search

IDRnD / ReDimNet

The official pytorch implemention of the Intespeech 2024 paper "Reshape Dimensions Network for Speaker Recognition"
MIT License
117 stars 5 forks source link

sphereface2 loss #9

Closed ngocson1804 closed 3 months ago

ngocson1804 commented 3 months ago

Thank you for your great work! Could you share with me the config of sphereface2 (such as alpha, s, m, t) that you used to train the best B2 model?

vanIvan commented 3 months ago

Hello! Here are the hyperparameters of sphereface2 loss (they are taken from wespeaker pipeline without change), please refer to wespeaker loss implementation:

# Loss/head(projection) definition
projection_args:
  do_lm: false
  easy_margin: false
  embed_dim: 192
  lanbuda: 0.7
  margin_type: C
  project_type: sphereface2
  scale: 32.0
  t: 3
ngocson1804 commented 3 months ago

Thanks for your useful information!

vanIvan commented 3 months ago

Hope it'll help, closing issue.