Code for A Closer Look at the Optimization Landscapes of Generative Adversarial Networks

This is the code for reproducing the experimental results in our paper A Closer Look at the Optimization Landscapes of Generative Adversarial Networks, Hugo Berard, Gauthier Gidel, Amjad Almahairi, Pascal Vincent, Simon Lacoste-Julien, 2019.

If you find this code useful please cite our paper:

@misc{berard2019closer,
    title={A Closer Look at the Optimization Landscapes of Generative Adversarial Networks},
    author={Hugo Berard and Gauthier Gidel and Amjad Almahairi and Pascal Vincent and Simon Lacoste-Julien},
    year={2019},
    eprint={1906.04848},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}

For any questions regarding the code please contact Hugo Berard (berard.hugo@gmail.com).

License

This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree.

Running the Code

We provide a conda environment to run the code: conda create -f mnist-exp_environment.yml

The code for computing the eigenvalues and the path-angle is in plot_path_tools.py.

To run the code for the Mixture of Gaussian experiment: python train_mixture_gan.py OUTPUT_PATH --deterministic --saving-stats

To run the code for the MNIST experiment: python train_mnist.py

The visualization of the results can be done with mnist_plots.ipynb