port FixedSizeCrop from detection references to prototype transforms

[pmeier]

import torch
from torchvision.prototype import features, transforms


image = features.EncodedImage.from_path("test/assets/fakedata/logos/rgb_pytorch.png").decode()

bounding_boxes = features.BoundingBox(
    [[60, 30, 15, 15]],
    format=features.BoundingBoxFormat.CXCYWH,
    image_size=(100, 100),
    dtype=torch.float,
)

segmentation_masks = torch.zeros((1, 100, 100), dtype=torch.bool)
segmentation_masks[..., 24:36, 55:66] = True
segmentation_masks = features.SegmentationMask(segmentation_masks)

target = dict(
    boxes=bounding_boxes,
    masks=segmentation_masks,
)
sample = image, target

transform = transforms.FixedSizeCrop(size=(40, 40))

torch.manual_seed(0)

for i in range(5):
    transformed_image, transformed_target = transform(sample)
    transformed_bounding_boxes = transformed_target["boxes"]
    transformed_segmentation_masks = transformed_target["masks"]

transform = transforms.FixedSizeCrop(size=(150, 150))

[datumbox]

LGTM, thanks!

[datumbox]

Why do we handle masks here like that?

@vfdev-5 do you need this on your PR?

[vfdev-5]

Yes, we need to apply that on one-hot encoded masks. Added that in my PR

found a potential issue

[datumbox]

LGTM, thanks!