GIN - additional MLP?

[vthost]

Hi!
I just came across this and wanted to mention it in case you missed it so far as well :)
In the original paper by Xu et al., the general equation in Cor.6 considers an additional function f before the aggregation summation. In GIN, this MLP is dropped since it comes with the below assumption. (If I am not missing anything) I think GINConv drops the MLP as well, however users may not use one-hot features. Maybe you can add the assumption in the comments or and additional version w/ a second MLP?

In the first iteration, we do not need MLPs before summation if input features are one-hot encodings as their summation alone is injective. 

[rusty1s]

That is correct. We follow the final GINConv definition here. If your features are not one-hot encoded, and you want to ensure maximal expressivity, you have to apply an MLP before-hand:

x = self.mlp(x)
x = self.conv1(x, edge_index)
...

[vthost]

Thank you for confirming!