A new theoretically motivated regularization method to stabilize the GAN training dynamics. (Please see more details in the paper: https://arxiv.org/abs/1806.09235)

• 64-bit Python 3.6 installation, Tensorflow 1.4 with GPU support
• Numpy 1.14.1, Matplotlib 2.2.0, Scipy 1.0.0, tqdm 4.12.0, imageio 2.8.0, six 1.13.0, opencv-python 3.4.9.31

The folder synthetic contains the code for experiments on synthetic data.

To run experiments on Isotropic Gaussian:

cd synthetic
python3 Affine_GAN.py

To run experiments on GMM:

cd synthetic
python3 GMM_GAN.py

The folder real contains the code for experiments on real data (CIFAR-10).

To run experiments on CIFAR-10, for example, we can do:

cd real/experiments
python3 jare.py 0 1

Here the first argument 0 represents the gpu_id (in the case of using multiple gpus), and the second argument 1 represents the job_id (0-5), each of which means one of six network settings.

To run baselines, for example ConOpt, we can do:

cd real/experiments
python3 conopt.py 0 1

Similarly, we can change the job_id in the script to run baselines on different network settings.

Note that in order to compute the FID score, we may need to first download the inception_frozen.zip and unzip it into the inception folder before training.

To cite this work, please use

@INPROCEEDINGS{Nie2019UAI,
  author = {Weili Nie and Ankit Patel},
  title = {Towards a Better Understanding and Regularization of GAN Training Dynamics},
  booktitle = {UAI},
  year = {2019}
}