15b training update

[bubblesld]

I train the 15b weight on 40b self-games from time to time. I will update the weight here when it beats the previous best in my test.

2019/2/3
edb61bc2
trained up to v202
60:40 over f438268e

2019/2/25
0a963117
trained up to v204
104:96 over edb61bc2

2019/5/7
da045f91
trained up to v224
227:173 over edb61bc2
211:189 in the official test.

2019/5/12
9006c708
trained up to v224 with elfv2
224:176 over da045f91
226:188 in the official test.

2019/5/17
3d7769e9
trained up to v225
223:177 over 9006c708
154:159 in the official test.

2019/5/23
40ea1748
trained up to v226
220:180 over 9006c708
209:190 in the official test.

2019/5/24
99a8d3a3
trained up to v226 (later version)
226:174 over 9006c708
115:132 in the official test.

2019/5/31
eef9843d
trained up to v226 (later version)
218:182 over 9006c708
190:176 in the official test.

2019/7/2
c9fd870f
trained up to v229
226:174 over 9006c708
233:189 in the official test.

2019/7/4
64e1f2ca
trained up to v229 (later version)
222:178 over c9fd870f
184:174 in the official test

2019/8/8
c11bc844
trained up to v234
235:165 over c9fd870f
232:185 in the official test

2019/8/9
5ee55afe
trained up to v234 (later version)
223:177 over c11bc844
133:148 in the official test

2019/9/1
53a5fe93
trained up to v238
223:177 over c11bc844
255:145 in the official test

2019/9/29
3d056bf4
trained up to v244
221:179 over 53a5fe93
137:143 in the official test

2019/10/3
2be1fc26
trained up to v245
220:180 over 53a5fe93
176:172 in the official test

2019/10/14
9c948dc4
trained up to v245 (another run)
216:184 over 53a5fe93
119:137 in the official test

2019/10/15
0d2694c5
trained up to v245 (later version)
229:171 over 53a5fe93
232:189 in the official test

2019/11/7
9273d548
trained up to v249
223:177 over 0d2694c5
217:192 in the official test

2019/11/10
a4b58a91
trained up to v249 (with more self-play games from v249)
219:181 over 0d2694c5
230:192 in the official test (220:180 at game 400).

2020/1/1
b4d5e8b6
trained up to v257
216:184 over a4b58a91
183:176 in the official test

2020/3/15
006eff16
trained up to v264
220:180 over a4b58a91
180:171 in the official test

2020/3/20
0c41abff
trained up to v266
220:180 over a4b58a91
183:178 in the official test

2020/3/21
da46e3b9
trained up to v266 (later version)
221:179 over a4b58a91
128:138 in the official test

2020/3/23
e912b007
trained up to v266 (later version)
218:182 over a4b58a91
169:166 in the official test

2020/3/28
ad865426
trained up to v266 (policy x4)
216:184 over a4b58a91
231:191 in the official test

2020/4/5
f5010dbd
trained up to v270
225:175 over a4b58a91
217:197 in the official test

2020/4/9
3353f154
trained up to v270 (later version)
227:173 over a4b58a91
202:179 in the official test

b7ccd6b4
trained up to v270 (later version)
221:179 over a4b58a91
176:174 in the official test

2020/4/10
0c4ade79
trained up to v270 (later version)
235:165 over a4b58a91
230:170 in the official test

[l1t1]

nice job, could you arrange a test match on http://zero.sjeng.org/? @roy7

[Friday9i]

It seems to be approximately on par vs LZ157 with half visits: I tested just a 3 games at 100 vs 200 visits and 1 game at 1000 vs 2000 visits, and they were all very balanced for a long time (which is probably more indicative of the balanced result that the brut score of 3 to 1 for this new net).
An official test against LZ157 and against the best @bjiyxo 15b net (f438268e) would be nice!

[wonderingabout]

i dont think its needed to test against lz157, we already know its stronger
just testing against the strongest 15b f438 is enough i think

[arondes]

Could anyone help put it in Baidu pan? Dropbox is banned in China.

[wonderingabout]

@arondes

@pangafu from leela master uses google drive, see : https://github.com/pangafu/LeelaMasterWeight

try google drive i suggest

[l1t1]

google also is band in cn.
https://userscloud.com can visit

[alreadydone]

https://pan.baidu.com/s/1Hx7l58DL5-yZpE7WAvZ-bg

[teraStars]

Thanks, at limited time ,It is stronger than LZ205

[l1t1]

10 days passed since last one, do you have a stronger weight?

[bubblesld]

0a963117
trained up to v204
104:96 over 93a5281b
probably only about the same strength.

[wonderingabout]

starting from lz 204, lz stops the 3-3 at move 3 (a lot less), and prefers to adjacent 4-4 at move 3 instead, like in normal human games

so i am looking forward to next 15b versions trained on lz 204 and higher
@bubblesld @bjiyxo

[l1t1]

@alreadydone pls help to upload to pan.baidu

[barrtgt]

@l1t1 is there another site that doesn't need cell phone?

