August 8, 2019

1164 words 6 mins read



Python package built to ease deep learning on graph, on top of existing DL frameworks.

repo name dmlc/dgl
repo link
language Python
size (curr.) 6373 kB
stars (curr.) 4185
created 2018-04-20
license Apache License 2.0

Deep Graph Library (DGL)

PyPi Latest Release Conda Latest Release Build Status Benchmark by ASV License

Documentation | DGL at a glance | Model Tutorials | Discussion Forum

DGL is an easy-to-use, high performance and scalable Python package for deep learning on graphs. DGL is framework agnostic, meaning if a deep graph model is a component of an end-to-end application, the rest of the logics can be implemented in any major frameworks, such as PyTorch, Apache MXNet or TensorFlow.

DGL News

03/02/2020: Check out this cool paper: Benchmarking Graph Neural Networks! It includes a DGL-based benchmark framework for novel medium-scale graph datasets, covering mathematical modeling, computer vision, chemistry and combinatorial problems. See repo here.

Using DGL

A data scientist may want to apply a pre-trained model to your data right away. For this you can use DGL’s Application packages, formally Model Zoo. Application packages are developed for domain applications, as is the case for DGL-LifeScience. We will soon add model zoo for knowledge graph embedding learning and recommender systems. Here’s how you will use a pretrained model:

from import Tox21, smiles_to_bigraph, CanonicalAtomFeaturizer
from dgl import model_zoo

dataset = Tox21(smiles_to_bigraph, CanonicalAtomFeaturizer())
model = model_zoo.chem.load_pretrained('GCN_Tox21') # Pretrained model loaded

smiles, g, label, mask = dataset[0]
feats = g.ndata.pop('h')
label_pred = model(g, feats)

Further reading: DGL is released as a managed service on AWS SageMaker, see the medium posts for an easy trip to DGL on SageMaker(part1 and part2).

Researchers can start from the growing list of models implemented in DGL. Developing new models does not mean that you have to start from scratch. Instead, you can reuse many pre-built modules. Here is how to get a standard two-layer graph convolutional model with a pre-built GraphConv module:

from dgl.nn.pytorch import GraphConv
import torch.nn.functional as F

# build a two-layer GCN with ReLU as the activation in between
class GCN(nn.Module):
    def __init__(self, in_feats, h_feats, num_classes):
        super(GCN, self).__init__()
        self.gcn_layer1 = GraphConv(in_feats, h_feats)
        self.gcn_layer2 = GraphConv(h_feats, num_classes)
    def forward(self, graph, inputs):
        h = self.gcn_layer1(graph, inputs)
        h = F.relu(h)
        h = self.gcn_layer2(graph, h)
        return h

Next level down, you may want to innovate your own module. DGL offers a succinct message-passing interface (see tutorial here). Here is how Graph Attention Network (GAT) is implemented (complete codes). Of course, you can also find GAT as a module GATConv:

import torch.nn as nn
import torch.nn.functional as F

# Define a GAT layer
class GATLayer(nn.Module):
    def __init__(self, in_feats, out_feats):
        super(GATLayer, self).__init__()
        self.linear_func = nn.Linear(in_feats, out_feats, bias=False)
        self.attention_func = nn.Linear(2 * out_feats, 1, bias=False)
    def edge_attention(self, edges):
        concat_z =[edges.src['z'], edges.dst['z']], dim=1)
        src_e = self.attention_func(concat_z)
        src_e = F.leaky_relu(src_e)
        return {'e': src_e}
    def message_func(self, edges):
        return {'z': edges.src['z'], 'e'['e']}
    def reduce_func(self, nodes):
        a = F.softmax(nodes.mailbox['e'], dim=1)
        h = torch.sum(a * nodes.mailbox['z'], dim=1)
        return {'h': h}
    def forward(self, graph, h):
        z = self.linear_func(h)
        graph.ndata['z'] = z
        graph.update_all(self.message_func, self.reduce_func)
        return graph.ndata.pop('h')

Performance and Scalability

Microbenchmark on speed and memory usage: While leaving tensor and autograd functions to backend frameworks (e.g. PyTorch, MXNet, and TensorFlow), DGL aggressively optimizes storage and computation with its own kernels. Here’s a comparison to another popular package – PyTorch Geometric (PyG). The short story is that raw speed is similar, but DGL has much better memory management.

