Use NN.Conv without edge_attr

[AdarshMJ]

Hi! Im trying to use
from torch_geometric.nn import NNConv
The documentation says -
nn (torch.nn.Module) – A neural network
that maps edge features edge_attr of shape [-1, num_edge_features] to shape [-1, in_channels * out_channels]

The dataset Im working with doesn't have edge_attr, is there a way to use this without edge_attr? Any ideas are appreciated! Thank you.

[rusty1s]

No, NNConv is defined to operate on edge features, and so edge features will be required to run this operator. As a workaround, you could think of creating edge features on the fly, e.g., by concatenating source and destination node features:

edge_attr = torch.cat([x[edge_index[0]], x[edge_index[1]], dim=-1)

[AdarshMJ]

Hi Matthias,
So I think this works, I have created the edge_attrs. Thank you again.
I have two questions slightly unrelated -

1)After incorporating the edge_attrs, I am getting an error - RuntimeError: "addmm_cuda" not implemented for 'Long' Pytorch has this issue highlighted here - pytorch/pytorch#44428
Im not sure what part is giving this error. Should I just convert all my tensors to float64 or something?

2)In PNA, how do you set the self.edge_emb = Embedding() My data looks like this. -
Number of graphs in the current batch: 64
DataBatch(x=[559, 1], edge_index=[2, 4911], y=[9600], edge_attr=[2388], batch=[559], ptr=[65])
Number of graphs in the current batch: 64
DataBatch(x=[562, 1], edge_index=[2, 4950], y=[9600], edge_attr=[2448], batch=[562], ptr=[65])

Thank you for your time.

[rusty1s]

1. All features need to be of type floating-point (at best float32). The error suggests that this is not the case for you.
2. Note that you only need self.edge_emb = Embedding() in case your edge_attr holds integer values. You can use the Embedding() layer to map them to floating-point types:

edge_emb = Embedding(num_categories, out_channels)
edge_attr = edge_emb(edge_attr)

[AdarshMJ]

Hi @rusty1s Took me some time, but it works. Thank you!
For creating the edge_attr, I am currently doing -

def transforms(data):
edge_attr = torch.cat([data.edge_index[0],data.edge_index[1]], dim=-1)

and then applying it to my dataset as -
dataset = [transforms(data) for data in dataset]

But the edge_attr created are of the form -

Number of graphs in the current batch: 64
DataBatch(x=[386, 1], edge_index=[2, 2364], y=[9600], edge_attr=[1488], batch=[386], ptr=[65])

Step 2:
=======
Number of graphs in the current batch: 64
DataBatch(x=[389, 1], edge_index=[2, 2415], y=[9600], edge_attr=[1500], batch=[389], ptr=[65])

Is there a way to create them in the shape ?
data.edge_attr: Edge feature matrix with shape [num_edges, num_edge_features]

Thank you in advance for your time :)

[rusty1s]

def transforms(data):
    edge_attr = torch.cat([data.x[data.edge_index[0]], data.x[data.edge_index[1]]], dim=-1)

should fix this.

[AdarshMJ]

Thank you!