retrain the last layer of TDNN chain model

[coas...@gmail.com]

Hi, all.

I have trained TDNN model on some data, and now I want to retrain the last layer on in-domain data.

I refer to run_nnet2_wsj.sh (in rm/ egs), but the nnet3 model differs to nnet2 model, and some command does not exists for nnet3 (e.g nnet-to-raw-nnet,which can extract all but the last layer).

Any one tried this and could you please offer some help?

Thanks in advance.

[Daniel Povey]

Look at

egs/rm/s5/local/chain/tuning/run_tdnn_wsj_rm_1?.sh

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
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+unsubscribe@googlegroups.com.
To post to this group, send email to kaldi...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/64ac9ba9-9b5c-47b1-add0-23cf28d35c4a%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Pegah Ghahremani]

You can check our transfer learning script for WSJ  to RM (egs/rm/s5/local/chain/run_tdnn_wsj_rm.sh) and also you can change the learning rate for other layer to 0 by setting  primary_lr_factor=0.0, although we found out that it degrades the performance.

You can also check our transfer learning paper for more details.

http://www.danielpovey.com/files/2017_asru_transfer_learning.pdf

[Armando]

I was reading the paper; it looks the case where all data (in domain, out of domain) all together for a normal training with no transfer learning is never considered as a baseline, even in case where the phone set stays the same across corpora (like Librispeech and wsj); or is it already obvious that it''d be worse than a baseline with only the target corpus?

[coas...@gmail.com]

Thank you, Pegah. I will try.

在 2018年3月27日星期二 UTC+8上午4:16:14，Pegah Ghahremani写道：

[coas...@gmail.com]

OK. 

Thank you, Dan.

在 2018年3月27日星期二 UTC+8上午1:35:22，Dan Povey写道：

Look at

egs/rm/s5/local/chain/tuning/run_tdnn_wsj_rm_1?.sh

On Sun, Mar 25, 2018 at 4:11 AM, <coas...@gmail.com> wrote:

Hi, all.

I have trained TDNN model on some data, and now I want to retrain the last layer on in-domain data.

I refer to run_nnet2_wsj.sh (in rm/ egs), but the nnet3 model differs to nnet2 model, and some command does not exists for nnet3 (e.g nnet-to-raw-nnet,which can extract all but the last layer).

Any one tried this and could you please offer some help?

Thanks in advance.

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
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.