anibali/wgan-cifar10

Unofficial implementation of WGAN in PyTorch for CIFAR-10

wgan-cifar10

An unofficial implementation of (improved) WGAN in PyTorch for CIFAR-10 image data.

Requirements

• A modern NVIDIA GPU
• Docker
• Docker Compose
• nvidia-docker
• nvidia-docker-compose

Usage

nvidia-docker-compose run --rm pytorch bin/train.py

Outputs from training are written to files in out/.

There are a number of command line options which can be used to configure the training process:

--epochs N      number of epochs to train (default=1000)
--gen-iters N   generator iterations per epoch (default=100)
--disc-iters N  discriminator iterations per generator iteration (default=5)
--batch-size N  input batch size (default=64)
--disc-lr LR    discriminator learning rate (default=2e-4)
--gen-lr LR     generator learning rate (default=2e-4)
--unimproved    disable gradient penalty and use weight clipping instead

References

• Wasserstein GAN, Arjovsky et al.
• Improved Training of Wasserstein GANs, Gulrajani et al.

Related repositories

• martinarjovsky/WassersteinGAN - Official repository for the original WGAN paper (PyTorch).
• igul222/improved_wgan_training - Official repository for the improved WGAN training paper (TensorFlow).
• caogang/wgan-gp - Unofficial partial port of the improved WGAN code (PyTorch).