Disconnected Manifold Learning for GANs

Find our paper at NeurIPS 2018 and ArXiv. Please cite the following if using the code:

@incollection{NIPS2018_7964,
  title = {Disconnected Manifold Learning for Generative Adversarial Networks},
  author = {Khayatkhoei, Mahyar and Singh, Maneesh K. and Elgammal, Ahmed},
  booktitle = {Advances in Neural Information Processing Systems 31},
  editor = {S. Bengio and H. Wallach and H. Larochelle and K. Grauman and N. Cesa-Bianchi and R. Garnett},
  pages = {7354--7364},
  year = {2018},
  publisher = {Curran Associates, Inc.},
  url = {http://papers.nips.cc/paper/7964-disconnected-manifold-learning-for-generative-adversarial-networks.pdf}
}

Running the code:

After installing the necessary python dependencies, simply run:

$ python run_dmgan.py -l logs -e 5000 -s 0

This code implements the line segments experiments from the paper. To change the number of generators, modify self.g_num from inside DMGAN.__init__ (default is 10 generators). To disable prior learning, uncomment the following line from inside DMGAN.step:

z_data = np.random.randint(low=0, high=self.g_num, size=batch_size)

To use modified GAN objective instead of WGAN, set the following from inside DMGAN.__init__ (default setting is for wgan with one sided gradient penalty):

self.d_loss_type = 'log'
self.g_loss_type = 'mod'