WGAN's underlying loss function

My question is about the WGAN loss:

WGAN loss:
A GAN loss that corresponds with image quality
A GAN loss that converges (decreasing loss actually means something), so you can actually tune your hyperparameters with something other than voodoo

How do we calculate this WGAN loss?

As indicated below it isn’t what we actually optimize because that doesn’t have the aforementioned properties. Is it W_1 that has those properties? What is W_1? How to calculate it?

1 Like