+NB#Y0sKijUrA;7jDiZ7iTnZIHpJwh%T zE5G2bGd^q-EF7o6w!Tv#s1p44B*__iJeY^%JS{ww7LsYSvqGDAB23}~zx7Rh(_WgU zLDvOKM11)O5mUtO0p5sHD@jJM1UfU(ShpHr@C@f4a{-jqx+)(Q7#DLNEiftB<5}uS z?1`^i|EpP#K5z&YL{qs9F{`kFM2nV^g{>YN zws#vn>z;R!Vt8~2T}>m|)*M`o!ti%1K`86T_o{xK4aPnYljz3rLOj=E(c&mptMzhK ztTEa=$Rw)zqvOZg^@@l|=&_8!XB}t0L+FK`G8}c$udX*F_Kr<)!00~|qc z>pt-!1w{7{|2sylc1mS6uhelqz0|;K(5Gd7JHqGPg;4WH14d_*EA5TeTjq2gjR*Y} zn62)qt`gECl6H_)-eK4*ejeaS-j#eiwc!(mn$faE^0E}Q*}?oP8D5$fD>^BVMeNp% zFBa-Ai>ge)F$@18tB=L|)kJ|UOdR4w-Bf&j6gc3t#MD5GWX-a0lc444{Y-hEa;3vY z%on?qHCMT96Vkpp*;3x$c~!4IR?YXf?K_$`lErs%GGJ@NE%%ztZ}#71(H0{ORexp> z`W~8xiud41zPp0*cl9zi# zMC6#SN<6{vlARs6RU%bXyXqxtiYbjKfWj?pXQ}TouZ_$)y8t>(08S?;5N3<94!@dJ zDjybUj+$_c3bYnz485i(K7I2neb{cY_|L^WW(Lq&x_8`}^3I3aJ2uchiJC1fcg{Pt zibBdL3>MwcTmTswFaZY=nR(m8ZSot({;x3|b2c3Z>1MMC3`l@7p+x>e zhXXXK5eby`E@B6zSY zRzj-J!Q@65CTa&1{^ZypMK0!bxYRG#Zv894WyHJSmp-6}c`-GU zQT=d-+x3yXgF%2qdWF)h1+MaJvAho75G&yRb<}vh-><^d>ysjLp823vY=F-|)WVKk z0>S?7Y~{*GCP9*Gct&7k-pAkvE+1)PLNNJXowl=W%NEw0|he2R=|J(J)9~QUX!i^YLFBZN7q( z4E*GHhrzwXq%$Ohl5E-gx5y9kVOY&xGWW7;SN?G&qXkXx4gDls%SJ9Zk(m8B{gzH? zSqtQRUFhzjfGW@Z{GBrp&xP= zK!{(j1i!`zhpS!>sk0-0ug#~S5*jFaLk*c-K0Kg6#g|Zg1YFZS8Xz-KTf&eJ>y80P zuQwyo?=qYZoOC$+`tYtqtI_DmD5o-8Dq66*2TI@Rxv2J>oOonNx`Z0i*2Hg)TW|** zOpkd-Xy}D}osZzlYj7+hrgJP5Q{cCq4{FZs88R^<-Me?zMmaI8EZIF7f;zN~Sn7!h zQ9glV1uk@f9XE^K#-05p@g(oQ`Kkp?osb&!#sO9tlex-eE!1+xtFpk}h|&69wLTmT zU&b(J+@d2@8chG`pHmB6j+mxh5Y)cG+S-R`I@ z%8j_KH(#`(ApuL+IX48haA;G{fMVcv_hsWn!r+awyqQ$Ne$~PkRGQr0^O*YDKu1Lw zrYY^k$al&(^X~QbFfoeqdZ1a`hE}4u)WNk`fmu1bVAMoc_Q&Z~txsK{VseA4H%E*6 zrS(&ISIWDAWOfk|*sUMt8@!SS)HK0HkQ1;9jpw_2M5PP62Y=lLFe8D16Q|Dy+Lo$Y zR*w(0eF;>_vgf^_NUGiuxO|%_vYuM!Oz=kL@Mvq-~xp=#x#Z;xC2Z{!V>OQ>$_02bs z1eRxvibBLHR=Wy(nAe~I@C{LMCB;vqLyPI~pRvbOXz zwJg2tBs%SPa|IfD;@`*gcP@9=7}_3CP-&1H>~1pAA!tiR7@KnDr^ zfI1_;wzN-TLeSaW6e}vXyoGx+c^$M*z2)Q7sSILrjZYG{j+c}8H}3B`aL=KuH+RZd z$yEoJ_;~Qja<~|VV<63Kzm}|d;DbGXT`&W&Ovs~5E}ngV-NL`-Jbz1p{`1%WEYtt$ z12R8xiiwD5&rLN&xl8|9+W*z1{m*j#XF2~?k47^5^0mCv%!1LS@*#p>YD!v)h4NPe F{tJI7qU!(v diff --git a/docs/sqlflow/doc/design/nav.md b/docs/sqlflow/doc/design/nav.md deleted file mode 100644 index 73ab749..0000000 --- a/docs/sqlflow/doc/design/nav.md +++ /dev/null @@ -1,7 +0,0 @@ ---- -title: Design -has_children: true -permalink: /doc/design -nav_order: 2 ---- -How we design SQLFLow diff --git a/docs/sqlflow/doc/design/pipe.md b/docs/sqlflow/doc/design/pipe.md deleted file mode 100644 index dd63433..0000000 --- a/docs/sqlflow/doc/design/pipe.md +++ /dev/null @@ -1,96 +0,0 @@ ---- -parent: Design -nav_order: 4 ---- - -# Piping Responses - - -## Streaming Responses - -As described in the [overall design](doc/design.md), a SQLFlow job could be a standard or an extended SQL statemnt, where an extended SQL statement will be translated into a Python program. Therefore, each job might generate up to the following data streams: - -1. standard output, where each element is a line of text, -1. standard error, where each element is a line of text, -1. data rows, where the first element consists of fields name/types, and each of the rest is a row of data, -1. status, where the element could be *pending*, *failed*, and *succeeded*. - -To create good user experience, we need to pipe these responses from SQLFlow jobs to Jupyter Notebook in real-time. - - -## Stages in the Pipe - -The pipe that streams outputs from SQLFlow jobs to the Jupyter Notebook consists of the following stages: - -``` -Web browser - ↑ - | HTTP - ↓ -Jupyter Notebook server - ↑ - | ZeroMQ streams: Shell, IOPub, stdin, Controls, Heartbeat - ↓ -iPython kernel - ↑ - | IPython magic command framework - ↓ -SQLFlow magic command for Jupyter - ↑ - | gRPC - ↓ -SQLFlow server - ↑ - | Go channels - ↓ -SQLFlow job manager (Go functions) - ↑ - | IPC with Go's standard library - ↓ -SQLFlow jobs -``` - -In the above figure, from the SQLFlow magic command to the bottom layer are our work. - - -## Streaming - -We have two alternative ideas: multiple streams and a multiplexing stream. -We decided to use a multiplexing stream because we had a unsuccessful trial with the multiple streams idea: we make the job writes to various Go channels and forward each Go channel to a streaming gRPC call, as the following: - -### Multiple streams - -The above figure shows that there are multiple streams between the Jupyter Notebook server and Jupyter kernels. According to the [document](https://jupyter-client.readthedocs.io/en/stable/messaging.html), there are five: Shell, IOPub, stdin, Control, and Heartbeat. These streams are ZeroMQ streams. We don't use ZeroMQ, but we can take the idea of having multiple parallel streams in the pipe. - - -```protobuf -service SQLFlow { - rpc File(string sql) returns (int id) {} - - rpc ReadStdout(int id) returns (stream string) {} - rpc ReadStderr(int id) returns (stream string) {} - rpc ReadData(int id) returns (stream Row) {} - rpc ReadStatus(int id) returns (stream int) {} -} -``` - -However, we realized that if the user doesn't call any one of the `SQLFlow.Read...