how to update only one or part of the nodes in the graph

[hellomuyi]

Hello, I would like to ask, how to update only one or part of the nodes in the graph in PyG's GAT. To be precise, I only need to get the attention weight of a node or some nodes (not all nodes) in the graph to its neighbors, I read the source code, that is, "torch_geometric.utils.softmax", the bottom layer uses operations such as torch.scatter_, but I still can't solve it.

If no such method exists, is there an efficient way to extract the attention weights of partial center nodes to their neighbors from all the attention weights (alpha: (num_edges, heads))?

[rusty1s]

You should be able to only pass in a single node and its neighbors to GATConv:

edge_index = torch.tensor([[1, 2, 3, 4, 5], [0, 0, 0, 0, 0]])  # bipartite graph only pointing to node 0
x = ...  # node feature matrix with shape [6, #features]

out_central = conv((x, x[:1]), edge_index)
# out_central has shape [1, #features]

Let me know if this resolves your issues :)

[hellomuyi]

Thank you. That's all right. But I'm a little confused about this example. I think node 0 should be updated like this:

edge_index = torch.tensor([[0, 1, 2, 3, 4], [0, 0, 0, 0, 0]])  # bipartite graph only pointing to node 0
x = ...  # node feature matrix with shape [6, #features]

out_central = conv((x[1:], x[:1]), edge_index)
# out_central has shape [1, #features]

or

edge_index = torch.tensor([[1, 2, 3, 4, 5], [0, 0, 0, 0, 0]])  # bipartite graph only pointing to node 0
x = ...  # node feature matrix with shape [6, #features]

out_central = conv((x, x[:1]), edge_index)
# out_central has shape [1, #features]

Did I misunderstand something?

[rusty1s]

No, you are correct. I fixed it in my response as well.