Crop or Pad transform.

[ryanpeach]

So, tensorflow has a transform function called resize_image_with_crop_or_pad. I recently needed this function for my own work. Moreover, it'd be nice to have all of the following:

• RandomPad - Makes an image a certain size by padding but performs random offset on the padding.
• CenterCropOrPad - Replicates Tensorflows function
• RandomCropOrPad - A conditional mix of RandomCrop and RandomPad, guaranteeing size but w random offset.

If I put in a pull request would you be open to these methods?

What are the standards for writing transforms?

[fmassa]

Isn't RandomCrop what you want?

[ryanpeach]

I tried RandomCrop and it threw an error when the desired size was bigger than the source. Is it supposed to be able to pad also?

[fmassa]

Yes, latest version of torchvision supports padding as well, check the link I sent you. There is an option for optionally adding padding if the input is too small

[ryanpeach]

Oh I see in the documentation it is, never mind, maybe the bug was something else then. Maybe there's a way to change the description of this method in the list of transforms so that it mentions padding as well? (Maybe that's the case and I missed it as well).

Sorry, just now learning the library. I'll try to be more careful before making issues next time.

[fmassa]

No worries.
We can definitely improve the documentation. Could you please send a PR?

[ryanpeach]

Yes I'll look into it.

[ryanpeach]

So I also learned that RandomCrop when you set pad_if_needed to True does not RandomPad. AKA, whereas random crop applies a crop at a random offset to the image, pad_if_needed performs a centered padding, which does not randomly place the image in the padded image. The goal for padding should be that the image is in a random spot. Imagine the following:

Imagine the image, whatever size, on an infinite canvas.
Apply a random offset, and a random crop, to that canvas, such that:
if the crop is bigger than the original image, all of the original image remains in the output.
if the crop is smaller than the original image, all of the output is part of the original image.

[ryanpeach]

The easy fix for this is to over-pad the image in RandomCrop.

[fmassa]

The random crop behavior was fixed by #564. @ryanpeach could you also send a PR improving the documentation?

[ryanpeach]

Yes sorry, I have limited time for that but next time I work on my project I'll get to it!

[ekagra-ranjan]

Can be closed via #791.

[soumith]

closed via #791