[add] A PairRandomCrop for both input and target.

[Time1ess]

RandomCrop does not support crop both image and target at same position, each time we call RandomCrop it will generate a new position which is not what we want.
This PairRandomCrop will remember the crop position of the image and use this same position to crop the target(vice versa).
It follows this idea:

Crop in the same worker process is sequential. Crop happens in different workers should not bother each other.

A dict is used to remember the crop position of the image, key is the os.getpid() and value is the position.
So each time we crop a image, PairRandomCrop will generate a new position, when it comes to the target, it will pop and reuse the position.

[fmassa]

Hi,
Thanks for the PR. We are discussing in #230 possible ways of improving the transforms in torchvision. It would be good to have your opinion on that, so that we can discuss on the best approach.

[alykhantejani]

Hi, this functionality can now be achieved in your own dataset using the new functional interface in torchvision.transforms i.e.

from torchvision.transforms import crop, RandomCrop
class MyDataset(Dataset):
    def __init__(self, crop_size):
         super(Dataset, self).__init__()
         self.crop_size = crop_size

    def __getitem__(self, idx):
        img = self.get_img(idx)
        target = self.get_target(idx)
        i, j, h, w = RandomCrop.get_params(img, self.crop_size)
        img = crop(img, i, j, h, w)
        target = crop(img, i, j, h, w)
        return img, target