Hands-On Generative Adversarial Networks with Keras

Mode analysis

When using datasets with known modes (such as a Gaussian mixture model or a labeled dataset), we can easily evaluate the fake samples for mode collapse and mode drop. In GANs, mode collapse loosely refers to the lack of diversity in generated samples, and mode drop refers to the absence of a mode that exists in the training data but does not exist in the generator data.

For example, when trained on a multimodal distribution such as MNIST, a dataset of handwritten digits with 10 modes, each representing a digit, the samples produced by the generator might not include all the modes of the distribution and drop some of the digits.

In this context, mode analysis is done by sampling fake images from the generator, plotting them, ...