Global parameter fine-tuning or partial fine-tuning?

Hi! I am very interested in Human Centric Visual Tasks, may I ask if the fine-tuning part for downstream tasks only implement at some layers or all layers?
Looking forward to your more awesome work!