[l1t1]

@barrtgt https://userscloud.com/
thanks

[l1t1]

I modified the minigo script (https://raw.githubusercontent.com/tensorflow/minigo/master/sgf/opening_stats.sh)

#!/bin/bash

set -e

DIR=$1
WORKDIR=`pwd`

cd $DIR

echo 'Total games'
ls -Rl * | grep .sgf | wc -l

echo 'Games with start at 5-5'
grep -l 'gz];B\[ee' ./*.sgf | wc -l

echo 'Games with start at 4-5'
R=`grep -l 'gz];B\[de' ./*.sgf | wc -l`
R=`expr $R + $(grep -l 'gz];B\[ed' ./*.sgf | wc -l)`
R=`expr $R + $(grep -l 'gz];B\[fe' ./*.sgf | wc -l)`
R=`expr $R + $(grep -l 'gz];B\[ef' ./*.sgf | wc -l)`
echo $R

echo 'Games with start at 4-4'
R=`grep -l 'gz];B\[dd' ./*.sgf | wc -l`
R=`expr $R + $(grep -l 'gz];B\[fd' ./*.sgf | wc -l)`
R=`expr $R + $(grep -l 'gz];B\[ff' ./*.sgf | wc -l)`
R=`expr $R + $(grep -l 'gz];B\[df' ./*.sgf | wc -l)`
echo $R

cd $WORKDIR

and use wget -r --limit-rate=300k -A.sgf http://zero.sjeng.org/self-plays to get sgf files
the result is

$ ./opening_stats.sh .
Total games
400
Games with start at 5-5
0
Games with start at 4-5
1
Games with start at 4-4
74

[l1t1]

if use code to rotate the board,

#include<cstdio>
#include<cstring>
char a[4000]="x";
char b[4000]="x";
char minb[4000]="x";
int sz=19; //19
char t[3]="";
char fn[100]="";
int main(int argc, char* argv[])
{
    FILE *fi,*fo;
    fi=fopen(argv[1],"rb");
    int c=fread(a,sizeof(char),3999,fi);
    memcpy(minb,a,c);
    fclose(fi);
    char x,y,x2,y2;
    for(int r=1; r<=7; r++) //find the minimal
    {
        int j=0; //b pos
        for(int i=0; i<c; i++)
        {
            if(a[i]==';' && (a[i+1]=='B'|| a[i+1]=='W') && a[i+2]=='[')
            {
                x=a[i+3],y=a[i+4];
                memcpy(b+j,a+i,3),j+=3;
                t[0]=x,t[1]=y;

                switch(r)
                {
                case 1: //up-down
                    x2=x,y2=sz-(y-'a')+'a'-1; //ok
                    break;
                case 2: //left-right
                    x2=sz-(x-'a')+'a'-1,y2=y; //ok
                    break;
                case 3:  //rotate a-clock 90
                    x2=y,y2=sz-(x-'a')+'a'-1;
                    break;
                case 4:  //rotate a-clock 180
                    x2=sz-(x-'a')+'a'-1,y2=sz-(y-'a')+'a'-1;
                    break;
                case 5:  //rotate clock 90
                    x2=sz-(y-'a')+'a'-1,y2=x;
                    break;
                case 6:  //rotate by line 135
                    x2=y,y2=x;
                    break;
                case 7:  //rotate by line 45
                    x2=sz-(y-'a')+'a'-1,y2=sz-(x-'a')+'a'-1;
                    break;
                }
                if(t[0]<'t')
                    t[0]=x2,t[1]=y2;

                memcpy(b+j,t,2),j+=2;

                i+=5;
                //memcpy(b+j,a+i,1),j+=1;
            }
            //else
                memcpy(b+j,a+i,1),j+=1;
        }
        if(strcmp(b,minb)<0)
            memcpy(minb,b,c);
    }
    sprintf(fn,"%s_min.sgf",argv[1]);
    fo=fopen(fn,"wb");
    fwrite(minb,sizeof(char),c,fo);
    fclose(fo);
    printf("%d\n",c);
}

the result

[sgf]$ g++ -O2 rtsgf4.cpp -o rtsgf
[sgf]$ ll rt*
-rwxr-xr-x 1 oracle dba 20771 Feb 26 08:51 rtsgf
-rw-r--r-- 1 oracle dba  2051 Feb 26 08:51 rtsgf4.cpp
[sgf]$ for f in rt*
> do
>   echo $f
> done
rtsgf
rtsgf4.cpp
[sgf]$ 

for f in *sgf
do
  ./rtsgf $f
done

[sgf]$ mkdir m
[sgf]$ mv *min.sgf m
[sgf]$ ./opening_stats.sh m
Total games
401
Games with start at 5-5
0
Games with start at 4-5
9
Games with start at 4-4
274

[alreadydone]

https://userscloud.com/t32auy11n3j0

[l1t1]

thanks @alreadydone

[l1t1]

i use sgfs of lz206 vs lz205 (http://zero.sjeng.org/match-games/5c754b51f06758029e912f22)

[ sgf]$ ./opst.sh n
Total games
414
Games with start at 4-4
360

[l1t1]

http://www.yss-aya.com/cgos/19x19/cross/lz15b0a96_1050ti.html very strong

[TFiFiE]

I'm going to bump this to second the request for test matches for these new 15-blocks. Apart from @gcp and @roy7, who else has the power to schedule them?

