How to train the Bert-net model?

Hi, when I directly specify squad_model as bert_qanet, I can only get an EM/F1 score of about 44, so I want to ask what the specific training steps of Bert-net are? How to implement the three-step layer-wise training you mentioned in article?