roywei commented on a change in pull request #11544: [MXNET-620]Fix flaky test
batchnorm training
URL: https://github.com/apache/incubator-mxnet/pull/11544#discussion_r200505080
##########
File path: tests/python/unittest/test_operator.py
##########
@@ -1527,7 +1526,7 @@ def check_batchnorm_training(stype):
test = mx.symbol.BatchNorm(data, fix_gamma=False,
use_global_stats=True, axis=chaxis)
check_numeric_gradient(test, in_location, xmean_std,
numeric_eps=1e-2, rtol=0.2, atol=0.01)
- stypes = ['row_sparse', 'default']
Review comment:
Without row_sprase, the flakyness come from difference between gradient
values of mkldnn/cuddn implementation of batchnorm and numeric calculation from
check_numeric_gradients
row_sparse makes it more flaky, and will fail at this part: [testing varying
channel
axis](https://github.com/apache/incubator-mxnet/blob/e870890ab93ae25d51da01a445c965e4db71069b/tests/python/unittest/test_operator.py#L1524).
Which is not in your test_mkddnn:test_batchnorm.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
