dmlc/dgl

Features

• Graph Message Passing Frameworks - Implements a generic compute pattern for aggregating and updating node and edge features across graph topologies.
• Message Passing Primitives - Implements a generic compute pattern for aggregating and updating node and edge features across graph topologies.
• Neighborhood Sampling - Reduces memory overhead by extracting small subsets of node neighborhoods for training on massive datasets.
• Deep Learning Frameworks - Integrates with existing deep learning engines to build and train graph-structured models.
• Distributed Training - Distributes model training across multiple machines to handle billion-sized datasets.
• Geometric Deep Learning Frameworks - Provides a toolkit for analyzing 3D point clouds and molecular structures by treating them as graphs.
• Graph Neighborhood Sampling - Reduces computational overhead by selecting a representative subset of neighbors when training large-scale networks.
• Heterogeneous Graph Modeling - Processes graphs with multiple node and edge types to model complex real-world entities.
• Large-Scale Model Training - Supports training models on massive, billion-edge datasets using distributed computing and neighbor sampling to exceed single-device capacity.
• Large-Scale Graph Training - Provides distributed training and neighbor sampling to handle graphs with billions of edges.
• Graph - Implements neural network layers that apply message passing to graph structures to learn node and edge representations.
• Aggregation Models - Processes graph convolutions by iteratively aggregating feature information from neighboring nodes across multiple layers.
• Pooling Mechanisms - Provides mechanisms to collapse node-level representations into a single vector for whole-graph classification.
• Heterogeneous Graph Modeling - Handles graphs with multiple node and edge types to learn representations across diverse entity relationships.
• Graph Representation Learning - Transforms nodes, edges, or entire graphs into low-dimensional vectors that preserve topological and feature information.
• Graph Neighborhood Sampling - Extracts subsets of nodes and edges using neighbor or layer-wise sampling to enable training on large-scale graphs.
• Graph Processing - Handles graphs containing multiple types of nodes and edges to model complex real-world entities.
• Distributed GPU - Distributes deep learning workloads across multiple graphics processors to process billion-sized graphs efficiently.
• Graph Partitioning Utilities - Implements utilities for splitting large-scale graphs across multiple machines to enable training on massive datasets.
• Heterogeneous - Uses typed node and edge identifiers to manage complex real-world entity relationships within a unified data structure.
• Data-Parallel Training - Distributes neural network workloads across several graphics processors using data parallelism to increase throughput.
• Custom Neural Architectures - Allows the assembly of custom neural network architectures using pre-defined layers and modules.
• Model Performance Benchmarking - Executes standardized benchmark suites to track execution time and accuracy across different hardware and versions.
• Link Prediction - Forecasts the likelihood of an edge existing between two nodes to complete missing connections or recommend relationships.
• Training Execution Loops - Manages the full training process from configuration to performance metric output.
• Node Classification - Predicts labels for individual nodes using semi-supervised or supervised learning techniques on graph-structured data.
• Spatio-Temporal Graph Neural Networks - Analyzes graphs that evolve over time to perform forecasting or activity monitoring.
• Framework Abstraction Layers - Provides a backend-agnostic tensor interface to ensure graph logic runs across different deep learning frameworks.
• Graph Classification - Assigns a category to a full graph structure by aggregating node and edge features into a global representation.
• Graph Generation - Creates new graph topologies or molecular structures using autoregressive models or variational autoencoders.
• Graph Pooling - Reduces the resolution of a graph by merging nodes or coarsening the structure to extract higher-level features.
• Knowledge Graph Embeddings - Implements specialized tools for learning low-dimensional embeddings of entities and relations in knowledge graphs.
• Property Prediction - Represents chemical structures as graphs to predict reactivity and quantum interactions for drug discovery and materials science.
• Spatio-Temporal Models - Provides neural network models for forecasting future states in spatial-temporal systems like traffic or biological networks.
• Graph Data Loaders - Imports structured graph data from various source formats into a representation suitable for deep learning models.
• Chunked File Storages - Persists large graph structures in a chunked binary format to allow efficient loading and partial access.
• Unified Memory Managers - Stores structural data and features in a unified object residing on CPU or GPU for efficient computation.
• Molecular Property Prediction - Predicts chemical reactivity and quantum interactions by representing molecules as graphs.
• Graph Learning - Framework for deep learning on graphs across multiple backends.
• Graph Machine Learning - Deep learning on graphs built on existing frameworks.
• Graph Neural Networks - Python package for deep learning on graphs.
• Machine Learning Libraries - Deep learning on graphs.
• Computation and Optimization - Scalable Python package for deep learning on graphs.

Star history