How many layers do you use for GCN based on DGL?

hbsun2113 · February 22, 2019, 3:45am

I note that the original paper and its implementation(https://github.com/tkipf/pygcn) use two GCN layers. Instead, you use three layers based on DGL. Do you compare the result based on a different number of layers?

zihao · February 22, 2019, 6:40am

Hi,
By default DGL use two layers for GCN, the default value for n_layers is one:

github.com

dmlc/dgl/blob/master/examples/pytorch/gcn/gcn_spmv.py#L78


             n_hidden,
             n_classes,
             n_layers,
             activation,
             dropout):
    super(GCN, self).__init__()
    self.layers = nn.ModuleList()
    # input layer
    self.layers.append(GCNLayer(g, in_feats, n_hidden, activation, 0.))
    # hidden layers
    for i in range(n_layers - 1):
        self.layers.append(GCNLayer(g, n_hidden, n_hidden, activation, dropout))
    # output layer
    self.layers.append(GCNLayer(g, n_hidden, n_classes, None, dropout))


def forward(self, features):
    h = features
    for layer in self.layers:
        h = layer(h)
    return h

qillbel · May 5, 2020, 12:06pm

Hey @zihao,

Does adding more layers improve the accuracy of node classification? and how do we know that we need to add more layers?

Thanks