[Conda] pillow 9.5.0 incompatible with any torchvision version due to switch from jpeg to libjpeg-turbo

[junpuf]

🐛 Describe the bug

Pillow no longer compatible with torchvision starting from version 9.5.0 due to issues discussed in conda-forge/libjpeg-turbo-feedstock#28

The core issue was that the jpeg conda package was "deprecated" by the conda-forge community and they want all users of jpeg to migrate to libjpeg-turbo.

Versions

all torchvision conda package

[malfet]

There are no pillow-9.5.0 on Anaconda https://anaconda.org/anaconda/pillow is there?

[junpuf]

The default channels are currently ok. But the conda-forge channel has already been updated. Since the conda-forge channel is more widely adopted (29M downloads) compare to anaconda (1.8M downloads), the impact is there.

[rymndhng]

We're seeing this impact our conda environment where we cannot upgrade dependencies b/c we have deps from conda-forge that depend on libjpeg-turbo and torchvision from the pytorch channel that depends on jpeg.

I'm wondering if we could unblock this scenario by removing jpeg as an explcit run dependency. Instead, could the torchvision package document & tell users to install jpeg or libjpeg-turbo into the environment themselves.

[NicolasHug]

Quick update on this: for conda builds, we're going to try to make torchvision depend on libjpeg-turbo instead of jpeg (like PIL). It's something we've been considering for a while anyway.

[rymndhng]

@NicolasHug Thanks for moving this along!

Do you have a timeline for when a new build will be pushed to conda?

[NicolasHug]

I believe the next release is planned for early October

[NicolasHug]

Hi all, this should be fixed now with the 0.16 release as we now ship with libjpeg-turbo. I'll close the issue, but LMK if there are still problems. Thanks