torch-geometric

SUMMARY

What it does

This Agent Skill provides procedural guidance for using PyTorch Geometric (PyG) to build and train graph neural networks. It covers core data structures (Data, HeteroData), dataset loading, common GNN layers (GCN, GAT, GraphSAGE, GIN, etc.), building custom message passing layers, task-specific patterns for node/graph/link classification, scaling to large graphs via neighbor sampling, heterogeneous graph support, custom dataset creation, and explainability. The skill includes installation instructions, code examples, and common pitfalls. It is intended for use with torch_geometric and is not for general NetworkX analytics or non-graph PyTorch models.