packaging: Remove pin for jpeg, numpy

[seemethere]

These may no longer be necessary due to the default anaconda channel
having the necessary packages now.

Signed-off-by: Eli Uriegas eliuriegas@fb.com

[fmassa]

Let's give this a try!

[kylevedder]

This fix impacted pytorch's conda channel for PyTorch 1.10, but not earlier versions. For those who want a work-around for earlier versions, e.g. opencv 3 and pytorch 1.8.1, this worked for me:

conda install pip python=3.8 pytorch::pytorch=1.8.1 cudatoolkit=11.1 cudatoolkit-dev=11.1 opencv=3.4.9 -c pytorch -c nvidia -c anaconda -c conda-forge -c defaults
# Manually specify all the requirements for pytorch except jpeg pegged.
# These requirements were found via `conda search --info torchvision=0.9 -c pytorch` and
# then matching the build against the python version (3.8) and CUDA version (11.1)
conda install ffmpeg>=4.2 libpng numpy >=1.11 pillow >=4.1.1
# Pro gamer move: force install without dependency checking via pip. conda 
# will *not* let you force install despite conflicts (--no-deps and --force don't work), see
# https://github.com/conda/conda/issues/2980
pip install torchvision==0.9.1 --no-deps

If you want a different version of PyTorch this should work, but make sure you update the torchvision version to correspond to the new pytorch version and double check the dependency versions.
NB: be very careful that these installs are not later disrupted by other packages.