Wasserstein GAN

This repository provides a Torch implementation of Wasserstein GAN as described by Arjovsky et. al. in their paper Wasserstein GAN.

Prerequisites

• Torch
• cutorch, cunn and cudnn to train the network on GPU. Training on CPU is supported but not recommended (very slow)

Please refer to the official Torch website to install Torch.

Usage

1. Choose a dataset and create a folder with its name (ex: mkdir celebA; cd celebA). Inside this folder create another folder (images for example) containing your images.
   Note: You can download the celebA dataset on the celebA web page. Extract the images and run

DATA_ROOT=celebA th data/crop_celebA.lua

2. Train the Wasserstein model

DATA_ROOT=<dataset_folder> name=<whatever_name_you_want> th main.lua

The networks are saved into the checkpoints/ directory with the name you gave.

3. Generate images

net=<path_to_generator_network> name=<name_to_save_images> th generate.lua

Example:

net=checkpoints/generator.t7 name=myimages display=2929 th generate.lua

The generated images are saved in myimages.png.

Display images in a browser

If you want, install the display package (luarocks install display) and run

th -ldisplay.start <PORT_NUMBER> 0.0.0.0

to launch a server on the port you chose. You can access it in your browser with the url http://localhost:PORT_NUMBER.

To train your network or for completion add the variable display=<PORT_NUMBER> to the list of options.

Optional parameters

In your command line instructions you can specify several parameters (for example the display port number), here are some of them:

• noise which can be either uniform or normal indicates the prior distribution from which the samples are generated
• batchSize is the size of the batch used for training or the number of images to reconstruct
• name is the name you want to use to save your networks or the generated images
• gpu specifies if the computations are done on the GPU or not. Set it to 0 to use the CPU (not recommended, too slow) and to n to use the nth GPU you have (1 is the default value)
• lr is the learning rate
• loadSize is the size to use to scale the images. 0 means no rescale
• niter is the number of epochs for training

References

• Original PyTorch implementation of Wasserstein GAN by the authors: https://github.com/martinarjovsky/WassersteinGAN
• The code is inspired from soumith's DCGAN implementation that can be found here: https://github.com/soumith/dcgan.torch