Advanced Deep Learning with Python

Building the generator and discriminator

First, we'll implement the build_generator function. The GAN models we've looked at so far started with some sort of latent vector. But here, the generator input is an image from one of the domains and the output is an image from the opposite domain. Following the paper's guidelines, the generator is a U-Net style network. It has a downsampling encoder, an upsampling decoder, and shortcut connections between the corresponding encoder/decoder blocks. We'll start with the build_generator definition:

def build_generator(img: Input) -> Model:

The U-Net downsampling encoder consists of a number of convolutional layers with LeakyReLU activations, followed by InstanceNormalization. The difference between ...