how can i training network with the order of dataset (audio has order with specific rule)

[Trương Trang]

hi everyone,

I want training my network in nnet3 with the order of utterance in dataset without shuffle when training. Because i want some data transfer will be train in last iter in each epoch.

But I dont't see any option for choosing for this. So I wonder how can I do this.

Thanks..

[Jan Yenda Trmal]

I think it's not possible. 

You might be able to do it with retraining on the selected subset using the previous final model as the initialization (you just need to provide new training data directories and one extra parameter --trainer.input-model to your steps/nnet3/chain/train.py

y.

--
Go to http://kaldi-asr.org/forums.html to find out how to join the kaldi-help group
---
You received this message because you are subscribed to the Google Groups "kaldi-help" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/f1d1af20-0999-4442-bade-9109f34aa4a8n%40googlegroups.com.

[Trương Trang]

yep, tks you for your suggest

[Bukai Sen]

Hi,

Please share the result of you experiment! I am interested to see if you get any improvement by doing so rather than pulling all of data at once.