Note

Click here to download the full example code

# Tutorial: Batched graph classification with DGL¶

**Author**: Mufei Li,
Minjie Wang,
Zheng Zhang.

In this tutorial, you learn how to use DGL to batch multiple graphs of variable size and shape. The tutorial also demonstrates training a graph neural network for a simple graph classification task.

Graph classification is an important problem with applications across many fields, such as bioinformatics, chemoinformatics, social network analysis, urban computing, and cybersecurity. Applying graph neural networks to this problem has been a popular approach recently. This can be seen in the following reserach references: Ying et al., 2018, Cangea et al., 2018, Knyazev et al., 2018, Bianchi et al., 2019, Liao et al., 2019, Gao et al., 2019).

## Simple graph classification task¶

In this tutorial, you learn how to perform batched graph classification with DGL. The example task objective is to classify eight types of topologies shown here.

Implement a synthetic dataset `data.MiniGCDataset`

in DGL. The dataset has eight
different types of graphs and each class has the same number of graph samples.

```
from dgl.data import MiniGCDataset
import matplotlib.pyplot as plt
import networkx as nx
# A dataset with 80 samples, each graph is
# of size [10, 20]
dataset = MiniGCDataset(80, 10, 20)
graph, label = dataset[0]
fig, ax = plt.subplots()
nx.draw(graph.to_networkx(), ax=ax)
ax.set_title('Class: {:d}'.format(label))
plt.show()
```

## Form a graph mini-batch¶

To train neural networks efficiently, a common practice is to batch multiple samples together to form a mini-batch. Batching fixed-shaped tensor inputs is common. For example, batching two images of size 28 x 28 gives a tensor of shape 2 x 28 x 28. By contrast, batching graph inputs has two challenges:

- Graphs are sparse.
- Graphs can have various length. For example, number of nodes and edges.

To address this, DGL provides a `dgl.batch()`

API. It leverages the idea that
a batch of graphs can be viewed as a large graph that has many disjointed
connected components. Below is a visualization that gives the general idea.

Define the following `collate`

function to form a mini-batch from a given
list of graph and label pairs.

```
import dgl
def collate(samples):
# The input `samples` is a list of pairs
# (graph, label).
graphs, labels = map(list, zip(*samples))
batched_graph = dgl.batch(graphs)
return batched_graph, torch.tensor(labels)
```

The return type of `dgl.batch()`

is still a graph. In the same way,
a batch of tensors is still a tensor. This means that any code that works
for one graph immediately works for a batch of graphs. More importantly,
because DGL processes messages on all nodes and edges in parallel, this greatly
improves efficiency.

## Graph classifier¶

Graph classification proceeds as follows.

From a batch of graphs, perform message passing and graph convolution for nodes to communicate with others. After message passing, compute a tensor for graph representation from node (and edge) attributes. This step might be called readout or aggregation. Finally, the graph representations are fed into a classifier \(g\) to predict the graph labels.

## Graph convolution¶

The graph convolution operation is basically the same as that for graph convolutional network (GCN). To learn more, see the GCN tutorial). The only difference is that we replace \(h_{v}^{(l+1)} = \text{ReLU}\left(b^{(l)}+\sum_{u\in\mathcal{N}(v)}h_{u}^{(l)}W^{(l)}\right)\) by \(h_{v}^{(l+1)} = \text{ReLU}\left(b^{(l)}+\frac{1}{|\mathcal{N}(v)|}\sum_{u\in\mathcal{N}(v)}h_{u}^{(l)}W^{(l)}\right)\)

The replacement of summation by average is to balance nodes with different degrees. This gives a better performance for this experiment.

