Download
Graph Nets, developed by Google DeepMind, is a Python library designed for constructing and training graph neural networks (GNNs) using TensorFlow and Sonnet. It provides a high-level, flexible framework for building neural architectures that operate directly on graph-structured data. A graph network takes graphs as inputs, consisting of edges, nodes, and global attributes, and produces updated graphs with modified feature representations at each level. This library implements the foundational ideas from DeepMind’s paper “Relational Inductive Biases, Deep Learning, and Graph Networks”, offering tools to explore relational reasoning and message-passing neural networks. Graph Nets supports both TensorFlow 1 and TensorFlow 2, working with CPU and GPU environments, and includes educational Jupyter demos for shortest path finding, sorting, and physical prediction tasks. The codebase emphasizes modularity, allowing users to easily define their own edge, node, and global update functions.
Features
- Framework for building graph neural networks using TensorFlow and Sonnet
- Supports graph-level, node-level, and edge-level feature learning
- Compatible with TensorFlow 1.x and 2.x, on both CPU and GPU setups
- Includes Colab and Jupyter demo notebooks for hands-on learning and experimentation
- Enables modular architecture design with customizable graph update functions
- Suitable for a range of tasks including physical simulation, sorting, and pathfinding
Project Samples
