EGANet

Efficient Global Attention Network via Multi-Granularity Dynamic Convolution for Lightweight Image Super-Resolution

Qian Wang, Jing Wei, Mengyang Wang, Yao Tang, Han Pan

💻Environment

• PyTorch >= 1.9
• Python 3.7
• Numpy
• BasicSR >= 1.3.4.9

🔧Installation

pip install -r requirements.txt

📜Data Preparation

The trainset uses the DIV2K (800). In order to effectively improve the training speed, images are cropped to 480 * 480 images by running script extract_subimages.py, and the dataloader will further randomly crop the images to the GT_size required for training. GT_size defaults to 128/192/256 (×2/×3/×4).

python extract_subimages.py

The input and output paths of cropped pictures can be modify in this script. Default location: ./datasets/DIV2K.

🚀Train

▶️ You can change the training strategy by modifying the configuration file. The default configuration files are included in ./options/train/EGANet. Take one GPU as the example.

### Train ###
### EGANet ###

python train.py -opt ./options/train/EGANet/train_eganet_x4.yml --auto_resume  # ×4

For more training commands, please check the docs in BasicSR

🚀Test

▶️ You can modify the configuration file about the test, which is located in ./options/test/EGANet. At the same time, you can change the benchmark datasets and modify the path of the pre-train model.

▶️ We have uploaded the pre-train weights for EGANet.Please change the path to the corresponding configuration file when testing.

### Test ###
### EGANet for Lightweight Image Super-Resolution ###
python basicsr/test.py -opt ./options/test/EGANet/test_eganet_x4.yml  # ×4

### EGANet for Large Image Super-Resolution ###
### Flicker2K  Test2K  Test4K  Test8K ###
python basicsr/test.py -opt ./options/test/EGANet/test_eganet_large.yml  # large image

🚩Results

Stay tuned.

📫Contact

If you have any questions, please feel free to contact us [email protected] and [email protected].