abylouw / idgan

Official PyTorch implementation on ID-GAN: High-Fidelity Synthesis with Disentangled Representation by Lee et al., 2020.

ID-GAN

Pytorch implementation on "High-fidelity Synthesis with Disentangled Representation" (https://arxiv.org/abs/2001.04296).
For ID-GAN augmented with Variational Discriminator Bottleneck (VDB) or VGAN, please refer to the vgan branch.

Usage

Prepare datasets

• Create data directory, and put the necessary datasets inside here.

mkdir data

• dSprites dataset.

cd data
git clone https://github.com/deepmind/dsprites-dataset.git
cd dsprites-dataset
rm -rf .git* *.md LICENSE *.ipynb *.gif *.hdf5

• CelebA dataset.

1. Go to the official website (link) and download img_align_celeba.zip file to data directory.

data
|- preprocess.py
|_ img_align_celeba.zip

2. Preprocess the data.

python data/preprocess.py celeba

• CelebA-HQ dataset.

1. Go to the google drive (link) and download data1024x1024.zip file to data directory.

data
|- preprocess.py
|_ data1024x1024.zip

2. Preprocess the data.

python data/preprocess.py celeba-hq

• 3D Chairs dataset.

1. Go to the official website (link) and download rendered_chairs.tar file to data directory.

data
|- preprocess.py
|_ rendered_chairs.tar

2. Preprocess the data.

python data/preprocess.py chairs 

• 3D Cars dataset.

1. Go to the official website (link) and download cars_train.tgz, cars_test.tgz, and car_devkit.tgz files to data directory.

data
|- preprocess.py
|- cars_train.tgz 
|- cars_test.tgz 
|_ car_devkit.tgz 

2. Preprocess the data.

python data/preprocess.py cars 

Train

• You can run pre-defined commands as follows

bash scripts/run_dsprites.sh
bash scripts/run_celeba.sh
bash scripts/run_chairs.sh
bash scripts/run_cars.sh

• Stage 1: Train VAEs.

python dvae_main.py --dataset [dataset_name] --name [dvae_run_name] --c_dim [c_dim] --beta [beta]

, where [dataset_name] can be one of dsprites, celeba, cars, and chairs. please refer to dvae_main.py for the details.

• Stage 2: Train ID-GAN through information distillation loss.

python train.py --config [config_name] --dvae_name [dvae_run_name] --name [idgan_run_name]

please refer to configs directory for [config_name].

Results

Results, including checkpoints, tensorboard logs, and images can be found in outputs directory.

Acknowledgement

This code is built on the repos as follows:

1. Beta-VAE: https://www.github.com/1Konny
2. GAN with R2 regularization: https://github.com/LMescheder/GAN_stability
3. VGAN: https://github.com/akanazawa/vgan

Citation

If you find our work useful for your research, please cite our paper.

@article{lee2020highfidelity, 
    title={High-Fidelity Synthesis with Disentangled Representation}, 
    author={Wonkwang Lee and Donggyun Kim and Seunghoon Hong and Honglak Lee}, 
    year={2020}, 
    journal={arXiv preprint arXiv:2001.04296}, 
}