Graph-RelNet

This repository contains the implementation for the paper:

"Message Passing Neural Networks for Sound Source Localization", Matija Marijan, Miloš Bjelić, presented at the 33rd Telecommunications Forum (TELFOR), Belgrade, Serbia, 2025.

This repository provides code for training graph neural networks (GNNs) on sound source localization tasks, supporting experiments with various microphone array geometries, synthetic and real audio signals, and deep learning models RelNet and Graph-RelNet. The RelNet model is adapted from GNN_SSL.

Directory Structure

• models/: Contains model definitions (Graph-RelNet and RelNet).
• utils/: Utility functions for datasets, geometry, signal processing, and general helpers.
• evaluation/: Scripts and notebooks for analyzing results.
• data/: Default location for generated or real datasets.
• create_dataset.py: Script for generating datasets.
• training.py: Main training script.
• prediction.py: Script for running model inference.
• run.sh: Example shell script for running experiments.

Installation

1. Clone the repository.
2. Create a conda environment using the provided environment.yml:

   conda env create -f environment.yml
   conda activate geometric

3. To download the TIMIT dataset, visit this link, unzip the archive, and place its contents under data/signals/timit/.

Usage

• Dataset Generation:

  python create_dataset.py --help

• Training:

  python training.py --help

Requirements

• Python 3.12
• PyTorch, PyTorch Geometric, and related dependencies (see environment.yml)