GMN application

[simba0626]

Hi author, The work is very well. I want to use this work for question answering. I have one question, when run codes: I use 200-d vector for representiong node and edge features. After running, I find the attention weights can not align nodes well when compute two graphs match. Like this:

I do not know where is it wrong. Is it features vector error or others ?

Thank you

ywsun

[yujiali]

Hi ywsun,

It may help if you play with the temperature of the attention logits a bit to get a more visible attention pattern. For example you could multiply the attention score s(h_i, h_j) with a factor of 1/T, and then change T to something small, potentially that will give you a clearer pattern.

[daddydrac]

@yujiali What is the shape of the inputs I can generate for training? For example, I have seen some tools take a tuple [ (0,1), (1,2), (2,1)...] where node 0 connects to node 1, and node 1 connects to node 2. I have also seen JSON representations of graphs as well as TSV files where their values make a graph.

I simply want to generate whatever data and shape is needed to train the model to test it out, and what arg takes the input for this. I hope this makes sense.

Thank you, joe