WGAN-GP-TTUR on CelebA

An PyTorch implementation of WGAN with gradient penalty and TTUR.

• WGAN: https://arxiv.org/abs/1701.07875
• Improved Training of Wasserstein GANs (gradient penalty): https://arxiv.org/abs/1704.00028
• GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium (TTUR): https://arxiv.org/abs/1706.08500

Environment

• OS: Ubuntu16.04
• Language: Python
• Packages: torch, torchvision, numpy, tensorflow (for tensorboard)

Prepare Data

1. Download CelebA from http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html (the Google Drive link)
   • Unzip them, and find the zip file with the name img_align_celeba.zip, which contains the training data to use
2. Create a folder: mkdir data/celeba/
3. Unzip the zip file: unzip img_align_celeba.zip and put the zipped files under the folder we just created: data/celeba/

Training

Run the following for training

python main.py --dataset celeba --dataroot data/celeba --batch_size 64 --image_size 128 --niter 10000 --exp celeba_experiment

and check the log in tensorboard with

tensorboard --logdir .

Acknowledgement

Implementation is hugely borrowed from

1. https://github.com/martinarjovsky/WassersteinGAN
2. https://github.com/igul222/improved_wgan_training
3. https://github.com/bioinf-jku/TTUR
4. https://github.com/LynnHo/WGAN-GP-DRAGAN-Celeba-Pytorch
5. https://github.com/tensorflow/tensor2tensor