Home

Welcome to IGNNITION​

Recent years have witnessed the vast potential of Graph Neural Networks (GNN) to be applied to a plethora of problems where data is structured as graphs (e.g., computer networks, chemistry, physics). Typically, each new application requires custom GNN models adapted to the problem environment. In this context, implementing a GNN model is a cumbersome task that currently requires Machine Learning (ML) experts with high skills in neural network programming. For instance, to translate the model to complex tensor-wise operations in languages such as TensorFlow or PyTorch. We claim that, in order to approach GNNs to real-world applications it is essential to count on tools that abstract users from the complexity behind the implementation of such models. This repository presents IGNNITION, a novel framework to fast prototype complex GNN models. IGNNITION targets users with little to no background on neural network programming, while still providing great flexibility to implement a broad variety of GNN architectures. This framework enables to define a GNN model as an intuitive human-readable description of the objects present in graphs and the type of relations they may have, and it automatically generates an efficient TensorFlow implementation of the model. In order to assist users, it provides advanced debugging mechanisms to easily identify and correct possible errors in GNN models. Moreover, GNN models generated by our framework do not incur performance loss compared to native implementations in TensorFlow.

IGNNITION is an open source project that loves your comments and contributions. Please join our community:

• How to contribute
• Report bugs
• Mailing Lists