anjishnu commented on issue #17298: [MXNET-1438] Adding SDML loss function URL: https://github.com/apache/incubator-mxnet/pull/17298#issuecomment-575405571 @haojin2 With the original confirguration I was getting around 1 in 100 or 1 in 200 failures, which is not acceptable for a unit test. Changing the hyperparameters (e.g. increasing dimensionality, lowering N) does make it more robust (tried up to MXNET_TEST_COUNT=2000 runs) but there's always a chance of failure Other options: - Go with the original precomputed values approach, but maybe over a larger range of precomputed values. - Set up a similar test ... but use correlated data instead of using purely random data: - Generate the positive samples that are actually correlated to the paired data rather than purely random. - Always use some set of samples that are known to converge successfully - Add a test to see loss reduction (e.g. to <0.05 as in many of the other tests) Which of these would be your preferred approach - I have tested them all to MXNET_TEST_COUNT=2000
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
