Class-Splitting Generative Adversarial Networks

Code for reproducing the experiments with splitting policy.

Full paper: https://arxiv.org/abs/1709.07359

This code is heavily based on the excellent work of Ishaan Gulrajani et al. "Improved Training of Wasserstein GANs" and his code https://github.com/igul222/improved_wgan_training.

Once defined the value for DATA_DIR, it should run without problems.

Some results

Generated images for the supervised CIFAR-10 experiment, with Inception score of 8.87.

Generated images for the unsupervised STL-10 experiment, with Inception score of 9.50.