` call, there would be no forwarding from the Go channel to Jupyter, thus the job would block forever at writing. - -## A Multiplexing Stream - -Another idea is multiplexing all streams into one. For example, we can have only one ZeroMQ stream, where each element is a polymorphic type -- could be a text string or a data row. - -```protobuf -service SQLFlow { - rpc Run(string sql) returns (stream Response) {} -} - -// Only one of the following fields should be set. -message Response { - oneof record { - repeated string head = 1; // Column names. - repeated google.protobuf.Any row = 2; // Cells in a row. - string log = 3; // A line from stderr or stdout. - } -} -``` diff --git a/docs/sqlflow/doc/design/sql_parser.md b/docs/sqlflow/doc/design/sql_parser.md deleted file mode 100644 index 6e55dc2..0000000 --- a/docs/sqlflow/doc/design/sql_parser.md +++ /dev/null @@ -1,70 +0,0 @@ ---- -parent: Design -nav_order: 2 ---- - -# Extended SQL Parser Design - -This documentation explains the technical decision made in building a SQL -parser in Go. It is used to parsed the extended SELECT syntax of SQL that -integrates TensorFlow Estimators. - -## Related Work - -### Lexer and Parser Generator - -In 2001, when I was in graduate school, I defined an extended SQL syntax for -querying distributed relational databases, as part of the course project of -Distributed RDBMS by Prof. Li-zhu Zhou. I wrote the parser using [`bison` (a -modern version of `yacc`) and `flex` (a modern version of -`lex`)](http://dinosaur.compilertools.net/). `yacc` and `lex` generate C code; -`bison` and `flex` generate C++ code. However, this time, I'd use Go. - -I surveyed [`goyacc`](https://godoc.org/golang.org/x/tools/cmd/goyacc), a -standard Go tool. The usage is very similar to that of `yacc` and `bison`. -However, the Go toolchain doesn't provide a tool like `lex`/`flex`. - -Google revealed [`golex`](https://github.com/cznic/golex), which is out of -maintenance. - -The [Medium -post](https://medium.com/@mhamrah/lexing-with-ragel-and-parsing-with-yacc-using-go-81e50475f88f) -recommends [Ragel](http://www.colm.net/open-source/ragel/), which is a C++ -program and could generate Go lexer; however, it lacks documents. - -### Handwritten Lexer and Parser - -Some documents, including [this -one](https://hackthology.com/writing-a-lexer-in-go-with-lexmachine.html) -recommends handwriting lexers. However, it doesn't explain how to write the -parser. - -GoAcademy always provides high-quality tech blog posts. [This -one](https://blog.gopheracademy.com/advent-2014/parsers-lexers/) is from the -author of [InfluxDB](https://github.com/influxdata/influxdb). However, I -stopped at where it explains wrapping a SQL statement as a string by an -`io.Reader`, because it is obvious that we should keep the string as a string so -that that token strings could refer to the same memory storage of the SQL -statement. - -Following a link in the above GoAcademy post, I found Rob Pike's excellent talk -on how to write a lexer in Go in 2011. Many works after that change Rob's -implementation somehow but always lead to longer and less comprehensible -codebases. - -## The Choice - -Therefore, I wrote the lexer and parser both following Rob Pike's idea. After -few days work, I realized that: - -1. I should borrow the idea from Rob to represent SQL statements as strings, but - not `io.Reader` as other work do, -1. but no need to use channels and goroutines at all, and -1. it is technically intractable to write a SQL lexer/parser manually. - -So, I switched to write a lexer manually, and to generate the parser using -goyacc. During my work, I referred to [this -example](https://github.com/golang-samples/yacc/blob/master/simple/calc.y) and -the official [`yacc` -manual](https://www.epaperpress.com/lexandyacc/download/yacc.pdf) for details -about operator association and precedence. diff --git a/docs/sqlflow/doc/design/syntax.md b/docs/sqlflow/doc/design/syntax.md deleted file mode 100644 index 1763bf4..0000000 --- a/docs/sqlflow/doc/design/syntax.md +++ /dev/null @@ -1,165 +0,0 @@ ---- -parent: Design -nav_order: 0 ---- - -# SQLFlow: Design Doc - -## What is SQLFlow - -SQLFlow is a bridge that connects a SQL engine, for example, MySQL, Hive, SparkSQL, Oracle, or SQL Server, and TensorFlow and other machine learning toolkits. SQLFlow extends the SQL syntax to enable model training and inference. - -## Related Work - -We could write simple machine