Chainer v7.0.0b1 をリリースしました

[Akifumi Imanishi]

Chainer v7.0.0b1 をリリースしました．リリースノートは以下の通りです．

This is the release note of v7.0.0b1. See here for the complete list of solved issues and merged PRs.

Highlights

• TabularDataset is added. This is a new dataset interface that supports rich manipulation in a tabular form (like pandas.DataFrame), e.g. loading only a specified subset of keys (columns), efficient slicing (with less transposition/concatenation), batch-wise preprocessing, etc. The API is still under development; we are adding more functionalities and widening its support in existing features where datasets are involved.

New Features

• Add interface to backprop from multiple variables (#5952)
• Option to show progress bar during evaluation (#6474, thanks @wkentaro!)
• Elementwise Power for ChainerX (#6496, thanks @dido1998!)
• Add chainerx.hstack, chainerx.vstack and chainerx.atleast_2d (#6886, thanks @kshitij12345!)
• Add TabularDataset (#7115)
• Add TabularDataset.concat/join (#7116)
• Add chainerx.expm1 and chainerx.exp2 (#7126, thanks @aksub99!)
• Add chainerx.log2 (#7139)
• Add TabularDataset.{transform/transform_batch} (#7150)
• Add chainerx.log1p (#7161, thanks @sky58!)
• Expose chainerx::AsContiguous as a public C++ API (#7166)
• Emit warning on chainerx import in debug mode (#7178)
• Add chainer.as_array for consistency with chainer.as_variable (#7252, thanks @tkerola!)
• Add chainerx.moveaxis (#7265, thanks @kshitij12345!)
• Add chainerx.leaky_relu (#7351, thanks @aksub99!)
• Add chainerx.dstack and chainerx.atleast_3d (#7353, thanks @kshitij12345!)
• Add Python operator __abs__ with chainerx.ndarray (#7364)
• Allow turning off the static subgraph optimizations using a config (#7369)
• Add NumPy constants to ChainerX (#7384)
• Add chainerx.erf (#7404, thanks @aksub99!)
• Add align_corners option to resize_images (#7429)
• Add nearest mode to resize_images (#7443)
• Add input_device to StandardUpdater (#7472)
• Add is_array_supported method on backend.Device (#7487)

Enhancements

• Refactor roi_max_align_2d and roi_average_align_2d (#6405, thanks @knorth55!)
• Support Tagged communication with MPI_Status. (#6696, thanks @y1r!)
• Support ChainerX in F.copy (#6982)
• Avoid unnecessary updates in F.batch_renormalization, and related fixes (#7104)
• Support ChainerX in Variable.addgrad (#7132)
• Fix cuda.DummyDevice inheritance (#7147)
• Add Device.name property (#7149)
• Link.serialize to support ChainerX (#7175)
• Fix typo in Variable.backward (#7196)
• Call require_grad() on ChainerX Variable.grad setter (#7198)
• Clear outputs in FunctionNode.unchain and raise error in ChainerX fallback mode (#7216)
• Support ChainerX in Variable.copydata (#7226)
• Support ChainerX in MNIST data parallel example (#7227)
• MultiprocessParallelUpdater to support new devices (#7245)
• Alias StackVector<int64_t, kMaxNdim> to Dims (#7258)
• Support bool dtypes in chainerx::{Max,Min}imum (#7261)
• Fix integral negative powers (#7262)
• Make chx.backward not cause error even if backprop is not required (#7287)
• Support None arguments in chainerx.clip and chainerx.ndarray.clip (#7296)
• Support scalar in chainerx::Where (#7325)
• F.clip function with None parameter to min/max (#7333)
• Support cudnn deterministic max pooling (#7390, thanks @anaruse!)
• Avoid transferring from a native device to another in Array::ToNative() (#7394)
• Add type hints to Variable (#7400)
• Improve get_device error message when ChainerX is not available (#7401)
• get_device to raise a more correct error types (#7421)
• Make EXEPECT_ARRAY_* macros able to used outside ChainerX (#7434)
• Add sequence support for ChainerX shape arguments (#7446)
• Check positive dilation in F.convolution_2d (#7448)
• Check positive dilation in F.deconvolution_2d (#7449)
• Explicit check of chainerx arrays on fallback functions (#7452)
• Support F.copy between non-ChainerX and ChainerX devices only if backprop is not required (#7473)

