[minor feature] move tranforms.RandomCrop.get_params to functional

[karandwivedi42]

For uniformity.

	# before
    i, j, h, w = transforms.RandomCrop.get_params(x, [224, 224])
    x = TF.crop(i, j, h, w)
	# after
    i, j, h, w = TF.get_random_crop_params(x, [224, 224])
    x = TF.crop(i, j, h, w)

or

x = TF.random_crop(x, [224, 224])

I can PR if it's a go-ahead from any maintainer.

[vfdev-5]

@karandwivedi42 can you explain more, please, what kind of uniformity you speak ?

[karandwivedi42]

Disclaimer: It's a very nit-picking issue and I am fine closing this issue.

Current code:

from torchvision import transforms # have to import this just for get_params
import torchvision.transforms.functional as TF
def transform_train(x):
    i, j, h, w = transforms.RandomCrop.get_params(x, [224, 224])
    x = TF.crop(i, j, h, w)
	...

What I would like:

import torchvision.transforms.functional as TF
def transform_train(x):
    x = TF.random_crop(x, [224, 224])
	...
# When I don't need i,j,h,w but still prefer using functional transforms instead of Compose

[fmassa]

I see your point @karandwivedi42 .
We have tried keeping all randomness outside of the functional interface.
While being a bit more verbose, I think it still gives a bit more flexibility to have the result of random transforms being a separate entity, specially if you are planning on applying the same random transforms to different inputs.

Thus, I'm not sure we would be adding a functional version of random_crop for the moment.
While I'm not 100% happy with the current get_params, I think I still prefer it to adding randomness in the functional interface.

I'm closing the issue, but feel free to comment if you think we should do things differently