as far as I know, and as can be seen here enter link description here it is quite clear that it is better to stop at the "turning point" where the validation loss starts growing.
What I do not understand is two things:
- As can be seen here: my best mAP is located quite far to the right of the turning point. Is it better to choose a model of option B (located in the overfit zone but the mAP is better), or it is better to choose model A, which is quite in the same area of the turning point, though it's mAP is smaller?
2)More theoretical question: As far as I understand the mAP and the validation loss both calculated on an unseen dataset - the validation dataset, Thus I do not quite get how technically happens that mAP gets better(grows) while the validation loss gets worse (grows)? Shouldnt they behave quite the same?