Doubt regarding the size parameter in self.propagate

[Sidharth3301]

I was trying to understand code in this repo and came across this line, which kind of confused me since they have used the size=(num_nodes, num_nodes) parameter. Does this mean that they are treating the node embedding tensor in a bipartite fashion and since the flow parameter is not explicitly specified, can I assume that it is doing flow="source_to_target" hence only nodes of one partite will get messages and are updated, correct? Then the input node embedding size should be [2*num_nodes, in_features] , but that isn't the case since num_nodes = x.size(0) according to L51 in the snippet shared above. I am a bit confused as to what exactly is happening here. This code is an official implementation of this paper, which is trying to solve the problem of edge feature imputation and then solving downstream node classification as well using GNN's.

[rusty1s]

The size argument just refers to the shape of the underlying adjacency matrix encoded by edge_index. In most cases, this is simply (edge_index[0].max() + 1, edge_index[1].max()). How we propagate information is independent of the size argument and is solely described by edge_index. Since we assume that an undirected graph holds both directions within edge_index, that also means that "both partites" will receive messages.

[Sidharth3301]

Ah that makes a lot of sense. Thanks!