ChebConv K's impact on shape

[LaurentBerder]

Hi,

I'm trying to model a phenomenon using a torch_geometric_temporal.nn.GConvLSTM, which itself is based on torch_geometric.nn.ChebConv, but I can't find a way to change the parameter K.

Here's my definition of the model:

from torch_geometric_temporal.nn import GConvLSTM

class DynGraphModel(torch.nn.Module):
    def __init__(self, input_dim, hidden_dim, output_dim, K):
        super(DynGraphModel, self).__init__()
        self.conv = GConvLSTM(input_dim, hidden_dim, K)
        self.linear = torch.nn.Linear(hidden_dim, output_dim)
    def forward(self, x, edge_index, edge_attr, H, C):
        h_0, c_0 = self.conv(x, edge_index, edge_attr, H, C)
        h = F.relu(h_0)
        out = self.linear(h)
        return h, h_0, c_0

I then initialize it

hidden_channels = 30
model = DynGraphModel(input_dim=nb_node_features, hidden_dim=hidden_channels, output_dim=1, K=1)

And then use it

snapshot = next(iter(train_dataset))

y_hat, H, C = model(x=snapshot.x, edge_index=snapshot.edge_index, edge_attr=snapshot.edge_attr, H=H, C=C)

However, whenever I try to modify the value for K, I get an error:

model = DynGraphModel(input_dim=nb_node_features, hidden_dim=hidden_channels, output_dim=1, K=3)

y_hat, H, C = model(x=snapshot.x, edge_index=snapshot.edge_index, edge_attr=snapshot.edge_attr, H=H, C=C)

>>> Traceback (most recent call last):
>>>   File "<stdin>", line 1, in <module>
>>>   File "python3.7/site-packages/torch/nn/modules/module.py", line 1130, in _call_impl
>>>     return forward_call(*input, **kwargs)
>>>   File "<stdin>", line 7, in forward
>>>   File "python3.7/site-packages/torch/nn/modules/module.py", line 1130, in _call_impl
>>>     return forward_call(*input, **kwargs)
>>>   File "python3.7/site-packages/torch_geometric_temporal/nn/recurrent/gconv_lstm.py", line 231, in forward
>>>     I = self._calculate_input_gate(X, edge_index, edge_weight, H, C, lambda_max)
>>>   File "python3.7/site-packages/torch_geometric_temporal/nn/recurrent/gconv_lstm.py", line 167, in _calculate_input_gate
>>>     I = self.conv_x_i(X, edge_index, edge_weight, lambda_max=lambda_max)
>>>   File "python3.7/site-packages/torch/nn/modules/module.py", line 1130, in _call_impl
>>>     return forward_call(*input, **kwargs)
>>>   File "python3.7/site-packages/torch_geometric/nn/conv/cheb_conv.py", line 174, in forward
>>>     Tx_1 = self.propagate(edge_index, x=x, norm=norm, size=None)
>>>   File "python3.7/site-packages/torch_geometric/nn/conv/message_passing.py", line 437, in propagate
>>>     out = self.message(**msg_kwargs)
>>>   File "python3.7/site-packages/torch_geometric/nn/conv/cheb_conv.py", line 189, in message
>>>     return norm.view(-1, 1) * x_j
>>> RuntimeError: The size of tensor a (29538) must match the size of tensor b (14769) at non-singleton dimension 0

I don't recognize the values for size of tensor, it's not a multiplier of any of my data object's dimenstions:

snapshot
>>> Data(x=[715, 3], edge_index=[2, 13339], edge_attr=[13339, 2], y=[715, 30])

[rusty1s]

I sadly cannot really help you with any problems related to PyG-Temporal since this library is not maintained by us. I assume the issue is that ChebConv cannot really handle multi-dimensional edge features, and so edge_attr should be restricted to be a vector. Does this fix your issues?

[LaurentBerder]

Thanks for your reply @rusty1s. Indeed, I know PyG-Temporal is a different library, but unhappily they're less reactive, and moreover, they're based on PyG.
Here, the issue does point to torch_geometric/nn/conv/cheb_conv.py in the stacktrace, that's why I thought you might be able to provide some insight.

Your solution does help. By reducing edge_attr to 1 dimension fixes the issue.
In PyG, is it not possible to have multiple dimensions for edge_attr?

[rusty1s]

It is, but not all GNN layers support it (by definition). You can take a look at the cheatsheet for information on that.