Acknowledgement

This repository is based on the original DeepSC implementation by H. Xie et al., "Deep Learning Enabled Semantic Communication Systems", IEEE Transactions on Signal Processing, 2021.

The original implementation is available at: https://github.com/13274086/DeepSC

• main.py, utils.py, transceiver.py updated
• train.py, eval.py added

This code extends the original framework with:

• representation space decomposition
• SNR-input gating mechanisms

Deep Learning Enabled Semantic Communication Systems

Huiqiang Xie, Zhijin Qin, Geoffrey Ye Li, and Biing-Hwang Juang

This is the implementation of Deep learning enabled semantic communication systems.

Requirements

• See the requirements.txt for the required python packages and run pip install -r requirements.txt to install them.

Bibtex

@article{xie2021deep,
  author={H. {Xie} and Z. {Qin} and G. Y. {Li} and B. -H. {Juang}},
  journal={IEEE Transactions on Signal Processing}, 
  title={Deep Learning Enabled Semantic Communication Systems}, 
  year={2021},
  volume={Early Access}}

Preprocess

mkdir data
wget http://www.statmt.org/europarl/v7/europarl.tgz
tar zxvf europarl.tgz
python preprocess_text.py

Train

python train.py 

Notes

• Please carefully set the $\lambda$ of mutual information part since I have tested the model in different platform, i.e., Tensorflow and Pytorch, same $\lambda$ shows different performance.
• I changed this part from main.py -> train.py (my own code)
• ex) python train.py --arch zsplit_sem --epochs 80 --channel Rician --checkpoint-path checkpoints/{arch}_80epoch ...

Evaluation

python eval.py

Notes

• If you want to compute the sentence similarity, please download the bert model.
• I changed this part from perfomance.py -> eval.py (my own code)
• BERT and sentence similarity, graph drawing, you can change your model check point directions inside eval.py