Counting 3,319 Big Data & Machine Learning Frameworks, Toolsets, and Examples...
Suggestion? Feedback? Tweet @stkim1

Author
Contributors
Last Commit
Nov. 21, 2018
Created
Aug. 31, 2018

Graph Nets DeepMind shortest path

Graph Nets library

Graph Nets is DeepMind's library for building graph networks in Tensorflow and Sonnet.

Contact graph-nets@google.com for comments and questions.

What are graph networks?

A graph network takes a graph as input and returns a graph as output. The input graph has edge- (E ), node- (V ), and global-level (u) attributes. The output graph has the same structure, but updated attributes. Graph networks are part of the broader family of "graph neural networks" (Scarselli et al., 2009).

To learn more about graph networks, see our arXiv paper: Relational inductive biases, deep learning, and graph networks.

Graph network

Installation

The Graph Nets library can be installed from pip.

This installation is compatible with Linux/Mac OS X, and Python 2.7 and 3.4+.

The library will work with both the CPU and GPU version of TensorFlow, but to allow for that it does not list Tensorflow as a requirement, so you need to install Tensorflow separately if you haven't already done so.

To install the Graph Nets library for CPU, run:

$ pip install graph_nets tensorflow tensorflow_probability

To install the Graph Nets library for GPU, run:

$ pip install graph_nets tensorflow_gpu tensorflow_probability_gpu

Usage example

The following code constructs a simple graph net module and connects it to data.

import graph_nets as gn
import sonnet as snt

# Provide your own functions to generate graph-structured data.
input_graphs = get_graphs()

# Create the graph network.
graph_net_module = gn.modules.GraphNetwork(
    edge_model_fn=lambda: snt.nets.MLP([32, 32]),
    node_model_fn=lambda: snt.nets.MLP([32, 32]),
    global_model_fn=lambda: snt.nets.MLP([32, 32]))

# Pass the input graphs to the graph network, and return the output graphs.
output_graphs = graph_net_module(input_graphs)

Demo Jupyter notebooks

The library includes demos which show how to create, manipulate, and train graph networks to reason about graph-structured data, on a shortest path-finding task, a sorting task, and a physical prediction task. Each demo uses the same graph network architecture, which highlights the flexibility of the approach.

Try the demos in your browser in Colaboratory

To try out the demos without installing anything locally, you can run the demos in your browser (even on your phone) via a cloud Colaboratory backend. Click a demo link below, and follow the instructions in the notebook.


Run "shortest path demo" in browser

The "shortest path demo" creates random graphs, and trains a graph network to label the nodes and edges on the shortest path between any two nodes. Over a sequence of message-passing steps (as depicted by each step's plot), the model refines its prediction of the shortest path.

Shortest path


Run "sort demo" in browser

The "sort demo" creates lists of random numbers, and trains a graph network to sort the list. After a sequence of message-passing steps, the model makes an accurate prediction of which elements (columns in the figure) come next after each other (rows).

Sort


Run "physics demo" in browser

The "physics demo" creates random mass-spring physical systems, and trains a graph network to predict the state of the system on the next timestep. The model's next-step predictions can be fed back in as input to create a rollout of a future trajectory. Each subplot below shows the true and predicted mass-spring system states over 50 steps. This is similar to the model and experiments in Battaglia et al. (2016)'s "interaction networks".

Physics


Run the demos on your local machine

To install the necessary dependencies, run:

$ pip install jupyter matplotlib scipy

To try the demos, run:

$ cd <path-to-graph-nets-library>/demos
$ jupyter notebook

then open a demo through the Jupyter notebook interface.