Langevin Cooling (L-Cool)

This code repository reproduces the results for the paper "Langevin Cooling for Domain Translation".

**L-Cool: Paper

This code has utilized majority of the code from CycleGAN and Tiramisu. The extension consists of implementation of Langevin dynamics which was written by Vignesh Srinivasan.

Requirements

pytorch 
python >= 3
cv2
imageio

To utilize Langevin Dynamics for CycleGAN, perform the following steps:

1. CycleGAN

Dataset

bash ./datasets/download_cyclegan_dataset.sh horse2zebra

Pre-trained Model

• Download the trained model from CycleGAN

bash ./scripts/download_cyclegan_model.sh horse2zebra

• Or train a model from scratch

python train.py --dataroot ./datasets/horse2zebra --name maps_cyclegan --model cycle_gan

2. Denoising Autoencoder (DAE)

Train a Tiramisu Model

python train_dae.py --dataroot ./datasets/horse2zebra --name horses_dae --model dae --display_id 0 --gaussian_noise 0.3  --netG tiramisu_67 --checkpoints_dir ./checkpoints/

The model checkpoint is stored in the directory checkpoints/horses_dae/. The noise added to the input of the DAE can be varied with --gaussian_noise.

3. L-Cool

Perform L-Cool

python test_dae_langevin.py --dataroot ./datasets/horse2zebra --name horses_dae --model dae --display_id 0 --gaussian_noise 0.3  --netG tiramisu_67 --checkpoints_dir ./checkpoints/ --langevin_steps 100 --step_size 0.005 --temp 0.001 --save_gifs

Hyperparameters for L-Cool

• --langevin_steps Number of steps
• --step_size Step size
• --temp Temperature

The results can be found in the directory results_dae_langevin.

GIF

Optionally, gifs can be saved by using --save_gifs.

Citations

When using this code for your research, please cite our paper

@article{srinivasan2020langevin,
  title={Langevin Cooling for Domain Translation},
  author={Srinivasan, Vignesh and M{\"u}ller, Klaus-Robert and Samek, Wojciech and Nakajima, Shinichi},
  journal={arXiv preprint arXiv:2008.13723},
  year={2020}
}
@inproceedings{srinivasan2020benign,
  title={Benign Examples: Imperceptible Changes Can Enhance Image Translation Performance.},
  author={Srinivasan, Vignesh and M{\"u}ller, Klaus-Robert and Samek, Wojciech and Nakajima, Shinichi}
  booktitle={AAAI},
  pages = {5842-5850},
  year={2020}     
}