Dataset Model Accuracy Time PyG    DGL Memory PyG    DGL
Cora GCN GAT 81.31 ± 0.88 83.98 ± 0.52 0.478    0.666 1.608    1.399 1.1    1.1 1.2    1.1
CiteSeer GCN GAT 70.98 ± 0.68 69.96 ± 0.53 0.490    0.674 1.606    1.399 1.1    1.1 1.3    1.1
PubMed GCN GAT 79.00 ± 0.41 77.65 ± 0.32 0.491    0.690 1.946    1.393 1.1    1.1 1.6    1.1
Reddit GCN 93.46 ± 0.06 OOM   28.6 OOM    11.7
Reddit-S GCN N/A 29.12    9.44 15.7    3.6

Table: Training time(in seconds) for 200 epochs and memory consumption(GB)

High memory utilization allows DGL to push the limit of single-GPU performance, as seen in below images.

Scalability: DGL has fully leveraged multiple GPUs in both one machine and clusters for increasing training speed, and has better performance than alternatives, as seen in below images.

Further reading: Detailed comparison of DGL and other Graph alternatives can be found here.

DGL Models and Applications

DGL for research

Overall there are 30+ models implemented by using DGL:

DGL for domain applications

DGL for NLP/CV problems

We are currently in Beta stage. More features and improvements are coming.


DGL should work on

  • all Linux distributions no earlier than Ubuntu 16.04
  • macOS X
  • Windows 10

DGL requires Python 3.5 or later.

Right now, DGL works on PyTorch 1.1.0+, MXNet nightly build, and TensorFlow 2.0+.

Using anaconda

conda install -c dglteam dgl           # cpu version
conda install -c dglteam dgl-cuda9.0   # CUDA 9.0
conda install -c dglteam dgl-cuda9.2   # CUDA 9.2
conda install -c dglteam dgl-cuda10.0  # CUDA 10.0
conda install -c dglteam dgl-cuda10.1  # CUDA 10.1

Using pip

Latest Nightly Build Version Stable Version
CPU pip install --pre dgl pip install dgl
CUDA 9.0 pip install --pre dgl-cu90 pip install dgl-cu90
CUDA 9.2 pip install --pre dgl-cu92 pip install dgl-cu92
CUDA 10.0 pip install --pre dgl-cu100 pip install dgl-cu100
CUDA 10.1 pip install --pre dgl-cu101 pip install dgl-cu101

Built from source code

Refer to the guide here.

DGL Major Releases

Releases Date Features
v0.4.2 01/24/2020 - Heterograph support - TensorFlow support (experimental) - MXNet GNN modules
v0.3.1 08/23/2019 - APIs for GNN modules - Model zoo (DGL-Chem) - New installation
v0.2 03/09/2019 - Graph sampling APIs - Speed improvement
v0.1 12/07/2018 - Basic DGL APIs - PyTorch and MXNet support - GNN model examples and tutorials

New to Deep Learning and Graph Deep Learning?

Check out the open source book Dive into Deep Learning.

For those who are new to graph neural network, please see the basic of DGL.

For audience who are looking for more advanced, realistic, and end-to-end examples, please see model tutorials.


Please let us know if you encounter a bug or have any suggestions by filing an issue.

We welcome all contributions from bug fixes to new features and extensions.

We expect all contributions discussed in the issue tracker and going through PRs. Please refer to our contribution guide.


If you use DGL in a scientific publication, we would appreciate citations to the following paper:

    title={Deep Graph Library: Towards Efficient and Scalable Deep Learning on Graphs},
    author={Wang, Minjie and Yu, Lingfan and Zheng, Da and Gan, Quan and Gai, Yu and Ye, Zihao and Li, Mufei and Zhou, Jinjing and Huang, Qi and Ma, Chao and Huang, Ziyue and Guo, Qipeng and Zhang, Hao and Lin, Haibin and Zhao, Junbo and Li, Jinyang and Smola, Alexander J and Zhang, Zheng},
    journal={ICLR Workshop on Representation Learning on Graphs and Manifolds},

The Team

DGL is developed and maintained by NYU, NYU Shanghai, AWS Shanghai AI Lab, and AWS MXNet Science Team.


DGL uses Apache License 2.0.

comments powered by Disqus