Hands-On Artificial Intelligence for Beginners

Generator network

You can think of the generator portion of the GAN as a reverse convolutional neural network. Like a VAE, it uses generic normal distribution, the only difference being that it up samples the distribution to form an image. This distribution represents our prior, and is updated during training as the GAN improves at producing images that the discriminator is unable to determine whether they are fake.

In between each layer, we utilize a ReLu activation function and batch _normalization to stabilize each layer's outputs. As the discriminator starts inspecting the outputs of generator, generator will continually adjust the distribution from which it's drawing to closely match the target distribution. The code will look fairly ...