Understanding the Message Passing class

[mdanb]

I'm trying to gain a better understanding of the Message Passing class so I'm looking at the source code in torch_geometric/nn/conv/message_passing.py. I'm struggling to understand why the __init__ does what it does. In particular:
1- What do the special_args represent semantically?
2- In:

self.inspector.inspect(self.message)
self.inspector.inspect(self.aggregate, pop_first=True)
self.inspector.inspect(self.message_and_aggregate, pop_first=True)
self.inspector.inspect(self.update, pop_first=True)

why pop_first for everything except self.message, and why do the popping in the first place?

[rusty1s]

1. special_args denote the arguments that are automatically filled and which can be additionally accessed in message, aggregate, update.
2. aggregate and update are defined to take the output of the previous call as first argument for input. Therefore, we drop these arguments from the argument list.

[mdanb]

Can you elaborate further?

1. I'm not quite sure what you mean by "automatically filled" and "can additionally access"
2. What do you mean by "previous call"?

[rusty1s]

Sure, sorry for being so fuzzy :)

1. The special_args holds all additional argument names one can access in message, aggregate, and update, e.g., you can access the index along which to aggregate, or the size_i of the output number of nodes (basically some additional information about the graph). For example, this allows you to do the following:

    self.propagate(edge_index, x=x)

def message(self, x_i, x_j, index, size_i):
    ...

2. The message passing flow we follow is: message -> aggregate -> update. With this, aggregate will receive the output of message as its first argument. The same holds true for update.