clovaai / voxceleb_trainer

In defence of metric learning for speaker recognition
MIT License
1.05k stars 272 forks source link

which epoch can get the best result on test set? #1

Closed wuqiangch closed 4 years ago

wuqiangch commented 4 years ago

@joonson ,Hi. The max epoch(=500) is too large for me to train my dataset because of my k80 gpu. So I want to know if the more epochs can get better result. Thanks!

joonson commented 4 years ago

Picture1

Picture2

These are plots of EER vs epoch of a network using two different loss functions, tested using 2-second speech segments. The training takes 5 days on our GPU.