This project is still in progress. If you are looking for the working code, use BEGAN-tensorflow.

• Python 2.7
• Pillow
• tqdm
• PyTorch
• torch-vision
• requests (Only used for downloading CelebA dataset)
• TensorFlow (Only used TensorBoard for logging)

First download CelebA datasets with:

$ apt-get install p7zip-full # ubuntu
$ brew install p7zip # Mac
$ python download.py

or you can use your own dataset by placing images like:

data
└── YOUR_DATASET_NAME
    ├── xxx.jpg (name doesn't matter)
    ├── yyy.jpg
    └── ...

To train a model:

$ python main.py --dataset=CelebA --num_gpu=1
$ python main.py --dataset=YOUR_DATASET_NAME --num_gpu=4 --use_tensorboard=True

To test a model (use your load_path):

$ python main.py --dataset=CelebA --load_path=./logs/CelebA_0405_124806 --num_gpu=0 --is_train=False --split valid

• BEGAN-tensorflow at least can generate human faces but BEGAN-pytorch can't.
• Both BEGAN-tensorflow and BEGAN-pytorch shows modal collapses and I guess this is due to a wrong scheuduling of lr (Paper mentioned that simply reducing the lr was sufficient to avoid them).

(in progress)

Taehoon Kim / @carpedm20