use Image.Resampling namespace for PIL mapping

[kaczmarj]

PIL version 9.1.0 shows a deprecation warning when accessing resampling constants via the Image namespace. The suggested namespace is Image.Resampling. This commit updates _pil_interpolation_to_str to use the Image.Resampling namespace.

/tmp/ipykernel_11959/698124036.py:2: DeprecationWarning: NEAREST is deprecated and will be removed in Pillow 10 (2023-07-01). Use Resampling.NEAREST or Dither.NONE instead.
  Image.NEAREST: 'nearest',
/tmp/ipykernel_11959/698124036.py:3: DeprecationWarning: BILINEAR is deprecated and will be removed in Pillow 10 (2023-07-01). Use Resampling.BILINEAR instead.
  Image.BILINEAR: 'bilinear',
/tmp/ipykernel_11959/698124036.py:4: DeprecationWarning: BICUBIC is deprecated and will be removed in Pillow 10 (2023-07-01). Use Resampling.BICUBIC instead.
  Image.BICUBIC: 'bicubic',
/tmp/ipykernel_11959/698124036.py:5: DeprecationWarning: BOX is deprecated and will be removed in Pillow 10 (2023-07-01). Use Resampling.BOX instead.
  Image.BOX: 'box',
/tmp/ipykernel_11959/698124036.py:6: DeprecationWarning: HAMMING is deprecated and will be removed in Pillow 10 (2023-07-01). Use Resampling.HAMMING instead.
  Image.HAMMING: 'hamming',
/tmp/ipykernel_11959/698124036.py:7: DeprecationWarning: LANCZOS is deprecated and will be removed in Pillow 10 (2023-07-01). Use Resampling.LANCZOS instead.
  Image.LANCZOS: 'lanczos',

[rwightman]

@kaczmarj how far back in PIL versions does the new enum exist?

[kaczmarj]

great question -- the new enum was added in python-pillow/Pillow#5954 (merged february 18 2022). only the most recent version of pillow (version 9.1.0) has this enum. torchvision requires pillow>=5.3.0, so this would break functionality on older pillow versions.

torchvision recently merged a pull request that used the new enums only if the pillow version was above 9.1.

what would you like to do? perhaps we could follow torchvision's example and use the enum only if the pillow version is above 9.1.

[adamjstewart]

I'm surprised to discover that timm imports numpy and pillow despite neither being listed as dependencies in setup.py. Even if these are automatically brought in by PyTorch/torchvision, they should still be explicitly listed as direct dependencies. For example, if one of these dependencies is removed (unlikely), or if the versions of pillow supported by torchvision and timm differ (already the case) then this would prevent a lot of issues.

[rwightman]

@adamjstewart @kaczmarj there's no difference in version of pillow supported from a compatbility standpoint, it's a deprecation warning. 9.1 is really new and there isn't even a Pillow-SIMD release of it yet(I only use that, normal Pillow is too slow to use for training).

I will put up with the warning for now, doing string parsing version checks is notoriously problematic. Checking for the existence of the new enum type would be more robust and I'd accept the PR if it used that approach.

[rwightman]

@adamjstewart @kaczmarj looking at the torchvision bug I just filed, there's the ans why you don't do this (string checks are hard to get right in all cases even when they appear straightforward)

[kaczmarj]

@rwightman - how prescient! i'll try to submit a pr this weekend that checks for the existence of the new resampling enum and uses it only if it exists.

[rwightman]

@rwightman - how prescient! i'll try to submit a pr this weekend that checks for the existence of the new resampling enum and uses it only if it exists.

Awesome, thx

[adamjstewart]

@rwightman thanks, I didn't even think about that, I made the same mistake in my library, will fix that.

[adamjstewart]

@rwightman - how prescient! i'll try to submit a pr this weekend that checks for the existence of the new resampling enum and uses it only if it exists.

This will prevent things from breaking, but it won't silence the warning. Personally, I would use packaging.version.Version like pytorch-lightning does: https://github.com/PyTorchLightning/pytorch-lightning/blob/master/pytorch_lightning/utilities/imports.py

[adamjstewart]

@rwightman corrected me, both are valid approaches.

[kaczmarj]

@rwightman - i added a commit to check for Image.Resampling. if this looks good to merge, i suggest squashing and merging to ignore the commit. i would squash myself, but then i'd have to force push and that doesn't feel right.

[adamjstewart]

Should this also be done in timm/data/auto_augment.py?

../../../anaconda/envs/adam-py37-min/lib/python3.7/site-packages/timm/data/auto_augment.py:41
  /anaconda/envs/adam-py37-min/lib/python3.7/site-packages/timm/data/auto_augment.py:41: DeprecationWarning: BILINEAR is deprecated and will be removed in Pillow 10 (2023-07-01). Use Resampling.BILINEAR instead.
    _RANDOM_INTERPOLATION = (Image.BILINEAR, Image.BICUBIC)

../../../anaconda/envs/adam-py37-min/lib/python3.7/site-packages/timm/data/auto_augment.py:41
  /anaconda/envs/adam-py37-min/lib/python3.7/site-packages/timm/data/auto_augment.py:41: DeprecationWarning: BICUBIC is deprecated and will be removed in Pillow 10 (2023-07-01). Use Resampling.BICUBIC instead.
    _RANDOM_INTERPOLATION = (Image.BILINEAR, Image.BICUBIC)