Skip to content

Latest commit

 

History

History
76 lines (60 loc) · 2.85 KB

README.md

File metadata and controls

76 lines (60 loc) · 2.85 KB
Raw

Meta-SR

Official implementation of Meta-SR: A Magnification-Arbitrary Network for Super-Resolution(CVPR2019)(PyTorch)

Paper

Our code is built on EDSR(PyTorch).

Attention

  • The master branch has fearful bug!!! we will check the code as soon as possible（2020-12-31）
  • I find an error in my camera-ready, the PSNR of our Meta-RDN on scale 1.2 is 40.04 not 40.40.

Requirements

  • Pytorch 0.4.0
  • Python 3.5
  • numpy
  • skimage
  • imageio
  • cv2
    *note that if you use another version of pytorch (>0.4.0), you can rewrite the dataloader.py

Install and run demo

  1. download the code
git clone https://github.com/XuecaiHu/Meta-SR-Pytorch.git
cd Meta-SR-Pytorch
  1. run training demo:
python main.py --model metardn --ext sep  --save metardn --lr_decay 200 --epochs 1000 --n_GPUs 1 --batch_size 1
  1. run test demo:
  • download the model from the BaiduYun fetch code: btc5.
  • put the model_1000.pt under the ./eperiment/metardn/model/
python main.py --model metardn --ext sep  --save metardn --n_GPUs 1 --batch_size 1 --test_only --data_test Set5 --pre_train  ./experiment/metardn/model/model_1000.pt  --save_results --scale 1.5

Train and Test as our paper

  1. prepare dataset
  • download the dataset DIV2K and test dataset fetch code: w3hk GoogleDrive
  • change the path_src = DIV2K HR image folder path and run /prepare_dataset/geberate_LR_metasr_X1_X4.m
  • upload the dataset
  • change the dir_data in option.py： dir_data = "/path to your DIV2K and testing dataset'(keep the training and test dataset in the same folder: test dataset under the benchmark folder and training dataset rename to DIV2K, or change the data_train to your folder name)
  1. pre_train model for test BaiduYun fetch code: btc5
    GoogleDrive

train

cd /Meta-SR-Pytorch 
python main.py --model metardn --save metardn --ext sep --lr_decay 200 --epochs 1000 --n_GPUs 4 --batch_size 16

test

python main.py --model metardn --save metardn --ext sep --pre_train ./experiment/metardn/model/model_1000.pt --test_only --data_test Set5  --scale 1.5 --n_GPUs 1

Citation

@article{hu2019meta,
  title={Meta-SR: A Magnification-Arbitrary Network for Super-Resolution},
  author={Hu, Xuecai and Mu, Haoyuan and Zhang, Xiangyu and Wang, Zilei  and Tan, Tieniu and Sun, Jian},
  booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
  year={2019}
}

Contact

Xuecai Hu (huxc@mail.ustc.edu.cn)