How to design the “the last few layers of the network be unfreezed for transfer learning or finetuning”?

[JenniferYingyiWu2020]

Hi,
I would like to make YOLO do transfer learning. If I stopped YOLO training after the “dota-yolov3-tiny_300000.weights” has been generated, then how to implement the “transfer learning” using my own data set and different “classes numbers”? Especially, how to design the “the last few layers of the network be unfreezed for transfer learning or finetuning”?
So, could you pls help to give me some suggestions? Thanks!