Fixed warning in torchvision.models.detection.transforms._resize_image_and_masks.

[maxfrei750]

torchvision.models.detection.transforms._resize_image_and_masks raises the following warning, when being run with torch>=1.6.0:

UserWarning: The default behavior for interpolate/upsample with float scale_factor changed in 1.6.0 to align with other frameworks/libraries, and now uses scale_factor directly, instead of relying on the computed output size. If you wish to restore the old behavior, please set recompute_scale_factor=True. See the documentation of nn.Upsample for details.

This was already mentioned in #2154, but apparently not fixed yet.

This PR fixes the warning by restoring the original behavior.

EDIT: I think, originally, @java-abhinav07 wanted to create a PR, so he might have some insights on whether the fix is good.

[datumbox]

@maxfrei750 Thanks for the PR.

If this is to be addressed like this we will probably need to also aligning the onnx method:

vision/torchvision/models/detection/transform.py

Line 30 in c864171

mask = F.interpolate(mask[:, None].float(), scale_factor=scale_factor)[:, 0].byte()

@fmassa Could you also have a look to confirm if we want to maintain the old behaviour here?

[fmassa]

Thanks for the PR!

@datumbox I would say that this PR keeps the old behavior (but without the warning), so this is good.
If we wanted to switch in the future to the new default behavior (recompute_scale_factor=False), we would need to perform a validation of the results to see if it affects the results a bit (I don't think it should affect them in a significant manner, but would still be good to validate)

@maxfrei750 can you address the other comment from @datumbox so that we can get this PR merged?

[maxfrei750]

@fmassa Thanks for the review.

I just made the requested change.

[fmassa]

Thanks!