search

HarryVolek / PyTorch_Speaker_Verification

PyTorch implementation of "Generalized End-to-End Loss for Speaker Verification" by Wan, Li et al.
BSD 3-Clause "New" or "Revised" License
575 stars 165 forks source link

perm action #28

Open czy97 opened 5 years ago

czy97 commented 5 years ago

Hi, sir. I don't think that the perm and unperm acitions in your code make any difference. Because the perm action is along the batch dim, in the forward process, the different data along the batch dim are unrelated. Or I misunderstood your purpose?

Oktai15 commented 4 years ago

https://github.com/HarryVolek/PyTorch_Speaker_Verification/issues/6

dkatsiros commented 3 years ago

@czy97 Before applying the perm action there is a torch.reshape(). This changes each batch from (N,M,sequence_length,features) to (N*M,sequence_length,features) so that speaker utterances are the one after another (e.g. U11,U12,U13,..,U1M,U21,U22,...,U2M,UN1,UN2,...,UNM). Perm action performs shuffling but in a way that is reversible.

baochi0212 commented 1 year ago

perm is unnecessary in training step