Decent: Unpaired Image-to-Image Translation with Density Changing Regularization (Neurips2022)

Basic Usage

• Training:

python train.py --dataroot=datasets/selfie2anime  

• Test:

python test.py --dataroot=datasets/selfie2anime

• Multi-GPU training:

python train.py --dataroot=datasets/selfie2anime --gpu=0,1,2,3 --batch_size=4  

• The Weight --lambda_var=0.01
• Compute density changing loss across images --var_all I have tested var_all=False.
• Number of Flow Blocks --flow_blocks=1
• Learning Rate of Flow --flow_lr=0.001
• Different flows --flow_type=bnaf BNAF works best for me. Feel free to experiment other flows.

Pretrained Models

• label2city
• horse2zebra
• cat2dog
• selfie2anime
• Summer2winter

Evaluation Script of label2city

Different Pretrained-DRN and evaluation protocols can cause big performance gaps. So, I created a repository to upload the evaluation script of label2city. Hope the script could make the future evaluation easier.

Citation

If you use this code for your research, please cite our paper:

@inproceedings{xieunsupervised,
  title={Unsupervised Image-to-Image Translation with Density Changing Regularization},
  author={Xie, Shaoan and Ho, Qirong and Zhang, Kun},
  booktitle={Advances in Neural Information Processing Systems},
year=2022,
}