philipperemy
commented
3 years ago We use a history (buffer) of triplets to train/test our models. And every history_every=100 batches, we re-generate the history (buffer). In your case, I'd say you just need to update the nb_speakers. history_length=4 controls how big your buffer is. nb_speakers is just the number of speakers you want to select from all the speakers. nb_per_speaker is just a constant and you don't need to care about it. Positive samples are made of 2 utterances of the same speaker and this 2 is equal to nb_per_speaker from what I remember.
So long story short, just update nb_speakers.
Hi, I have some question about LazyTripletBatcher configuration. self.nb_per_speaker = 2 self.nb_speakers = 640 self.history_length = 4 self.history_every = 100 I'm training deep-speaker model on my own dataset, maybe I need to know what these parameters mean. Thanks.