roywei commented on issue #11544: [MXNET-620][WIP][Do not merge]Fix flaky test batchnorm training URL: https://github.com/apache/incubator-mxnet/pull/11544#issuecomment-402819447 @zheng-da Hi, I have modified your unit test for mkldnn batchnorm, increased atol to 1e-2 as its flaky at 1e-3 and 1e-4. Reasons listed in PR description, all batchnorm implementation (mkldnn, cudnn) gradient is slightly off from numeric gradient. Using 1e-2 is more stable and passed 10000 runs.
---------------------------------------------------------------- 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
