Hands-On Generative Adversarial Networks with Keras

Increasing variation using minibatch standard deviation

In their paper, Tero Karras et al. propose a replacement for minibatch discrimination that uses no learnable parameters or new hyperparameters. Minibatch discrimination adds a layer toward the end of the discriminator that computes statistics across the minibatch. These obtained statistics are concatenated to the discriminator's output, allowing the discriminator to use the statistics explicitly.

Tero's replacement is simple and efficient and is also introduced toward the end of the discriminator. In their paper, they describe a solution that uses the average of the standard deviations computed for each feature on each spatial location in the minibatch. This operation produces a single ...