Awesome

ECCV2022 Paper - Dynamic Dual Trainable Bounds for Ultra-low Precision Super-Resolution Networks paper

Dependence

• Python 3.6
• 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_limitrange_incremental.py
datasets
  benchmark
  DIV2K

Usage

1: train full-precision models:

An example:

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

Please refer to 'baseline.sh' for more commands.

2: train quantized models:

An example:

python main_limitrange_incremental.py --scale 4 \
--k_bits 4 --model EDSR \
--pre_train ./pretrained/edsr_baseline_x4.pt --patch_size 192 \
--data_test Set14 \
--dynamic_ratio 0.3 \
--save "output/edsrx4/4bit" --dir_data ./datasets --print_every 10

Please refer to 'run.sh' for more commands.

3: test quantized models

An example:

python3 main_limitrange_incremental.py --scale 4 --model EDSR \
--k_bits 4 --save_results --test_only \
--data_test Set5+Set14+B100+Urban100  \
--save "../experiment/output/edsrx4/4bit" --dir_data ./datasets

Please refer to 'test.sh' for more commands.

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: here

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

Results of pre-trained models are shown below:

EDSR

RDN

SRResNet

Contact

For any question, be free to contact: viper.zhong@gmail.com. The github issue is also welcome.