liuguoyou / wdsr-pytorch

PyTorch implementation of Wide Activation for Efficient and Accurate Image Super-Resolution (CVPR Workshop 2018)

WDSR

This repository is implementation of the "Wide Activation for Efficient and Accurate Image Super-Resolution".

Requirements

• PyTorch 1.1.0
• TensorBoard 1.14.0
• Numpy 1.15.4
• Pillow-SIMD 5.3.0.post1
• h5py 2.8.0
• tqdm 4.30.0

Prepare dataset

To prepare dataset used in experiments, first download dataset files from this link and organize it as shown below.

/YOUR_STORAGE_PATH/DIV2K
├── DIV2K_train_HR
├── DIV2K_train_LR_bicubic
│   └── X2
│   └── X3
│   └── X4
├── DIV2K_valid_HR
├── DIV2K_valid_LR_bicubic
│   └── X2
│   └── X3
│   └── X4
├── DIV2K_train_HR.zip
├── DIV2K_train_LR_bicubic_X2.zip
├── DIV2K_train_LR_bicubic_X3.zip
├── DIV2K_train_LR_bicubic_X4.zip
├── DIV2K_valid_HR.zip
├── DIV2K_valid_LR_bicubic_X2.zip
├── DIV2K_valid_LR_bicubic_X3.zip
└── DIV2K_valid_LR_bicubic_X4.zip

By default, we use "0001-0800.png" images to train the model and "0801-0900.png" images to validate the training. All experiments also use images with BICUBIC degradation on RGB space.

Training

WDSR Baseline (=WDSR-A) Example

python train.py --dataset-dir "/YOUR_STORAGE_PATH/DIV2K" \
                --output-dir "/YOUR_STORAGE_PATH/output" \
                --model "WDSR-A" \
                --scale 2 \
                --n-feats 32 \
                --n-res-blocks 16 \
                --expansion-ratio 4 \
                --res-scale 1.0 \
                --lr 1e-3

WDSR-B Example

python train.py --dataset-dir "/YOUR_STORAGE_PATH/DIV2K" \
                --output-dir "/YOUR_STORAGE_PATH/output" \
                --model "WDSR-B" \
                --scale 2 \
                --n-feats 32 \
                --n-res-blocks 16 \
                --expansion-ratio 6 \
                --low-rank-ratio 0.8 \
                --res-scale 1.0 \
                --lr 1e-3

If you want to modify more options, see the core/option.py file.

Evaluation

Trained model is evaluated on DIV2K validation 100 images. If you want to use self-ensemble for evaluation, add --self-ensemble option.

python eval.py --dataset-dir "/YOUR_STORAGE_PATH/DIV2K" \
               --checkpoint-file "/YOUR_STORAGE_PATH/output/WDSR-A-f32-b16-r4-x2-best.pth.tar"

Model	Scale	Residual Blocks	Parameters	PSNR	Pre-trained Model
WDSR Baseline 1)	x2	16	1.19M	34.67 dB	Download
WDSR Baseline+ 2)	x2	16	1.19M	34.77 dB	-

¹⁾ filters=32, res-blocks=16, expansion-ratio=4, res-scale=1.0
²⁾ filters=32, res-blocks=16, expansion-ratio=4, res-scale=1.0, with self-ensemble

References

• Yu, Jiahui, et al. "Wide activation for efficient and accurate image super-resolution." arXiv preprint arXiv:1808.08718 (2018).
• Fan, Yuchen, Jiahui Yu, and Thomas S. Huang. "Wide-activated deep residual networks based restoration for BPG-compressed images." Proc. IEEE Conf. Comput. Vis. Pattern Recognit. Workshops. 2018.
• Lim, Bee, et al. "Enhanced deep residual networks for single image super-resolution." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. 2017.