yjchoe / alphagan

A PyTorch implementation of alpha-GAN

alphaGAN

A PyTorch implementation of alpha-GAN (https://arxiv.org/abs/1706.04987) with a sample run on MNIST.

Dependencies

• PyTorch v0.1+ with CUDA support
• torchvision v0.1.8+
• TensorFlow v1.2+ (for TensorBoard only)

Usage

train_mnist.py contains sample code that runs the package on MNIST data. On the command line, run

$ python train_mnist.py --output_path YOUR_SAVED_PATH

While or after running, you will able to monitor the training progress on TensorBoard. Run

$ tensorboard --logdir=YOUR_SAVED_PATH/logs/ --port=6006

and access https://localhost:6006 (or the corresponding server URL) on your browser.

References

• Paper
  • Rosca, M., Lakshminarayanan, B., Warde-Farley, D., & Mohamed, S. (2017). Variational Approaches for Auto-Encoding Generative Adversarial Networks. arXiv preprint arXiv:1706.04987.
• Code
  • PyTorch VAE implementation
  • Martin Arjovsky's DCGAN implementation
  • Yunjey Choi's TensorBoard tutorial for PyTorch

Generated MNIST Images