About the module pipeline

[lonelygoatherd]

Hi, I have a question on the module in your paper, which says "apply a Diffpool layer after two Graphsage layers", "A total of 2 Dffpool layers are used ", "After each Diffpool layer, 3 layers of graph convolutions are performed ". Ain't these paradox? The "graph convolutions" here are not Graphsage layers?
If so, is the pipeline 2 Graphsage+diffpool1+3 graph convolutions+fc1+2 Graphsage+diffpool2 +3 graph convolutions +fc2+fc3?
Is the Graphgsage used as edbedding GNN and "graph convolutions" as pooling GNN? why use different GNNs?
And Why add a prediction layer( fc layer) after each graph convolution in the code?
Also in the encoders.py, I see sometimes classes call each other's methods and attributes without raising an error? How does that work?
Could you give me some guidance on these? Have confused me for several days. Appreciate.

[phucdoitoan]

I have the same question after reading the paper. Hope someone can help clearing this point out.