learning prediction (or scoring) algorithms in SQL using operators like [`DOT_PRODUCT`](https://thenewstack.io/sql-fans-can-now-develop-ml-applications/). However, this requires copy-n-pasting model parameters from the training program into SQL statements. - -Some proprietary SQL engines provide extensions to support machine learning. - -### Microsoft SQL Server - -Microsoft SQL Server has the [machine learning service](https://docs.microsoft.com/en-us/sql/advanced-analytics/tutorials/rtsql-create-a-predictive-model-r?view=sql-server-2017) that runs machine learning programs in R or Python as an external script: - -```sql -CREATE PROCEDURE generate_linear_model -AS -BEGIN - EXEC sp_execute_external_script - @language = N'R' - , @script = N'lrmodel <- rxLinMod(formula = distance ~ speed, data = CarsData); - trained_model <- data.frame(payload = as.raw(serialize(lrmodel, connection=NULL)));' - , @input_data_1 = N'SELECT [speed], [distance] FROM CarSpeed' - , @input_data_1_name = N'CarsData' - , @output_data_1_name = N'trained_model' - WITH RESULT SETS ((model varbinary(max))); -END; -``` - -A challenge to the users is that they need to know not only SQL but also R or Python, and they must be capable of writing machine learning programs in R or Python. - -### Teradata SQL for DL - -Teradata also provides a [RESTful service](https://www.linkedin.com/pulse/sql-deep-learning-sql-dl-omri-shiv), which is callable from the extended SQL SELECT syntax. - -```sql -SELECT * FROM deep_learning_scorer( - ON (SELECT * FROM cc_data LIMIT 100) - URL('http://localhost:8000/api/v1/request') - ModelName('cc') - ModelVersion('1') - RequestType('predict') - columns('v1', 'v2', ..., 'amount') -) -``` - -The above syntax couples the deployment of the service (the URL in the above SQL statement) with the algorithm. - -### Google BigQuery - -Google [BigQuery](https://cloud.google.com/bigquery/docs/bigqueryml-intro) enables machine learning in SQL by introducing the `CREATE MODEL` statement. - -```sql -CREATE MODEL dataset.model_name - OPTIONS(model_type='linear_reg', input_label_cols=['input_label']) -AS SELECT * FROM input_table; -``` - -Currently, BigQuery only supports two simple models: linear regression and logistic regression. - -## Design Goal - -None of the above meets our requirement. - -First of all, we want to build an open source software. Also, we want it to be extensible: - -- We want it extensible to many SQL engines, instead of targeting any one of them. Therefore, we don't want to build our syntax extension on top of user-defined functions (UDFs); otherwise, we'd have to implement them for each SQL engine. - -- We want the system extensible to support sophisticated machine learning models and toolkits, including TensorFlow for deep learning and [xgboost](https://github.com/dmlc/xgboost) for trees. - -Another challenge is that we want SQLFlow to be flexible enough to configure and run cutting-edge algorithms, including specifying [feature crosses](https://www.tensorflow.org/api_docs/python/tf/feature_column/crossed_column). At the same time, we want SQLFlow easy to learn -- at least, no Python or R code embedded in the SQL statements, and integrate hyperparameter estimation. - -We understand that a key to address the above challenges is the syntax of the SQL extension. To craft a highly-effective and easy-to-learn syntax, we need user feedback and fast iteration. Therefore, we'd start from a prototype that supports only MySQL and TensorFlow. We plan to support more SQL engines and machine learning toolkits later. - -## Design Decisions - -As the beginning of the iteration, we propose an extension to the SQL SELECT statement. We are not going a new statement way like that BigQuery provides `CREATE MODEL`, because we want to maintain a loose couple between our system and the underlying SQL engine, and we cannot create the new data type for the SQL engine, like `CREATE MODEL` requires. - -We highly appreciate the work of [TensorFlow Estimator](https://www.tensorflow.org/guide/estimators), a high-level API for deep learning. The basic idea behind Estimator is to implement each deep learning model, and related training/testing/evaluating algorithms as a Python class derived from `tf.estimator.Estimator`. As we want to keep our SQL syntax simple, we would make the system extensible by calling estimators contributed by machine learning experts and written in Python. - -The SQL syntax must allow users to set Estimator attributes (parameters of the Python class' constructor, and those of `train`, `evaluate`, or `predict`). Users can choose to use default values. We have a plan to integrate our hyperparameter estimation research into the system to optimize