Training GAN can be baffling

Training GAN can be baffling. For example, the generator and the discriminator just don't "learn" at the same scale sometimes. Would you try to balance the generator loss and discriminator loss by hand? Soumith Chintala ( @ FAIR ) put together this...