Chainer v5.0.0b4をリリースしました

[nis...@preferred.jp]

Chainer v5.0.0b4 をリリースしました！ リリースノートは以下の通りです。

Highlights

Changes without compatibility

• Change the initial avg_var of L.BatchNormalization to 1 (#4742)
• Fix backward computation in F.forget (#5179). In this fix, the double backprop capability of F.forget is removed, since it did not work correctly in some cases.

New Features

• Add new functions:
  • Add F.rrelu, Randomized Leaky ReLU (RReLU) activation function (#3059, thanks @raven38!)
  • Add F.erfcx, scaled complementary error function (#5195)
  • Add F.erfcinv, inverse complementary error function (#5202)
  • Add F.ndtr, normal cumulative distribution function (#5237)
  • Add F.log_ndtr (#5239)
  • Add F.ndtri, the inverse of ndtr (#5247)
  • Add F.roi_average_align_2d (#5070, thanks @wkentaro!, #5259)
  • Add F.cumprod (#5074)
• Add new distributions:
  • D.MultivariateNormal (#4899)
  • D.Beta (#5088)
  • D.Categorical (#5028)
  • D.Uniform (#5123)
  • D.LogNormal (#5124)
• Support default dtype in some links:
  • L.GoogLeNet (#5099)
  • L.ResNetLayers (#5101)
  • L.VGG16Layers (#5107)
• Support all float dtypes in some functions:
  • F.absolute_error (#5145)
  • F.contrastive (#5152)
  • F.cross_covariance (#5158)
  • F.decov (#5174)
  • F.hinge (#5175)
  • F.huber_loss (#5176)
  • F.squared_error (#5212)
  • F.triplet (#5214)
  • F.batch_l2_norm_squared (#5235)
  • F.mean_squared_error (#5052)
• Implemented dataset using pickle (#4581)
• Add WarmupShift and MultistepShift extensions (#4935, thanks @mingxiaoh!)
• Improve initializer support in L.Maxout (#5068)
• Support n_batch_axis in L.Linear(#5103)
• Add raw kernel function (#5106)
• Support axis argument for F.log_softmax (#5215)

Enhancements

• Improve type check messages for some functions (#5189, #5200, #5224, #5248)
• Detect stalled datasets in MultiprocessIterator (#4607)
• New backward_accumulate (#4772)
• Avoid numpy.ascontiguousarray when iDeep is used (#5063)
• Use automatic broadcasting in distributions (#5086)
• Use F.cumprod in backward of F.prod (#5094)
• Return params and children as ordered by name in Link and Chain (#5119)
• Use cuda.get_array_module in fused function (#5120)
• Fix L.Convolution2D error message (#5138, thanks @fiarabbit!)
• Remove cuda_fusion.py (#5144)
• Implement eps_inside_sqrt option to RMSprop (#5150)
• Normalize chainer.config.dtype in chainer.get_dtype() (#5167)
• Use collections.abc to avoid DeprecationWarning in Python 3.7 (#5172)
• Minor fixes to D.MultivariateNormal (#5173)
• Avoid collections.Iterable (#5180)
• Make imports in alphabetical order (#5181)
• Avoid keyword arguments in FunctionHook callbacks (#5191)
• Retain outputs in F.erfinv (#5199)
• Use xp.einsum in F.bilinear (#5207)
• Minor fixes to D.Beta (#5219)
• Minor fixes to D.Uniform (#5225)
• Check eps < CUDNN_BN_MIN_EPSILON in FixedBatchNormalization (#5232, thanks @cycentum!)
• Use ndtr and log_ndtr in normal distribution (#5240)
• Use erfcinv for Normal.icdf (#5242)
• Use ndtri in normal distribution (#5254)
• Use normcdfinv in F.ndtri (#5260)

Bug Fixes

• Move backends.cuda.copyto to backends.copyto and make it work with iDeep (#5095)
• Fix the condition for the switching of cuDNN in F.deconvolution_nd (#5129, thanks @fiarabbit!)
• Fix test failure of TestResNetLayers (#5133)
• Fix backward compatibility of Link.__call__ MRO (#5141)
• Flush the output stream after PrintReport reports (#5146)
• Support old numpy in F.split_axis (#5157)
• Avoid cancellation in D.Normal (#5185)
• Support 0-dim input in F.logsumexp (#5190, thanks @cadenacchi!)
• Fix cpu codes of indexing in F.softmax_cross_entropy (#5238)
• Comment out extreme test of D.Categorical (#5261)

Documentation

• Add iDeep to backend docs (#5121)
• Improve iterator description in Chainer at a glance documentation (#5132, thanks @fiarabbit!)
• Avoid use of ideep in doctest (#5148)
• Fix grammar in PR template (#5178)
• Fix docstring of F.erfinv (#5201)
• Fix Sphinx issues in the reference of probability distributions (#5203)
• Change iDeep in tips.rst to Chainer Backend for Intel Architecture (#5208, thanks @mingxiaoh!)
• Fix toc level in iterator documentation(#5257)

Examples

• Rename VAE hyperparameter C to beta in the example (#5135, thanks @Evanc123!)
• Override Link.forward in MNIST model parallel example (#5159)

Tests

• Simplify and stabilize softmax_cross_entropy test (#3409)
• Ignore float warnings if testing extreme value (#5122)
• Trivial fix for parameterized test case of F.contrastive (#5147)
• Fix a class name in test_erfinv (#5165)
• Fix doctests of open_pickle_dataset (#5182)
• Fix test failure on Windows (#5186)
• Add sphinx to doctest requirements (#5187)
• Fix occasional test failure of l2normalize (#5210)
• Fix occasional test failure of contrastive (#5218)
• Ignore Theano warnings in Python 3.7 (#5223)
• Ignore DeprecationWarnings at importing Theano (#5230)
• Adjust tolerance of TestMatMul (#5236)
• Add .pytest_cache/ to .gitignore (#5193)