the default values. - -Though estimators derived from `tf.estimator.Estimator` run algorithms as TensorFlow graphs; SQLFlow doesn't restrict that the underlying machine learning toolkit has to be TensorFlow. Indeed, as long as an estimator provides methods of `train`, `evaluate`, and `predict`, SQLFlow doesn't care if it calls TensorFlow or xgboost. Precisely, what SQLFlow expect is an interface like the following: - -```python -class AnEstimatorClass: - __init__(self, **kwargs) - train(self, **kwargs) - evaluate(self, **kwargs) - predict(self, **kwargs) -``` - -We also want to reuse the [feature columns API](https://www.tensorflow.org/guide/feature_columns) from Estimator, which allows users to columns of tables in a SQL engine to features to the model. - - -## Extended SQL Syntax - -Again, just as the beginning of the iteration, we propose the syntax for training as - -```sql -SELECT * FROM kaggle_credit_fraud_training_data -LIMIT 1000 -TRAIN DNNClassifier /* a pre-defined TensorFlow estimator, tf.estimator.DNNClassifier */ -WITH layers=[100, 200], /* a parameter of the Estimator class constructor */ - train.batch_size = 8 /* a parameter of the Estimator.train method */ -COLUMN *, /* all columns as raw features */ - cross(v1, v9, v28) /* plus a derived (crossed) column */ -LABEL class -INTO sqlflow_models.my_model_table; /* saves trained model parameters and features into a table */ -``` - -We see the redundancy of `*` in two clauses: `SELECT` and `COLUMN`. The following alternative can avoid the redundancy, but cannot specify the label. - -```sql -SELECT * /* raw features or the label? */ - corss(v1, v9, v28) /* derived featuers */ -FROM kaggle_credit_fraud_training_data -``` - -Please be aware that we save the trained models into tables, instead of a variable maintained by the underlying SQL engine. To invent a new variable type to hold trained models, we'd make our system tightly integrated with the SQL engine, and harms the extensibility to other engines. - -The result table should include the following information: - -1. The estimator name, e.g., `DNNClassifier` in this case. -1. Estimator attributes, e.g., `layer` and `train.batch_size`. -1. The feature mapping, e.g., `*` and `cross(v1, v9, v28)`. - -Similarly, to infer the class (fraud or regular), we could - -```sql -SELECT * FROM kaggle_credit_fraud_development_data -PREDICT kaggle_credit_fraud_development_data.class -USING sqlflow_models.my_model_table; -``` - -## System Architecture - -### A Conceptual Overview - -In the prototype, we use the following architecture: - -``` -SQL statement -> our SQL parser --standard SQL-> MySQL - \-extended SQL-> code generator -> execution engine -``` - -In the prototype, the code generator generates a Python program that trains or predicts. In either case, - -1. it retrieves the data from MySQL via [MySQL Connector Python API](https://dev.mysql.com/downloads/connector/python/), -1. optionally, retrieves the model from MySQL, -1. trains the model or predicts using the trained model by calling the user specified TensorFlow estimator, -1. and writes the trained model or prediction results into a table. - -### Working with Jupyter Notebook and Kubernetes - -The following figures shows the system components and their runtime environment. The left part shows how to run the system on a PC/laptop, the right part shows how to run it on a Kubernetes cluster. - -![](doc/figures/sqlflow-arch.png) diff --git a/docs/sqlflow/doc/design/walkthrough.md b/docs/sqlflow/doc/design/walkthrough.md deleted file mode 100644 index f297ca1..0000000 --- a/docs/sqlflow/doc/design/walkthrough.md +++ /dev/null @@ -1,76 +0,0 @@ ---- -parent: Design -nav_order: 1 ---- - -# SQLFlow: Code Walkthrough - -## User Experience - -SQLFlow allows users to write SQL programs with extended syntax in Jupyter Notebook or a command-line tool. - -The following SQL statements train a TensorFlow model named `DNNClassifier`, which is a Python class derived from [`tf.estimator.Estimator`](https://www.tensorflow.org/api_docs/python/tf/estimator/Estimator): - -```sql - -SELECT * FROM a_table TRAIN DNNClassifier WITH learning_rate=0.01 INTO sqlflow_models.my_model; -``` - -And the following statement uses the trained model for prediction. - -```sql -SELECT * FROM b_table PREDICT b_table.predicted_label USING sqlflow_models.my_model; -``` - -Please be aware that the part in the above statements before the extended keyword TRAIN and PREDICT is a standard SQL statement. This feature simplifies the implementation of the SQLFlow system. - -## System Implementation - -If a SQL statement is of the standard syntax, SQLFlow throws it to the SQL engine and redirects the output to the user; otherwise, SQLFlow