search

KaleidoZhouYN / Angular-Triplet-Loss

ATL(Angular Triplet Loss):A new loss function based on sphereface & triplet loss
184 stars 37 forks source link

Training from scratch #5

Open ghost opened 6 years ago

ghost commented 6 years ago

Why do think we should do the training using pre-trained model not a new training from scratch?! "method 1:step step 1: train a pre-trained model use sphereface"

KaleidoZhouYN commented 6 years ago

It's hard to train triplet loss directly,so do angular triplet,always need lots of iteration to reach the final converge status. while finetune on sphereface model won't have this problem.

ghost commented 6 years ago

I think this will decrease the ability of the generalization as you know, how you can solve this?

KaleidoZhouYN commented 6 years ago

yes，but the result show that sphereface constrain is too strong