It seems like a bug?

[Luciennnnnnn]

KernelGAN/imresize.py

Line 164 in 7218a37

for channel in range(np.ndim(im)):

If we want to get the size of last dimension(the number of channels), we should use im.shape[-1], np.ndim(im) is the number of dimension of im, in RGB case just equal to im.shape[-1], but for gray image that is wrong.