h-DJSCC: Deep Joint Source-Channel Coding for Hybrid Mobile Multi-hop Networks

This repository contains the source code for the paper:

A Deep Joint Source-Channel Coding Scheme for Hybrid Mobile Multi-hop Networks
Chenghong Bian, Yulin Shao, Deniz Gunduz
Published in IEEE Journal on Selected Areas in Communications (JSAC), 2025
IEEE Xplore Link

---

Overview

h-DJSCC introduces a novel Deep Joint Source-Channel Coding (DJSCC) framework tailored for C-RAN (cloud radio access network). The key innovation is to efficiently and effectively convert the analog DeepJSCC codewords into bit sequences to be transmitted over the fibers/backhaul links.

Key features include:

• DeepJSCC codeword Compression: Instead of naive vector quatization schemes, we consider learning based image compression algorithm to compress the analog DeepJSCC codewords.
• SNR-adaptive: SNR-adaptive transmission for DeepJSCC has been introduced in the literature, yet how to seamlessly combine it with the compression module is unexplored.
• Variable rate compression: To achieve different rate-distortion trade-offs using a single model, we consider using the technics in the image compression literature.

---

Repository Structure

• dataset.py: Handles data loading and preprocessing for CIFAR10 and CelebA datasets.
• entropy_models.py: Compression modules copied from the CompressAI.
• get_args.py: Hyper parameters for the h-DJSCC framework.
• modules.py: DeepJSCC encoder and decoders.
• models/: The folder for the checkpoints.
• multihop_full_adapt.py: Implementations for h-DJSCC framework which integrates both DeepJSCC and fully-adaptive compression models.
• run_full_finetune_hyper.py: Script for training and evaluating the h-DJSCC model.
• utils.py: Utility functions.

---

Getting Started

Training

To train the h-DJSCC model over AWGN and fading channel (with CSIT):

python run_full_finetune_hyper.py --resume False --fading False --snr_min 1 --snr_max 9
python run_full_finetune_hyper.py --resume False --fading True --snr_min 5 --snr_max 15

Also adjust the variable lamdas in the run_full_finetune_hyper.py file as needed.

Evaluation

After training, evaluate the model's performance for the AWGN channel:

python multihop_full_adapt.py --resume True --fading False --snr_min 1 --snr_max 9

Checkpoints

Download some trained models from the google drive link. Place it under the models/ folder.

Citation

If you find this work useful in your research, please cite:

@article{hdjscc,
  title={A Deep Joint Source-Channel Coding Scheme for Hybrid Mobile Multi-hop Networks},
  author={Bian, Chenghong and Shao, Yulin and Gunduz, Deniz},
  journal={IEEE Journal on Selected Areas in Communications},
  year={2025},
  volume={43},
  number={5},
  pages={1234--1245},
  doi={10.1109/JSAC.2025.10960356}
}