NeighborSampler vs. NeighborLoader

[pintonos]

Hello!
What is the difference between the NeighborSampler and the NeighborLoader? I am currently working in an inductive setting and the only example in the repo uses NeighborSampling in combination with SageConv. However, for my specific use case (additional edge weights + GCNConv) generating batches via the NeighborLoader would be more convenient. So I was wondering if they differ in the inductive case?

Thanks for the help!

[rusty1s]

We recommend the usage of NeighborLoader, which streamlines the interface of NeighborSampler: NeighborLoader will just return a single data object (not a list of bipartite graphs), which makes it much more broadly applicable and easier-to-use. The functionality/output embeddings should be the same in either way :)

[rusty1s]

No, it will not guarantee that you only sample training nodes. To apply it in an inductive learning scenario, you first have to filter out non-training nodes and only push a "subgraph" data object only holding training nodes to the loader:

train_loader = NeighborLoader(data.subgraph(data.train_mask), num_neighbors=[10, 10], shuffle=True,
                              input_nodes=None, batch_size=512)

Now, input_nodes can be None (it will sample neighbors for all nodes in the graph). Note that data.subgraph(...) is a rather new method (you may need to install from master or manually filter out your data object via torch_geometric.utils.subgraph).

[pintonos]

This code seems to work fine on CPU, but gets a Segmentation Fault on the GPU:

train_edge_index, train_edge_weight = subgraph(data.train_mask, edge_index=data.edge_index, edge_attr=data.edge_weight)
train_graph = Data(data.x, train_edge_index, edge_weight=train_edge_weight)
train_loader = NeighborLoader(train_graph, num_neighbors=[10, 10], shuffle=True,
                              input_nodes=None, batch_size=512)

Any idea why?

[rusty1s]

NeighborLoader can only process CPU tensors, it does not have a GPU implementation. As such, this error is to be expected (although it might be great to add a more meaningful error message).

[AntonCoon]

Is it correct to understand that, for GPU usage, NeighborSampler is the sole option without any alternatives?

[EdisonLeeeee]

No, NeighborSampler also does not have GPU support.