How would the identity loss fit when both generators have different input and output channels, e.g. image colorization? The output and input shape should match for the generator while calculating the loss.
Suppose generator AB goes from 1-channel to 3-channel image, when calculating the identity loss you would need it to output a 1-channel image to compare to the identity image. How do you achieve that?