How do I get RandomRotation to work on 'L' images

vision
1

I’m trying to RandomRotate a greyscale images using transform. But it won’t let me!
can someone please help?
Note: It works fine with RGB based tensors

Code and error message below

> img = torch.randn(1,224,224)
> 
> tfms = Compose([ToPILImage(),RandomRotation(20)])
> 
> tfms(img)
> .
> .
> .

TypeError: function takes exactly 1 argument (3 given)

2

It looks like this might be a bug in torchvision.transforms.functional.rotate(), but you can work around it by defining your RandomRotation transform like so: RandomRotation(20, fill=(0,)). The problem is that by default rotate() passes to PIL the three-channel fill value (0, 0, 0) which of course breaks on one-channel images.

3

It looks like this was fixed in torchvision master seven days ago.

4

it’s worth visiting new kornia.augmentation API
https://kornia.readthedocs.io/en/latest/augmentation.html

same transforms as torchvision but for torch tensors