PatchGAN loss calculated incorrectly (perhaps) about pytorch-cyclegan HOT 3 OPEN

@alexander-soare I was wondering whether if it makes too much of a difference in the results? Have you tried comparing the two?

from pytorch-cyclegan.

@naga-karthik it's been a while since but I think I remember determining that the impact was equivalent to a scalar multiple of the gradient, therefore it only affects the learning rate parameter. Don't take me word for it though as it's been a while.

from pytorch-cyclegan.

@naga-karthik it's been a while since but I think I remember determining that the impact was equivalent to a scalar multiple of the gradient, therefore it only affects the learning rate parameter. Don't take me word for it though as it's been a while.

Late to the party here but you are correct. This really needs to be changed. The current implementation misses the point of the patchGAN paper entirely. If you are going to average the patches before performing mse, you might has well add a dense layer and make a global discriminator because you just obscured all the local information.

Unfortunately, this is not just a matter of scaling the learning rate. This definitely affects the gradient propogation.

from pytorch-cyclegan.