Capsule GAN

Link to the paper

Code for my Master thesis on "Capsule Layer as a Discriminator in Generative Adversarial Networks". In order to reproduce results, follow the "capsule_gan" Jupyter notebook that contains:

• Dataset loading and preprocessing
• Both Discriminator and Generator structures
• Training, loss functions
• Image outputs
• Metrics visualization

But first you may want to install Miniconda and corresponding dependencies from environment.yml: conda env create -f environment.yml within the project directory as well as install required tools for GPU computing. If no GPU is going to be used - delete the tensorflow-gpu line from environment.yml.

Generated images

All generated MNIST images over 30k epochs All generated CIFAR10 images over 30k epochs

Generator weights for MNIST Generator weights for CIFAR10

Thanks to @eriklindernoren (https://github.com/eriklindernoren/Keras-GAN) who I borrowed the Keras implementation of DCGAN from and @XifengGuo (https://github.com/XifengGuo/CapsNet-Keras) who I took the squashing function from.