GAUDI

By Mirja Granfors, Jesús Pineda, Blanca Zufiria Gerbolés, Joana B. Pereira, Carlo Manzo, and Giovanni Volpe.

GAUDI is an unsupervised geometric deep learning framework for analyzing complex graph-structured data. GAUDI's hourglass architecture, with multiple hierarchical pooling and upsampling steps, maps graphs into a structured latent space, capturing their underlying parameters.

This repository contains an implementation of GAUDI, introduced in Global graph features unveiled by unsupervised deep learning.

The data folder contains the Single-Molecule Localization Microscopy (SMLM) simulations and the script used to generate the Vicsek model simulations.

Getting started

An example of how GAUDI is trained on Watts-Strogatz small-world graphs can be found here:

Training GAUDI on Watts-Strogatz Small-World Graphs

To run the example, first download or clone this GitHub repository, then run the notebook from within the repository folder.

Running the example takes about 5 minutes on a standard laptop.

Dependencies

To use this implementation, ensure you have the following dependencies installed:

• deeplay (tested with 0.1.3)
• torch (tested with 2.6.0 and 2.7.0)

You can install them using:

pip install deeplay

pip install torch

Additional dependencies for the example

If you also want to run the provided Watts-Strogatz example, you’ll need these extra packages:

• PyGSP (tested with 0.5.1)
• networkx (tested with 3.4.2)
• torch-geometric(tested with 2.5.2, 2.6.1)

You can install them using:

pip install PyGSP

pip install networkx

pip install torch-geometric

Citation

If you use GAUDI in your project, please cite us: http://iopscience.iop.org/article/10.1088/2632-2153/ae8d7f

"Global graph features unveiled by unsupervised deep learning"
Mirja Granfors, Jesús Pineda, Blanca Zufiria-Gerbolés, Joana B. Pereira, Carlo Manzo and Giovanni Volpe
Machine Learning: Science and Technology (2026).