Can anyone provide the documentation for the GATConv code? The code and terminology are slightly different than the Graph Attention Networks.

[arake001]

Can anyone provide the documentation for the GATConv code? The code and terminology are slightly different than the Graph Attention Networks.

[rusty1s]

I updated the GATConv in order to make it more readable and added some more comments/documentation.

Here, self.lin refers to the weight matrix W in the paper, and self.lin(x) refers to W * h. The parameters used for computing attention coefficients are given by self.att_src and self.att_dst. Note that we separate this computation into two steps: We first compute a scalar value individually for both node-level source and target node representations, and sum them up later in the message function. This is equivalent to first concatenating source and target node representations, and apply the dot product afterwards. Let me know if there is anything still unclear to you.