Why the implementation of gcn use 3 pgl.nn.GCNConvs?

[yunlong315]

I was confused by the implementation of gcn in examples/gcn(showed as below),where 3 pgl.nn.GCNConvs was used to build the model,while the original paper(Semi-Supervised Classification with Graph Convolutional Networks) and the author's implementation just use 2.

class GCN(nn.Layer):
    """Implement of GCN
    """

    def __init__(self,
                 input_size,
                 num_class,
                 num_layers=1,
                 hidden_size=64,
                 dropout=0.5):
        super(GCN, self).__init__()
        self.num_class = num_class
        self.num_layers = num_layers
        self.hidden_size = hidden_size
        self.dropout = dropout
        self.gcns = nn.LayerList()
        for i in range(self.num_layers):
            if i == 0:
                self.gcns.append(
                    pgl.nn.GCNConv(
                        input_size,
                        self.hidden_size,
                        activation="relu",
                        norm=True))
            else:
                self.gcns.append(
                    pgl.nn.GCNConv(
                        self.hidden_size,
                        self.hidden_size,
                        activation="relu",
                        norm=True))
            self.gcns.append(nn.Dropout(self.dropout))
        self.gcns.append(pgl.nn.GCNConv(self.hidden_size, self.num_class))

    def forward(self, graph, feature):
        for m in self.gcns:
            if isinstance(m, nn.Dropout):
                feature = m(feature)
            else:
                feature = m(graph, feature)
        return feature

I hope someone can help me resolve this confusion.

[yunlong315]

Sorry,i misunderstood the loop.The implementation is right.