GeNet

This repo is the implementation of the paper "GeNet: A Graph Neural Network-based Anti-noise Task-Oriented Semantic Communication Paradigm": https://arxiv.org/abs/2403.18296.

Architecture

General semantic communication paradigm: GeNet semantic communication paradigm:

Installation

And all experiments are conduct either on a computer equipped with a single NVIDIA GeForce RTX3090 GPU (Ubuntu 20.04, CUDA 12.0, Python 3.10.13) or one with two NVIDIA GeForce RTX3090 GPUs (Ubuntu 22.04, CUDA 11.6, Python 3.10.13).

git clone https://github.com/chunbaobao/GeNet
cd GeNet
pip install -r requirements.txt

Dataset Preparation

The prepare_dataset.py script is used to convert images from CIFAR-10, MNIST, and FashionMNIST datasets into graph structures.

python prepare_dataset.py # for all datasets
python prepare_dataset.py --dataset $DATASET_NAME # for a specific dataset

The $DATASET_NAME.pkl files will be saved in the ./data directory by default.

Training

The train.py script is used to train the GeNet model.

python train.py --out $OUTPUT_DIR --dataset_name $DATASET_NAME --model_name $MODEL_NAME

For loops within Python can cause process killing for unknown reasons, so bash script are provided for training all backbone GNN models of GeNet on all datasets.

bash run.sh

Evaluation

The eval.py provides the evaluation of the trained GeNet model and baseline models.

You may need modify slightly to evaluate the GeNet or baseline models in __main__ function. And for GeNet evaluation, you need to specify the path to the trained model in eval_model function.

python eval.py

All training and evaluation results are saved in the ./out directory by default. The ./out directory may contain the structure as follows:

./out
├── checkpoint # trained models
│   ├── $MODELNAME_$DATASETNAME_$TIMES_on_$DATE_$HOST
│       ├── epoch_$num.pth
│       ├── ...
│   ├── GAT_CIFAR10_03h57m06s_on_Mar_15_2024_PC
│   ├── GATEDGCN_CIFAR10_06h53m54s_on_Mar_15_2024_PC
│   ├── ...
├── configs # training configurations
│   ├── $MODELNAME_$DATASETNAME_$TIMES_on_$DATE_$HOST.yaml
│   ├── ...
├── logs # training logs
│   ├── $MODELNAME_$DATASETNAME_$TIMES_on_$DATE_$HOST
│       ├── tensorboard logs
│   ├── ...
├── eval # evaluation results
│   ├── n_sp # for number of superpixels evaluation
│   ├── SNR # for SNR evaluation
│   ├── rotation # for rotation evaluation
│   ├── cross # for cross evaluation between n_sp and SNR
│       ├── tensorboard logs
│   ├── ...

Visualization

The ./visualization directory contains the scripts for visualization of the training and evaluation results.

• dataset_visualization.ipynb is used to get visualizations of the dataset properties.
• painted_visualization.ipynb is used to get visualizations of the painted images of different rotation angles.
• pipeine_visualization.ipynb is used to get visualizations of the pipeline of GeNet. And the materials of ./demo/pipeline.pdf are generated by this script.
• superpixels_visualization.ipynb is used to get visualizations of the superpixels of the images.
• plot_visualization.ipynb is used to get visualizations of the training and evaluation results.

All figures are saved in the ./demo directory by default.

Acknowledgement

This repo refers the graph generation and nn modules from benchmarking-gnns and the superpixel generation from graph_attention_pool