Using NeighborLoader

[loisbijing]

Hi all,
I'm trying to use NeighborLoader to obtain subgraphs from a large graph. I was testing how to change the settings so that I can control the size of the subgraph. I wrote a test code such that I could sample all nodes and edges of a large graph, meaning the subgraph is same as the entire graph. However, I don't seem to be able to increase my subgraph size above 17242 nodes. My code is below:

for data in train_loader:
    print(data)
    # print out: DataBatch(x=[283214, 3], edge_index=[2, 1655208], edge_attr=[1655208, 4], y=[283214], pos=[283214, 3], batch=[283214], ptr=[2])
    sub_loader = NeighborLoader(
        data,
        num_neighbors=[-1] * 280000,
        input_nodes=torch.arange(280000),
        shuffle=False
    )
    sub_batch_counter = 0
    for sub_batch in sub_loader:
        T.ToUndirected()(sub_batch)
        print(sub_batch)
        # print out: DataBatch(x=[17242, 3], edge_index=[2, 100196], edge_attr=[100196, 4], y=[17242], pos=[17242, 3], ptr=[2], n_id=[17242], e_id=[100196], num_sampled_nodes=[280001], num_sampled_edges=[280000], input_id=[1], batch_size=1)

How may I increase my subgraph size? e.g., node number to 283214? (My torch_geometric version is 2.3.1 )
Thanks a lot!

[EdisonLeeeee]

I don't seem to be able to increase my subgraph size above 17242 nodes.

This might be because not all nodes in the graph are connected together, and the largest connected component in the graph consists of a maximum of 17,242 nodes. So you are not able to increase the subgraph size.

[loisbijing]

It does seem my graph has disconnected regions/disconnected graphs. I wrote a simple code to test the number of nodes it could reach from node 0 and it turns out it could only reach 17242 nodes. You are totally right on what happened here. Thanks a lot!