graph_nets

Why we included this project

DeepMind's graph_nets library gives researchers and engineers a practical way to work with graph-structured data in TensorFlow and Sonnet. Instead of hand-coding message passing, you define edge, node, and global-level attribute updates, and the library handles the rest. That abstraction fits many real problems that don't behave like flat vectors, such as shortest path routing, physical simulations, or sorting tasks. The repo ships Jupyter notebooks for each of those three cases, so you can see whether the API matches your own problem before committing to it. Teams already on TensorFlow will find this the most direct route into graph-based modeling, though note the project has been in maintenance mode since around 2022.