Why is there a disagreement between edge_index dimension and the x feature matrix dimension?

[mcneela]

Has anyone dealt with an example in pytorch geometric where there is a disagreement between the edge_index (sparse form of the adj matrix) and the number of rows in the feature matrix x? I am currently working with the PPI dataset that comes bundled with pytorch geometric. The first graph in that dataset has x with size 1767x50 but the edge_index only has dimension 1750 which means that when I create the adjacency matrix by calling to_dense_adj(adj) it ends up being size 1750x1750. This causes issues, because I have an attention weight adjacency matrix based on the dimensionality of x which has size 1767x1767 and I cannot multiply them. I assume this is because 17 nodes (1767-1750) are fully disconnected and not connected to any other nodes in the graph, but why would pytorch geometric not account for this when creating the adjacency matrix and add an extra 17 rows/columns with all 0s and one 1?

[rusty1s]

Yes, this is correct. As such, most of our functions that just expect an edge_index matrix as input also support max_num_nodes or num_nodes arguments, see here:

adj = to_dense_adj(data.edge_index, max_num_nodes=data.num_nodes)

[mcneela]

Got it, thank you!