aszot/ggnn

Gated Graph Sequential Neural Networks in PyTorch. Easy to read and understand.

ggnn's Issues

Graph Level Output

I want to ask that, is it possible to get the graph level output using this code. As you indicated in your code

Equation 7 of section 3.3
      output = self.out(torch.cat((hidden_state, annotation), 2))
      output = output.sum(2)

This must evaluate the graph level representation of the graph. But i think this outputs the node scores using the following equation described in the paper
o = g(h; x)
and then passing these scores in cross entropy loss where first softmax is applied and then the loss is evaluated. I am confused, can you please guide a bit on this aspect.
Moreover, if it provides a graph level representation vector, then for graph classification task i can take this vector as input to another neural network that will output the class the graph belongs to and that thing is not yet available with the current code?