The official repository with Pytorch

Our paper can be downloaded from EQSR.

EQSR is designed to pursue scale-equivariance image super-resolution. We compare the PSNR degradation rate of our method and ArbSR. Taking the SOTA fixed-scale method HAT as reference, our model presents a more stable degradation as the scale increases, reflecting the equivariance of our method.

Clone this repo:

git clone https://github.com/neuralchen/EQSR.git
cd EQSR

Dependencies:

• python3.7+
• pytorch
• pyyaml, scipy, tqdm, imageio, einops, opencv-python
• cupy

(Note: Please do not directly use "pip install" to install basicsr. It might lead to some issues due to version differences.)

We divide the training into two stages. The first stage involves pretraining on the ImageNet dataset, and the second stage entails fine-tuning on the DF2K dataset.

You can modify parameters such as batch size, iterations, learning rate, etc. in the configuration files.

• Phase 1 Modify the dataset path in options/train/train*.xml, and run the following command to train.

CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4 --master_port=4320 train.py -opt train_EQSR_ImageNet_from_scratch --launcher pytorch

• Phase 2 Modify the paths of the datasets and the location of the pretrained model in the configuration file.

CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4 --master_port=4320 train.py -opt train_EQSR_finetune_from_ImageNet_pretrain --launcher pytorch

TODO

TODO

• Baidu Netdisk (百度网盘)：https://pan.baidu.com/s/1ui-GSbAQLuTyOmxBlAQZVg
• Extraction Code (提取码)：lspg

Modify the dataset path and pre-trained model path in options/test/test*.xml, and run the following command to test. If GPU memory is limited, you can consider adding PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:32 before these commands.

python test.py -opt options/test/testx234.yml
python test.py -opt options/test/testx6.yml
python test.py -opt options/test/testx8.yml

This code is built based on HAT and BasicSR. We thank the authors for sharing the codes.