Skip to content

GraphStorm v0.1 release
Compare
Choose a tag to compare
@classicsong classicsong released this 23 May 06:54
· 1 commit to v0.1 since this release
0.1
c31b35f

V0.1 is the first official release of GraphStorm which provides an end2end user experiences of graph machine learning (GML) model training and inference starting from graph construction from data in Parquet or JSON format, to GML model training and inference. It provides a highly optimized and scalable pipeline for graph construction pipeline, model training and inference. Users can build graph data, train or inference a built-in GML model with a single command without writing any code. GraphStorm can scale GML training and inference on graphs with billions of nodes with multiple GPUs or multiple machines.

Major features

  • A single-machine data loading pipeline that accepts data stored in Parquet, JSON or HDF5 format, to build graph data for GML model training and inference. Users can build graph data with a single command without writing any code.
  • End-to-end training and inference pipelines that support common GML tasks including node classification, node regression, edge classification, edge regression and link prediction.
  • A collection of built-in GML models including RGCN, RGAT and LM-GNN. GraphStorm provides HGT as an custom model example.
  • A large collection of model configurations and training/inference configurations that allow users to tune GML model training without writing any code.
  • Scale GML training and inference on graphs with billions of nodes with multiple GPUs or multiple machines.
  • Support custom GML models. GraphStorm can scale user-defined models to billion-scale graphs with multiple GPUs and multiple machines.
  • Complete tutorial of graph construction, GML model training and GML model inference. (#Link to tutorial).
  • Huggingface BERT-GNN co-training and Graph-aware Huggingface BERT fine-tuning. Users can combine GML with Language models to either improve the performance of GML tasks or extend the expressiveness of LMs with graph information.
  • AWS native support. We provide a guideline to build GraphStorm Docker images for AWS EC2.(https://github.com/awslabs/graphstorm/tree/main/docker)

Contributors

Assets 2
Loading