remove fill blending for bilinear affine

[pmeier]

Fixes #8083 and addresses the corresponding part in #6517.

cc @vfdev-5

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/vision/8098

• 📄 Preview Python docs built from this PR

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

❌ 17 New Failures, 20 Unrelated Failures

As of commit c6e81a4 with merge base f69eee6 ():

NEW FAILURES - The following jobs have failed:

• CMake / linux (linux.12xlarge, cpu) / linux-job (gh)
• CMake / linux (linux.g5.4xlarge.nvidia.gpu, cuda, 11.8) / linux-job (gh)
• Docs / build / linux-job (gh)
• Lint / python-types / linux-job (gh)
• Prototype tests on Linux / unittests-prototype (3.10, linux.12xlarge, cpu) / linux-job (gh)
• Prototype tests on Linux / unittests-prototype (3.11, linux.12xlarge, cpu) / linux-job (gh)
• Prototype tests on Linux / unittests-prototype (3.8, linux.12xlarge, cpu) / linux-job (gh)
• Prototype tests on Linux / unittests-prototype (3.8, linux.g5.4xlarge.nvidia.gpu, cuda, 11.8) / linux-job (gh)
• Prototype tests on Linux / unittests-prototype (3.9, linux.12xlarge, cpu) / linux-job (gh)
• Tests / onnx / linux-job (gh)
• Tests / unittests-extended / linux-job (gh)
• Tests / unittests-linux (3.10, linux.12xlarge, cpu) / linux-job (gh)
• Tests / unittests-linux (3.11, linux.12xlarge, cpu) / linux-job (gh)
• Tests / unittests-linux (3.8, linux.12xlarge, cpu) / linux-job (gh)
• Tests / unittests-linux (3.8, linux.g5.4xlarge.nvidia.gpu, cuda, 11.8) / linux-job (gh)
• Tests / unittests-linux (3.9, linux.12xlarge, cpu) / linux-job (gh)
• Tests / unittests-windows (3.8, windows.g5.4xlarge.nvidia.gpu, cuda, 11.8) / windows-job (gh)

FLAKY - The following jobs failed but were likely due to flakiness present on trunk:

• Build Linux Conda / pytorch/vision / conda-py3_8-cpu (gh)
• Build Linux Conda / pytorch/vision / conda-py3_8-cuda11_8 (gh)
• Build Linux Conda / pytorch/vision / conda-py3_8-cuda12_1 (gh)
• Build M1 Conda / pytorch/vision / conda-py3_8-cpu (gh)
• Build Macos Conda / pytorch/vision / conda-py3_8-cpu (gh)
• Build Windows Conda / pytorch/vision / conda-py3_8-cpu (gh)
• Build Windows Conda / pytorch/vision / conda-py3_8-cuda11_8 (gh)
• Build Windows Conda / pytorch/vision / conda-py3_8-cuda12_1 (gh)
• CMake / macos (macos-12) / macos-job (gh)
• CMake / macos (macos-m1-12) / macos-job (gh)
• CMake / windows (windows.4xlarge, cpu) / windows-job (gh)
• Tests / unittests-macos (3.10, macos-12) / macos-job (gh)
• Tests / unittests-macos (3.11, macos-12) / macos-job (gh)
• Tests / unittests-macos (3.8, macos-12) / macos-job (gh)
• Tests / unittests-macos (3.8, macos-m1-12) / macos-job (gh)
• Tests / unittests-macos (3.9, macos-12) / macos-job (gh)
• Tests / unittests-windows (3.10, windows.4xlarge, cpu) / windows-job (gh)
• Tests / unittests-windows (3.11, windows.4xlarge, cpu) / windows-job (gh)
• Tests / unittests-windows (3.8, windows.4xlarge, cpu) / windows-job (gh)
• Tests / unittests-windows (3.9, windows.4xlarge, cpu) / windows-job (gh)

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

[NicolasHug]

This is only equivalent to the 'bilinear' block if mask was a boolean mask, right?

Is that the case? And if it is, why was there even a distinction between bilinear and nearest mode in the first place?

[pmeier]

This is only equivalent to the 'bilinear' block if mask was a boolean mask, right?

Yes.

Is that the case?

No.

And if it is, why was there even a distinction between bilinear and nearest mode in the first place?

I would need to dig, but maybe @vfdev-5 knows? My guess is that the blend strategy we implemented was to have a smooth transition from fill to image. And that works well if we fill with zeros, but creates a shadow artifact if we use a different fill color with bilinear interpolation. See #8083.

Meaning, this PR is BC breaking, but I consider it a bug fix since the blending behavior seems to be "nice to have" while the shadow is an actual issue.

[vfdev-5]

I think it was an incorrect implementation. At the moment of PR (#2904), I thought it could be ok as assumption to linearly interpolate around the boundaries: #2904 (comment)