Distribution-Flexible Subset Quantization for Post-Quantizing Super-Resolution Networks link

Dependence

• Python 3.8
• PyTorch >= 1.7.0

Datasets

Please download DIV2K datasets.

Then, create a directory 'datasets' and re-organise the downloaded dataset directory as follows:

option.py
main_setq.py
datasets
  benchmark
  DIV2K

Additionally,you need to create some directory:'data'、'log'、'result' as follows:

option.py
main_setq.py
data
log
result
datasets
  benchmark
  DIV2K

Usage

1: train full-precision models:

An example:

python main_ori.py --scale 4 \
--model edsr
--save edsr_baseline_x4 \
--patch_size 192 \
--epochs 300 \
--decay 200 \
--gclip 0 \
--dir_data ./datasets

2: get activation map for cluster:

An example:

python main_setq.py --scale 4 \
--model edsr \
--pre_train path/fp_model --patch_size 192 \
--w_bits 4 --a_bits 4 \
--quant_file "edsr_4x_4bit" \
--data_test "Set14+Set5+B100+Urban100" \
--dir_data ./datasets

The result will be saved in data/edsr_4x_4bit/

3: cluster for quantization paramters:

python cluster_process.py 

4: test quantized models

An example:

python main_setq.py --scale 4 \
--w_bits 4 --a_bits 4 \
--model edsr \
--pre_train path/fp_model --patch_size 192 \
--data_test "Set14+Set5+B100+Urban100" \
--quant_file "edsr_4x_4bit" --calib \
--dir_data ./datasets

calculate PSNR/SSIM

After saving the images, modify path inmetrics/calculate_PSNR_SSIM.m to generate results.

matlab -nodesktop -nosplash -r "calculate_PSNR_SSIM('$dataset',$scale,$bit);quit"

refer to metrics/run.sh for more details.

Trained FP models and quantized models' cluster results(quantization paramters):

here Download these model. Then use the commands above to obtain the reported results of the paper.

Acknowledgments

Code is implemented based on DDTB and PAMS