how can I use edge attribution with GATConv layer(s)

[Byun-jinyoung]

Dear pyg users,
I wanted to use GATConv layer(s) when I performed node classification task.
I didn't encounter any error(s) when I didn't use edge_attr with GATConv layer.
However, I've encountered the error as below:

x = layer(x, edge_index, edge_attr)
  File "/home/byun/anaconda3/envs/pyg_torch/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/byun/anaconda3/envs/pyg_torch/lib/python3.7/site-packages/torch_geometric/nn/conv/gat_conv.py", line 242, in forward
    size=size)
  File "/home/byun/anaconda3/envs/pyg_torch/lib/python3.7/site-packages/torch_geometric/nn/conv/message_passing.py", line 317, in propagate
    out = self.message(**msg_kwargs)
  File "/home/byun/anaconda3/envs/pyg_torch/lib/python3.7/site-packages/torch_geometric/nn/conv/gat_conv.py", line 274, in message
    assert self.lin_edge is not None
AssertionError

Could you tell me what I'm missing? How can I solve this problem??

Best regards :)

[rusty1s]

I think you need to specify the edge_dim attribute of GATConv.

[Byun-jinyoung]

I specified the edge_dim=2, but I got another error.
Ummm... I have to know how the attention is implemented...

    x = layer(x, edge_index, edge_attr)
  File "/home/byun/anaconda3/envs/pyg_torch/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/byun/anaconda3/envs/pyg_torch/lib/python3.7/site-packages/torch_geometric/nn/conv/gat_conv.py", line 245, in forward
    size=size)
  File "/home/byun/anaconda3/envs/pyg_torch/lib/python3.7/site-packages/torch_geometric/nn/conv/message_passing.py", line 317, in propagate
    out = self.message(**msg_kwargs)
  File "/home/byun/anaconda3/envs/pyg_torch/lib/python3.7/site-packages/torch_geometric/nn/conv/gat_conv.py", line 278, in message
    edge_attr = self.lin_edge(edge_attr)
  File "/home/byun/anaconda3/envs/pyg_torch/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/byun/anaconda3/envs/pyg_torch/lib/python3.7/site-packages/torch_geometric/nn/dense/linear.py", line 109, in forward
    return F.linear(x, self.weight, self.bias)
  File "/home/byun/anaconda3/envs/pyg_torch/lib/python3.7/site-packages/torch/nn/functional.py", line 1848, in linear
    return torch._C._nn.linear(input, weight, bias)
RuntimeError: mat1 and mat2 shapes cannot be multiplied (618992x7 and 2x64)

[rusty1s]

Can you say something about your input shapes and the way you initialize GATConv?

[Byun-jinyoung]

Thank you for your fast reply !!
I posted my pyg model code and I used use_edge_dim=True and edge_dim=2 with the GATConv layer
I wonder whether the edge dim has at least 7 as my edge_attr has shape of [xxx, 7]

class MyPygNet(nn.Module):
    def __init__(self,
                 model_type, input_size, num_classes, hidden_size_list,
                 use_bn, use_dropout, use_edge_attr, edge_dim):
        super().__init__()

        self.model_type, self.input_size = model_type, input_size
        self.hidden_size_list = hidden_size_list
        self.use_bn, self.use_dropout = use_bn, use_dropout
        self.use_edge_attr, self.edge_dim = use_edge_attr, edge_dim

        # add first layer                                                                                                                        
        self.pyg_layers = []
        print(f'[INFO ] My pyg model use {model_type} layer(s).')
        if model_type == 'GCNConv':
            self.pyg_layers.append( GCNConv(self.input_size, self.hidden_size_list[0]) )
            self.myGCNConv()

        elif model_type == 'GATConv':
            if self.use_edge_attr and self.edge_dim:
                self.pyg_layers.append( GATConv(self.input_size, self.hidden_size_list[0], edge_dim=self.edge_dim) )
            else:
                self.pyg_layers.append( GATConv(self.input_size, self.hidden_size_list[0]) )
            self.myGATConv()  

        self.module_list = nn.ModuleList(self.pyg_layers)
        self.lin1 = Linear(hidden_size_list[-1], num_classes) # 21: num of classes                                                                                                                                               

    def myGATConv(self):
        for idx, hidden_size in enumerate(self.hidden_size_list[1:]):
            if self.use_edge_attr and self.edge_dim:
                self.pyg_layers.append(GATConv(self.hidden_size_list[idx], hidden_size, edge_dim=self.edge_dim)
 
            else:
                self.pyg_layers.append(GATConv(self.hidden_size_list[idx], hidden_size))

    def forward(self, data):
        x, edge_index, edge_attr, batch = data.x, data.edge_index, data.edge_attr, data.batch
        for layer in self.module_list:
            if self.use_edge_attr and self.edge_dim:
                x = layer(x, edge_index, edge_attr)
            else:
                x = layer(x, edge_index)
            x = F.gelu(x)
            x = F.dropout(x, training=self.training)

        
        x = F.gelu(self.lin1(x))

My data are such below:

....
Data(x=[169, 15], y=[169], pos=[169, 3], edge_index=[2, 1744], edge_attr=[1744, 7])
Data(x=[169, 15], y=[169], pos=[169, 3], edge_index=[2, 1744], edge_attr=[1744, 7])
Data(x=[188, 15], y=[188], pos=[188, 3], edge_index=[2, 4524], edge_attr=[4524, 7])
Data(x=[188, 15], y=[188], pos=[188, 3], edge_index=[2, 4524], edge_attr=[4524, 7])
Data(x=[95, 15], y=[95], pos=[95, 3], edge_index=[2, 1114], edge_attr=[1114, 7])
Data(x=[95, 15], y=[95], pos=[95, 3], edge_index=[2, 1114], edge_attr=[1114, 7])
Data(x=[713, 15], y=[713], pos=[713, 3], edge_index=[2, 14458], edge_attr=[14458, 7])
Data(x=[713, 15], y=[713], pos=[713, 3], edge_index=[2, 14458], edge_attr=[14458, 7])
Data(x=[496, 15], y=[496], pos=[496, 3], edge_index=[2, 11844], edge_attr=[11844, 7])
Data(x=[496, 15], y=[496], pos=[496, 3], edge_index=[2, 11844], edge_attr=[11844, 7])
....

[rusty1s]

If your edge_attr has 7 features per edge, you should pass edge_dim=7.

[Byun-jinyoung]

Is any other edge dim not acceptable?

[rusty1s]

I'm not sure what you mean. edge_dim refers to the edge feature dimensionality of edge_attr, i.e., edge_attr.size(1).