Faster ElasticTransform using separable Gaussian filters

[Callidior]

Fix

Fixes #9577 by using separable Gaussian filters for smoothing the displacement field in torchvision.transforms.v2.ElasticTransform.

Runtime Comparison of ElasticTransform.make_params

Image Size 1008 x 1008

Sigma	Old Implementation	New Implementation
1.0	19 ms	35 ms
3.0	28 ms	35 ms
5.0	49 ms	35 ms
8.0	96 ms	35 ms
10.0	142 ms	36 ms
20.0	575 ms	37 ms
30.0	1,210 ms	43 ms
50.0	3,310 ms	47 ms
100.0	31,900 ms	77 ms

Image Size 256 x 256

Sigma	Old Implementation	New Implementation
1.0	0.6 ms	0.7 ms
3.0	1.4 ms	0.7 ms
5.0	2.7 ms	0.7 ms
10.0	9 ms	1 ms
20.0	34 ms	1 ms
30.0	82 ms	1 ms
50.0	276 ms	2 ms

Tests

Larger image sizes and different variants for sigma have been added to the test test_transforms_v2.py::TestElastic::test_transform. Run with:

pytest test/test_transforms_v2.py::TestElastic::test_transform

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/vision/9578

• 📄 Preview Python docs built from this PR

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

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