GraphSAINTNodeSampler with CUDA

[sdesabbata]

I am trying to run a graph autoencoder using the GraphSAINTNodeSampler. The code works fine on CPU but not on CUDA due to the assertion below in the constructor of GraphSAINTSampler, which returns an AssertionError.

pytorch_geometric/torch_geometric/loader/graph_saint.py

Line 56 in 55fe439

assert not data.edge_index.is_cuda

I was wondering why the assertion is there and if there is a way to use GraphSAINTNodeSampler on CUDA (also considering that NeighborLoader doesn't work either as it returns NotImplementedError: Could not run 'pyg::neighbor_sample' with arguments from the 'CUDA' backend).

Thanks in advance for any help and suggestion. 😊

[sdesabbata]

My question above seems to be the same as the one discussed in this issue on GraphSAINTRandomWalkSampler, which, however, doesn't outline an example solution. I guess the sampler needs to be instantiated using as input a graph dataset on CPU, and then the device conversion (e.g., to(device)) needs to be applied to each mini-batch.

For instance

loader = GraphSAINTNodeSampler(
        graph,
        batch_size=128,
        num_steps=100,
        sample_coverage=100
    )

and then

for mini_batch in loader:
    mini_batch = mini_batch.to(device)

Would that be the intended usage?

[rusty1s]

Yes, this is the intended usage.