Scale-wise Convolution for Image Restoration

AAAI 2020 [Arxiv]

This package is adapted from WDSR @ 3aeff43.

Performance

Networks	Parameters	DIV2K (val)	Set5	B100	Urban100	Pre-trained models	Training command
WDSR x2	1,190,100	34.76	38.08	32.23	32.34	Download	details `python trainer.py --dataset div2k --eval_datasets div2k set5 bsds100 urban100 --model wdsr --scale 2 --job_dir ./wdsr_x2`
WDSR x3	1,195,605	31.03	34.45	29.14	28.33	Download	details `python trainer.py --dataset div2k --eval_datasets div2k set5 bsds100 urban100 --model wdsr --scale 3 --job_dir ./wdsr_x3`
WDSR x4	1,203,312	29.04	32.22	27.61	26.21	Download	details `python trainer.py --dataset div2k --eval_datasets div2k set5 bsds100 urban100 --model wdsr --scale 4 --job_dir ./wdsr_x4`

Usage

Dependencies

conda install pytorch torchvision -c pytorch
conda install tensorboard h5py scikit-image
pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" git+https://github.com/NVIDIA/apex.git

Evaluation

python trainer.py --dataset div2k --eval_datasets div2k set5 bsds100 urban100 --model wdsr --scale 2 --job_dir ./wdsr_x2 --eval_only

Datasets

DIV2K dataset: DIVerse 2K resolution high quality images as used for the NTIRE challenge on super-resolution @ CVPR 2017
Benchmarks (Set5, BSDS100, Urban100)

Download and organize data like:

wdsr/data/DIV2K/
├── DIV2K_train_HR
├── DIV2K_train_LR_bicubic
│   └── X2
│   └── X3
│   └── X4
├── DIV2K_valid_HR
└── DIV2K_valid_LR_bicubic
    └── X2
    └── X3
    └── X4
wdsr/data/Set5/*.png
wdsr/data/BSDS100/*.png
wdsr/data/Urban100/*.png