translates the statement of extended syntax into a Python program. Currently, it generates a program that throws the standard-syntax part of SELECT to MySQL, reads the results in the train-loop of a TensorFlow program. We will talk about how to extend SQLFlow to connect more SQL engines like Oracle, Hive, and SparkSQL, and generates more types of machine learning programs that calls distributed TensorFlow, PyTorch, and xgboost later. Before that, let us explain the system components. - -### SQLFlow as a gRPC Server - -SQLFlow is a gRPC server, which can connect with multiple clients. A typical client is [pysqlflow](https://github.com/sql-machine-learning/pysqlflow), the SQLFlow plugin for Jupyter Notebook server. Another once is a text-based client [/cmd/sqlflowserver/main.go](/cmd/sqlflowserver/main.go). - -``` -Jupyter Notebook ---(SQL statements)--> SQLFlow gRPC server -(SQLFlow magic command) <--(a stream of messages)-- -``` - -The protobuf definition of the gRPC service is at [/server/proto/sqlflow.proto](/server/proto/sqlflow.proto). The return of the method `SQLFlow.Run` is a stream of `Reponse`s, where each represents either a table header, a row, or a log message. The header and rows are usually from a standard SQL statement, for example, SELECT or DESCRIBE, and the log messages are usually from the running of a generated Python program. - -### SQLFlow in the gRPC Server - -Once the SQLFlow server receives a batch of SQL statements via a gRPC call, it runs the following steps for each statement: - -1. the [parser](/sql/sql.y) to generate parsing result, -2. the [verifier](/sql/verifier.go) to verify the semantics given the parsing result, -3. the [code generator](/sql/codegen.go) to generate a Python program, or the *submitter*, from the parsing result, -4. the [executor](/sql/executor.go) that runs the submitter locally. - -Step 3. and 4. are only for a SQL statement of extended syntax; otherwise, SQLFlow server proxies the standard-syntax statement to the SQL engine. - -The executor calls Go's [standard package](https://godoc.org/os/exec) that captures the stdout and stderr from the submitter process and passing the result back to the gRPC client. Therefore, it is the responsibility of the submitter to print log messages to its stderr and stdout. - -### Minimal Viable Product - -In the minimal viable product (MVP) of SQLFlow, the code generator generates a Python program consists of two parts: - -1. throw the standard SELECT part in the extended-syntax statement to MySQL via ODBC, and -1. a loop that reads outputs from the run of the SELECT statement and trains the model (or, using a trained model to predict). - -The training part calls TensorFlow to update the parameters of the model specified in the TRAIN clause. - -### Extensibility - -By writing more code generators, we could extend SQLFlow to support more SQL engines, e.g., Hive and Oracle, and use machine learning toolkits, e.g., PyTorch and xgboost, in addition to TensorFlow, on various computing platforms. You are welcome to add more code generators such as - -- `codegen_distributed_tf.go` to generate a submitter program similar to the MVP but runs a distributed TensorFlow training job. -- `codegen_kubernetes_tf.go` to launch a distributed TensorFlow job on a Kubernetes cluster, other than running locally, in the same container as where SQLFlow gRPC server resides. -- `codegen_gcloud_pytorch.go` to launch a submitter that calls PyTorch instead of TensorFlow for training on the Google Cloud. - -### Job Management - -The current design of the gRPC interface assumes that the connection between the client, e.g., the Jupyter Notebook, and the SQLFlow server keeps alive during the running of the training program. This assumption is reasonable because even if the user closes her/his Web browser and disconnect to the Jupyter Notebook server, the connection between Jupyter to SQLFlow server might keep alive. However, this might not be robust enough if the Jupyter Notebook server runs on a user's laptop and gets killed. In such a case, the gRPC server cannot stream the messages back to the client and would cause the failure of the submitter. - -A solution is to change the gRPC interface of SQLFlow server to have a method that files a job and returns immediately, and another method to get a batch of recent messages given a job ID. We will make a design for that soon. diff --git a/docs/sqlflow/doc/quickstart.html b/docs/sqlflow/doc/quickstart.html index 994be44..2c6e170 100644 --- a/docs/sqlflow/doc/quickstart.html +++ b/docs/sqlflow/doc/quickstart.html @@ -1,416 +1,406 @@ - - - - - - Quick start - SQLFLOW - - + + + + Quick start - SQLFLOW + + + + + + + - - - - - - + -

