Triplet loss `invert_best=True` -> `False`

DIVA-DIA / DeepDIVA

⛔️ DEPRECATED <Python Framework for Reproducible Deep Learning Experiments>

https://diva-dia.github.io/DeepDIVAweb

GNU Lesser General Public License v3.0

32 stars 31 forks source link

Triplet loss `invert_best=True` -> `False` #12

Closed ashlaban closed 5 years ago

ashlaban commented 5 years ago

AFAIU, triplet training uses the MAP metric which ranges from 0 (bad) to 1 (good). I.e. the best value is the one with the maximum value. Hence I suspect this line of code:

https://github.com/DIVA-DIA/DeepDIVA/blob/68061031c186eda346f125cac4779e1df61c4938/template/runner/triplet/triplet.py#L130

should read

invert_best=False,

instead of True.

Renthal commented 5 years ago

You are indeed correct. This is a legacy bug for when we were using another metric; if you'd make a PR I accept it