Not all prototype kernels support arbitrary batch sizes

[pmeier]

There are currently five image kernels that don't support arbitrary batch sizes:

• perspective
  

  vision/test/prototype_transforms_kernel_infos.py

  Lines 906 to 907 in 7f4c55b

  	# FIXME: kernel should support arbitrary batch sizes
  	extra_dims=[(), (4,)],

  
  

  vision/test/prototype_transforms_kernel_infos.py

  Lines 930 to 931 in 7f4c55b

  	# FIXME: kernel should support arbitrary batch sizes
  	extra_dims=[(), (4,)],

• elastic
  

  vision/test/prototype_transforms_kernel_infos.py

  Lines 974 to 975 in 7f4c55b

  	# FIXME: kernel should support arbitrary batch sizes
  	extra_dims=[(), (4,)],

• gaussian_blur
  

  vision/test/prototype_transforms_kernel_infos.py

  Lines 1125 to 1126 in 7f4c55b

  	# FIXME: kernel should support arbitrary batch sizes
  	extra_dims=[(), (4,)],

• equalize
  

  vision/test/prototype_transforms_kernel_infos.py

  Lines 1152 to 1153 in 7f4c55b

  	# FIXME: kernel should support arbitrary batch sizes
  	extra_dims=[(), (4,)],

• adjust_sharpness
  

  vision/test/prototype_transforms_kernel_infos.py

  Lines 1302 to 1303 in 7f4c55b

  	# FIXME: kernel should support arbitrary batch sizes
  	extra_dims=[(), (4,)],

cc @vfdev-5 @datumbox @bjuncek

[pmeier]

The situation is a little more nuanced:

• All kernel from above except adjust_sharpness don't support 5d+ inputs or a 4d inputs with a degenerate batch dimension, i.e. (0, C, H, W)

• adjust_sharpness also doesn't support 5d+ inputs, but does support 4d inputs with a degenerate batch dimension iff the input only has a single channel, i.e. (0, 1, H, W). In addition, if the shortcut

  vision/torchvision/transforms/functional_tensor.py

  Lines 842 to 843 in 7eb5d7f

  	if img.size(-1) <= 2 or img.size(-2) <= 2:
  	return img

  is triggered, arbitrary batch sizes are supported.

[pmeier]

#6667 adds xfails for these cases.

Edit: The PR ended up fixing equalize and adjust_sharpness.