Nonlinear Transform Source-Channel Coding for Semantic Communications

Pytorch Implementation of JSAC 2022 Paper "Nonlinear Transform Source-Channel Coding for Semantic Communications"

Arxiv Link: https://arxiv.org/abs/2112.10961

Project Page: https://semcomm.github.io/ntscc/

Usage

Pretrained Models

Pretrained models (optimized for MSE) trained from scratch using randomly chose 300k images from the OpenImages dataset.

Other pretrained models will be released successively.

Note: We reorganize code and the performances are slightly different from the paper's.

RD curves

Citation

If you find the code helpful in your resarch or work, please cite:

@ARTICLE{9791398,
  author={Dai, Jincheng and Wang, Sixian and Tan, Kailin and Si, Zhongwei and Qin, Xiaoqi and Niu, Kai and Zhang, Ping},
  journal={IEEE Journal on Selected Areas in Communications}, 
  title={Nonlinear Transform Source-Channel Coding for Semantic Communications}, 
  year={2022},
  volume={40},
  number={8},
  pages={2300-2316},
  doi={10.1109/JSAC.2022.3180802}
  }

Acknowledgements

The NTSCC model is partially built upon the Swin Transformer and CompressAI. We thank the authors for sharing their code.