Advanced Deep Learning with Python

Deep Convolutional GAN

In this section, we'll implement the Deep Convolutional GAN (DCGAN, Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks, https://arxiv.rg/abs/1511.06434). In the original GAN framework proposal, the authors only used fully-connected networks. In contrast, in DCGANs both the generator and the discriminator are CNNs. They have some constraints that help stabilize the training process. You can think of these as general guidelines for GAN training and not just for DCGANs:

• The discriminator uses strided convolutions instead of pooling layers.
• The generator uses transpose convolutions to upsample the latent vector, , to the size of the generated image.
• Both networks use batch normalization. ...