Tensorflow implementation of f-GAN (NIPS 2016) - f-GAN: Training Generative Neural Samplers Using Variational Divergence Minimization.

• make these divergences work (welcome the suggestions)
  • Kullback-Leibler with tricky G loss
  • Reverse-KL with tricky G loss
  • Pearson-X2 with tricky G loss
  • Squared-Hellinger with tricky G loss
  • Jensen-Shannon with tricky G loss
  • GAN with tricky G loss
• test more divergence

• Using tricky G loss (see Section 3.2 in the paper)

  Kullback-Leibler	Reverse-KL	Pearson-X2
  Squared-Hellinger	Jensen-Shannon	GAN
  NaN

• Using theoretically correct G loss

  Kullback-Leibler	Reverse-KL	Pearson-X2
  Squared-Hellinger	Jensen-Shannon	GAN
  NaN

• Prerequisites

  • tensorflow 1.7 or 1.8
  • python 2.7

• Examples of training

  • training

    CUDA_VISIBLE_DEVICES=0 python train.py --dataset=mnist --divergence=Pearson-X2 --tricky_G

  • tensorboard for loss visualization

    CUDA_VISIBLE_DEVICES='' tensorboard --logdir ./output/mnist_Pearson-X2_trickyG/summaries --port 6006

If you find f-GAN useful in your research work, please consider citing:

@inproceedings{nowozin2016f,
  title={f-GAN: Training Generative Neural Samplers Using Variational Divergence Minimization},
  author={Nowozin, Sebastian and Cseke, Botond and Tomioka, Ryota},
  booktitle={Advances in Neural Information Processing Systems (NIPS)},
  year={2016}
}