Vision: F.to_pil_image for two dimensional arrays

[IssamLaradji]

Would it be possible to have to_pil_image work for 2D arrays and convert them to an L PIL Images?

Currently, it converts 3 dimensional arrays to RGB if they have 3 channels and to L if they have one channel.

However, like in the code below, to_pil_image complains if it is called on a 2D array.

Would it be simpler if 2D arrays are assumed to be single channel images so that to_pil_image converts them to L PIL images ? Thanks.

ipdb> import torchvision.transforms.functional as FT

ipdb> FT.to_pil_image(_points[:,:,0]) 
*** IndexError: tuple index out of range

ipdb> FT.to_pil_image(_points[:,:,:1])
<PIL.Image.Image image mode=L size=640x480 at 0x7FD4947341D0>

[alykhantejani]

@IssamLaradji I guess this could be made possible, but when you convert back to a tensor you will get a [1xHxW] tensor as the generic fromat in torch is batch x channels x height x width for images.

[IssamLaradji]

I see! I guess it's better to keep it consistent this way. Thanks ! :)

[vfdev-5]

@alykhantejani may be torchvision should report something else than IndexError: tuple index out of range ?

--> 108     if npimg.shape[2] == 1:
    109         expected_mode = None
    110         npimg = npimg[:, :, 0]

for example, check npimg.ndim == 3 here and raise about what kind of input we need.