Enhanced Fourier-Mixture Transformer for High-Performance Image Super-Resolution (EFMSR), The Visual Computer

DOI

Dependencies

• Ubuntu 24.04
• Python 3.12
• PyTorch 2.8.0 + cu129
• NVIDIA GPU + CUDA

git clone https://github.com/MnTeriri/EFMSR.git
cd EFMSR
conda create -n EFMSR python=3.12
conda activate EFMSR
pip3 install torch==2.8.0 torchvision==0.23.0 --index-url https://download.pytorch.org/whl/cu129
pip install -r requirements.txt
pip install -e . --no-build-isolation

Training

• Download training (DF2K) and testing (Set5, Set14, BSD100, Urban100, Manga109) datasets, place them in datasets/.

• Run the following scripts. The training configuration is in options/train/.

• The training experiment is in experiments/.

• For multiple GPUs

  # EFMSR 8 GPUs
  torchrun --nproc-per-node=8 --master-port=4321 basicsr/train.py -opt options/train/train_EFMSR_x2.yml --launcher pytorch
  torchrun --nproc-per-node=8 --master-port=4321 basicsr/train.py -opt options/train/train_EFMSR_x3.yml --launcher pytorch
  torchrun --nproc-per-node=8 --master-port=4321 basicsr/train.py -opt options/train/train_EFMSR_x4.yml --launcher pytorch
  
  # EFMSR-light 8 GPUs
  torchrun --nproc-per-node=8 --master-port=4321 basicsr/train.py -opt options/train/train_EFMSR_Light_x2.yml --launcher pytorch
  torchrun --nproc-per-node=8 --master-port=4321 basicsr/train.py -opt options/train/train_EFMSR_Light_x3.yml --launcher pytorch
  torchrun --nproc-per-node=8 --master-port=4321 basicsr/train.py -opt options/train/train_EFMSR_Light_x4.yml --launcher pytorch

• For single GPU

  # EFMSR single GPU
  torchrun --nproc-per-node=1 --master-port=4321 basicsr/train.py -opt options/train/train_EFMSR_x2.yml --launcher pytorch
  torchrun --nproc-per-node=1 --master-port=4321 basicsr/train.py -opt options/train/train_EFMSR_x3.yml --launcher pytorch
  torchrun --nproc-per-node=1 --master-port=4321 basicsr/train.py -opt options/train/train_EFMSR_x4.yml --launcher pytorch
  
  # EFMSR-light single GPU
  torchrun --nproc-per-node=1 --master-port=4321 basicsr/train.py -opt options/train/train_EFMSR_Light_x2.yml --launcher pytorch
  torchrun --nproc-per-node=1 --master-port=4321 basicsr/train.py -opt options/train/train_EFMSR_Light_x3.yml --launcher pytorch
  torchrun --nproc-per-node=1 --master-port=4321 basicsr/train.py -opt options/train/train_EFMSR_Light_x4.yml --launcher pytorch

Testing

• We provide EFMSR and EFMSR-light with scale factors: x2, x3, x4.

• Download testing (Set5, Set14, BSD100, Urban100, Manga109) datasets, place them in datasets/.

• Run the following scripts.

  # EFMSR
  python basicsr/test.py -opt options/test/test_EFMSR_x2.yml
  python basicsr/test.py -opt options/test/test_EFMSR_x3.yml
  python basicsr/test.py -opt options/test/test_EFMSR_x4.yml
  
  # EFMSR-light
  python basicsr/test.py -opt options/test/test_EFMSR_Light_x2.yml
  python basicsr/test.py -opt options/test/test_EFMSR_Light_x3.yml
  python basicsr/test.py -opt options/test/test_EFMSR_Light_x4.yml

Acknowledgements

This code is built on BasicSR. The smm-cuda is derived from PFT-SR.