torch-geometric

Graph neural networks enable learning on graph-structured data using layers like GCN, GAT, and GraphSAGE. PyTorch Geometric provides pre-built GNN layers, standard data objects (node features, edge indices, labels), and mini-batch processing to train models for node classification, graph classification, link prediction, and molecular property prediction tasks.

PyTorch Geometric includes GCNConv, GATConv, GraphSAGE, GINConv, TransformerConv, and ChebConv layers. These primitives cover attention-based, spatial, and spectral approaches, enabling flexible architecture design for diverse graph learning tasks from citation networks to 3D point clouds.

Yes. PyTorch Geometric integrates with PyTorch's multi-GPU training capabilities. The library supports mini-batch processing and standard PyTorch data objects, allowing distributed training across GPUs for large-scale graph datasets.

PyTorch Geometric uses standard Data objects with node features (x), edge indices (edge_index), edge attributes (edge_attr), and labels (y). The library includes built-in datasets for common benchmarks and templates in scripts/ to adapt your own graph data.

PyTorch Geometric supports node classification, graph classification, link prediction, molecular property prediction, social and citation network analysis, and 3D geometric data processing. Pre-built architectures and datasets accelerate implementation for these common tasks.