    -
    - SQLFLOW - -
    - - +
    +
    + SQLFLOW + +
    + + +
    +
    +

    This site uses Just the Docs, a documentation theme for Jekyll.

    +
    +
    +
    +
    +
    + +
    +
    + + Search
    -
    -

    This site uses Just the Docs, a documentation theme for Jekyll.

    -
    +
    +
    + + + +
    -
    -
    -
    - -
    -
    - - Search -
    -
    -
    - - - - -
    -
    -
    - - - -
    - - - - -

    Quick start

    - -

    SQLFlow is currently under active development. For those who are interested in trying it out, we have provided several demos. Play around with it. Any bug report and issue are welcomed. :)

    - -

    Setup

    - -
      -
    1. Install Docker.
      2. -
    2. Set up a MySQL server following example/datasets/README.md.
      3. -
    3. Pull the latest SQLFlow Docker image: docker pull sqlflow/sqlflow:latest.
      4. -
    - -

    Demo 1: Jupyter Notebook

    - -
      -
    1. -

      Start a Docker container that runs sqlflowserver and Jupyter Notebook. If you are using Docker for Linux, please change host.docker.internal:3306 to localhost:3306.

      - -
      - 
      docker run --rm -it -p 8888:8888 sqlflow/sqlflow:latest \
+
      +
      + + + +
      + + + + +

      Quick start

      + +

      SQLFlow is currently under active development. For those who are interested in trying +it out, we have provided several demos. Play around with it. Any bug report and +issue are welcomed. :)

      + +

      Setup

      + +
        +
      1. Install Docker.
        2. +
      2. Set up a MySQL server following example/datasets/README.md.
        3. +
      3. Pull the latest SQLFlow Docker image: docker pull sqlflow/sqlflow:latest.
        4. +
      + +

      Demo 1: Jupyter Notebook

      + +
        +
      1. +

        Start a Docker container that runs sqlflowserver and Jupyter Notebook. If you are +using Docker for Linux, please change host.docker.internal:3306 to localhost:3306.

        + + 
        docker run --rm -it -p 8888:8888 sqlflow/sqlflow:latest \
 bash -c "sqlflowserver --db_user root --db_password root --db_address host.docker.internal:3306 &
 SQLFLOW_SERVER=localhost:50051 jupyter notebook --ip=0.0.0.0 --port=8888 --allow-root"
-
        -
        +
      -

      If you are using Docker for Mac, please be aware the option --db_address host.docker.internal:3306 where - host.docker.internal translates to the host ip address as recommended here.

      +

      If you are using Docker for Mac, please be aware the option --db_address host.docker.internal:3306 where +host.docker.internal translates to the host ip address as recommended here.

