YOLOE full fine-tune vs linear probing best practices

Discussion
1

YOLOE supports 2 modes of training, “full finetuning” and “linear probing”.

  1. what are the best practices for choosing which mode to use, is it simply a choice based on if we are compute restricted?

  2. I was thinking that for domain adaptation, (COCO style camera angle domain to AERIAL domain for example) full fine-tuning would be required to update the backbone to handle the new domain.

  3. In particular though I was curious if first fine-tuning for a small amount of epochs (10, 20, 50, 100?) and then switching to Linear probing is a good practice?, or is it a bad idea for some reason.

There is some work on Linear Probing to Finetuning in LLM, but from what I can tell it does not apply to detection models like YOLOE

2
  1. The train usage section explains what each does, and when linear probing may be preferable. YOLOE: Real-Time Seeing Anything - Ultralytics YOLO Docs
  2. Probably.
  3. I would do it the other way. If linear probing doesn’t get decent performance, then more of the model needs to be unfrozen and fine-tuned.
1 Like