benedekrozemberczki/SimGNN

Visualizing Attention

sajjadriaj opened this issue · 2 comments

Hi, I want to visualize the attention with respect to the input graph (like Figure 5 in the paper). Can you please guide me how to visualize the attention weight with the input graphs?

In every epoch you get the weights/activations and turn hem into Numpy arrays. You can use a heatmap.

Could you please provide more clarification on this? The activation weights are on the output of the convolutional pass. so they are vectors of length 32. How do you get a heatmap on the graph nodes from this? Sorry I am quite new to attention so I may not fully understand how to do this.