Using customized GNN

Sometimes you would like to design your own GNN module or use GNN for other purposes. In this chapter, we introduce how to use GNN layer in CogDL to write your own GNN model and how to write a GNN layer from scratch.

GNN layers in CogDL to Define model

CogDL has implemented popular GNN layers in cogdl.layers, and they can serve as modules to help design new GNNs. Here is how we implement Jumping Knowledge Network (JKNet) with GCNLayer in CogDL.

JKNet collects the output of all layers and concatenate them together to get the result:

\begin{gather*} H^{(0)} = X \\ H^{(i+1)} = \sigma(\hat{A} H^{(i)} W^{(i)} \\ OUT = CONCAT([H^{(0)},...,H^{(L)}) \end{gather*}
import torch
from cogdl.models import BaseModel

class JKNet(BaseModel):
    def __init__(self, in_feats, out_feats, hidden_size, num_layers):
        super(JKNet, self).__init__()
        shapes = [in_feats] + [hidden_size] * num_layers
        #
        self.layers = nn.ModuleList([
            GCNLayer(shape[i], shape[i+1])
            for i in range(num_layers)
        ])
        self.fc = nn.Linear(hidden_size * num_layers, out_feats)

    def forward(self, graph):
        graph.add_remaining_self_loops()
        graph.sym_norm()
        h = graph.x
        out = []
        for layer in self.layers:
            h = layer(x)
            out.append(h)
        out = torch.cat(out, dim=1)
        return self.fc(out)

Define your GNN Module

In most cases, you may build a layer module with new message propagation and aggragation scheme. Here the code snippet shows how to implement a GCNLayer using Graph and efficient sparse matrix operators in CogDL.

import torch
from cogdl.utils import spmm

class GCNLayer(torch.nn.Module):
    """
    Args:
        in_feats: int
            Input feature size
        out_feats: int
            Output feature size
    """
    def __init__(self, in_feats, out_feats):
        super(GCNLayer, self).__init__()
        self.fc = torch.nn.Linear(in_feats, out_feats)

    def forward(self, graph, x):
        # symmetric normalization of adjacency matrix
        graph.sym_norm()
        h = self.fc(x)
        h = spmm(graph, h)
        return h

spmm is sparse matrix multiplication operation frequently used in GNNs.

\[H = AH = SpMM(A, H)\]

Sparse matrix is stored in Graph and will be called automatically. Message-passing in spatial space is equivalent to matrix operations. CogDL also supports other efficient operators like edge_softmax and multi_head_spmm, you can refer to this page for usage.

Use Custom models with CogDL

Now that you have defined your own GNN, you can use dataset/task in CogDL to immediately train and evaluate the performance of your model.

data = dataset.data
# Use the JKNet model as defined above
model = JKNet(data.num_features, data.num_classes, 32, 4)
experiment(model=model, dataset="cora", mw="node_classification_mw", dw="node_classification_dw")