dgl

Why we included this project

If graph-structured data is a regular part of your machine learning work, DGL gives you a way to train neural networks on it without reinventing message passing. You get a graph object that can sit on CPU or GPU, holding the structure together with node and edge features, plus efficient, customizable message-passing primitives that keep GNN code readable. The library is framework agnostic, so the surrounding application can be written in PyTorch or TensorFlow and the graph model plugs in as one component. It also ships a large collection of example implementations spanning many GNN model families, and DGL-Go provides a command-line interface for training and studying state-of-the-art models without boilerplate. For anyone building real graph deep learning systems rather than a single toy model, this is a mature and well-documented base to start from.