What's the difference between torch_geometric.sampler.NeighborSampler and torch_geometric.loader.NeighborSampler?

[jeyendranbalakrishnan]

• Which one should I use?
• It appears that torch_geometric.sampler.NeighborSampler is used by torch_geometric.loader.NeighborLoader. Is this correct?
  Thanks a lot!

[rusty1s]

• You should use NeighborLoader - loader.NeighborSampler is deprecated and will be removed soon.
• Yes, NeighborSampler is an internal class in NeighborLoader that is used for graph sampling (and calls the specific C++ low-level routines).

[jeyendranbalakrishnan]

Thanks a lot! That helps me understand better.
One follow up question:

• Can we view NeighborLoader as a sort of "improved" version of DataLoader that adds neighbor subsampling capability for nodes in a minibatch, whereas DataLoader copies all the neighbors into each minibatch?

[rusty1s]

Mh, no, they serve different purposes. NeighborLoader enables mini-batch sampling on single large graphs (e.g., for learning on a single social network), while DataLoader corresponds to the traditional scenario of grouping multiple (small) graphs into a mini-batch (e.g., for learning on a large dataset holding multiple molecules).

[jeyendranbalakrishnan]

Thanks for the clarification, @rusty1s.