/black
thanks, for the typing error, a simple workaround would be removing -> Union[np.ndarray, torch.Tensor] in the newly added classes

Thanks for comments. I have:

• Fixed the typing error (I think!)
• Used @rijobro's cleaner code for generating the channels. I've decided to stick with the spatial_channels argument starting from 1 rather than 0 but can change if you feel stronlg they should start from 0. I looked for some precedent in other transforms but couldn't find any.
• Finally, I've renamed the transform to AddCoordinateChannels, and provided a bit of motivation for when you might want to use it in the docs. It isn't correct to call it CoordConv because CoordConv changes the convolution layer itself to append spatial channels for each convolution, while my code only appends these channels to the input data as a transform. I think my version has particular use for medical imaging examples where e.g. you might train on patches of 3D volumes that are all aligned to a common space. Appending the coordinate frame to the patch might allow the network to understand where the patch is in the image - CoordConv only appends the local reference coordinates of the input feature map.

Thanks, looks great. Since the output is a concatenation of the channels and the original image, in the unit test it might be worth checking that the first part of the image matches the original image?

I suggested another update to the calculation of the coords (sorry for playing a bit of code golf!) if you want to add that too.

Thanks - I've taken your suggestion for meshgrid (had to add indexing='ij' to get the first two dimensions right), and added that unit test.

Looks good to me, happy to merge once the checks are passing!