graph_nets

Released alongside DeepMind's foundational paper on relational inductive biases in deep learning, Graph Nets provides production-quality implementations of the message-passing neural network paradigm that underlies modern GNN architectures.

The library's core abstraction is the GraphsTuple data structure a batch-friendly representation of graphs with attributes on nodes, edges, and the global graph level and the GraphNetwork module, which implements the general GNN update function covering node updates, edge updates, and global updates in a composable, differentiable framework.

This flexibility allows Graph Nets to implement GNN variants including graph convolutional networks, graph attention networks, message-passing neural networks, and interaction networks by configuring the update functions appropriately.

Graph Nets is open-source under the Apache 2.0 license and is primarily used in scientific ML research molecular property prediction, physics simulation, combinatorial optimization, social network analysis, and knowledge graph reasoning where graph structure is the natural representation of the data.

The library served as an educational reference implementation for GNN concepts and influenced the design of more recent GNN frameworks including PyTorch Geometric and DGL (Deep Graph Library).

While newer GNN frameworks have largely supplanted it for new projects, Graph Nets remains a well-documented reference for understanding foundational GNN design patterns.