Mode Collapse Evaluation in GANs

Code for training Generative Adversarial Networks (GANs), and evaluating the models' mode collapse. The code was taken and adapted from https://github.com/carpedm20/DCGAN-tensorflow .

The following models are supported:

• DCGAN (https://arxiv.org/pdf/1511.06434.pdf)
• ALI (Adversarially Learned Inference, https://arxiv.org/pdf/1606.00704.pdf)
• Unrolled GAN (https://arxiv.org/pdf/1611.02163.pdf)

Evaluating mode collapse using the following 2 methods, introduced in https://arxiv.org/pdf/1611.02163.pdf :

• Number of modes covered with the Stacked Mnist dataset (section 3.3.1, Discrete Mode Collapse)
• Inference via Optimisation (section 3.4.1)

Requirements:

• Tensorflow
• SciPy
• Keras (only used to import keras.optimizers.Adam )

Datasets

• Stacked MNIST - can be created by running StackedMNIST_data_setup.py
• CIFAR10 - download the python version from http://www.cs.toronto.edu/~kriz/cifar.html and place the content in /data/cifar10

Running

The Run_*.py files contain preset configurations for each model/dataset.