[MRG] Gromov_Wasserstein2 not performing backward properly on GPU

[ncassereau]

Types of changes

The backpropagation is not working in gromov_wasserstein2 if the given tensors are located on a GPU. This is due to the fact that part of the computation is performed with numpy and the device was forgotten when casting back to torch.

Motivation and context / Related issue

Resolves #351

How has this been tested (if it applies)

PR checklist

• I have read the CONTRIBUTING document.
• The documentation is up-to-date with the changes I made (check build artifacts).
• All tests passed, and additional code has been covered with new tests.
• I have added the PR and Issue fix to the RELEASES.md file.

[codecov]

Codecov Report

Merging #352 (2d045de) into master (1781472) will decrease coverage by 0.03%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #352      +/-   ##
==========================================
- Coverage   93.63%   93.60%   -0.04%     
==========================================
  Files          22       22              
  Lines        5591     5595       +4     
==========================================
+ Hits         5235     5237       +2     
- Misses        356      358       +2     

[tbng]

Thanks for the quick PR! One thing I wonder though is that how to check if the build on Circle CI does run the test with GPU? https://github.com/PythonOT/POT/runs/5389073677?check_suite_focus=true

Because if you do this on the test file POT/test/testgromov.py:

https://github.com/ncassereau-idris/POT/blob/c0b2eda0ccdd6da7e7f5296c9abf4e0e9720ec0e/test/test_gromov.py#L185-L186

if torch.cuda.is_available():
    devices.append(torch.device("cuda"))

This line might not be hit at all during running the test file

[ncassereau]

No you're right, tests performed here do not have access to a GPU. But I ran it on my machine which is equipped with V100s and the test passes. As always with GPU tests, they should be run locally to detect any issue.

[tbng]

No you're right, tests performed here do not have access to a GPU. But I ran it on my machine which is equipped with a V100 and the test passes. As always with GPU tests, they should be run locally to be detected.

The tests passed on my local run as well, thanks.

[rflamary]

Thank you to both of you this was a quickly found and squashed bug, we need to do a release soon