How pyg-team/pytorch_geometric works
PyTorch Geometric is a graph and geometric deep-learning library. Its primary user workflow is to obtain graph data, represent and batch it, optionally transform or sample it, run graph neural network operators and models, then train or explain the results. It also supports heterogeneous and large-scale/distributed graph workflows, graph transformations, benchmark datasets, and GraphGym experiment orchestration. The README documents the workflow; sampled entry-point code confirms GraphGym orchestration. Other subsystem links are limited to explicit imports or documented relationships, since much of the runtime was not sampled.
Main entry points: data.py Graph Data · main.py GraphGym CLI · conv GNN Operators · loader.py GraphGym Data
Components 27
Graph Data
- Datasets
torch_geometric/datasets - Graph Data
torch_geometric/data/data.py - Heterogeneous Data
torch_geometric/data/hetero_data.py - Batching
torch_geometric/data/batch.py - Graph Transforms
torch_geometric/transforms - Mini-batch Loaders
torch_geometric/loader
GNN Compute
- GNN Operators
torch_geometric/nn/conv - GNN Models
torch_geometric/nn/models - Hetero Conversion
torch_geometric/nn/to_hetero_with_bases_transformer.py
Learning Workflows
- GraphGym Training
torch_geometric/graphgym/train.py - GraphGym Data
torch_geometric/graphgym/loader.py - GraphGym Config
torch_geometric/graphgym/config.py - Component Registry
torch_geometric/graphgym/register.py - GNN Explanations
torch_geometric/explain/config.py - Link Metrics
torch_geometric/metrics/link_pred.py - GraphGym CLI
graphgym/main.py
Scale and Storage
- Graph Store
torch_geometric/data/graph_store.py - Feature Store
torch_geometric/data/feature_store.py - Neighbor Sampling
torch_geometric/sampler/neighbor_sampler.py - Distributed Runtime
torch_geometric/distributed - Data Database
torch_geometric/data/database.py
Core Platform
- Edge Index
torch_geometric/edge_index.py - Graph Utilities
torch_geometric/utils - Backend Typing
torch_geometric/typing.py - Logging
torch_geometric/logging.py - Warnings
torch_geometric/warnings.py
Other components
- Library User
Connections 31
- Library User → Datasets: loads data
- Library User → GNN Operators: builds model
- Datasets → Graph Data: provides graphs
- Graph Data → Batching: batches graphs
- Graph Data → Graph Transforms: transforms graph
- Graph Data → Mini-batch Loaders: feeds loader
- Mini-batch Loaders → Neighbor Sampling: samples neighbors
- Graph Data → Heterogeneous Data: represents graph data
torch_geometric/data/__init__.py - Graph Data → Graph Store: exposes store API
torch_geometric/data/__init__.py - Graph Data → Feature Store: exposes store API
torch_geometric/data/__init__.py - Graph Data → Data Database: exposes database API
torch_geometric/data/__init__.py - GNN Operators → Edge Index: operates on edges
- GNN Operators → GNN Models: builds models
- GNN Models → Library User: produces predictions
- Heterogeneous Data → Hetero Conversion: supports conversion
- Library User → GNN Explanations: configures explanations
- Library User → Link Metrics: evaluates links
- Library User → Distributed Runtime: runs at scale
- Library User → GraphGym CLI: runs experiment
graphgym/main.py - GraphGym CLI → GraphGym Config: loads configuration
graphgym/main.py - GraphGym CLI → GraphGym Data: creates data module
graphgym/main.py - GraphGym CLI → GNN Models: creates model
graphgym/main.py - GraphGym CLI → GraphGym Training: starts training
graphgym/main.py - GraphGym CLI → Logging: sets printing
graphgym/main.py - GraphGym Config → Component Registry: uses registry
torch_geometric/graphgym/config.py - GraphGym Data → GraphGym Config: reads config
torch_geometric/graphgym/loader.py - GraphGym Data → Component Registry: uses registry
torch_geometric/graphgym/loader.py - Graph Data → Graph Utilities: uses graph operations
- Neighbor Sampling → Graph Store: reads graph store
torch_geometric/sampler/neighbor_sampler.py - Graph Data → Backend Typing: uses backend types
- Data Database → Warnings: emits warnings
torch_geometric/data/database.py