C1_W3_WGAN-GP_Assignment_Critic with Batch Norm or Instance Norm?

In the WGAN-GP paper, the authors emphasized on the essence and benefit of using instance normalization over batch normalization:

However, class Critic(nn.Module) is constructed using torch.nn.BatchNorm2d.

Shouldn’t we rather use torch.nn.InstanceNorm2d in this assignment? If so, could someone kindly explain the primary idea of such normalization in the first place?


1 Like

Good question, @mrgransky. I suspect the author of the assignment just used nn.BatchNorm2d because it was familiar, and good enough for the simple case of the MNIST character example in this assignment. You can try experimenting with nn.InstanceNorm2d and see what you find.

Also, FYI, I noticed you put this in the week 1 folder and put W1 in the title, but the WGAN assignment is in week 3, so I’m updating the title and moving the post to the week 3 folder.