GELU Operator

[haojin2]

Description

Address #12984, operator from https://arxiv.org/abs/1606.08415.

Checklist

Essentials

Please feel free to remove inapplicable items for your PR.

• The PR title starts with [MXNET-$JIRA_ID], where $JIRA_ID refers to the relevant JIRA issue created (except PRs with tiny changes)
• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage:
• Unit tests are added for small changes to verify correctness (e.g. adding a new operator)
• Nightly tests are added for complicated/long-running ones (e.g. changing distributed kvstore)
• Build tests will be added for build configuration changes (e.g. adding a new build option with NCCL)
• Code is well-documented:
• For user-facing API changes, API doc string has been updated.
• For new C++ functions in header files, their functionalities and arguments are documented.
• For new examples, README.md is added to explain the what the example does, the source of the dataset, expected performance on test set and reference to the original paper if applicable
• Check the API doc at http:https://mxnet-ci-doc.s3-accelerate.dualstack.amazonaws.com/PR-$PR_ID/$BUILD_ID/index.html
• To the my best knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change

Changes

• GELU operator (sym/nd.LeakyReLU(act_type='gelu'))
• Unit test
• Gluon interface (gluon.nn.GELU)

Comments

Flakiness also checked:

MXNET_TEST_COUNT=10000 nosetests tests/python/unittest/test_operator.py:test_gelu
[INFO] Setting module np/mx/python random seeds, use MXNET_MODULE_SEED=1909868054 to reproduce.
.
----------------------------------------------------------------------
Ran 1 test in 314.370s

OK
MXNET_TEST_COUNT=10000 nosetests tests/python/gpu/test_operator_gpu.py:test_gelu
[INFO] Setting module np/mx/python random seeds, use MXNET_MODULE_SEED=2117426005 to reproduce.
.
----------------------------------------------------------------------
Ran 1 test in 446.883s

OK

[haojin2]

@eric-haibin-lin @szha for review

[hendrycks]

Often xxx is cheaper than x ** 3. Also, if one takes inspiration from https://en.wikipedia.org/wiki/Horner's_method , which is often more numerically stable, then g(x) could be

return TWO_OVER_PI * x * (1 + CUBE_CONSTANT * x * x)

[haojin2]

This is not the actual implementation so performance is not critical here... I'll change the internal mshadow math kernel in C++ based on the link you provided.

[haojin2]

Please check this part for the backend implementation: https://github.com/apache/incubator-mxnet/blob/3bf1279fcd5efb80c0d098583ae213cbd2867b66/src/operator/mshadow_op.h#L134-L147. Thanks!

[hendrycks]

This looks correct.

[haojin2]

@hendrycks Thanks for your prompt reply! I'll wait for @eric-haibin-lin and @szha to give a review before we merge this in.

[karan6181]

@mxnet-label-bot add [Feature request, Operator, pr-awaiting-review]

[eric-haibin-lin]

Great contribution! One comment about the macro

[szha]

LGTM. Remember to also add test for the new gluon block

[haojin2]

@szha Gluon test added, also fixed a problem with existing SELU test.

[haojin2]

@szha Build finally passed, good for merge?