when fine-tuning a pre-trained model, do i always need to train the new classification head first?

Question

Let's say i want to fine-tune the last 50 layers of a mobilenetV2 model. First thing i do is i add a new classification head. I sometimes read that this new head should be trained exclusively before unfreezing the 50 layers that i want to fine-tune.

Is that really necessary, or can i just re-train these 50 layers plus the new classification head simultaneously right away?

Answer 1

It's better to train mobilbetv2 body and classification head simultaneously. In this way the model will be able to generalize you data as accurately as possible