Issue using GENConv

[loisbijing]

I created a graph dataset with:

data = Data(x=torch.from_numpy(features), y=torch.from_numpy(y), edge_index=edge_index.t().contiguous(), edge_attr=edge_dis.unsqueeze(1), pos=pos)

My data has following dimension:

Data(x=[253080, 5], edge_index=[2, 100320], edge_attr=[100320, 1], y=[253080, 3], pos=[253080, 3])

I got error about node and edge feature dimensions are not matching:

assert x[0].size(-1) == edge_attr.size(-1)

However, I did change both my node feature and edge attr to the same dimension using a linear layer, I also printed x[0].size(-1) and edge_attr.size(-1) and they show the same value. It seems I got this error whenever my GENConv has different in_channels size and out_channels size. I can't figure out why.

Here is my code:

class DeeperGCN(torch.nn.Module):
    def __init__(self, hidden_channels, num_layers):
        super(DeeperGCN, self).__init__()

        self.node_encoder = Linear(5, hidden_channels)
        self.edge_encoder = Linear(1, hidden_channels)

        self.layers = torch.nn.ModuleList()
        for i in range(1, num_layers+1):
            conv = GENConv(hidden_channels, hidden_channels, aggr='softmax',
                           t=1.0, learn_t=True, num_layers=2, norm='layer')
            norm = LayerNorm(hidden_channels, elementwise_affine=True)
            act = LeakyReLU(inplace=True)

            layer = DeepGCNLayer(conv, norm, act, block='res+')
            self.layers.append(layer)
            
        self.conv2 = GENConv(hidden_channels, 3, aggr='softmax',
                           t=1.0, learn_t=True, num_layers=2, norm='layer')
    
    def forward(self, data):
        x, y, edge_index, edge_attr, pos = data.x, data.y, data.edge_index, data.edge_attr, data.pos
        
        out = self.node_encoder(x)
        edge_attr = self.edge_encoder(edge_attr)
        
        
        for layer in self.layers[1:]:
            out = layer(out, edge_index, edge_attr)

         out = self.conv2(out, edge_index, edge_attr)
        
        return out

[wsad1]

Hi @loisbijing ,
The issue is that edge_dim is not set. Set edge_dim attribute while initialising GENConv to hidden_channels.

[loisbijing]

Cool, thanks a lot! This solved the issue!