rusty1s/pytorch_geometric
Pytorch Geometric
PyTorch Geometric is a library for building and training machine learning models on graph-structured data. It provides a framework for developing graph neural networks, including a specialized system for implementing node-to-node information exchange via customizable message passing, aggregation, and update functions.
The library includes a sparse data processing toolkit that utilizes accelerated CPU and CUDA kernels to perform efficient reductions on large sparse datasets. It supports the creation of specialized architectures for structured data such as 3D meshes and point clouds.
The project covers a broad data pipeline for loading, transforming, and sampling massive heterogeneous graphs. This includes capabilities for mini-batch graph loading and sub-graph sampling to enable training on datasets that exceed available system memory.
Features
- Graph Neural Networks - Provides a comprehensive framework for building and training graph neural networks on structured data.
- Graph Message Passing Frameworks - Implements a graph message passing framework for defining how nodes exchange and aggregate information.
- Heterogeneous Graph Modeling - Provides specialized support for representing and learning from graphs with multiple types of nodes and edges.
- Sparse Data Processing - Provides accelerated CPU and CUDA kernels to perform efficient mathematical reductions on large sparse datasets.
- Dataset Batch Loading - Provides mechanisms for loading graph datasets using mini-batch loaders and a library of common benchmarks.
- Graph Batching Optimizations - Implements graph batching optimizations to group multiple small graphs into a single disjoint graph for GPU processing.
- Graph Neighborhood Sampling - Implements algorithms for sampling representative sub-graphs to enable training on datasets that exceed system memory.
- Custom GNN Layers - Supports the creation of specialized GNN architectures via custom message-passing and aggregation methods.
- Graph Dataset Pipelines - Ships a complete toolset for loading, transforming, and sampling massive graphs and 3D point clouds.
- Graph Transformation Engines - Includes graph transformation engines to process arbitrary graphs, point clouds, and 3D meshes for ML tasks.
- Large-Scale Graph Training - Runs learning processes on massive datasets with millions of nodes or complex heterogeneous structures.
- Sparse Computing Kernels - Provides accelerated CPU and CUDA kernels for performing efficient mathematical reductions on sparse datasets.
- Graph Neighborhood Sampling - Extracts representative sub-sections of massive graphs to enable training on datasets that exceed system memory.
- Geometric Deep Learning: Graph & Irregular Structures - Listed in the “Geometric Deep Learning: Graph & Irregular Structures” section of the The Incredible Pytorch awesome list.