The self edges added in the dataset initialization allows you to include the original node feature \(h_{v}^{(l)}\) when taking the average.

```
import dgl.function as fn
import torch
import torch.nn as nn
# Sends a message of node feature h.
msg = fn.copy_src(src='h', out='m')
def reduce(nodes):
"""Take an average over all neighbor node features hu and use it to
overwrite the original node feature."""
accum = torch.mean(nodes.mailbox['m'], 1)
return {'h': accum}
class NodeApplyModule(nn.Module):
"""Update the node feature hv with ReLU(Whv+b)."""
def __init__(self, in_feats, out_feats, activation):
super(NodeApplyModule, self).__init__()
self.linear = nn.Linear(in_feats, out_feats)
self.activation = activation
def forward(self, node):
h = self.linear(node.data['h'])
h = self.activation(h)
return {'h' : h}
class GCN(nn.Module):
def __init__(self, in_feats, out_feats, activation):
super(GCN, self).__init__()
self.apply_mod = NodeApplyModule(in_feats, out_feats, activation)
def forward(self, g, feature):
# Initialize the node features with h.
g.ndata['h'] = feature
g.update_all(msg, reduce)
g.apply_nodes(func=self.apply_mod)
return g.ndata.pop('h')
```

## Readout and classification¶

For this demonstration, consider initial node features to be their degrees. After two rounds of graph convolution, perform a graph readout by averaging over all node features for each graph in the batch.

In DGL, `dgl.mean_nodes()`

handles this task for a batch of
graphs with variable size. You then feed the graph representations into a
classifier with one linear layer to obtain pre-softmax logits.

```
import torch.nn.functional as F
class Classifier(nn.Module):
def __init__(self, in_dim, hidden_dim, n_classes):
super(Classifier, self).__init__()
self.layers = nn.ModuleList([
GCN(in_dim, hidden_dim, F.relu),
GCN(hidden_dim, hidden_dim, F.relu)])
self.classify = nn.Linear(hidden_dim, n_classes)
def forward(self, g):
# For undirected graphs, in_degree is the same as
# out_degree.
h = g.in_degrees().view(-1, 1).float()
for conv in self.layers:
h = conv(g, h)
g.ndata['h'] = h
hg = dgl.mean_nodes(g, 'h')
return self.classify(hg)
```

## Setup and training¶

Create a synthetic dataset of \(400\) graphs with \(10\) ~ \(20\) nodes. \(320\) graphs constitute a training set and \(80\) graphs constitute a test set.

```
import torch.optim as optim
from torch.utils.data import DataLoader
# Create training and test sets.
trainset = MiniGCDataset(320, 10, 20)
testset = MiniGCDataset(80, 10, 20)
# Use PyTorch's DataLoader and the collate function
# defined before.
data_loader = DataLoader(trainset, batch_size=32, shuffle=True,
collate_fn=collate)
# Create model
model = Classifier(1, 256, trainset.num_classes)
loss_func = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=0.001)
model.train()
epoch_losses = []
for epoch in range(80):
epoch_loss = 0
for iter, (bg, label) in enumerate(data_loader):
prediction = model(bg)
loss = loss_func(prediction, label)
optimizer.zero_grad()
loss.backward()
optimizer.step()
epoch_loss += loss.detach().item()
epoch_loss /= (iter + 1)
print('Epoch {}, loss {:.4f}'.format(epoch, epoch_loss))
epoch_losses.append(epoch_loss)
```

