Hands-On Generative Adversarial Networks with Keras

Distribute latent vector

In his paper Sampling Generative Networks, Tom White claims that sampling from a spherical distribution prevents diverging from a model's prior distribution and produces sharper samples. The guidelines are as follows:

Do not sample from a uniform distribution; sample from a spherical distribution.

Sampling a spherical distribution in keras is simple and is equal to sampling a normal distribution, as follows:

z = keras.random_normal(0, 1, z_dim)