Improve transforms

[Oktai15]

Hi there!

At this moment, torchvision is so poor; in particular, transforms. It really restricts our ability for data augmentation. However, people have already done a few nice libraries for it. For instance:
https://github.com/mdbloice/Augmentor (already has compatibility with PyTorch)
https://github.com/aleju/imgaug

What about contact with authors and merged these great jobs to torchvision.transforms?

[fmassa]

Hi,

Thanks for the feedback!

I'm in the process of adding new functionality to torchvision, to extend it to work with other data types (like bounding boxes).
I'll have a look at the links you provided.

[Oktai15]

@fmassa, okay, thanks, it will cool :) Actually, I strongly recommend to look at this libraries instead of write new code, but, of course, you know better :)

By the way, don't forget about segmentation!

[fmassa]

Don't worry, I won't forget about instance segmentation :-)

[Oktai15]

@fmassa sorry for offtop: I would like to notice that first provided library uses excellent way to transform two or more images. All transforms take list of images rather than just one image like transform in torchvision. For example:

def perform_operation(self, images):
#init random values
for image in images:
    augmented_images.append(do(image))
return augmented_images

I suppose it's great idea ^^

[fmassa]

It's indeed much simpler in some cases, but also more restrictive in more general setups.

For the record, this is something that has been bugging me for a long time, see #9 and #230 for some context.

In many cases, you don't want to pass all transforms to all data (no color augmentation for segmentation masks for example), and the approach you mentioned doesn't easily allows that (without making the API overly complex).

[nnop]

The codes of imaug is somehow far more complicated.

[fmassa]

@nnop yes, I saw imaug and agree that it is very complex. I would rather have a simpler API, even if it requires having to write a bit more code.