generative-adversarial-networks experimenting with GANs on vast.ai notebook contains code for training a GAN on aligned celeba images architecture includes: hinge loss spectral norm on all layers dense self attention exponential moving average two timescale update rule