Fill arg and _apply_grid_transform improvements

[vfdev-5]

Few years ago we introduced non-const fill value handling in _apply_grid_transform using mask approach:

vision/torchvision/transforms/functional_tensor.py

Lines 550 to 568 in 0d69e35

	# Append a dummy mask for customized fill colors, should be faster than grid_sample() twice
	if fill is not None:
	dummy = torch.ones((img.shape[0], 1, img.shape[2], img.shape[3]), dtype=img.dtype, device=img.device)
	img = torch.cat((img, dummy), dim=1)
	img = grid_sample(img, grid, mode=mode, padding_mode="zeros", align_corners=False)
	# Fill with required color
	if fill is not None:
	mask = img[:, -1:, :, :] # N * 1 * H * W
	img = img[:, :-1, :, :] # N * C * H * W
	mask = mask.expand_as(img)
	len_fill = len(fill) if isinstance(fill, (tuple, list)) else 1
	fill_img = torch.tensor(fill, dtype=img.dtype, device=img.device).view(1, len_fill, 1, 1).expand_as(img)
	if mode == "nearest":
	mask = mask < 0.5
	img[mask] = fill_img[mask]
	else: # 'bilinear'
	img = img * mask + (1.0 - mask) * fill_img

There are few minor problems with this approach:

1. if we pass fill = [0.0, ], we would expect to have a similar result as fill=None. This is not exactly true for bilinear interpolation mode where we do linear interpolation:
   

   vision/torchvision/transforms/functional_tensor.py

   Lines 567 to 568 in 0d69e35

   	else: # 'bilinear'
   	img = img * mask + (1.0 - mask) * fill_img

Most probably, we would like to skip fill_img creation for all fill values that has sum(fill) == 0 as grid_sample pads with zeros.

- if fill is not None:
+ if fill is not None and sum(fill) > 0:

2. Linear fill_img and img interpolation may be replaced by directly applying a mask:

         mask = mask < 0.9999
         img[mask] = fill_img[mask] 

That would match better PIL Image behaviour.

vision/torchvision/transforms/functional_tensor.py

Lines 567 to 568 in 0d69e35

	else: # 'bilinear'
	img = img * mask + (1.0 - mask) * fill_img

cc @datumbox

[pmeier]

Since we have another report in #8083, do we want to tackle this? IMO, we should just align the two branches

vision/torchvision/transforms/v2/functional/_geometry.py

Lines 588 to 594 in f69eee6

	if mode == "nearest":
	bool_mask = mask < 0.5
	float_img[bool_mask] = fill_img.expand_as(float_img)[bool_mask]
	else: # 'bilinear'
	# The following is mathematically equivalent to:
	# img * mask + (1.0 - mask) * fill = img * mask - fill * mask + fill = mask * (img - fill) + fill
	float_img = float_img.sub_(fill_img).mul_(mask).add_(fill_img)

with something like

bool_mask = mask < 1
float_img[bool_mask] = fill_img.expand_as(float_img)[bool_mask] 

This removes the blending and in turn the "shadow" for bilinear interpolation. Plus, this is equivalent for nearest interpolation, since the mask in that case only contains 0.0 and 1.0 entries.

[vfdev-5]

@pmeier the value 0.9999 for mask was sort of on purpose. In the description example affine rotation by 50 degrees with bilinear mode creates a rotated mask with unique values:

tensor([0.00000000, 0.02883029, 0.02883148, 0.10955429, 0.10955477, 0.11125469,
         0.11125565, 0.19197845, 0.19197917, 0.19367909, 0.19367981, 0.27440262,
         0.27440357, 0.35512805, 0.35512924, 0.35682678, 0.35682797, 0.43755341,
         0.43755519, 0.43925095, 0.43925512, 0.51997960, 0.51998138, 0.60240537,
         0.60240555, 0.68312985, 0.68313217, 0.68482971, 0.68482977, 0.76555562,
         0.76555634, 0.76725388, 0.76725554, 0.84798002, 0.84798050, 0.92870331,
         0.92870587, 0.93040466, 0.93040580, 0.99999994, 1.00000000]))

and 0.99999994 can appear inside the mask:

plt.imshow(((mask > 0.999) & (mask < 1.0))[0, 0, ...], interpolation="none")

so, using mask < 1 gives:

[amanikiruga]

this hasn't been fixed