Dataloader for mini-batch is only for inductive learning with multiple graph?

[Jeriousman]

Not the new nodedataloader, but the DataLoader described in ADVANCED MINI-BATCHING in PyG documentation. This is only for explicit multiple graphs for inductive learning? That is how I understood. For transductive learning, as it is a huge one graph, we cant use the dataloader but use NeighborLoader (for sampling) to fit into GPU if I am correct? Can you please verify if I am right? If I am wrong can you please specify some details?

And for inductive learning, I guess we can use NeighborLoader as well if we want to sample even from a relatively small graph of many graphs?

Thank you in advance.

[rusty1s]

Yes, for learning on large graphs, have a look at NeighborLoader, ClusterLoader, GraphSAINTLoader, etc (all loaders that take in a data object instead of a dataset). If you want to perform inductive learning, you may first obtain a subgraph only containing training nodes: train_data = data.subgraph(train_mask).

[Jeriousman]

Thank you for the verification!