InfoGAN

Keras implementation of InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets

Requirements

python modules

• keras, theano or tensorflow backend
• h5py
• matplotlib
• opencv 3
• numpy
• tqdm
• parmap

Part 1. Processing the data

Follow these instructions.

Part 2. Running the code

Follow these instructions

Part 3. Example results

MNIST example results

Note 1 The figures below were obtained with a slight modification of the original InfoGAN paper: supervised categorical cross entropy loss for the discriminator and simple MSE loss for the continuous variables. Credits to @burisuriburi for the original idea.

Note 2 The code in this repository matches OpenAI's original implementation, without the trick of Note 1.

Varying the categorical code along each row:

Varying the continuous code along rows and columns: