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

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

Update notes

• 2019/12/23: fix a bug in https://github.com/XuecaiHu/Meta-SR-Pytorch/blob/f2cf094248defef242973282627ac8ea50d2e806/trainer.py#L107 , since the zeros in the double data type isnot a real zero.

• 2019/12/06: I rewrite the input_matrix_wpn function in trainer.py. Since the offset is repeated, there are many repeated weight prediction. I remove them. In the metardn, we use repeated_weights to extend a small matrix to a matrix with same size of the feature maps. This version use less memory and less inference times.

• todo provide code for pytorch 1.0, 1.1 and 1.2

Install and run demo

1. download the code

git clone https://github.com/XuecaiHu/Meta-SR-Pytorch.git
cd Meta-SR-Pytorch

2. run training demo:

python main.py --model metardn --ext sep  --save metardn --lr_decay 200 --epochs 1000 --n_GPUs 1 --batch_size 1

3. 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: ev7u 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)

2. 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)