In the documentation for fastai.basic_train.freeze() it is written that the function will “freeze up to last layer”. What actually happens is that the function freezes up to last layer group. The docstring in the source code is likewise misleading.
In the documentation for fastai.basic_train.freeze_to(n:int) it is written that the function will “freeze layers up to layer n”. What actually happens is that the function freezes up to layer group n. The docstring in the source code is likewise misleading.
Correcting this would cause a lot less confusion.