jpeg <0.0.0a run constrain breaks pillow torchvision compatibility

[junpuf]

Comment:

Hi @hmaarrfk and all

There is an compatibility issue introduced between libjpeg-turbo 2.1.5.1 (via pillow 9.5.0) and any torchvison version due to the addition of jpeg <0.0.0a run constrain in 6cee832.

May i ask the rational of this breaking change in an update between 2.1.4 build 0 to 2.1.4 build 1?

Also, further investigation found that jpeg <0.0.0a doesn't map to any existing jpeg conda package versions in conda-forge, which means this constrain will eliminate the compatibility between libjpeg-turbo and any other conda package that depend on jpeg.

The following packages are incompatible
├─ pillow 9.5.0**  is installable and it requires
│  └─ libjpeg-turbo >=2.1.5.1,<3.0a0 , which requires
│     └─ jpeg <0.0.0a , which can be installed;
└─ torchvision   is uninstallable because there are no viable options
   ├─ torchvision [0.15.1|0.15.2] would require
   │  └─ jpeg  , which conflicts with any installable versions previously reported;
   └─ torchvision [0.10.0|0.10.1|...|0.9.1] conflicts with any installable versions previously reported.

Reference:

• Pillow's update on libjpeg-turbo Updated libjpeg-turbo to 2.1.5.1 python-pillow/Pillow#6929

[hmaarrfk]

Please see conda-forge/conda-forge.github.io#673

The last comment is me mostly admitting the difficulty in fixing the problem you are hitting.

My solution has been to

conda uninstall --force --offline --yes jpeg
mamba update --all --yes

generally gets you through.

conda-forge/conda-forge.github.io#1894

[junpuf]

Thanks for providing the insight @hmaarrfk. Forcing uninstall may work for conda environment users. However, for software maintainers who currently depend on jpeg like torchvision , they will have to migrate to libjpeg-turbo and deprecate jpeg.

I read through conda-forge/conda-forge.github.io#673 and understand that there are valid reasons to force switch to libjpeg-turbo and deprecate jpeg. But like you expected, it is a painful experience for those who depend on jpeg. I will create a separate issue to torchvision maintainers and ask them to switch to libjpeg-turbo.

[hmaarrfk]

software maintainers

As I understand it torchvision, the recipe you linked to, is not dependent on the conda-forge ecosystem.
https://github.com/pytorch/vision/blob/main/packaging/torchvision/conda_build_config.yaml#L2

while there used to be a big alignment between defaults and conda-forge in the past, the two channels are, today, largely incompatible.

The torchvision package from conda-forge should be compatible
https://github.com/conda-forge/torchvision-feedstock/

on windows, we could use help in the form of packing, or donation of your "time" to manually build windows from time to time.

see the many attempts at building windows:
https://github.com/conda-forge/pytorch-cpu-feedstock/pulls

[hmaarrfk]

I forgot to say, thank you for being so understanding of the tradeoffs.