Training of the refinement stage

[Lancial]

Hi

Thank you for this great work!

In your original paper, you mentioned about a second refinement stage will improve registration result. If I'm not mis-understanding it, you have a new seperate regularization network that refines the first network's result. And later, two displacement fields will be composited together to obtained a final displacement. I wonder for the first level network, do you update its parameter while second network is trained?