sczhou / IGNN

[NeurIPS 2020] Cross-Scale Internal Graph Neural Network for Image Super-Resolution

IGNN

Code repo for "Cross-Scale Internal Graph Neural Network for Image Super-Resolution"   [paper] [supp]

Prepare datasets

1 Download training dataset and test datasets from here.

2 Crop training dataset DIV2K to sub-images.

python ./datasets/prepare_DIV2K_subimages.py

Remember to modify the 'input_folder' and 'save_folder' in the above script.

Dependencies and Installation

The denoising code is tested with Python 3.7, PyTorch 1.1.0 and Cuda 9.0 but is likely to run with newer versions of PyTorch and Cuda.

1 Create conda environment.

conda create --name ignn
conda activate ignn
conda install pytorch=1.1.0 torchvision=0.3.0 cudatoolkit=9.0 -c pytorch

2 Install PyInn.

pip install git+https://github.com/szagoruyko/pyinn.git@master

3 Install matmul_cuda.

bash install.sh

4 Install other dependencies.

pip install -r requirements.txt

Pretrained Models

Downloading the pretrained models from this link and put them into ./ckpt

Training

Use the following command to train the network:

python runner.py
        --gpu [gpu_id]\
        --phase 'train'\
        --scale [2/3/4]\
        --dataroot [dataset root]\
        --out [output path]

Use the following command to resume training the network:

python runner.py 
        --gpu [gpu_id]\
        --phase 'resume'\
        --weights './ckpt/IGNN_x[2/3/4].pth'\
        --scale [2/3/4]\
        --dataroot [dataset root]\
        --out [output path]

You can also use the following simple command with different settings in config.py:

python runner.py

Testing

Use the following command to test the network on benchmark datasets (w/ GT):

python runner.py \
        --gpu [gpu_id]\
        --phase 'test'\
        --weights './ckpt/IGNN_x[2/3/4].pth'\
        --scale [2/3/4]\
        --dataroot [dataset root]\
        --testname [Set5, Set14, BSD100, Urban100, Manga109]\
        --out [output path]

Use the following command to test the network on your demo images (w/o GT):

python runner.py \
        --gpu [gpu_id]\
        --phase 'test'\
        --weights './ckpt/IGNN_x[2/3/4].pth'\
        --scale [2/3/4]\
        --demopath [test folder path]\
        --testname 'Demo'\
        --out [output path]

You can also use the following simple command with different settings in config.py:

python runner.py

Visual Results (x4)

For visual comparison on the 5 benchmarks, you can download our IGNN results from here.

Some examples

Citation

If you find our work useful for your research, please consider citing the following papers :)

@inproceedings{zhou2020cross,
title={Cross-scale internal graph neural network for image super-resolution},
author={Zhou, Shangchen and Zhang, Jiawei and Zuo, Wangmeng and Loy, Chen Change},
booktitle={Advances in Neural Information Processing Systems},
year={2020}
}

Contact

We are glad to hear from you. If you have any questions, please feel free to contact [e-mail].

License

This project is open sourced under MIT license.