deepgnn-ge 0.1.49

Graph engine - distributed graph engine to host graphs.

DeepGNN Overview

DeepGNN is a framework for training machine learning models on large scale graph data. DeepGNN contains all the necessary features including:

• Distributed GNN training and inferencing on both CPU and GPU.
• Custom graph neural network design.
• Online Sampling: Graph Engine (GE) will load all graph data, each training worker will call GE to get node/edge/neighbor features and labels.
• Automatic graph partitioning.
• Highly performant and scalable.

Usage

Install pip package:

python -m pip install deepgnn-torch

If you want to build package from source, see instructions in CONTRIBUTING.md.

Train and evaluate a graphsage model with pytorch on cora dataset:

cd examples/pytorch/graphsage
./run.sh

Training other models

Examples folder contains various models one can experiment with DeepGNN. To train models with Tensorflow you need to install deepgnn-tf package, deepgnn-torch package contains packages to train pytorch examples. Each model folder contains a shell script run.sh that will train a corresponding model on a toy graph, a README.md file with a short description of a model, reference to original paper, and explanation of command line arguments.