v1.3.1

This is a revision release that contains the following changes:

bug #381 #382 #386 #387 #389 #390 #391 #394 #398 #400 #401 #403 #404 #410 #411 #413 #415
enhancement #383
test #392 #395 #396 #406 #407 #414
document #379 #395 #397 #405

Summary:

• Fixes of many bugs relating CuPy including
  • issues on multi-GPU (device switching (Thanks Amit), accumulate_grads)
  • restration of the environment variable CHAINER_SEED. You can fix the random seed for each device by setting this variable.
• FunctionSet can contain attributes other than Function and FunctionSet (Thanks @choni).

v1.3.0

This is a minor release that contains following changes:

feature #266 #331 #343 #351 #356 #358 #372 #375
bug #332 #334 #339 #340 #341 #346 #353 #357 #360 #362 #367 #369 #373 #374
enhancement #333 #335 #337 #347 #356 #359 #364 #365 #370 #371 #375
test #344 #361
document #338 #342 #348 #349 #350 #352 #366 #376

Summary:

• CUDA array backend is now switched to CuPy. This is an ndarray implementation on CUDA with NumPy-like interface. See CuPy document for details. Chainer package includes CuPy.
• Due to the change of CUDA array backend, users do not have to install chainer-cuda-deps package anymore to enable CUDA support. CUDA support is automatically enabled once users make the PATH and LD_LIBRARY_PATH environment variables correctly.
• The chainer.functions package is reorganized into several subpackages. It does not change users’ code that only uses symbols directly under the chainer.functions package.
• Type checking can be disabled by setting the Function.type_check_enable attribute to False or by setting the CHAINER_TYPE_CHECK environment variable to ‘0’. It removes overheads for the type checking. Note that the type check is enabled by default.
• Adam optimizer is updated to the latest version of the arXiv paper (@kashif thanks!). It might change the learning behavior a bit.
• Sum now supports summation along one or more axes.

For Windows users:
We do not support Windows officially, though we want to make it run on Windows as well as Linux. We made CuPy installable to Windows, but it might have some incorrect behaviors. It is appreciated to send us reports of errors on Windows!

v1.2.0

This is a minor release containing the following updates:

feature #170 #171 #177 #180 #184 #204 #209 #211 #212 #213 #217 #245 #247 #259 #262 #264 #289 #290 #307
bug #293 #303 #304 #308 #310 #312 #313 #316 #317 #319 #320 #321
enhancement #123 #230 #246 #311
test #294 #295 #309 #314 #322
document #147 #292 #296 #322
example #147 #168
other #300

Many new functions are added:

• Bilinear (binary operation counterpart of Linear)
• clipped_relu (thanks @jnishi!)
• cross_covariance (thanks @jheymann85!)
• sin/cos (thanks @jfsantos!)
• softplus
• spatial_pyramid_pooling_2d (thanks @mitmul!)
• Variational AutoEncoder support (gaussian, bernoulli_nll, gaussian_nll, and gaussian_kl_divergence)

In addition, softmax and softmax_cross_entropy now supports replicated softmax that repeats softmax over spatial locations (thanks @rezoo!), and BinaryHierarchicalSoftmax now supports GPU mode.

FunctionSet objects are now directly passed to Optimizer.setup without calling collect_parameters (thanks @sinhrks!). The collect_parameters method is deprecated as of this version.

The type checking on backward is removed. Existing user-defined functions that implement the check_type_backward method are not corrupted by this change, though the check is skipped during the backward computation.

We now have a list of external examples. If you want to add an example to this list, notify us at our mailing list.

v1.1.2

This is a revision release containing the following updates:

bug #253 #254 #255 #256 #257 #258 #261 #267 #269 #271 #276 #279
enhancement #232 #272 #284 #285
test #68 #273 #281 #282 #283 #286
document #33 #205 #263 #265 #275 #287 #288
example #270
other #260 #268

We added a word2vec example at examples/word2vec directory. Note that the hierarchical softmax option does not run on GPU: it will be supported by the next release v1.2.0.

