Open-world GAN Discovery

PyTorch implementation of Towards Discovery and Attribution of Open-world GAN Generated Images

@inproceedings{girish2021towards,
  title={Towards discovery and attribution of open-world gan generated images},
  author={Girish, Sharath and Suri, Saksham and Rambhatla, Sai Saketh and Shrivastava, Abhinav},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={14094--14103},
  year={2021}
}

Installation

Create new virtualenv/conda environment. The project uses the ThunderSVM library which requires CUDA 9.0 if installed via pip. To install all the required libraries, run the following command:

pip install -r requirements.txt

Data

Setup data for train and eval in Pytorch's ImageFolder format:

data/train/real_celeba/xxx.png
data/train/attgan_celeba/xxy.png

data/eval/real_celeba/yyy.png
data/eval/began_celeba/yyy.png

Example train commands

Hyperparameters are defined in yaml files in the cfgs folder. An example run command with the default config would look like:

python main.py --config cfgs/config.yaml

Hyperparameters can additionally be overriden with command-line arguments which are dot separated. For e.g. running the pipeline for 3 iterations instead of 4 would look like:

python main.py --config cfgs/config.yaml --common.num_iters 3

Evaluation

Clusters can be evaluated at every stage of the pipeline for a given run as follows:

python eval_run.py --config cfgs/config.yaml

License

This project is released under the MIT License. Please review the License file for more details.