NOMRG Show libjpeg version when building

[NicolasHug]

No description provided.

[NicolasHug]

From the logs it looks like linux is using jpeg-9b-h024ee3a_2, macos is using jpeg-9b-he5867d9_2 and windows is using jpeg-9b-hb83a4c4_2.

So they're all 9b versions clearly, but looking at https://anaconda.org/anaconda/jpeg/files?sort=time&sort_order=desc&version=9b the windows binary is 1.5 years older than the linux and the macos ones... which seems a bit fishy. It's possible that the linux/macos versions contains patches or fixes that the windows version doesn't have.

CC @fmassa , any thought?

[NicolasHug]

That being said I still don't know how to verify which version is used by torchvision and by PIL 🤔

[fmassa]

From the logs it looks like linux is using jpeg-9b-h024ee3a_2, macos is using jpeg-9b-he5867d9_2 and windows is using jpeg-9b-hb83a4c4_2.

Thanks for checking this!

the windows binary is 1.5 years older than the linux and the macos ones... which seems a bit fishy

that is interesting. Yeah, that definitely looks suspicious.

That being said I still don't know how to verify which version is used by torchvision and by PIL 🤔

We could probably add a method like

vision/torchvision/csrc/vision.cpp

Lines 25 to 30 in 21426dd

	int64_t cuda_version() {
	#ifdef WITH_CUDA
	return CUDA_VERSION;
	#else
	return -1;
	#endif

that prints the jpeg version.

Something like

torchvision.io._print_jpeg_version()

But there might be an easier way if we ssh into the CI

EDIT: Looks like we might want to use JPEG_LIB_VERSION

[NicolasHug]

Thanks a lot for the help @fmassa . As we suspected, the libjpeg version on Windows CI jobs is different for PIL and for torchvision:

python -c "from PIL import _imaging; print(_imaging.jpeglib_version)"
8.0
python -c "import torch; import torchvision; print(torch.ops.torchvision._jpeg_version())"
90

on the linux job, it is the same:

(/root/project/env) [root@4d1d2d6edd3d project]# python -c "from PIL import _imaging; print(_imaging.jpeglib_version)"
9.0
(/root/project/env) [root@4d1d2d6edd3d project]# python -c "import torch; import torchvision; print(torch.ops.torchvision._jpeg_version())"
90

EDIT: interestingly enough: In both windows and linux we get the exact same pil version

conda list | grep pil
pillow                    8.2.0                    pypi_0    pypi

So the same pil 8.2.0 version uses a different libjpeg backend version for linux and for windows

[fmassa]

Awesome, thanks for the investigation @NicolasHug !

This means that we are consistent on our side. Maybe we can open an issue in PIL to mention that they use different versions of libjpeg depending on the OS?

[NicolasHug]

Oh:

On windows:

python -c "from PIL import features; print(features.check_feature('libjpeg_turbo'))"
True

On linux:

python -c "from PIL import features; print(features.check_feature('libjpeg_turbo'))"
False

[NicolasHug]

Closing this, I followed up in #3913