GAN is just a Generator followed by (or connected to) a discriminator, so when you want to train the generator, you fix the discriminator (aka. don’t apply the delta weights on the discriminator part), and you backpropagate through discriminator to the generator as if you had 2 regular layers following each others.