Performance Improvements

• In FunctionNode ChainerX fallback, reuse ChainerxDevice taken from inputs to create outputs (#7397)

Bug Fixes

• Fix type check of F.where (#6872)
• Fix a bug in Bernoulli.log_prob (#7064, thanks @seiyab!)
• Fix uncopyable MultiNodeBatchNormalization (#7106)
• Bugfix: MultiNodeChainList should not assume float32 (#7165)
• Fix initialization of L.Linear when called with n_batch_axes (#7167)
• Fix float16 and Tensor Core related issue in ChainerX (#7189, thanks @anaruse!)
• Fix recomputation of L.BatchRenormalization (#7256)
• Fix F.absolute_error for ChainerX (#7281, thanks @crcrpar!)
• Fix a bug that root is ignored in scatter_dataset and bcast (#7289)
• Fix condition to invoke cuDNN dropout (#7293, thanks @crcrpar!)
• Improve type check in _values_to_dicts so it works with unicode of python 2 too (#7316)
• Fix DtypeError in chainerx.square (#7321)
• Fix mypy errors (#7423)
• Make WeightDecay aware of loss scale (#7491)
• Fix GradientMethod ChainerX fallback for uninitialized parameters (#7492)
• Bugfix for pytest 2x2 (#7509)
• Fix AdamW update rule regression on CPU (#7512)

Code Fixes

• Split binary functions from math.cc (#7128)
• Avoid using cuda.DummyDevice and cuda.get_device_from_array (#7148)
• Fix pointless comparison compiler warning in ChainerX (#7160)
• Remove backslashes to continue lines of link targets (#7170)
• Split trigonometric/hyperbolic routines from math.cc (#7171)
• Remove duplicated code in logic.cc (#7176)
• Consistent cases for Inplace (#7181)
• Improve code in testing.backend.BackendConfig (#7212)
• Split ChainerX statistics routines from math.cc (#7222)
• Fix code style for long expressions (#7231)
• Check device instance using xp when possible (#7234)
• Move declaration of AMax and AMin to statistics routines (#7269)
• Split reduction routines from math.cc (#7270)
• Use _ for private classes under chainer.dataset.tabular (#7275)
• Remove unused using declaration (#7284)
• Split misc routines from math.cc (#7298)
• Fix wrong comment in ChainerX backward implementation (#7311)
• Split explog routines from math.cc (#7317)
• Fix style on imports (#7338)
• Split rounding routines (#7407)
• Split arithmetic ops from routines/math.h (#7415)
• Put comments in FindCuDNN.cmake (#7419)
• DRY optimizer test parameterizations (#7437)
• Split logic routines from math (#7444)
• Qualify some arguments of pool kernels const& (#7453)
• Include cuda_fp16.h instead of cuda_fp16.hpp (#7480)
• Use py::arg literal in ChainerX python binding (#7490)
• Remove rounding kernels from math (#7497)
• Rename and move activation routines from math.h (#7501)
• Remove ChainerX AsTypeKernel (#7522, thanks @kshitij12345!)
• Split python binding math routines (#7527)
• Use absolute namespace in macros (#7536)

Documentation

• Improve contribution guide (#6140)
• Fix dead sphinx links (#6450)
• Fix F.normalize documentation (#7062, thanks @crcrpar!)
• Document F.copy view behavior (#7135)
• Improve device documentation (#7162)
• Document backend.get_device_from_array (#7163)
• Remove chainerx.md (#7179)
• Add optimizers.MSVAG to documentation (#7183)
• Fix grammatical errors in documentation (#7186)
• Fix capitalization of F.relu in doc (#7188)
• Add missing doc entry for CommunicatorBase.allgather (#7192)
• Fix invalid escape sequences in ChainerX routine docstrings (#7214)
• Fix typos in chainer.utils.type_check (#7249, thanks @ktns!)
• Document observe_value and observe_lr trigger interval (#7266)
• Fix robots.txt to allow indexing root (#7306)
• Avoid installing ChainerX when building docs of other projects on ReadTheDocs (#7363, thanks @knorth55!)
• Improve F.normalize documentation (#7371, thanks @crcrpar!)
• Fix format of static_graph.rst (#7389)
• Change Deformable Convolution 2D docs to match arguments (#7402, thanks @higumachan!)
• Avoid setting test_iter.epoch manually in the tutorial of training loop (#7405)
• Remove "Comparison with other frameworks" from docs (#7417)
• Fix documentation for shape in generate_array (#7450)
• Remove test coverage from ChainerX contribution guide (#7462)
• Correct Embed ID documentation (#7484)
• Fix typo in tabular_dataset.py (#7495, thanks @nai62!)