[roy7]

I think it's just me, gcp, and bjiyxo. Can you let me know which specific network(s) you want tested and vs what hashes? Thanks.

[TFiFiE]

I'd say test both of bubblesld's networks (93a5281b and 0a963117) against f438268e.

[hwj-111]

I put 0a96 on my best pc(2x1070+1x1060), lz15b0a96_3ghwj, perform better than all 40b networks using the same hardware on CGOS.

[l1t1]

who knows the config of Rnz-3.2.7-d3gp on cgos

[roy7]

@TFiFiE Tests scheduled.

[l1t1]

why the hash code changed?

[l1t1]

-p 0 -v 1600 -r 5 -m 0 -t 1 -d --noponder parameter seems at low visits to 15b weight

[bubblesld]

2be1fc26
trained up to v245
220:180 over 53a5fe93
176:172 in the official test

[l1t1]

build KataGo client on windows is no proble, do you mean the server?

[bubblesld]

My server is with ubuntu.

[bubblesld]

9c948dc4
trained up to v245 (another run)
216:184 over 53a5fe93
119:137 in the official test

[bubblesld]

0d2694c5
trained up to v245 (later version)
229:171 over 53a5fe93
232:189 in the official test

[aproprospro]

I was able to compile KataGo. Would you be interested in a compiled version for Ubuntu to try it out?

[bubblesld]

I was able to compile KataGo. Would you be interested in a compiled version for Ubuntu to try it out?

I know little about linux. I thought that it has to be built for each machine. But I can certainly give it a try.

[bubblesld]

9273d548
trained up to v249
223:177 over 0d2694c5
217:192 in the official test

[bubblesld]

a4b58a91
trained up to v249 (with more self-play games from v249)
219:181 over 0d2694c5
230:192 in the official test

Since it was 220:180 at game 400 in the official test, I will treat it as passed.

[bubblesld]

b4d5e8b6
trained up to v257
216:184 over a4b58a91
183:176 in the official test

[Marcin1960]

"b4d5e8b6
trained up to v257
216:184 over a4b58a91"

I made this net available for play on KGS, as LeelaZeroT

[Dorus]

For the record: On ogs, 15bTurboLeela also runs this net.

[melox1]

bubblesld, do you still train 15b nets?
There were many promotions on 40b net since 15b last update.

[bubblesld]

bubblesld, do you still train 15b nets?
There were many promotions on 40b net since 15b last update.

currently I am training 40b, SE-40b, 20b. Hope to get a stronger 20b this week, and switch 20b to 15b.

[bubblesld]

006eff16
trained up to v264
220:180 over a4b58a91
180:171 in the official test

[wind68]

Maybe the another one， for example 15b-264-600k , can win over 55% in the official test.
I suggest you try more times.

[wind68]

it is so strict that the new net should pass both in official test and your own test.

[bubblesld]

0c41abff
trained up to v266
220:180 over a4b58a91
183:178 in the official test

[bubblesld]

da46e3b9
trained up to v266 (later version)
221:179 over a4b58a91
128:138 in the official test

[bubblesld]

e912b007
trained up to v266 (later version)
218:182 over a4b58a91
169:166 in the official test

[bubblesld]

ad865426
trained up to v266 (policy x4)
216:184 over a4b58a91
231:191 in the official test

[bubblesld]

f5010dbd
trained up to v270
225:175 over a4b58a91
217:197 in the official test

[bubblesld]

2020/4/9
3353f154
trained up to v270 (later version)
227:173 over a4b58a91
202:179 in the official test

b7ccd6b4
trained up to v270 (later version)
221:179 over a4b58a91
176:174 in the official test

2020/4/10
0c4ade79
trained up to v270 (later version)
235:165 over a4b58a91
230:170 in the official test

[melox1]

The last one 0c4ade79 seems really nice, setting quite a gap for its rivals. Have anyone tested it for blind spots?
Also interesting how good it is vs Elf v2 at time parity.

[Marcin1960]

@melox1 "The last one 0c4ade79 seems really nice, setting quite a gap for its rivals. "

Started the run on KGS as LeelaZeroT

[cryptsport]

I can not find the best network 6b (for Leela Zero), strongest network trained like 15b.
for other networks there is a similar issue, like 15b training update, 20b training update?

[portkata]

Is it theoretically possible apply this same method with katago? training the s167 15b using the s509 40b

[Marcin1960]

Is it time to train for v278?

[balazsbme]

Are the versions available which are referred by this project: https://github.com/inclement/LazyBaduk (see: "[...] 15x192 network, from July 2018, in order to try to be as fast as practical on a phone") ?

[portkata]

@bubblesld wonsik made this script to train katago using a linux server https://colab.research.google.com/drive/1VsolGGLI-NsDHaLc-AzFJ_LuU95w8EPa?usp=sharing
would it be possible to modify the script and config file to train the 15b katago net on the new s663 40b data? one could do it without a gpu, using just the free google colab if a script was written.