We also released the tutorial on type checking and a quick table of framework comparison.
http://docs.chainer.org/en/latest/tutorial/type_check.html
http://docs.chainer.org/en/latest/comparison.html

v1.1.1

This is a revision release containing following changes:

Bug fixes #157 #194(#195) #197 #199 #206 #215
Enhancement #79 #109(#241) #139(#163) #173 #198(#200) #201(#208) #207 #210 #123(WIP, done: #188 #189 #190 #191 #192 #202 #203 #219 #220 #221 #222 #223 #224 #225 #226 #229 #231(#239) #233(#249) #234(#244) #236(#243) #237(#242) #238(#240) #248)
Test #137 #150(#179) #160(#161) #162 #186(#187)
Document #152 #154(#227) #167(#235) #175 #176 #182 #218(#250)
Example #88 #158 #164 #165
Other #155(#159) #196

Be careful that scikits.cuda is now renamed to scikit-cuda. The existing CUDA-enabled environment may be broken by the update. Be sure to uninstall the old scikits.cuda and update the chainer-cuda-deps package for CUDA-enabled Chainer.

The issue #123 and related issues/PRs add type checking routines for many Function implementations. It is still on going, and may be done by the next release.

v1.1.0

This is a minor release including new features:

• Caffe model import #3(#110 #148)
• Function type checking system #1(#95 #121 #122 #124 #131 #132 #151)
• Computational graph visualization #2(#102 #129 #144 #152)
• New functions:
  • sigmoid_cross_entropy #57 (Thank you @rezoo!)
  • matmul/batch_matmul #8(#116)
  • linear/convolution_2d #9(#117): These are non-parameterized versions of Linear/Convolution2D
  • abs #11(#111)
  • split_axis #10(#119)
  • InceptionBN #45(#128)
  • NegativeSampling #14
• New optimizer: RMSpropGraves #82
• Python3 support #22(#67)
• Version number of the chainer package #29(#106)
• Remove dependency on chainer-cuda-requirements #59(#104)

and other modifications:

• Bug fixes #44(#76) #84 #96(#101) #97(#100) #103 #107(#108) #120(#125) #126(#127) #140
• Test #85 #113 #133(#135) #145
• Document #26 #83(#89) #149
• Example #87 #90 #91 #93 #114(#115) #146
• Other #35 #86 #134 #141

We have also released following things for our development.

• Contribution Guide: From this release, all the contributions should follow this document, and all the codes must follow the style guidelines in this document.
• Continuous Integration: We have started using Travis CI to test all PRs on CPU (#23). It runs CPU tests and coding style checks. Be careful that the CI does not check GPU code.
• CI environment: We have released testing environment at https://github.com/pfnet/chainer-test. It includes test scripts and dockerfiles to run tests on clean environments. Note that these scripts are different from those used in Travis CI: test scripts of chainer-test include GPU tests, too.

v1.0.1

This is a revision release including following updates:

Bug fix #18 #44(#76) #62
Improvement #7 #43(#77) #46(#49) #51 #56
Test #4(#54) #5(#63) #20 #21 #64 #65 #66 #80
Document #12 #15 #16 #28 #31 #61
Example #19 #30 #38(#52)
Other #40 #55

Summary:

• Fixed the GPU routine of optimizer.Adam. Thanks @y-tag!
• New installation way of CUDA-related dependencies: pip install chainer-cuda-deps
• functions.BatchNormalization gets faster
• Added many tests

v1.0.0

This is the first release of Chainer! The first version includes:

• Computational graph manipulation (core functionality of Chainer)
• Many functions
• Popular optimizers
• CUDA support
• CuDNN v2 support
• Four examples (MLP, LSTM, ConvNets, Recursive Nets)
• Tutorial