GAN-pruning

Code for our ICCV 2019 paper, Co-Evolutionary Compression for unpaired image Translation

This paper proposes a co-evolutionary approach for reducing memory usage and FLOPs of generators on image-to-image transfer task simultaneously while maintains their performances.

Description

  • GAN pruning search/finetune/test code for image to image translation task.

Files description

Requirements: Python3.6, PyTorch0.4

  • search.py is the search script ultilizing Genetic Algorithem for GAN pruning.
  • finetune.py is the script for finetuning searched pruned architectures.
  • test.py is the script for testing pruned architectures.
  • models.py defines original architecture of generators and discriminators.
  • models_prune.py defines searched pruned architecture with binary channel mask.
  • GA.py defines evolutionary operations .

Dataset

Image to image translation dataset, like horse2zebra, summer2winter_yosemite, cityscapes.

Performance

Performance on cityscapes compared with conventional pruning method:

Citation

@inproceedings{GAN pruning,
	title={Co-Evolutionary Compression for Unpaired Image Translation},
	author={Shu, Han and Wang, Yunhe and Jia, Xu and Han, Kai and Chen, Hanting and Xu, Chunjing and Tian, Qi and Xu, Chang},
	booktitle={ICCV},
	year={2019}
}

GitHub