Gan with Feature Loss giving NaNs

Hey, everybody!

I’m trying to use the GAN and feature loss shown as shown in Lesson 7 together for super-resolution. Both the generator and critic function well, but I cannot get the GAN to work, it keeps giving me NaNs for the losses as shown in the image below.

The only difference currently with the setup are the weights and Adam optimizer betas, as I was trying to see if the Deoldify parameters used would work. Would appreciate any input on how I should proceed with debugging this. This is my first time trying to get a GAN to work, so I do not know a lot about them yet.
Thanks in advance!

1 Like