yjn870 / drrn-pytorch

PyTorch implementation of Image Super-Resolution via Deep Recursive Residual Network (CVPR 2017)

DRRN

This repository is implementation of the "Image Super-Resolution via Deep Recursive Residual Network".

Requirements

• PyTorch 1.0.0
• Numpy 1.15.4
• Pillow 5.4.1
• h5py 2.8.0
• tqdm 4.30.0

Prepare

The images for creating a dataset used for training (291-image) or evaluation (Set5) can be downloaded from the paper author's implementation.

You can also use pre-created dataset files with same settings as the paper.

Dataset	Scale	Type	Link
291-image	2, 3, 4	Train	Download
Set5	2	Eval	Download
Set5	3	Eval	Download
Set5	4	Eval	Download

Generate training dataset

python generate_trainset.py --images-dir "BLAH_BLAH/Train_291" \
                            --output-path "BLAH_BLAH/Train_291_x234.h5" \
                            --patch-size 31 \
                            --stride 21

Generate test dataset

python generate_testset.py --images-dir "BLAH_BLAH/Set5" \
                           --output-path "BLAH_BLAH/Set5_x2.h5" \
                           --scale 2

Train

Model weights will be stored in the --outputs-dir after every epoch.

python train.py --train-file "BLAH_BLAH/Train_291_x234.h5" \
                --outputs-dir "BLAH_BLAH/DRRN_B1U9" \
                --B 1 \
                --U 9 \
                --num-features 128 \
                --lr 0.1 \
                --clip-grad 0.01 \
                --batch-size 128 \
                --num-epochs 50 \
                --num-workers 8 \
                --seed 123

You can also evaluate using --eval-file, --eval-scale options during training after every epoch. In addition, the best weights file will be stored in the --outputs-dir as a best.pth.

python train.py --train-file "BLAH_BLAH/Train_291_x234.h5" \
                --outputs-dir "BLAH_BLAH/DRRN_B1U9" \
                --eval-file "BLAH_BLAH/Set5_x2.h5" \
                --eval-scale 2 \
                --B 1 \
                --U 9 \
                --num-features 128 \
                --lr 0.1 \
                --clip-grad 0.01 \
                --batch-size 128 \
                --num-epochs 50 \
                --num-workers 8 \
                --seed 123

Evaluate

The pre-trained weights can be downloaded from the following links.

Model	Link
DRRN_B1U9	Download

python eval.py --weights-file "BLAH_BLAH/DRRN_B1U9/best.pth" \
               --eval-file "BLAH_BLAH/Set5_x2.h5" \
               --eval-scale 2 \
               --B 1 \
               --U 9 \
               --num-features 128               

Results

The our model was learned and evaluated on the Y(luminance) channel.

For performance, we modified the original implementation as follows.

• Batch normalization was removed from the residual unit.
• No bias was used in the convolution layer.

Performance comparision on the Set5

Eval. Mat	Scale	DRRN_B1U9 (Paper)	DRRN_B1U9 (Ours)
PSNR	2	37.66	37.62
PSNR	3	33.93	33.86
PSNR	4	31.58	31.52

References

1. https://github.com/tyshiwo/DRRN_CVPR17

drrn-pytorch's Issues

no BN layers?

hi,

I love your work!

just one point: I think the original paper and code applies BN layer, so you can also include BN layer in your code.

thanks