gerritgr / gina

Neural Graph Inference From Independent Snapshots of Interacting Systems

GINA

Unsupervised relational inference using masked reconstruction

Copyright: 2023, Gerrit Großmann, Group of Modeling and Simulation at Saarland University

Official implementation of Unsupervised relational inference using masked reconstruction

Version: 0.1 (Please note that this is proof-of-concept code.)

Overview

Gina takes observations of interacting systems and infers/reconstructs the (latent) underlying interaction graph (contact network).

Here, we see the successful reconstruction of a 7x7 grid graph.

Method

GINA considers each snapshot (observation of all components) individually and tries to predict the observable state of each node, given the measurements of each adjacent node. GINA optimizes the interaction graph in order to maximize the accuracy of this prediction.

Usage

Possibility I: With Standard Python

Install Python3, Pip3, and Jupyter notebook. Install Python-dependencies with:

pip install -r requirements.txt

Start Jupyter notebook

jupyter notebook

and select the GINA.ipynb file.

Possibility II: With Colab

You can upload the GINA.ipynb file to Google Colab or click on the open in colab badge at the top.

Possibility III: Within a Conda Environment

To use GINA in a Conda environment: Fist, install Miniconda. Then,

conda create -n ginaenv python=3.6
conda install -n ginaenv pip
conda activate ginaenv
conda install nb_conda
pip install -r requirements.txt
jupyter lab

and open the GINA.ipynb file.

Possibility IV: Use Docker

You can use the GINA docker image.

docker pull gerritgr/gina:2022-07-17--13-11

Run with:

docker run -p 8888:8888 gerritgr/gina:2022-07-17--13-11

and open the GINA.ipynb file in Jupyter lab.