A problem about prototypical loss.

clovaai / voxceleb_trainer

In defence of metric learning for speaker recognition

MIT License

1.03k stars 272 forks source link

A problem about prototypical loss. #52

Closed XinhaoMei closed 4 years ago

XinhaoMei commented 4 years ago

Hello, if M = 2, then is it a one-shot learning using the prototypical loss? The prototype of a class is just an embedding of a utterance? thank you very much!

XinhaoMei commented 4 years ago

And if the batch size is 200, is N also 200?

joonson commented 4 years ago

If you are referring to M and N in our paper, the answer is yes to both questions.