search

davidsandberg / facenet

Face recognition using Tensorflow
MIT License
13.71k stars 4.8k forks source link

Triplet loss finetune #1075

Open tangchen2 opened 5 years ago

tangchen2 commented 5 years ago

Hi, when i use FaceNet train CASIA-Webface, i first use Softmax Loss, and after 90 epoch, the model gets 98% accuracy and 95% validation rate. Then, i want to use Triplet Loss to finetune the model, however, i can't find suitable parameters to improve the model. I have try the learning rate from 0.001 to 1e-7, the model performance all decreased (especially the validate rate) and I am really confused about that, how do you guys finetune the model with Triplet Loss?

xlphs commented 4 years ago

I have written some comments in this thread to help. https://github.com/davidsandberg/facenet/issues/1012#issuecomment-490119259