Out:

```
Epoch 0, loss 2.0559
Epoch 1, loss 1.9484
Epoch 2, loss 1.8944
Epoch 3, loss 1.7909
Epoch 4, loss 1.7366
Epoch 5, loss 1.6587
Epoch 6, loss 1.5899
Epoch 7, loss 1.5323
Epoch 8, loss 1.4729
Epoch 9, loss 1.3548
Epoch 10, loss 1.2930
Epoch 11, loss 1.2424
Epoch 12, loss 1.2033
Epoch 13, loss 1.1713
Epoch 14, loss 1.1371
Epoch 15, loss 1.1176
Epoch 16, loss 1.0801
Epoch 17, loss 1.0558
Epoch 18, loss 1.0102
Epoch 19, loss 0.9844
Epoch 20, loss 0.9519
Epoch 21, loss 0.9472
Epoch 22, loss 0.9255
Epoch 23, loss 0.9479
Epoch 24, loss 0.9492
Epoch 25, loss 0.8991
Epoch 26, loss 0.8748
Epoch 27, loss 0.8541
Epoch 28, loss 0.8414
Epoch 29, loss 0.8355
Epoch 30, loss 0.8199
Epoch 31, loss 0.8132
Epoch 32, loss 0.8036
Epoch 33, loss 0.8111
Epoch 34, loss 0.8224
Epoch 35, loss 0.8645
Epoch 36, loss 0.8290
Epoch 37, loss 0.7651
Epoch 38, loss 0.7651
Epoch 39, loss 0.7564
Epoch 40, loss 0.7541
Epoch 41, loss 0.7420
Epoch 42, loss 0.7425
Epoch 43, loss 0.7255
Epoch 44, loss 0.7163
Epoch 45, loss 0.7470
Epoch 46, loss 0.7510
Epoch 47, loss 0.7478
Epoch 48, loss 0.7440
Epoch 49, loss 0.7233
Epoch 50, loss 0.7041
Epoch 51, loss 0.6958
Epoch 52, loss 0.6701
Epoch 53, loss 0.6851
Epoch 54, loss 0.7178
Epoch 55, loss 0.7004
Epoch 56, loss 0.6891
Epoch 57, loss 0.6726
Epoch 58, loss 0.6584
Epoch 59, loss 0.6591
Epoch 60, loss 0.6493
Epoch 61, loss 0.6521
Epoch 62, loss 0.6532
Epoch 63, loss 0.6322
Epoch 64, loss 0.6159
Epoch 65, loss 0.6475
Epoch 66, loss 0.6713
Epoch 67, loss 0.6223
Epoch 68, loss 0.6303
Epoch 69, loss 0.6146
Epoch 70, loss 0.6431
Epoch 71, loss 0.6255
Epoch 72, loss 0.6252
Epoch 73, loss 0.6046
Epoch 74, loss 0.6121
Epoch 75, loss 0.6166
Epoch 76, loss 0.5900
Epoch 77, loss 0.5818
Epoch 78, loss 0.5718
Epoch 79, loss 0.5707
```

The learning curve of a run is presented below.

The trained model is evaluated on the test set created. To deploy the tutorial, restrict the running time to get a higher accuracy (\(80\) % ~ \(90\) %) than the ones printed below.

```
model.eval()
# Convert a list of tuples to two lists
test_X, test_Y = map(list, zip(*testset))
test_bg = dgl.batch(test_X)
test_Y = torch.tensor(test_Y).float().view(-1, 1)
probs_Y = torch.softmax(model(test_bg), 1)
sampled_Y = torch.multinomial(probs_Y, 1)
argmax_Y = torch.max(probs_Y, 1)[1].view(-1, 1)
print('Accuracy of sampled predictions on the test set: {:.4f}%'.format(
(test_Y == sampled_Y.float()).sum().item() / len(test_Y) * 100))
print('Accuracy of argmax predictions on the test set: {:4f}%'.format(
(test_Y == argmax_Y.float()).sum().item() / len(test_Y) * 100))
```

Out:

```
Accuracy of sampled predictions on the test set: 67.5000%
Accuracy of argmax predictions on the test set: 78.750000%
```

The animation here plots the probability that a trained model predicts the correct graph type.

To understand the node and graph representations that a trained model learned, we use t-SNE, for dimensionality reduction and visualization.

The two small figures on the top separately visualize node representations after one and two layers of graph convolution. The figure on the bottom visualizes the pre-softmax logits for graphs as graph representations.

While the visualization does suggest some clustering effects of the node features, you would not expect a perfect result. Node degrees are deterministic for these node features. The graph features are improved when separated.

## What’s next?¶

Graph classification with graph neural networks is still a new field. It’s waiting for people to bring more exciting discoveries. The work requires mapping different graphs to different embeddings, while preserving their structural similarity in the embedding space. To learn more about it, see How Powerful Are Graph Neural Networks? a research paper published for the International Conference on Learning Representations 2019.

For more examples about batched graph processing, see the following:

- Tutorials for Tree LSTM and Deep Generative Models of Graphs
- An example implementation of Junction Tree VAE

**Total running time of the script:** ( 0 minutes 23.176 seconds)