BatchNorm can not converge with scale=False

[nttstar]

Description

BatchNorm operator with scale=False can not converge.

Error Message

No error message, but loss value and training accuracy is abnormal comparing with scale=True BatchNorm.

To Reproduce

We can try https://github.com/nttstar/arcface.np to train arcface. Add one BatchNorm op with scale=False after final embedding layer

What have you tried to solve it?

1. Set Scale=True, it can work but with slightly worse test accuracy.

Environment

----------Python Info----------
Version : 3.6.9
Compiler : GCC 7.3.0
Build : ('default', 'Jul 30 2019 19:07:31')
Arch : ('64bit', '')
------------Pip Info-----------
Version : 19.3.1
Directory : /root/anaconda2/envs/py36/lib/python3.6/site-packages/pip
----------MXNet Info-----------
Version : 2.0.0
Directory : /root/anaconda2/envs/py36/lib/python3.6/site-packages/mxnet
Num GPUs : 8
Hashtag not found. Not installed from pre-built package.
----------System Info----------
Platform : Linux-3.10.0-327.el7.x86_64-x86_64-with-centos-7.5.1804-Core
system : Linux
node : gpu06
release : 3.10.0-327.el7.x86_64
version : #1 SMP Thu Nov 19 22:10:57 UTC 2015

[wkcn]

Hi @nttstar , there was a bug in BatchNorm in the previous version of MXNet. The bug was reported in #18373 , and it was fixed in PR #18377 . Could you please try the latest version of MXNet?

[sxjscience]

@wkcn I communicated with @nttstar offline and #18373 should not be the casue. Would you help try with scale=False and see if there is anything wrong?

[wkcn]

@sxjscience I'm sorry that I don't have any machine with GPU to check it recently.

I read the code of batch norm and its unittest.
There is a gradient check when fix_gamma=True:

https://github.com/apache/incubator-mxnet/blob/master/tests/python/unittest/test_operator.py#L1777,

but no output check when fix_gamma=True:

https://github.com/apache/incubator-mxnet/blob/master/tests/python/unittest/test_operator.py#L1882

[wkcn]

I try to test the batch norm with fix_gamma=True.
The result on CPU is right, but that on GPU is wrong when fix_gamma=True, axis = 1 and cudnn_off=False. The bn_beta.grad is the only wrong value.

Here are the failure cases when fix_gamma=True.

operator	shape	axis	cudnn_off	output_mean_var
BatchNorm	(24, 2)	1	False	False
BatchNorm	(24, 2)	1	False	True
BatchNorm	(24, 3, 4)	1	False	False
BatchNorm	(24, 3, 4)	1	False	True
BatchNorm	(24, 4, 4, 4)	1	False	False
BatchNorm	(24, 4, 4, 4)	1	False	True
BatchNorm	(24, 8, 4, 4)	1	False	False
BatchNorm	(24, 8, 4, 4)	1	False	True
BatchNorm	(24, 5, 6, 4, 4)	1	False	False
BatchNorm	(24, 5, 6, 4, 4)	1	False	True

[wkcn]

I'm fixing the bug and I will submit a PR later.

[wkcn]

Hi @nttstar , the bug of BatchNorm has been fixed in #18500 .

Thank you for the report!

[nttstar]

@wkcn Thanks! I will check it in the next pip package release.