SPR²Q: Static Priority-based Rectifier Routing Quantization for Image Super-Resolution

This project is developed based on the PaddlePaddle framework.

Environment Setup

cd SPR2Q
conda create -n SPR2Q python=3.9
conda activate SPR2Q
python -m pip install paddlepaddle-gpu==3.2.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu118/
pip install -r requirements.txt

Dataset Preparation

This repository uses DF2K for training and Set5, Set14, B100, Urban100, and Manga109 for testing. Please modify the dataset paths in the YAML configuration files located in options/train/, options/test/, and options/finetune/.

Training Instructions

For example, to train 2× super-resolution with 4-bit quantization, modify the pretrain_network_FP in the YAML file to point to the full-precision model (.pdparams), then run:

python basicsr/train.py -opt options/train/train_mamba_quant_x2.yml --force_yml bit=4 name=train_x2_bit4

The training results will be saved in the experiments folder.

Fine-Tuning for Weight Factor Adjustment

Modify pretrain_network_FP to the full-precision model (.pdparams) and pretrain_network_Q to the trained quantized model (.pdparams) you want to fine-tune. Then run:

python basicsr/train.py -opt options/finetune/finetune_mamba_quant_x2.yml --force_yml bit=1 name=finetune_x2_bit4

The fine-tuned results will be saved in the experiments folder.

Testing Instructions

Modify pretrain_network_Q in the YAML file to the trained quantized model (.pdparams). Then run:

python basicsr/test.py -opt options/test/test_mamba_quant_x2.yml --force_yml bit=4 name=test_x2_bit4

The test results will be saved in the results folder.