Can't update nn.dense.linear object inside nn.conv.nn_conv using .data

[swedenfish]

Hi, I have a GNN model which has conv like this:

    nn = Sequential(Linear(6, 36), ReLU())
    self.conv1 = NNConv(1, 36, nn, aggr='mean')

And I am trying to update this model by manually updating its fields as below:
model.conv1.nn[0].weight.data = main_model.conv1.nn[0].weight.data.clone()
model.conv1.nn[0].bias.data = main_model.conv1.nn[0].bias.data.clone()
model.conv1.bias.data = main_model.conv1.bias.data.clone()
model.conv1.lin.data = main_model.conv1.lin.data.clone()

The first three works fine but the fourth one doesn't work, it is complained that Linear object doesn't have attributes data.
To be specific, Linear object here is the "lin" above, which is a torch_geometric.nn.dense.linear object.
And it is an attribute declared inside NNConv.

Could I please have some helps on how I can update the model's conv1.lin? Thank you!!!

[wsad1]

This is expected. Instead try lin.weight.data.

[swedenfish]

Thanks for answering! But as before, I think I am trying to update the whole Linear object with this line.
model.conv1.lin.data = main_model.conv1.lin.data.clone()
Is that equivalent to just updating the weight? What about other properties inside the Linear object?

[wsad1]

model.conv1.lin should have weight and bias you'll need to update both weight.data and bias.data in that case.

[swedenfish]

I think bias should be fine? Here is the source code for NNConv

if root_weight: self.lin = Linear(in_channels[1], out_channels, bias=False, weight_initializer='uniform')

In each NNConv object, it declares its inner Linear object with a default of not learning bias.
I was trying to think about other parameters I would need to update in my codes.

[wsad1]

In that case there aren't any more parameters.

[swedenfish]

Thank you so much for your help!