pytorch-geometric

Overview

PyTorch Geometric (PyG) is built on top of PyTorch to simplify the implementation of Graph Neural Networks. It treats graphs as Data objects containing node features and edge indices, and provides a powerful MessagePassing base class for custom layer development.

When to Use

Use PyG for data that is naturally represented as a graph, such as social networks, molecular structures, or point clouds. Use it when you need to perform node classification, edge prediction, or graph-level regression.

Decision Tree

1. Do you have a list of small graphs?
   • USE: torch_geometric.loader.DataLoader to create a single giant disjoint graph.
2. Do you need to pool node features into a graph-level feature?
   • USE: global_mean_pool or global_max_pool using the batch vector.
3. Are you building a custom convolution?
   • INHERIT: From torch_geometric.nn.MessagePassing.

Workflows

1. Defining a Custom GNN Layer

   1. Inherit from torch_geometric.nn.MessagePassing.
   2. Set the aggregation scheme (aggr='add', 'mean', or 'max') in __init__.
   3. Implement the forward pass using self.propagate().
   4. Define the message() function to compute the transformation for each edge.
   5. Optionally define the update() function to transform aggregated results.

2. Mini-batching Large Graphs

   1. Use torch_geometric.loader.DataLoader instead of the standard PyTorch version.
   2. PyG automatically creates a single giant disjoint graph from a list of small graphs.
   3. The batch vector in the resulting Data object tracks which original graph each node belongs to.
   4. Use global pooling (e.g., global_mean_pool) to aggregate node features into graph-level representations.

3. Constructing Graphs from Point Clouds

   1. Represent point clouds as node features in a tensor [N, F].
   2. Apply knn_graph() to dynamically compute an edge_index based on spatial proximity.
   3. Pass the results into an EdgeConv or DynamicEdgeConv layer for feature extraction.

Non-Obvious Insights

• Auto-Indexing: The _i and _j notation in MessagePassing methods automatically handles indexing into source and destination nodes during propagation without manual slice logic.
• Disjoint Representation: PyG batches multiple graphs by stacking them into a single block-diagonal adjacency matrix. This allows standard sparse matrix operations to process multiple graphs in parallel without zero-padding.
• Modular Aggregation: Aggregation is a first-class principle in PyG; users can swap simple sum/mean with advanced learnable schemes like SoftmaxAggregation by simply changing the aggr parameter.

Evidence

• "PyG provides the MessagePassing base class, which helps in creating such kinds of message passing graph neural networks by automatically taking care of message propagation." (https://pytorch-geometric.readthedocs.io/en/latest/tutorial/create_gnn.html)
• "Tensors passed to propagate() can be mapped to the respective nodes i and j by appending _i or _j to the variable name." (https://pytorch-geometric.readthedocs.io/en/latest/tutorial/create_gnn.html)

Scripts

• scripts/pytorch-geometric_tool.py: Template for a custom GNN layer and graph data loader.
• scripts/pytorch-geometric_tool.js: Node.js script to run PyG training experiments.

Dependencies

• torch
• torch-geometric
• torch-scatter / torch-sparse (optional but recommended)

References

• PyTorch Geometric Reference