Elastic CV-CUDA Backend #9293
Elastic CV-CUDA Backend #9293
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/vision/9293
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: ❌ 5 New FailuresAs of commit 95cbb8b with merge base aa35ca1 ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
justincdavis
force-pushed
the
feat/elastic_cvcuda
branch
from
29df2ef to
be9183a
Compare
| @pytest.mark.needs_cvcuda | ||
| @needs_cuda |
There was a problem hiding this comment.
We can just use @needs_cvcuda, check:
https://github.com/pytorch/vision/pull/9305/changes#diff-9c2dde92db86c123fee225e39b7c1ef96e08a3e79a9dcc9a2d68b21ed51a81d0R6849-R6850
|
|
||
| # checking properties that are not the pixel values | ||
| # see note below on pixel-value differences | ||
| assert_close(result, expected, atol=get_max_value(torch.uint8), rtol=0) |
There was a problem hiding this comment.
get_max_value(torch.uint8) is 255. This assertion will pass for any pixel values (since uint8 range is 0-255). The comment says "checking properties that are not the pixel values" - does this mean it's checking shape/dtype? If so, consider making that explicit:
assert result.shape == expected.shape
assert result.dtype == expected.dtype| # visually, the results are identical, however the underlying computations are different | ||
| # we can define an mae_threshold based on the interpolation mode | ||
| # the primary difference is along the borders where pixels appear to be shifted in location | ||
| # by up to 1, causing potentially up to a diff of 255 on a single pixel | ||
| # this could be because one has fill of 0 and CV-CUDA is shifted and has value with some color | ||
| # thresholds decrease as image size gets larger since shifted pixels are fewer in comparision |
There was a problem hiding this comment.
Need some more investigation about this. @NicolasHug
| # this could be because one has fill of 0 and CV-CUDA is shifted and has value with some color | ||
| # thresholds decrease as image size gets larger since shifted pixels are fewer in comparision | ||
| mae = (expected.float() - result.float()).abs().mean() | ||
| mae_threshold = 30.0 if interpolation is transforms.InterpolationMode.NEAREST else 20.0 |
There was a problem hiding this comment.
Enum comparison is preferred to use "==".
| """ | ||
| Get the CV-CUDA interpolation mode for a given interpolation mode. | ||
| CV-CUDA has the two following differences (evaluated in tests) comapred to TorchVision/PIL: |
2 participants
Summary
Implements Elastic using CV-CUDA backend kernel with
cvcuda.remapTesting
python3 -m pytest test/test_transforms_v2.py::TestElastic