Add RandomTranslation transform

[daavoo]

Using PIL.Image.transform and PIL.Image.AFFINE

[fmassa]

Hey, thanks for the PR!

Quick question: is this equivalent to crop + scale? I'm not sure how PIL affine behave on the boundaries of the image, could you expand on that?

Thanks!

[daavoo]

Hi! I had never thought about it but after some tests, yes. You can achieve the same results using either translate or crop. It appears that regarding the borders the affine transform behaves just like in crop, it fills with 0s the outside of the image. For example the following transforms are equivalent:

import numpy as np
from torchvision.transforms.functional import to_pil_image

image = to_pil_image((np.random.rand(100, 100, 1)* 255).astype(np.uint8))

horizontal = 10
vertical = 10

t = image.transform(image.size, Image.AFFINE, (1, 0, horizontal, 0, 1, vertical))

c = image.crop((0 + horizontal, 0 + vertical, image.width + horizontal, image.height + vertical))

np.testing.assert_equal(np.array(t), np.array(c))

So I guess that translate could use crop internally or should we still use the affine transform?