pytorch_xvectors icon indicating copy to clipboard operation
pytorch_xvectors copied to clipboard

Published 20 hours ago • verified publisher icon manojpamk

training data

Open mystlee opened this issue 4 years ago • 0 comments

Hi,

I'm tying to train the speaker verification with voxceleb1&2 using 'pytorch_run.sh' However, index error occurred in train_xent.py(stage 6) I think the default script need 80 archives(numArchives), however I can get only 75 archives.

I can take only 1483600743 frames in stage 5, so only 75 archives were generated in sid/nnet3/xvector/get_egs.sh(stage 6) "num_train_archives=$[($num_train_frames*$num_repeats)/$frames_per_iter + 1]" in sid/nnet3/xvector/get_egs.sh line 129

I wonder if it is correct to simply modify numArchives 80 to 75 is correct for reproducing.

thank you

mystlee avatar Jul 25 '21 11:07 mystlee