Multiple edge attributes

From now on, we recommend using our discussion forum (https://github.com/rusty1s/pytorch_geometric/discussions) for general questions.

❓ Questions & Help

I use GraphConv layers in my model. My graph is represented as follows:

Data(edge_attr=[1357, 2], edge_index=[2, 1357], test_mask=[943], train_mask=[943], val_mask=[943], x=[943, 7], y=[943])

From the edge_attr we see that I have 2 features per edge. When I propagate my data through the model, I receive an error:

RuntimeError                              Traceback (most recent call last)
<ipython-input-102-8bd3fc340e4f> in <module>
      1 model = GraphConvClass(data_for_train, hidden_channels=16)
----> 2 out = model(data_for_train.x, data_for_train.edge_index, data_for_train.edge_attr)
      3 
      4 visualize(out, color=data_for_train.y)

~\anaconda3\lib\site-packages\torch\nn\modules\module.py in _call_impl(self, *input, **kwargs)
    725             result = self._slow_forward(*input, **kwargs)
    726         else:
--> 727             result = self.forward(*input, **kwargs)
    728         for hook in itertools.chain(
    729                 _global_forward_hooks.values(),

<ipython-input-82-1825ffb98b68> in forward(self, x, edge_index, edge_weight)
     15 
     16     def forward(self, x, edge_index, edge_weight): # Define forward propagation path
---> 17         x = self.conv1(x, edge_index, edge_weight)
     18         x = x.relu()
     19         x = self.conv2(x, edge_index, edge_weight)

~\anaconda3\lib\site-packages\torch\nn\modules\module.py in _call_impl(self, *input, **kwargs)
    725             result = self._slow_forward(*input, **kwargs)
    726         else:
--> 727             result = self.forward(*input, **kwargs)
    728         for hook in itertools.chain(
    729                 _global_forward_hooks.values(),

~\anaconda3\lib\site-packages\torch_geometric\nn\conv\graph_conv.py in forward(self, x, edge_index, edge_weight, size)
     60 
     61         # propagate_type: (x: OptPairTensor, edge_weight: OptTensor)
---> 62         out = self.propagate(edge_index, x=x, edge_weight=edge_weight,
     63                              size=size)
     64         out = self.lin_l(out)

~\anaconda3\lib\site-packages\torch_geometric\nn\conv\message_passing.py in propagate(self, edge_index, size, **kwargs)
    235 
    236             msg_kwargs = self.inspector.distribute('message', coll_dict)
--> 237             out = self.message(**msg_kwargs)
    238 
    239             # For `GNNExplainer`, we require a separate message and aggregate

~\anaconda3\lib\site-packages\torch_geometric\nn\conv\graph_conv.py in message(self, x_j, edge_weight)
     71 
     72     def message(self, x_j: Tensor, edge_weight: OptTensor) -> Tensor:
---> 73         return x_j if edge_weight is None else edge_weight.view(-1, 1) * x_j
     74 
     75     def message_and_aggregate(self, adj_t: SparseTensor,

RuntimeError: The size of tensor a (2714) must match the size of tensor b (1357) at non-singleton dimension 0

I guess the error says that the number of edge attributes > number of edges. In other words, I do not understand how to code 2 features per edge explicitly . My edge_attr looks like this:

tensor([[1.5888e-04, 3.3484e-01],
        [1.3956e-04, 4.6317e-01],
        [3.2826e-04, 2.5242e-01],
        ...,
        [2.5134e-04, 3.2950e-01],
        [1.1220e-03, 1.7618e-01],
        [1.5108e-04, 2.3872e-01]])

Can somebody point me to the relevant example or to give a hint how to solve my problem?

Thanks!

[rusty1s]

I don’t think that issue has anything to do with edge_attr, as it failsl on transforming node features in query = self.lin_query(x_i).view(-1, self.heads, self.out_channels). It looks like your node features have 20 features, while in_channels was set to 2. Can you confirm?

[RostyslavUA]

@rusty1s Exactly! This solves my problem, thank you very much!