zhuzhui-2000 / mobilesuperresolution

[ECCV2022] Compiler-Aware Neural Architecture Search for On-Mobile Real-time Super-Resolution

Pytorch Implementation of Paper [Arxiv]

Usage

Dependencies

conda create -n sr-nas python=3.8
conda activate sr-nas
conda install -y pytorch==1.9.1 torchvision==0.10.1 cudatoolkit=10.2 tensorboard h5py scikit-image -c pytorch

Train

• Configuration
  • dataset (default: div2k): train dataset
  • eval_datasets (set5/set14/urban100/bsds100...): evaluation dataset
  • scale (4): scale factor
  • num_blocks (default: 16): number of blocks in wdsr
  • num_residual_units (default 24): number of residual units in wdsr

1. Download the dataset, put them in folder data
2. Prepare dataset using prepare_dataset.py before distributed training
3. Training
   1. Pretrain
      1. Using pretraining.bash to train a pretrained model. Two pretrained weights have already offered here
   2. Search
      • You may modify the weights here to fine-tune search results
      • Multi GPUs Training
        • modify configuration in train.bash then run bash train.bash <log path (optional)>

• Speed model
  • Here are several trained speed models provided here for different feature size and platform

Datasets

DIV2K dataset: DIVerse 2K resolution high quality images as used for the NTIRE challenge on super-resolution @ CVPR 2017

Benchmarks (Set5, BSDS100, Urban100)

Download and organize data like:

./data/DIV2K/
├── DIV2K_train_HR
├── DIV2K_train_LR_bicubic
│   └── X2
│   └── X3
│   └── X4
├── DIV2K_valid_HR
└── DIV2K_valid_LR_bicubic
    └── X2
    └── X3
    └── X4
./data/Set5/*.png
./data/BSDS100/*.png
./data/Urban100/*.png

Acknowledgements

https://github.com/ychfan/wdsr

Citation

If you find this code useful for your research, please cite our paper

@misc{https://doi.org/10.48550/arxiv.2207.12577,

  author = {Wu, Yushu and Gong, Yifan and Zhao, Pu and Li, Yanyu and Zhan, Zheng and Niu, Wei and Tang, Hao and Qin, Minghai and Ren, Bin and Wang, Yanzhi},
  title = {Compiler-Aware Neural Architecture Search for On-Mobile Real-time Super-Resolution},
  doi = {10.48550/ARXIV.2207.12577},
  url = {https://arxiv.org/abs/2207.12577},
  publisher = {arXiv},
  year = {2022},
}