Where is the sampling for SAGEConv implemented?

[OliEfr]

Hi all!

According to the paper GraphSAGE (SAGEConv) implements a neighbourhood sampling for the nodes. I couldn`t find this implementation detail in the code for pyg. Could someone please hint me to the respective LOC in the code?

Thank you :)

[rusty1s]

We have seperated GNN layer and neighbor sampling routine in PyG:

• SAGEConv implements the GNN layer
• NeighborLoader implements the neighborhood sampling routine

[OliEfr]

Ok, so when I create a NN with a SAGEConv layer the NeighborLoader is called somewhere during the forward pass (without the need for me to implement anything), correct? Could you point me to the LOC where this happaning? Sorry, I can not seem to find it (stepping through the code with the debugger also didn't help me).

A lot of thanks in advance, and sorry for my inexperience.

[OliEfr]

Ok I found this example, which uses the NeighborLoader as the dataloader (as opposed to "normal" DataLoader). Is that what I need to do?

However, I have been using the SAGEConv module in a heterogenous graph with the "normal" DataLoader. What would that mean for the sampling. Would just all neighbor nodes be sampled in that case?

[rusty1s]

Ok I found #7371, which uses the NeighborLoader as the dataloader (as opposed to "normal" DataLoader). Is that what I need to do?

Yes, exactly.

Would just all neighbor nodes be sampled in that case?

Yes, there is no sampling inside SAGEConv.

[OliEfr]

Ok thank you. Maybe its a good idea to add a hint in the documentation regarding that.

[arunraja-hub]

@rusty1s does that mean that just SAGEConv with its default parameters is just a vanilla message passing layer?