Question about the process of parsing the arguments input to propagate() and lifting tensors?

[alantes]

Hi folks, thank you so much for developing PyG.
When I read the source code in propagate() and __collect__() in message_passing.py and set breakpoints along the way, I found self.__user_args__ really making me confused.
For example, if I input an tensor x (shape: [2190, 2]) into propagate() as a kwarg, when executing coll_dict = self.__collect__(self.__user_args__, edge_idex, size, kwargs), the self.__user_args__ already contains x_j. But I do not find a place in MessagePassing that you change self.__user_args__, except for __init__(). But when executing __init__(), x is not even introduced.
I wonder where in the program do you parse the arguments input to propagate(), and how do you judge whether or not an argument can be lifted to _i and _j components? Through dimensions?
Thank you in advance. I will appreciate your help a lot.

[rusty1s]

__user_args__ collects the arguments of message, aggregate and update and collects them. By default, message is implemented as

def message(self, x_j: Tensor):
     return x_j

That's why it will contain x_j without any custom message definition.

[alantes]

That's very helpful. Thank you for your reply.