Tuple form of input x?

[Nokimann]

Modules in torch_geometric.nn.conv can get the tuple form of input x.
Each means x_src and x_tgt (?)

1. Can we use them, with different dimensions of x_src and x_tgt, if we just set the tuple of in_channels?

2. If the above true, does [j, i] pair in edge_index means "from j-th row of x_src to i-th row of x_tgt"?

[rusty1s]

1. Yes, you can see here which operators support "bipartite" message passing flow.
2. Yes :)

[Nokimann]

Thank you as always.

[ZIKEYUAN]

I have a question about input as tuple. I would like to ask what is the difference between input x as tuple and input as tensor? I found that some GNN models do not need to convert x to tuple but some do. I am very confused about this.

[rusty1s]

A tuple of node features, e.g., (x_src, x_dist) represents a bipartite message passing flow where source node features x_src are passed to destination nodes and aggregated to x_dst to compute a new destination node representation. If it is passed as a tensor, we assume that both x_src and x_dst are equal.

[ZIKEYUAN]

Thank you very much!