f-GANs in an Information Geometric Nutshell

Pytorch implementation of f-GANs in an Information Geometric Nutshell

Prerequisites

• Python 2.7
• Pytorch 0.1.12
• numpy 1.12.1

Usage

Put both mnist and lsun in a folder called DATA_ROOT. Download lsun with https://github.com/fyu/lsun. MNIST will be downloaded automatically in the first run.

$ python download.py -o <DATA_ROOT> -c tower

Assume all experimental results are put in EXPERIMENTAL_RESULTS.

Evaluate a feedforward network with wasserstein GAN loss and mu-ReLU as the activation of hidden layers of the generator:

$ python main.py --dataset mnist --dataroot <DATA_ROOT> --cuda -D wgan -A mlp -H murelu --experiment <EXPERIMENTAL_RESULTS> --task mu

Evaluate DCGAN with GAN as the loss, and mu-ReLU as the activation of hidden layers of the generator:

$ python main.py --dataset lsun --subset tower --dataroot <DATA_ROOT> --cuda -D gan -A dcgan -H murelu --experiment <EXPERIMENTAL_RESULTS> --task mu

Author

Lizhen Qu / @qulizhen