Size argument missing in TransformerConv.forward(...)

[aabbas90]

Hi,

I am trying to use TransformerConv layer on bipartite graph G = ((V_l, V_r), E). The goal is to send information from right vertex set V_r and edges E to V_l and then vice-versa. The code is:

left_updater= TransformerConv((num_node_feat, num_node_feat), num_node_feat, edge_dim = num_edge_feat, flow="source_to_target")
right_updater = TransformerConv((num_node_feat, num_node_feat), num_node_feat, edge_dim = num_edge_feat, flow="target_to_source")

left_updated_features = left_updater((left_features, right_features), edge_index_left_right, edge_features)
right_updated_features = right_updater((left_features, right_features), edge_index_left_right, edge_features)

However, I am getting the following error:

File "~/anaconda3/envs/learnDA/lib/python3.7/site-packages/torch_geometric/nn/conv/message_passing.py", line 127, in lift
return src.index_select(self.node_dim, index)
IndexError: index out of range in self

I guess it might be because I am not passing the size argument to give the size of nodes in V_l, V_r because the TransformerConv.forward(...) does not accept it as an argument. Would be happy to get your feedback on how to use TransformerConv on bipartite graphs. Thanks!

[rusty1s]

According to our tests, TransformerConv should work on bipartite graphs just fine, e.g.,

conv = TransformerConv((8, 16), 32)
conv((x1, x2), edge_index)

I guess this is induced by the flow="target_to_source" modification. Note that in this case, you need to flip the (left_features, right_features) formulation, i.e.

right_updated_features = right_updater((right_features, left_features), edge_index_left_right, edge_features)