port test_accimage_xxx in test_transforms to pytest #3985
Conversation
test/test_transforms.py
Outdated
| assert np.abs((expected_output - output).mean()) < 1e-3 | ||
| assert (expected_output - output).var() < 1e-5 | ||
| # note the high absolute tolerance | ||
| self.assertTrue(np.allclose(output.numpy(), expected_output.numpy(), atol=5e-2)) |
There was a problem hiding this comment.
I think this should be torch.testing.assert_close(output.numpy(), expected_output.numpy())\maybe you'll need to adjust atol and rtol too
There was a problem hiding this comment.
Though I've made the changes as suggested, since I don't have support for accimage these tests are skipped on my local machine and I'm not sure how to verify the atol and rtol values. Any ideas?
There was a problem hiding this comment.
You did well using the numpy's default for rtol. Looks like the CI doesn't even run those tests anyway... that's not good lol.
There was a problem hiding this comment.
Exactly! I had the same doubt and I checked that CI was also skipping them so I used the numpy default rtol and atol as before.
Reviewed By: fmassa Differential Revision: D29097728 fbshipit-source-id: 14f243179f6617fc0ede63ee5bb68f94954a62b9
Closed #3982 in favour of this PR. This fixes the Group B in #3945.
Extremely sorry for the confusion @NicolasHug.