      -

      If you are running MySQL on remote, please be aware that MySQL only allows connections from localhost by default. Fix can be found here.

      -
      2. -
    2. -

      Open a Web browser and direct to localhost:8888 and input the token. Then you can create notebooks. In a cell, you should be able to type

      +

      If you are running MySQL on remote, please be aware that MySQL only allows connections from localhost +by default. Fix can be found here.

      +
      3. +
    3. +

      Open a Web browser and direct to localhost:8888 and input the token. Then you +can create notebooks. In a cell, you should be able to type

      -
      - 
      %%sqlflow
+    
      %%sqlflow
 select 1
-
      
-                            
      
-
      4. -
    4. -

      Explore more examples at example.ipynb

      -
      5. -
    +
    + +
  • +

    Explore more examples at example.ipynb

    +
    • + -

    Demo 2: Command Line Prompt

    +

    Demo 2: Command Line Prompt

    -

    Start a Docker container that runs SQLFlow command line prompt. If you are using Docker for Linux, please change host.docker.internal:3306 to localhost:3306.

    +

    Start a Docker container that runs SQLFlow command line prompt. If you are using +Docker for Linux, please change host.docker.internal:3306 to localhost:3306.

    -
    - 
    docker run -it --rm --net=host sqlflow/sqlflow:latest demo \
+
    docker run -it --rm --net=host sqlflow/sqlflow:latest demo \
 --db_user root --db_password root --db_address host.docker.internal:3306
-
    
-                    
    
+
    -

    You should be able to see the following prompt.

    +

    You should be able to see the following prompt.

    -
    - 
    sqlflow>
-
    -
    +
    sqlflow>
+
    -

    Example

    +

    Example

    -
      -
    • Select data -
      - 
      sqlflow> select * from iris.train limit 2;
+
        
+  
      • Select data
+    
        sqlflow> select * from iris.train limit 2;
 -----------------------------
 +--------------+-------------+--------------+-------------+-------+
 | SEPAL LENGTH | SEPAL WIDTH | PETAL LENGTH | PETAL WIDTH | CLASS |
@@ -418,12 +408,10 @@ 
        Example
        
 |          6.4 |         2.8 |          5.6 |         2.2 |     2 |
 |            5 |         2.3 |          3.3 |           1 |     1 |
 +--------------+-------------+--------------+-------------+-------+
-
        
-                            
        
-                        
        • 
-                        
      • Train a Tensorflow DNNClassifier
-                            
        
-                                
        sqlflow> SELECT *
+
            
        
+  
        • 
+  
      • Train a Tensorflow DNNClassifier
+    
        sqlflow> SELECT *
 FROM iris.train
 TRAIN DNNClassifier
 WITH n_classes = 3, hidden_units = [10, 20]
@@ -434,36 +422,30 @@ 
        Example
        
 ...
 Training set accuracy: 0.96721
 Done training
-
        
-                            
        
-                        
        • 
-                        
      • Prediction using a trained model
-                            
        
-                                
        sqlflow> SELECT *
+
            
        
+  
        • 
+  
      • Prediction using a trained model
+    
        sqlflow> SELECT *
 FROM iris.test
 predict iris.predict.class
 USING sqlflow_models.my_dnn_model;
-
        
-                            
        
-                        
        • 
-                        
      • Checkout prediction result
-                            
        
-                                
        sqlflow> select * from iris.predict limit 10;
-
        
