Recipe improvements - Githubissues

conda-forge / torchvision-feedstock

A conda-smithy repository for torchvision.

BSD 3-Clause "New" or "Revised" License

2 stars 16 forks source link

Recipe improvements #22

Open h-vetinari opened 3 years ago

h-vetinari commented 3 years ago

Some open points remaining from (OP of) #12:

[x] build with cuda
[x] build with ffmpeg support
[ ] turn into multi-output recipe (e.g. with/without ffmpeg/image support)?
[ ] package c++ lib separately?
[x] build for osx
[ ] build for win
[ ] double-check PTX vs. cuda-arches, see this comment
[ ] depend on pillow-simd, see https://github.com/conda-forge/pillow-feedstock/issues/103
[ ] Split test files into a separate packages to reduce download size from 7MB to 1MB

CC @hmaarrfk @isuruf

hmaarrfk commented 3 years ago

Thoughts:

build with ffmpeg support Sure
turn into multi-output recipe (e.g. with/without ffmpeg/image support)? Yes
package c++ lib separately? Maybe just create a split package and have the lib package not depend on python https://github.com/conda-forge/opencv-feedstock/blob/master/recipe/meta.yaml#L1
build for osx Thanks!!!!
build for win

Didn't know this was dropped but good to keep track!

h-vetinari commented 3 years ago

package c++ lib separately?

Maybe just create a split package and have the lib package not depend on python https://github.com/conda-forge/opencv-feedstock/blob/master/recipe/meta.yaml#L1

That's (another aspect of) what I meant by making the recipe have multiple outputs

build for osx

Thanks!!!!

That credit belongs to @isuruf

build for win

Didn't know this was dropped but good to keep track!

When I picked up the feedstock again, linux was the only build variant... Not sure how much effort this would be, but the pytorch channel has windows builds (even with cuda).

hmaarrfk commented 3 years ago

Also cross referencing this issue: https://github.com/conda-forge/pytorch-cpu-feedstock/issues/40#issuecomment-800611656

Tobias-Fischer commented 2 years ago

It would also be great if the correct version of pytorch is being pinned. Currently there can be a mismatch between the pytorch and torchvision versions. The compatibility table is shown in their GitHub repo

hmaarrfk commented 2 years ago

As test files are being copied into the recipe archive, the many binary files are actually increasing the size of said archive: