Awesome
DAQ: Channel-Wise Distribution-Aware Quantization for Deep Image Super-Resolution Networks
Official implementation of our WACV 2022 paper.
Conda Environment setting
conda env create -f environment.yml --name DAQ
conda activate DAQ
conda install -c anaconda scikit-image
Dependencies
- Python 3.6
- PyTorch == 1.1.0
- coloredlogs >= 14.0
- scikit-image
Codes
Our implementation is based on EDSR(PyTorch).
Train
sh train_EDSR_x4.sh
Pretrained model to start training from can be accessed from Google Drive.
Test
sh test.sh edsr_baseline 2 2 4 (edsr_baseline w2a2qq4)
sh test.sh edsr_baseline 3 3 4 (edsr_baseline w3a3qq4)
sh test.sh edsr_baseline 4 4 4 (edsr_baseline w4a4qq4)
sh test.sh edsr_full 2 2 8 (edsr_full w2a2qq8)
Our pretrained model can be accessed from Google Drive.
Additional Results
Our model achieves the following performance (PSNR / SSIM) when trained for 60 epochs :
| Model | Precision (w/a) | Set5 | Set14 | B100 | Urban100 |
|---|---|---|---|---|---|
| EDSR-baseline (x4) | 32 / 32 | 32.10 / 0.894 | 28.58 / 0.781 | 27.56 / 0.736 | 26.04 / 0.785 |
| EDSR-baseline-DAQ | 4 / 4 | 31.85 / 0.887 | 28.38 / 0.776 | 27.42 / 0.732 | 25.73 / 0.772 |
| EDSR-baseline-DAQ | 3 / 3 | 31.66 / 0.884 | 28.19 / 0.771 | 27.28 / 0.728 | 25.40 / 0.762 |
| EDSR-baseline-DAQ | 2 / 2 | 31.01 / 0.871 | 27.89 / 0.762 | 27.09 / 0.719 | 24.88 / 0.740 |
| Model | Precision (w/a) | Set5 | Set14 | B100 | Urban100 |
|---|---|---|---|---|---|
| RDN (x4) | 32 / 32 | 32.24 / 0.896 | 28.67 / 0.784 | 27.63 / 0.738 | 26.29 / 0.792 |
| RDN-DAQ | 4 / 4 | 31.91 / 0.889 | 28.38 / 0.775 | 27.38 / 0.733 | 25.81 / 0.779 |
| RDN-DAQ | 3 / 3 | 31.57 / 0.883 | 28.18 / 0.771 | 27.27 / 0.728 | 25.47 / 0.765 |
| RDN-DAQ | 2 / 2 | 30.71 / 0.866 | 27.61 / 0.755 | 26.93 / 0.715 | 24.71 / 0.731 |
| Model | Precision (w/a) | Set5 | Set14 | B100 | Urban100 |
|---|---|---|---|---|---|
| SRResNet (x4) | 32 / 32 | 32.07 / 0.893 | 28.50 / 0.780 | 27.52 / 0.735 | 25.86 / 0.779 |
| SRResNet-DAQ | 4 / 4 | 31.85 / 0.889 | 28.41 / 0.777 | 27.45 / 0.732 | 25.70 / 0.772 |
| SRResNet-DAQ | 3 / 3 | 31.81 / 0.889 | 28.35 / 0.776 | 27.40 / 0.733 | 25.63 / 0.772 |
| SRResNet-DAQ | 2 / 2 | 31.57 / 0.886 | 28.19 / 0.773 | 27.30 / 0.729 | 25.39 / 0.765 |
Citation
@inproceedings{hong2022daq,
title={DAQ: Channel-Wise Distribution-Aware Quantization for Deep Image Super-Resolution Networks},
author={Hong, Cheeun and Kim, Heewon and Baik, Sungyong and Oh, Junghun and Lee, Kyoung Mu},
booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
pages={2675--2684},
year={2022}
}
Contact
Email : cheeun914@snu.ac.kr