Question about MessagePassing Class

[hkim716]

Hi, Matt.
I'm trying to understand MessagePassing Class used in PyG.
This page helped me a lot, but I have some questions.
When I see each code of convolution layers in PyG, they define forward function, and in the forward function, there is propagate() function, which internally calls the message(), aggregate() and update() functions.
I found most conv layers in PyG have message() functions there, but I cannot easily find aggregate() and update() functions. For example, in spline_conv.py, there is forward function, which calls propagate(). And propagate() will call message().

Here are my questions.

1. Are aggregate() and update() hidden by default?
2. I found some conv layers have message_and_aggregate() function. Does this function take the output from message() by default?
3. In my understanding, the entire MessagePassing operation has some orders like this :
   forward()->propagate()->message()->( message_and_aggregate() )->aggregate()->update()
   Am I right?

[rusty1s]

1. aggregate() and update() have default implementations. For example, by default, aggregate will aggregate incoming messages according to the aggr attribute ("add", "mean" or "max") of the MessagePassing class. update() just returns the input by default.
2. message_and_aggregate() fuses message and aggregate computation into a single step, which is useful in order to allow for more memory-efficient GNN operators. This post tries to explain it in more detail.
3. That is mostly correct. If message_and_aggregate is not implemented, the order of operations is:

forward -> propagate -> message -> aggregate -> update

If message_and_aggregate is implemented, the order of operations is:

forward -> propagate -> message_and_aggregate -> update