search

clovaai / voxceleb_trainer

In defence of metric learning for speaker recognition
MIT License
1.03k stars 272 forks source link

Experimental results #40

Closed yy835055664 closed 4 years ago

yy835055664 commented 4 years ago

Hello, thank you very much for your extensive experiment. I have a question for you. In the experimental results, both GE2E and Prototypical Angular Prototypical use M (2, 3, 4 5） for comparison test, how do you set this M in the code?

joonson commented 4 years ago

M in the paper is nSpeakers in the code for metric learning based loss functions

yy835055664 commented 4 years ago

M in the paper is nSpeakers in the code for metric learning based loss functions

Thank you very much, I still need to ask some questions.

  1. When adjusting the M value, are other parameters (m, s) set to 0 or the default value in the code is selected? (Such as GE2E, Prototypical, Angular Prototypical)
  2. In Triplet loss, does CHNM directly choose hard_prob=0.5,hard_rank=10?
joonson commented 4 years ago
  1. The values of m and s don't matter for metric learning objectives.
  2. No, curriculum hard negative mining is not implemented in this code. You will need to do this by stopping and re-starting the experiment with the new parameters.
yy835055664 commented 4 years ago
  1. The values of m and s don't matter for metric learning objectives.
  2. No, curriculum hard negative mining is not implemented in this code. You will need to do this by stopping and re-starting the experiment with the new parameters.

thank you very much