I wonder why Pytorch uses PIL not the cv2

1

I wonder why Pytorch uses PIL not the cv2.
As the article(https://www.kaggle.com/vfdev5/pil-vs-opencv) says, cv2 is three times faster than PIL.
I did an experiment and I really think so.
But why does Pytorch use PIL?

3 Likes
2

If PyTorch can remove it’s dependency on PIL and use OpenCV 3 instead, it would help me work with multi-channel images that are greater than 3-channels. At the moment, if you pass a 10-ch image or 10-ch mask, PIL truncates it to a 3-ch image.

See these issues:


1 Like
3

The only thing that uses PIL is the torchvision package. Rewriting it to support cv2 probably it’s not such a huge effort, since you just have to replace functions call here and there - if you have experience with cv2 you could probably try that yourself!

2 Likes
4

If you want a performance increase, you could install Pillow-SIMD instead of Pillow. This is a drop-in replacement without any code changes.

4 Likes
5

Thanks for the heads up, I’ll try that out.

6

If you finish it, will you make it public like on github? That will be beneficial to many others.

7

Yes, I’ll will. I ran into this issue with PIL last week, trying to work with multi-channel images. Replacing the TorchVision’s dependency on PIL sounds like the right thing to do. It looks like TorchVision library already supports multiple imaging backends, so it should be a matter of adding a cv2 backend.

I’ll let you know once I’m done.

3 Likes
8

Thank you very much!

9

I have rewritren the “transforms” in torchvision package with cv2.

4 Likes
11

I also have the same question.

12

May be because cv2 is a heavier module, and PIL is a lighter module. Not going for a heaver module, if we could get our work done just by a lighter one.

2 Likes
14

With the recent addition of torchvision.io, now might be a good time to consider adding an option to use opencv as the image backend.