search

clovaai / voxceleb_trainer

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

how do I train angular-prototypical loss for M=2? #41

Closed zeek-han closed 4 years ago

zeek-han commented 4 years ago

I tried nSpeakers=2 to make gsize_dict['angleproto'] == 2, but lowest VEER is about 2.58% until the 400 Epoch with training batch size 400 (--nSpeakers=2, --batch_size=200)

how do I train angular-prototypical loss for M=2?? Thank you...

joonson commented 4 years ago

For best performance, try:

  1. Larger batch size (e.g. --nSpeakers=2, --batch_size=400)
  2. Longer max_frames (e.g. 400 frames) during test phase
zeek-han commented 4 years ago

Thank you, do I need extra setting for M=2?

@ actually, --batch_size=200 is due to my GPU memory..;;;; but thank you

joonson commented 4 years ago

No extra settings are required.

zeek-han commented 4 years ago

thank you so much.