Installation

• Fix ChainerX compilation with MSVC (#7108, thanks @durswd!)
• Allow CUDNN_LIBNAME to be specified by environment variable (#7243)
• Use external $MAKEFLAGS instead if set in Travis CI script (#7331)
• In FindCuDNN.cmake, prioritize explicit variables over environment variables (#7441)
• Add ChainerX build option to use cuDNN from CuPy installation (#7442)
• Pin typing == 3.6.6 (#7562)
• Fix typing requirements (#7564)

Examples

• Add CIFAR example to ChainerMN (#6839, thanks @ai-kase!)
• Support device specifiers in MNIST data parallel example (#6857)
• Support device specifiers in PTB example (#7055)
• Support device specifiers in pix2pix example (#7076)
• Support device specifiers in static graph example (#7153)
• Support device specifiers in ImageNet data parallel example (#7164)
• Support ChainerX in MNIST inference example (#7169)
• Support device specifier in image captioning example (#7204)
• Support device specifier in image captioning example (predict.py) (#7206)
• Remove PlotReport.available() check in glance example (#7209)
• Minor fix in DCGAN example README (#7210)
• Fix sentiment example test (#7215)
• Support device specifiers in MNIST model parallel example (#7225)
• Use Agg backend in examples with plot functionality (#7247)
• Support ChainerX in PTB gentxt example (#7314)
• Support ChainerX in MNIST model parallel example (#7330)
• Warn NaN in FP16 mode in dcgan example (#7344)
• Warn NaN in FP16 mode in memnn example (#7345)
• Warn NaN in FP16 mode in pix2pix example (#7346)
• Warn NaN in FP16 mode in pos example (#7354)
• Warn NaN in FP16 mode in reinforcement learning examples (#7355)
• Warn NaN in FP16 mode in sentiment example (#7356)
• Warn NaN in FP16 mode in static_graph_optimizations/cifar example (#7357)
• Warn NaN in FP16 mode in static_graph_optimizations/mnist example (#7358)
• Warn NaN in FP16 mode in vae example (#7362)
• Warn NaN in FP16 mode in word2vec example (#7366)
• Fix typo in wavenet example requirements (#7367)
• Warn NaN in FP16 mode in wavenet example (#7372)
• Support ChainerX in static subgraph optimization examples (#7431)
• Implement reset method in the PTB example (#7533)

Tests

• Add FP16 test to multi_node_chain_list (#6575)
• [chainerx] Fix skipped_backward tests to return as PASS (#6815, thanks @kshitij12345!)
• Add configuration of new CI system (#6843)
• Simplify F.tensordot test (#6968, thanks @ishanrai05!)
• Simplify F.cumprod test (#6978, thanks @hikjik!)
• Simplify F.average test (#6995, thanks @hikjik!)
• Move test_cuda.py to backends_tests (#7144)
• Fix missing cuda in chainerx.swapaxes test (#7184, thanks @kshitij12345!)
• Split Variable.grad and Variable.grad_var tests (#7191)
• Refactor Variable.zerograd test (#7199)
• Add Tensor Core test for chainerx.conv and chainerx.conv_transpose (#7203)
• Move TestTanh from test_math.py to test_trigonometric_hyperbolic.py (#7207)
• Refactor Variable.copydata test (#7224)
• Add a test to reproduce the bcast deadlock problem (#7257)
• Add float16 comparison test (#7260)
• Use CUDA_VISIBLE_DEVICES in ChainerX tests (#7290)
• Add chainer.as_array test (#7318)
• Rewrite StandardUpdater tests with pytest style assertion (#7326)
• Change 0 to 0.0 for python2 (#7373)
• Add missing parameter dstack to invalid_shape test (#7457, thanks @kshitij12345!)
• Use pytest.mark.xfail instead of unittest.expectedFailure (#7488)

Others

• Remove "Research projects using Chainer" from README (#7416)