Skip to content

Latest commit

 

History

History
110 lines (98 loc) · 4.52 KB

README.md

File metadata and controls

110 lines (98 loc) · 4.52 KB
Raw

DivCo: Diverse Conditional Image Synthesis via Contrastive Generative Adversarial Network

Pytorch implementation for our DivCo. We propose a simple yet effective regularization term named latent-augmented contrastive loss that can be applied to arbitrary conditional generative adversarial networks in different tasks to alleviate the mode collapse issue and improve the diversity.

Contact: Rui Liu (ruiliu@link.cuhk.edu.hk)

Paper

DivCo: Diverse Conditional Image Synthesis via Contrastive Generative Adversarial Network
Rui Liu, Yixiao Ge, Ching Lam Choi, Xiaogang Wang, and Hongsheng Li
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021
[arxiv]

Citing DivCo

If you find DivCo useful in your research, please consider citing:

@inproceedings{Liu_DivCo,
  author = {Liu, Rui and Ge, Yixiao and Choi, Ching Lam and Wang, Xiaogang and Li, Hongsheng},
  booktitle = {IEEE Conference on Computer Vision and Pattern Recognition},
  title = {DivCo: Diverse Conditional Image Synthesis via Contrastive Generative Adversarial Network},
  year = {2021}
}

Framework

Usage

Prerequisites

Install

  • Clone this repo:
git clone https://github.com/ruiliu-ai/DivCo.git

Training Examples

Download datasets for each task into the dataset folder

mkdir datasets

Label-conditoned Image Generation

cd DivCo/DivCo-DCGAN
python train.py --dataroot ./datasets/Cifar10

Paried Image-to-image Translation

  • Paired Data: facades and maps
  • Baseline: BicycleGAN

You can download the facades and maps datasets from the BicycleGAN [Github Project].
We employ the network architecture of the BicycleGAN and follow its training process.

cd DivCo/DivCo-BicycleGAN
python train.py --dataroot ./datasets/facades

Unpaired Image-to-image Translation

  • Unpaired Data: Yosemite (summer <-> winter) and Cat2Dog (cat <-> dog)
  • Baseline: DRIT

You can download the datasets from the DRIT [Github Project].
Specify --concat 0 for Cat2Dog to handle large shape variation translation

cd DivCo/DivCo-DRIT
python train.py --dataroot ./datasets/cat2dog --concat 0 --lambda_contra 0.1
python train.py --dataroot ./datasets/yosemite --concat 1 --lambda_contra 1.0

Pre-trained Models

Download and save them into

./models/

Evaluation

For BicycleGAN, DRIT and MSGAN, please follow the instructions of corresponding github projects of the baseline frameworks for more evaluation details.

Testing Examples

DivCo-DCGAN 

python test.py --dataroot ./datasets/Cifar10 --resume ./models/DivCo-DCGAN/00199.pth

DivCo-BicycleGAN 

python test.py --dataroot ./datasets/facades --checkpoints_dir ./models/DivCo-BicycleGAN/facades --epoch 400

python test.py --dataroot ./datasets/maps --checkpoints_dir ./models/DivCo-BicycleGAN/maps --epoch 400

DivCo-DRIT 

python test.py --dataroot ./datasets/yosemite --resume ./models/DivCo-DRIT/yosemite/01199.pth --concat 1

python test.py --dataroot ./datasets/cat2dog --resume ./models/DivCo-DRIT/cat2dog/01199.pth --concat 0

Reference

Quantitative Evaluation Metrics