RankSRGAN

Paper | Supplementary file | Project Page

RankSRGAN: Generative Adversarial Networks with Ranker for Image Super-Resolution

By Wenlong Zhang, Yihao Liu, Chao Dong, Yu Qiao

Dependencies

Python 3 (Recommend to use Anaconda)
PyTorch >= 0.4.0
NVIDIA GPU + CUDA
Python packages: pip install numpy opencv-python lmdb
[option] Python packages: pip install tensorboardX, for visualizing curves.

Codes

This version is under testing. We will provide more details of RankSRGAN in near future.

How to Test

Clone this github repo.

git clone https://github.com/WenlongZhang0724/RankSRGAN.git
cd RankSRGAN

Place your own low-resolution images in ./LR folder.
Download pretrained models from Google Drive or Baidu Drive. Place the models in ./experiments/pretrained_models/. We provide three Ranker models and three RankSRGAN models (see model list).
Run test. We provide RankSRGAN (NIQE, Ma, PI) model and you can config in the test.py.

python test.py -opt options/test/test_ranksrgan.json

The results are in ./results folder.

How to Train

Train Ranker

Download DIV2K and Flickr2K from Google Drive or Baidu Drive
Generate rank dataset ./datascripts/generate_rankdataset/
Run command:

python train.py -opt options/train/Ranker.json

Train RankSRGAN

We use a PSNR-oriented pretrained SR model to initialize the parameters for better quality.

Prepare datasets, usually the DIV2K dataset.
Prerapre the PSNR-oriented pretrained model. You can use the SRResNet_bicx4_in3nf64nb16.pth as the pretrained model that can be downloaded from Google Drive or Baidu Drive.
Modify the configuration file options/train/RankSRGAN_NIQE.json
Run command:

python train.py -opt options/train/RankSRGAN_NIQE.json

python train_PI.py -opt options/train/RankSRGAN_NIQE.json

Using the train.py can output the convergence curves with NIQE and PSNR; Using the train_PI.py can output the convergence curves with NIQE, Ma, PI and PSNR.

Acknowledgement

Part of this codes was done by Yihao Liu.
This codes are based on BasicSR.

zhuangzhong/RankSRGAN