-                            
        
-                        
        • 
-                    
      
+
      +
      • +
    • Checkout prediction result + 
      sqlflow> select * from iris.predict limit 10;
+
      +
      • +
    - -
    -
    +
    +
    +
    - - \ No newline at end of file + diff --git a/docs/sqlflow/doc/quickstart.md b/docs/sqlflow/doc/quickstart.md deleted file mode 100644 index 75a00a6..0000000 --- a/docs/sqlflow/doc/quickstart.md +++ /dev/null @@ -1,93 +0,0 @@ -# Quick start - -SQLFlow is currently under active development. For those who are interested in trying -it out, we have provided several demos. Play around with it. Any bug report and -issue are welcomed. :) - -## Setup - -1. Install [Docker](https://docs.docker.com/install/). -1. Set up a MySQL server following [example/datasets/README.md](/example/datasets/README.md). -1. Pull the latest SQLFlow Docker image: `docker pull sqlflow/sqlflow:latest`. - -## Demo 1: Jupyter Notebook - -1. Start a Docker container that runs sqlflowserver and Jupyter Notebook. If you are - using Docker for Linux, please change `host.docker.internal:3306` to `localhost:3306`. - - ``` - docker run --rm -it -p 8888:8888 sqlflow/sqlflow:latest \ - bash -c "sqlflowserver --db_user root --db_password root --db_address host.docker.internal:3306 & - SQLFLOW_SERVER=localhost:50051 jupyter notebook --ip=0.0.0.0 --port=8888 --allow-root" - ``` - - If you are using Docker for Mac, please be aware the option `--db_address host.docker.internal:3306` where - `host.docker.internal` translates to the host ip address as recommended [here](https://docs.docker.com/docker-for-mac/networking/). - - If you are running MySQL on remote, please be aware that MySQL only allows connections from localhost - by default. Fix can be found [here](https://stackoverflow.com/questions/14779104/how-to-allow-remote-connection-to-mysql). - -1. Open a Web browser and direct to `localhost:8888` and input the token. Then you -can create notebooks. In a cell, you should be able to type - - ``` - %%sqlflow - select 1 - ``` - -1. Explore more examples at [example.ipynb](/example/jupyter/example.ipynb) - -## Demo 2: Command Line Prompt - -Start a Docker container that runs SQLFlow command line prompt. If you are using -Docker for Linux, please change `host.docker.internal:3306` to `localhost:3306`. - -``` -docker run -it --rm --net=host sqlflow/sqlflow:latest demo \ ---db_user root --db_password root --db_address host.docker.internal:3306 -``` - -You should be able to see the following prompt. - -``` -sqlflow> -``` - -### Example - -- Select data -```sql -sqlflow> select * from iris.train limit 2; ------------------------------ -+--------------+-------------+--------------+-------------+-------+ -| SEPAL LENGTH | SEPAL WIDTH | PETAL LENGTH | PETAL WIDTH | CLASS | -+--------------+-------------+--------------+-------------+-------+ -| 6.4 | 2.8 | 5.6 | 2.2 | 2 | -| 5 | 2.3 | 3.3 | 1 | 1 | -+--------------+-------------+--------------+-------------+-------+ -``` -- Train a Tensorflow [DNNClassifier](https://www.tensorflow.org/api_docs/python/tf/estimator/DNNClassifier) -```sql -sqlflow> SELECT * -FROM iris.train -TRAIN DNNClassifier -WITH n_classes = 3, hidden_units = [10, 20] -COLUMN sepal_length, sepal_width, petal_length, petal_width -LABEL class -INTO sqlflow_models.my_dnn_model; ------------------------------ -... -Training set accuracy: 0.96721 -Done training -``` -- Prediction using a trained model -```sql -sqlflow> SELECT * -FROM iris.test -predict iris.predict.class -USING sqlflow_models.my_dnn_model; -``` -- Checkout prediction result -```sql -sqlflow> select * from iris.predict limit 10; -``` \ No newline at end of file From c8ff5279a76dffc0ab3bf7360524d66e0142ff92 Mon Sep 17 00:00:00 2001 From: mxz96102 Date: Sat, 20 Apr 2019 11:42:42 +0800 Subject: [PATCH 9/9] Add /pysqlflow/index.html --- docs/feed.xml | 2 +- docs/pysqlflow/index.html | 12 ++++++++++++ docs/sitemap.xml | 4 ++++ pysqlflow/index.html | 12 ++++++++++++ 4 files changed, 29 insertions(+), 1 deletion(-) create mode 100644 docs/pysqlflow/index.html create mode 100644 pysqlflow/index.html diff --git a/docs/feed.xml b/docs/feed.xml index 963dd98..842fe89 100644 --- a/docs/feed.xml +++ b/docs/feed.xml @@ -1 +1 @@ -Jekyll2019-04-20T11:30:34+08:00/feed.xmlSQLFLOW \ No newline at end of file +Jekyll2019-04-20T11:40:20+08:00/feed.xmlSQLFLOW \ No newline at end of file diff --git a/docs/pysqlflow/index.html b/docs/pysqlflow/index.html new file mode 100644 index 0000000..0eaa06b --- /dev/null +++ b/docs/pysqlflow/index.html @@ -0,0 +1,12 @@ + + + + + + + +

    Redirecting to https://github.com/sql-machine-learning/pysqlflow

    + + \ No newline at end of file diff --git a/docs/sitemap.xml b/docs/sitemap.xml index 273a020..39003c7 100644 --- a/docs/sitemap.xml +++ b/docs/sitemap.xml @@ -100,4 +100,8 @@ /gohive/vendor/git.apache.org/thrift.git/lib/go/thrift/ 2019-04-20T11:08:15+08:00 + +/pysqlflow/ +2019-04-20T11:40:14+08:00 + diff --git a/pysqlflow/index.html b/pysqlflow/index.html new file mode 100644 index 0000000..0eaa06b --- /dev/null +++ b/pysqlflow/index.html @@ -0,0 +1,12 @@ + + + + + + + +

    Redirecting to https://github.com/sql-machine-learning/pysqlflow

    + + \ No newline at end of file