Generative Models#
A generative model learns to produce new samples that resemble its training data. A generative adversarial network (GAN) does so by training two networks against each other: a generator that turns random noise into images, and a discriminator that tries to tell generated images from real ones.
The notebooks train a deep convolutional GAN (DCGAN) on MNIST for 50 epochs and follow a fixed set of noise vectors through training, so the same samples can be compared from epoch to epoch. They also explain why the GAN losses, unlike the loss of a classifier, say little about sample quality.