Hey, I believe there's a minor bug in NLayerDiscriminator: padding is set as padw = int(np.ceil((kw-1.0)/2)), but it should be floor, not ceil (assuming you want to have same padding, similar to pix2pix).

As a result, the output of NLayerDiscriminator is a tiny bit (5 pixels or so) bigger than pix2pix patchgan, but MSEloss/BCEloss will average over all outputs anyways, so I guess it shouldn't matter much performance-wise.

EDIT: fixed it and made a pull request in case you want to change it.

Best regards,
Felix