Hands-On Generative Adversarial Networks with Keras

Pixelwise feature vector normalization in the generator

As the name suggests, the authors normalize the feature vectors to the unit norm after each convolution layer in the generator. They claim that, by applying feature vector normalization, the escalation of signal magnitudes is prevented.

In practice, they use an adaptation of the local response normalization approach by Krizhevsky and others, which was published in 2012. Taking into account all feature maps, N, the feature vectors, , are, therefore, normalized according to the following equation: