[Feature Request] Pass multiple images to randomCrop, randomFlip, etc.

[skief]

Recently I've worked with some semantic segmentation algorithms and there I've needed to randomly crop the images and apply the same crop to the mask/label image(to make sure that they are still aligned) and I couldn't do this with the current torchvision transformations.
Therefore I wanted to ask if anybody else would be interested in an implementation of these transforms for multiple images. If you are interested in a feature like this I could implement this :)

[Naman-ntc]

Why don't you just use functional transforms? In my opinion it gives you better control as well for augmenting targets instead of passing to transforms.

[Randl]

I believe joint transformation of input and target is pretty common, and thus probably worth inclusion in vision

[fmassa]

Indeed, they are very common and would deserve a place here.

Unfortunately, all the alternatives I've seem / came up with were not generic nor good enough to so that I could add it to torchvision.

In the end, it just seemed easier to have the transforms be written by the user, leveraging the functional interface. That's the most generic way, even though it is a bit more verbose.

[sotte]

This feature request keeps popping up. I think we should improve the docs and the tutorial (especially the data loader tutorial) to cover this case better.

[fmassa]

@sotte that's a great point! Would you be willing to improve the documentation?
Maybe not in the beginner's tutorial, but maybe in an intermediate tutorial?

[sotte]

@fmassa sure, I'll try to do it on the weekend.

Maybe I would add a few words to the api docs of functional transforms along the lines "functional transforms give you fine grained control...yada yada") plus a short tutorial for for functional transforms. Sounds good?

[fmassa]

@sotte yes, please!

[sotte]

I think (and I'm clearly biased) that the ddb145b / #602 improved the multiple images transformation situation. Maybe we don't need the tutorial.

@skief you asked the initial question. What do you think? Is the improvement good enough or is a tutorial still needed?

[fmassa]

I think the current situation is better now than what it was before, thanks to the note from @sotte
As I mentioned before, I think it's best for more complex cases to be handled using the functional interface, so I'm closing this issue now. But please fell free to comment if you disagree.