wsyCUHK / Discrete-TaskOriented-JSCC

DT-JSCC: discrete joint source-channel coding for task-oriented communication with digital modulation

Discrete Task-oriented Joint Source-Channel Coding (DT-JSCC)

This is a Pytorch implementation of DT-JSCC for task-oriented communication with digital modulation, as proposed in the paper Robust Information Bottleneck for Task-Oriented Communication with Digital Modulation.

Requirements

The codes are compatible with the packages:

• pytorch 1.8.0

• torchvision 0.9.0a0

• numpy 1.23.1

• tensorboardX 2.4

The code can be run on the datasets such as MNIST and CIFAR-10, etc. One should download the datasets in a directory (e.g., ./data/) and change the root parameter in datasets/dataloader.py, e.g.,

root = r'./data/

Run experiments

Training the DT-JSCC model

1. Training the DT-JSCC on the MNIST dataset

   python main.py --dataset MNIST --channels 1 --lam 1e-3 --lr 1e-3 --epoches 400 --latent_d 64 --num_latent 16 --num_embeddings 16 --psnr 4

2. Training the DT-JSCC on the CIFAR-10 dataset

   python main.py --dataset CIFAR10 --mod psk --lam 1e-3 --lr 1e-3 --epoches 320 --num_embeddings 16 --psnr 4

The parameter num_embeddings is the size of trainable codebook $K$, the parameter latent_d is the length of codeword $D$, the num_latent is the dimension $d$ of encoded representation $\mathbf{z}$ and the psnr is the PSNR of AWGN channel. In the experiments, $Dd = 1024$ for MNIST dataset and $D=512$ and $d=16$ for CIFAR-10 dataset.

Evaluating the trained DT-JSCC model

python evaluate.py --dataset CIFAR10 --save_root ./results --name CIFAR10-num_e16-num_latent4-modpsk-snr10.0-lam0.0

The parameter name is the trained model.

Citation

@article{xie2022robust,
  title={Robust Information Bottleneck for Task-Oriented Communication with Digital Modulation},
  author={Xie, Songjie and Wu, Youlong and Ma, Shuai and Ding, Ming and Shi, Yuanming and Tang, Mingjian},
  journal={arXiv preprint arXiv:2209.10382},
  year={2022}
}