Awesome

Self-Calibrated Efficient Transformer for Lightweight Super-Resolution (offical)

Introduction

We have implemented our SCET method through the mmediting and mim algorithm framework. Next, we will describe the main processes of training and testing.

• Paper The SCET has been accepted by CVPRW2022, you can read the paper here.
• Model

Datasets

We use DIV2K and Flickr2K as our training datasets. First, we need to crop our training set by cropping each image in the dataset to a 480x480 size patch.

How to use the code to train SCET network.

1. Installation environment

conda create -n open-mmlab python=3.8 pytorch=1.10 cudatoolkit=11.3 torchvision -c pytorch -y
conda activate open-mmlab
pip3 install openmim
mim install mmedit

2. Modify the configuration file configs/SCETx2.py as follows:

# train
gt_folder='${dataset_workspace}/dataset/DF2K_train_HR_sub' # your train data path
lq_folder='${dataset_workspace}/dataset/DIV2K_train_LR_bicubic/X4_sub',
gt_folder='${dataset_workspace}/dataset/DF2K_train_HR_sub',
ann_file='${dataset_workspace}/dataset/meta_info.txt',
# Testing
lq_folder='${dataset_workspace}/dataset/validation/lr_x4' # your test data LR path
gt_folder='${dataset_workspace}/dataset/validation/gt' # your test data HR path

ps: Please refer to the official mmediting instructions for specific instructions on the configuration file.

3. Modify the bash file train.sh as follows:

# modify the number of gpus, config path and outdir path.
PYTHONPATH=$PWD:$PYTHONPATH mim train mmedit ./config/SCETx2.py --gpus 1 --work-dir {Your save ckpt path}

4. train SCET network, as follows:

cd SCET

bash train.sh

How to use the code to test SCET network.

1. Modify the bash file test.sh as follows:

# modify the config path, checkpoint path and outdir path.
PYTHONPATH=$PWD:$PYTHONPATH mim test mmedit ./config/SCET_x2.py --checkpoint ./weights/SCETx2.pth --save-path {Your save image path}

2. train SCET network, as follows:

cd SCET

bash test.sh

If you find this repo useful for your research, please consider citing the papers.

@article{zou2022self,
  title={Self-Calibrated Efficient Transformer for Lightweight Super-Resolution},
  author={Zou, Wenbin and Ye, Tian and Zheng, Weixin and Zhang, Yunchen and Chen, Liang and Wu, Yi},
  journal={arXiv preprint arXiv:2204.08913},
  year={2022}
}