Hands-On Generative Adversarial Networks with Keras

Image variety

Another important aspect of GAN evaluation is image variety. When training GANs, we want to make sure that the Generator is producing varied samples and not just memorizing a few samples. In such cases when the Generator produces only a few images, we . say that the generator is producing only a few modes in the distribution and suffers from mode collapse.

The samples on the image below were generated at an intermediary step of the training process that is later than the step we shared previously. Although the images are considerably less blurry at later iterations, the three red covers seem suspicious. An explanation for these repeated images could be mode collapse, or that the data has a strong bias from images that are repeated ...