Using edge features with NeighborSampler and SageConv

[a-ism256]

Hello,

I'm trying to make NeighborSampler and SageConv work for my dataset (node classification) which has both node and edge features. Each graph is a complete graph with no self loops.

So far I'm trying to use the SageConv modification proposed here which was further implemented by the person asking the question here.

However, I'm running into issues using the NeighborSampler with my data for the following reasons:

1. How can I properly use e_id to get the corresponding edge features? Does e_id correspond to the index of the corresponding edge in the original edge_index or to the actual id of the edge (I don't understand the latter case since an edge in edge_index is defined by a pair of node ids).
2. My dataset has multiple graphs and NeighborSampler does not seem to be able to sample from multiple graphs. So far, the only solution I see is looping over the graphs in my dataset and creating a neighbor sampler for each. Or loading everything into a single Data graph however only nodes originally belonging to the same graph would be connected.

I would highly appreciate if someone could help me figure these out.

[wsad1]

1. Yes e_id corresponds to the index of the edge in the full(original) graph. So edge_attr[ed_id] should give you the edge attributes for the edges in the sample. Check this for more documentation and examples..
2. Yes, creating on large graph with multiple disconnected components as described here seems like the best option.

[rusty1s]

Yes, this looks good. Self-loops are not necessary anymore :)

[a-ism256]

Hello again!

While the above method seemed to work I ran into an OOM error since the single giant graph turned out to be too big. Therefore I don't think I can set the batch_size to len(dataset) but will have to rely on mini-batches.

My question is how would the NeighborSampler be used in this case? Should I create a sampler for each batch as follow?

for batch in loader:
    sampler = NeighborSampler(batch.edge_index, node_idx=batch.train_mask, sizes=nbr_sizes, batch_size=1024, shuffle=True, num_workers=12)
    for batch_size, n_id, adjs in sampler:
        ...

I'm worried about using NeighborSampler in this case and ending up sampling across different graphs.

Honestly at this point I'm wondering if there could be some other framework to suit my application: inductive node classification using node and edge features with a multi-graph dataset. Since I'm trying to do this using SAGEConv I'm having to do a lot of workarounds and I'm starting to go in circles.

Appreciate and advice you might provide.

[rusty1s]

How big is your total dataset? And how big is each individual graph? Are you talking about CPU or GPU OOM errors? Since you are already holding dataset in CPU memory, creating a single giant graph from it should fit in CPU memory as well.

[a-ism256]

It's a CPU OOM issue. I actually ran into the OOM error when I was trying to use the InMemoryDataset so I switched to Dataset where each graph is saved as an individual file to disk.

However, if I were still to do what I mentioned above loading all the individual graphs into one giant graph by setting batch_size = len(dataset) would again cause the OOM error.

[rusty1s]

I see. I think your best bet is to create a sampler for each individual graph object, and iterate over each sampler for each graph.