large scale heterogeneous graph classification

[leethu2012]

I am building a classifier for heterogeneous graphs. The graphs are too big and cause GPU OOM problem during training.
Can you give some advice for this problem?
I wonder whether sampling methods, such as HGTLoader or NeighborLoader, which are used for node classification, can be used for large scale heterogeneous graph classification?
Thanks for any advice!

[rusty1s]

Do your graph don‘t even fit with a batch size of 1? One way you could solve it is to run data.subgraph on a random node subset before moving data to the GPU. This way, you can reduce the graph size.

[rusty1s]

Yes, you would be responsible for creating the subset_dict on your own, e.g., via

subset_dict = {}
for node_type in data.node_types:
    perm = torch.randperm(data[node_type].num_nodes
    subset = perm[:10000]  # take 10,000 random nodes per node type
    subset_dict[node_type] = subset

[rusty1s]

Will also work on HeteroData support for the random node loader.

[jiaqian]

ok, I actually could not provide subset_dict since I don't want to randomly split the graph. But I think I can make the graph homogeneous in the first place and use SparseTensor.partition() to create a list of subgraphs ([A_1,A_2,...,A_n]) and take each of them as a heterogeneous graph later to run models. Thank you so much.

[rusty1s]

Yes, this works as well.

[rusty1s]

Integrated into RandomNodeLoader via #6007.