Confused about relation between paper and code

[wishforgood]

Hi, I am confused about the relation between "propagation_net", "node update net" in your code and the propagation model in your paper. It seems that the "node update net" is the propagation model in your paper, and the parameter "self.prop_net_h_sizes" is {} in "propagation_net" when used in "BaseGGNN".
This is not a program error but I can't find anyone to ask to, you can close it if necessary.

[yujiali]

The propagation_net in the code refers to the module that computes the messages passed on the edges, i.e. the a vectors in eq(2) in the paper. In the implementation you can specify prop_net_h_sizes to use a MLP instead of a linear transformation as described in the paper.

The node update net refers to the other parts that takes these messages and the current node states as input and then update the node states.

[wishforgood]

Thanks for your answer. But the eq(2) in the paper uses the adjacency matrices to be the weights of the linear transformation while in the code the weights seem to have nothing to do with the adjacency matrices. On the contrary, the "node update net" in the code uses adjacency matrix(adjavency_list and self.a_list in the code) while in the paper it seems to be just a GRU-like net. So I'm still so confused, can you explain it please?

[wishforgood]

I guess the "propagation net" computes a representation for every 2*edge types of origin hidden states as the h in eq(2) and the rest part of eq(2) is then conducted in "node propagation net"?