gan

This does the GAN training.

This code is based on carpedm20/DCGAN-tensorflow. Hsiao-Yu Tung ([email protected]) did most of the modifications for the paper; feel free to contact her and/or Dougal Sutherland ([email protected]) with any questions. (Issues here are fine too.)

Basic Usage:

First download the MNIST files with data/mnist/fetch.sh.

• The original MMD generative model: ./run_mmd.sh.

• The model maximizing the MMD/variance t-statistic: ./run_tmmd.sh.

• MMD generative model with adversarially optimized kernel function using GAN loss: ./run_mmd_fm.sh.

To visualize e.g. the MMD results with tensorboard: tensorboard --logdir=logs_mmd --port=1234.

You can get more samples from the trained models by e.g. ./run_mmd.sh sample, which will save one image per sample into official_samples.