yhjo09/SRFlow-DA

SRFlow-DA: Super-Resolution Using Normalizing Flow with Deep Convolutional Block

NTIRE 2021 Learning the Super-Resolution Space Challenge.

[Paper]

• Challenge results of our SRFlow-DA model.

Upscale	LR-PSNR	LPIPS	Diversity
X4	50.70 (1st)	0.121 (3rd)	23.091 (4th)
X8	50.86 (1st)	0.266 (3rd)	23.320 (4th)

Dependency

• Python 3.6 (anaconda, miniconda or pyenv is recommended)
• PyTorch 1.7
• Other dependencies in requirements.txt,

  pip install -r requirements.txt
  

• Because the file (requirements.txt) contains the information of abstract dependencies, you can install other compatible versions referring to the file when you have a problem with the above command.
• Most of the code from the original SRFlow repository.

First things to do

1. Clone this repo.

git clone https://github.com/yhjo09/SRFlow-DA
cd SRFlow-DA

2. Download datasets and baseline models.

sh ./prepare.sh

Test

1. Download SRFlow-DA models and unzip it.

unzip ./experiments.zip

2. Run.

cd ./code
python test.py ./confs/SRFlow-DA_DF2K_4X.yml        # SRFlow-DA 4X SR
python test.py ./confs/SRFlow-DA_DF2K_8X.yml        # SRFlow-DA 8X SR
python test.py ./confs/SRFlow-DA-R_DF2K_4X.yml      # SRFlow-DA-R 4X SR
python test.py ./confs/SRFlow-DA-R_DF2K_8X.yml      # SRFlow-DA-R 8X SR
python test.py ./confs/SRFlow-DA-S_DF2K_4X.yml      # SRFlow-DA-S 4X SR
python test.py ./confs/SRFlow-DA-S_DF2K_8X.yml      # SRFlow-DA-S 8X SR
python test.py ./confs/SRFlow-DA-D_DF2K_4X.yml      # SRFlow-DA-D 4X SR
python test.py ./confs/SRFlow-DA-D_DF2K_8X.yml      # SRFlow-DA-D 8X SR

• If your GPU memory lacks, please try with prefix CUDA_VISIBLE_DEVICES=-1 (CPU only).
• You may check dataroot_LR of the configuration file for the test.

3. Check your results in ./results.

Train

1. You may have to modify some variables (e.g. directories) in a config file ./confs/*.yml.

2. Run.

cd ./code
python train.py -opt ./confs/SRFlow-DA_DF2K_4X.yml        # SRFlow-DA 4X SR
python train.py -opt ./confs/SRFlow-DA_DF2K_8X.yml        # SRFlow-DA 8X SR
python train.py -opt ./confs/SRFlow-DA-R_DF2K_4X.yml      # SRFlow-DA-R 4X SR
python train.py -opt ./confs/SRFlow-DA-R_DF2K_8X.yml      # SRFlow-DA-R 8X SR
python train.py -opt ./confs/SRFlow-DA-S_DF2K_4X.yml      # SRFlow-DA-S 4X SR
python train.py -opt ./confs/SRFlow-DA-S_DF2K_8X.yml      # SRFlow-DA-S 8X SR
python train.py -opt ./confs/SRFlow-DA-D_DF2K_4X.yml      # SRFlow-DA-D 4X SR
python train.py -opt ./confs/SRFlow-DA-D_DF2K_8X.yml      # SRFlow-DA-D 8X SR

• If your GPU memory lacks, please try with lower batch size or patch size.

3. Training logs, model parameters, and validation result images will be stored in ./experiments.

BibTeX

@InProceedings{jo2021srflowda,
   author = {Jo, Younghyun and Yang, Sejong and Kim, Seon Joo},
   title = {SRFlow-DA: Super-Resolution Using Normalizing Flow with Deep Convolutional Block},
   booktitle = {The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
   month = {June},
   year = {2021}
}