Author: Guo-Jun Qi, Date: 1/9/2017, most recent update: 2/8/2017

Questions about the source codes can be directed to Dr. Guo-Jun Qi at [email protected]. All copyrights reserved.

Please cite the following paper when referring to the following algorithms (LS-GAN and CLS-GAN)

Guo-Jn Qi. Loss-Sensitive Generative Adversarial Networks on Lipschitz Densities. arXiv:1701.06264 [pdf]

We are keeping updating this repository of source codes, and more results and algorithms will be released soon.

• We now have a new project generalizing LS-GAN to a more general form, called Generalized LS-GAN (GLS-GAN). It unifies Wasserstein GAN as well as this LS-GAN in an integrated framework, along with a super class of GLS-GANs including various new GAN members to explore. Check the github project at https://github.com/guojunq/glsgan, and the Appendix D of the newly updated prepint

• We also show an incomplete map of GANs at http://www.cs.ucf.edu/~gqi/GANs.htm, which plots the terrority of basic GANs in our view. We will update the map as more GAN territories are found.

• gamma: this is the coefficient for loss-minimization term (the first term in the objective for optimizing L_\theta).
• lambda: the scale of margin. This controls the desired margins between real and fake samples. However, we found this is a hyperparameter not very sensitive to the generation performance.

1.Setup and download dataset

`mkdir celebA; cd celebA

Download img_align_celeba.zip from http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html under the link "Align&Cropped Images".

unzip img_align_celeba.zip; cd ..
DATA_ROOT=celebA th data/crop_celebA.lua

2.Training the LS-GAN

DATA_ROOT=celebA dataset=folder th lsgan.lua

1. Please download bedroom_train_lmdb from http://lsun.cs.princeton.edu

2. Prepare the dataset following the instructions below

3. Install LMDB in your system: sudo apt-get install liblmdb-dev

4. Install torch packages: luarocks install lmdb.torch luarocks install tds

5. Once downloading bedroom_train_lmdb, unzip the dataset and put it in a directory lsun/train

6. Create an index file : Copy lsun_index_generator.lua to lsun/train, and run cd lsun/train DATA_ROOT=. th lsun_index_generator.lua Now you should have bedroom_train_lmdb_hashes_chartensor.t7 in lsun/train

7. Now return to the parent direcotry of lsun, and you should be ready to run lsgan.lua: DATA_ROOT=lsun th lsgan.lua

To display images during training and generation, we will use the display package.

• Install it with: luarocks install https://raw.githubusercontent.com/szym/display/master/display-scm-0.rockspec
• Then start the server with: th -ldisplay.start
• Open this URL in your browser: http://localhost:8000

1. Download and prepare datasets

2. MNIST: run torch-rocks install https://raw.github.com/andresy/mnist/master/rocks/mnist-scm-1.rockspec

3. CIFAR10: run th ./data/Cifar10bintoTensor.lua

4. SVHN: run th ./data/svhn.lua

5. Now you should be able to run clsgan.lua now. Select the dataset you want to use. For example you want to run MNIST to generate handwritten digits according to ten digit classes. Then you can run the following command line

   dataset=mnist th clsgan.lua
   

   For the other parameters you can set, please refer to the script in clsgan.lua.

1. parts of codes are reused from DCGAN at https://github.com/Newmu/dcgan_code

2. the code downloading cifar10 is available at https://github.com/soumith/cifar.torch

3. the code downloading SVHN: http://ufldl.stanford.edu/housenumbers/