This is a PyTorch implementation of the paper "Integrated Communication and Learned Recognizer with Customized RIS Phases and Sensing Durations" in IEEE Transactions on Communications. Its conference version has been presented in FCN 2024 and received the Best Student Paper Award, Valletta, Malta, Nov. 2024. Arxiv link: https://arxiv.org/abs/2503.02244

This paper realizes high-accuracy and fast target recognition in RIS-aided ISAC systems using wireless signals. Specifically, the RIS phases are customized according to the scene, task, quantization, and target priors, and the sensing duration is dynamically determined based on the previously captured information.

Packages

• python==3.11.5
• pytorch==2.0.1
• numpy==1.26.1
• wandb==0.16.0

Training

The training scripts come with several options. An example for training is:

python main.py --is_train True --num_glimpses 5 --learned_start True --wandb_project 'CRISense'  

Testing

The codes rely on the .csv files exported from wandb to conduct tests. An example file name is:

wandb_export_2025-03-01T12_00_00.000+08_00.csv  

An example for testing is:

python main.py --is_train False --test_index 1 --test_wandb_data 'wandb_export_2025-03-01T12_00_00.000+08_00.csv'  

Citation

@article{huang2025integrated,
  title={Integrated Communication and Learned Recognizer with Customized {RIS} Phases and Sensing Durations},  
  author={Huang, Yixuan and Yang, Jie and Wen, Chao-Kai and Jin, Shi},
  journal={IEEE Trans. Commun.},
  volume={73},
  number={9},
  pages={7332--7348},
  year={Sep. 2025},
  publisher={IEEE}
}

@inproceedings{huang2024learned,
  title={Learned intelligent recognizer with adaptively customized RIS phases in communication systems},
  author={Huang, Yixuan and Yang, Jie and Wen, Chao-Kai and Xia, Shuqiang and Li, Xiao and Jin, Shi},
  booktitle={Proc. Int. Conf. Future Commun. Netw. (FCN)},
  pages={1--6},
  year={Nov. 2024}
}

References

This work is motivated by Recurrent Models of Visual Attention and Dynamic Computational Time for Recurrent Attention Model in the computer vision filed.

The codes for CRISense are built on recurrent-visual-attention and DT-RAM using wireless communication and sensing knowledge.