Support SyncBatchNorm5D

[wkcn]

Description

Hi! there.
Currently, SyncBatchNorm doesn't support 5+D input.
In this PR, I fix it.

Checklist

Essentials

Please feel free to remove inapplicable items for your PR.

• tiny change. 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

• modify the code of SyncBatchNorm to support 2+D inputs.
• Remove the default value key for SyncBatchNorm
• Using CUDNN to compute 5+D inputs for BatchNorm
• modify the testcase test_sync_batchnorm
• Add an operator-test for BatchNorm and SyncBatchNorm
• reformat tests/python/gpu/test_gluon_gpu.py

Comments

• If this change is a backward incompatible change, why must this change be made.
• Interesting edge cases to note here

[zhreshold]

So it will support only 3D to 5D? What's the limitation here?

[zhreshold]

looks good

[wkcn]

There is a bug when the dimension of inputs is 2.

======================================================================

FAIL: test_gluon_gpu.test_sync_batchnorm

----------------------------------------------------------------------

Traceback (most recent call last):

  File "/usr/local/lib/python2.7/dist-packages/nose/case.py", line 197, in runTest

    self.test(*self.arg)

  File "/work/mxnet/tests/python/gpu/../unittest/common.py", line 177, in test_new

    orig_test(*args, **kwargs)

  File "/work/mxnet/tests/python/gpu/test_gluon_gpu.py", line 373, in test_sync_batchnorm

    num_devices=ndev, cuda=cuda)

  File "/work/mxnet/tests/python/gpu/test_gluon_gpu.py", line 352, in _check_batchnorm_result

    atol=1e-3, rtol=1e-3)

  File "/work/mxnet/python/mxnet/test_utils.py", line 495, in assert_almost_equal

    raise AssertionError(msg)

AssertionError: 

Items are not equal:

Error 1.597678 exceeds tolerance rtol=0.001000, atol=0.001000.  Location of maximum error:(0,), a=0.912201, b=0.909150

 a: array([0.9122006, 0.906133 ], dtype=float32)

 b: array([0.9091504, 0.9045997], dtype=float32)

-------------------- >> begin captured logging << --------------------

common: INFO: Setting test np/mx/python random seeds, use MXNET_TEST_SEED=868530648 to reproduce.

[wkcn]

There seems to be a bug in SyncBatchNorm, when spatial_shape is 1x1, 1xn or nx1. I am checking it.

[wkcn]

It seems that the bug has been addressed, although I do not know the specific reason yet.

I will add a test for multi-output.

[wkcn]

@zhreshold @szha Hi! I have updated the PR and add adquate unittests.
Would you mind reviewing it? Thank you!

[zhreshold]

If @szha has no complaint, I can merge it in 24hr

[zhreshold]

thanks @wkcn , this is merged!

[szha]

This line requires having GPU when CUDA is installed, or it would throw this error:

======================================================================
ERROR: test_gluon.test_sync_batchnorm
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/usr/local/lib/python2.7/dist-packages/nose/case.py", line 197, in runTest
    self.test(*self.arg)
  File "/home/travis/build/dmlc/mxnet-distro/mxnet-build/tests/python/unittest/common.py", line 177, in test_new
    orig_test(*args, **kwargs)
  File "/home/travis/build/dmlc/mxnet-distro/mxnet-build/tests/python/unittest/test_gluon.py", line 693, in test_sync_batchnorm
    num_gpus = mx.context.num_gpus()
  File "/home/travis/build/dmlc/mxnet-distro/mxnet/context.py", line 258, in num_gpus
    check_call(_LIB.MXGetGPUCount(ctypes.byref(count)))
  File "/home/travis/build/dmlc/mxnet-distro/mxnet/base.py", line 254, in check_call
    raise MXNetError(py_str(_LIB.MXGetLastError()))
MXNetError: [11:47:54] include/mxnet/base.h:427: Check failed: e == cudaSuccess (30 vs. 0) :  CUDA: unknown error
Stack trace:
  [bt] (0) /home/travis/build/dmlc/mxnet-distro/mxnet/libmxnet.so(+0x4b60fb) [0x7f8d608830fb]
  [bt] (1) /home/travis/build/dmlc/mxnet-distro/mxnet/libmxnet.so(+0x2440eec) [0x7f8d6280deec]
  [bt] (2) /home/travis/build/dmlc/mxnet-distro/mxnet/libmxnet.so(MXGetGPUCount+0x19) [0x7f8d6280df79]
  [bt] (3) /usr/lib/x86_64-linux-gnu/libffi.so.6(ffi_call_unix64+0x4c) [0x7f8d9a2e1c7c]
  [bt] (4) /usr/lib/x86_64-linux-gnu/libffi.so.6(ffi_call+0x1fc) [0x7f8d9a2e15ac]
  [bt] (5) /usr/lib/python2.7/lib-dynload/_ctypes.x86_64-linux-gnu.so(_ctypes_callproc+0x48e) [0x7f8d9a4f85fe]
  [bt] (6) /usr/lib/python2.7/lib-dynload/_ctypes.x86_64-linux-gnu.so(+0x15f9e) [0x7f8d9a4f9f9e]
  [bt] (7) /usr/bin/python(PyEval_EvalFrameEx+0x965) [0x4c84a5]
  [bt] (8) /usr/bin/python(PyEval_EvalCodeEx+0x2ac) [0x4cfedc]
-------------------- >> begin captured logging << --------------------
common: INFO: Setting test np/mx/python random seeds, use MXNET_TEST_SEED=1179889124 to reproduce.
--------------------- >> end captured logging << ---------------------
----------------------------------------------------------------------

Can you please move this test to tests/python/gpu/test_gluon_contrib_gpu.py? @wkcn @zhreshold

[wkcn]

I don't know why a unknown CUDA error was raised.
https://github.com/apache/incubator-mxnet/blob/master/include/mxnet/base.h#L424

[szha]

I was testing it on a platform without GPU, with CUDA installed. In any case, the test seems misplaced.