Thanks. I guess that is one variation that you could use especially if you could set a metric when the discriminator should freeze weights. In this example, both generator and discriminator are learning simultaneously. I also observed other variations wherein the discriminator is pre-trained before it is plugged to the generator-discriminator model.