Fix flaky test test_operator_gpu.test_spatial_transformer_with_type (…

[ddavydenko]

…#7645)

Description

Modified test_operator_gpu.test_spatial_transformer_with_type test in order to remove its flakiness

Checklist

Essentials

Please feel free to remove inapplicable items for your PR.

• The PR title starts with [MXNET-$JIRA_ID], where $JIRA_ID refers to the relevant JIRA issue created (except PRs with tiny changes)
• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage:
• Unit tests are added for small changes to verify correctness (e.g. adding a new operator)
• Nightly tests are added for complicated/long-running ones (e.g. changing distributed kvstore)
• Build tests will be added for build configuration changes (e.g. adding a new build option with NCCL)
• Code is well-documented:
• For user-facing API changes, API doc string has been updated.
• For new C++ functions in header files, their functionalities and arguments are documented.
• For new examples, README.md is added to explain the what the example does, the source of the dataset, expected performance on test set and reference to the original paper if applicable
• Check the API doc at http://mxnet-ci-doc.s3-accelerate.dualstack.amazonaws.com/PR-$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

• removed predefined seed from the test
• updated precision of comparison object from float32 to float64

Comments

I confirmed that original test code was causing test to consistently fail even with 100 iterations of running it with the error specified in original issue. After code modification I tested on 100,000 iterations and didn't notice single test failure.
PYTHONPATH=./python/ MXNET_TEST_COUNT=100000 nosetests —verbose tests/python/gpu/test_operator_gpu.py:test_spatial_transformer_with_type was used to run the test.

[marcoabreu]

Hi, thanks a lot for fixing this test!

I'm a bit concerned that changing the data type leads to the code taking a different path and that the issue persists with float32. Could another reviewer please help out here, please? @szha @eric-haibin-lin

[szha]

@marcoabreu this operator has type template in it so it's the same code with different type (and thus numeric precision)

[anirudh2290]

What is the reason for failure with float32 ?

[ddavydenko]

Probably CO, but anyway... My guess is that there are some cumulative arithmetic operations happening where due to round up precision of the result deteriorates slowly. With increased precision of data type this loss is not able to accumulate fast enough to cause surpassing of the threshold of error when comparing results and test data.