Question about model details in the paper

[minhquoc0712]

Hi,

Thank you for your interesting research. I have two questions regarding the implementation of the model in the paper.

1. The graph encoder has two modules: GNN and Transformer. Each module takes the computational graph as input. How do you obtain a single embedding from many node embedding from EACH module before concatenating them together? Are you averaging them?
2. There are two losses mentioned in the paper: The CL loss and the regression loss. Are the graph encoder and the prediction head trained together? If so, then how did you combine these two losses? Or did you first train the graph encoder and then freeze it?

Thank you.