Getting a discrepancy when using to_undirected()

[scigeek72]

Hello,

I am building a graph data from a .csv file. Once the data is built (it is directed at this point), I am using to_undirected to convert it to an undirected graph. Below is what I am getting depending on whether I am using num_nodes parameter in to_undirected.

c = to_undirected(edge_index=ei_tr, edge_attr=ea_tr)

c[0].shape
Out[224]: torch.Size([2, 33282])

is_undirected(c[0])
Out[225]: True

now with num_nodes

n_nodes_tr = 3786
d = to_undirected(edge_index=ei_tr, edge_attr=ea_tr, num_nodes= n_nodes_tr)

d[0].shape
Out[227]: torch.Size([2, 33258])

is_undirected(d[0])
Out[228]: False

Note that the number of directed edges is 16641. So, without num_nodes, the code seems to produce the correct shape.

When I am using num_nodes, the shape of the edge_index changes (my guess is that this is causing is_undirected() to be False as for some edges, the inverse is not getting included in the output of the function).

I looked at the source code for to_undirected() (here) but couldn't resolve this behavior.

Can num_nodes change the behavior of to_undirected()?

Your help is greatly appreciated. Thanks.

[EdisonLeeeee]

Hi @scigeek72
In your final example, I guess num_nodes is smaller than the actual number of nodes in your graph, i.e., num_nodes < ei_tr.max() + 1, which results in unexpected behaviors.

[scigeek72]

Thanks for your reply @EdisonLeeeee, but that bothers me a bit (although I may be wrong). num_nodes == ei_tr.max()+1 may not always be true, for example, if the graph has isolated nodes. We may not have that information prior to building the dataset. Certainly, in my case, I need to first build the graph and pytorch_geometric data before I know that there are isolated nodes.

Also, you are assuming that nodes are renamed between 0 and range(len(number of nodes)). But is it a fair assumption?

Again, my view of the situation is through my own problem. May be in the larger context, your assumptions are correct.

Thanks for taking the time to respond.

[EdisonLeeeee]

num_nodes == ei_tr.max()+1 may not always be true, for example, if the graph has isolated nodes.

I agree with that there are cases with isolated nodes would result in unexpected behaviors. In that cases, PyG assumes that users must be responsible for their inputs and they should make sure num_nodes > max_node_id. One can use torch_geometric.utils.remove_isolated_nodes to remove isolated nodes and reorder them.

edge_index = ... # build graph
edge_index = remove_isolated_nodes(edge_index)[0]
edge_index = to_undirected(edge_index)[0]

Also, you are assuming that nodes are renamed between 0 and range(len(number of nodes)). But is it a fair assumption?

In fact, nodes should be renamed between 0 and range(max_node_id), so PyG uses num_nodes == ei_tr.max()+1 to figure out the number of nodes by default. Not ideally, but that would make things simple in most standard cases.