Fix ElasticTransform displacement normalization for non-square images

[raimbekovm]

Fixes #9299

The displacement normalization was swapped - dx was divided by height and dy by width, but it should be the other way around since grid coordinates in _create_identity_grid are normalized by their respective dimensions (x by width, y by height).

Tested on a 100x200 image with alpha=[50, 50]:

Current:  dx=1.3px, dy=0.3px (ratio 4.0)
Fixed:    dx=0.7px, dy=0.7px (ratio 1.0)

Equal alpha values should produce ~equal displacement in both directions. The current code gives 4x stronger horizontal displacement on a 2:1 image.

Square images are unaffected (both implementations identical), which explains why this wasn't caught before.

The failing CI is the v1 compatibility check - v1 has the same bug. Happy to fix v1 as well or skip the check for non-square images, whichever you prefer.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/vision/9331

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 8 New Failures

As of commit 512f717 with merge base 1e53952 ():

NEW FAILURES - The following jobs have failed:

• Tests / unittests-linux (3.10, linux.12xlarge, cpu) / linux-job (gh)
  test/test_transforms_v2.py::TestElastic::test_transform[cpu-size2-make_image_pil]
• Tests / unittests-linux (3.10, linux.g5.4xlarge.nvidia.gpu, cuda, 12.6) / linux-job (gh)
  test/test_transforms_v2.py::TestElastic::test_transform[cuda-size2-make_image_pil]
• Tests / unittests-linux (3.11, linux.12xlarge, cpu) / linux-job (gh)
  test/test_transforms_v2.py::TestElastic::test_transform[cpu-size2-make_image_pil]
• Tests / unittests-linux (3.12, linux.12xlarge, cpu) / linux-job (gh)
  test/test_transforms_v2.py::TestElastic::test_transform[cpu-size2-make_image_pil]
• Tests / unittests-linux-cvcuda (3.10, linux.g5.4xlarge.nvidia.gpu, cuda, 12.6) / linux-job (gh)
  test/test_transforms_v2.py::TestElastic::test_transform[cuda-size2-make_image_pil]
• Tests / unittests-windows (3.10, windows.4xlarge, cpu) / windows-job (gh)
  test/test_transforms_v2.py::TestElastic::test_transform[cpu-size2-make_image_pil]
• Tests / unittests-windows (3.11, windows.4xlarge, cpu) / windows-job (gh)
  test/test_transforms_v2.py::TestElastic::test_transform[cpu-size2-make_image_pil]
• Tests / unittests-windows (3.12, windows.4xlarge, cpu) / windows-job (gh)
  test/test_transforms_v2.py::TestElastic::test_transform[cpu-size2-make_image_pil]

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[raimbekovm]

The failing tests are v1 compatibility checks for non-square images. v1 has the same bug so the outputs now differ.

Should I fix v1 as well or skip the v1 check for non-square images?

[zy1git]

@raimbekovm Thank you so much for this PR! And really appreciate your investigations on the test errors!

Since author of the original issue #9299 also submit a PR #9300 , we might just follow up on that PR. However, I will discuss with my tech lead whether it would be good for you two to work together on that PR.

I’ll keep you updated once I have more clarity. Thanks again for your contributions!

[raimbekovm]

No worries, glad I could help with the investigation at least!

[NicolasHug]

Looks like OP followed-up on #9300 already, so I think we'll just go ahead with that one.

@raimbekovm thank you so much for your help with this PR and with the other ones you opened! We really appreciate your help 🙏

[raimbekovm]

Happy to help! Feel free to ping me if you need anything.