steps/nnet3/train_dnn.py training/validation loss output

[jeyhou...@gmail.com]

I'm using tedlium recipe to train chain/tdnn models and was wondering if there is a simple way of asking the training script (steps/nnet3/train_dnn.py) to print the training loss value for each iteration?

(also on some dev set as well)

[Daniel Povey]

That script may not even work any more.  You must be using an outdated recipe.  You should be focusing on scripts like local/chain/run_tdnn.sh.

On Mon, Feb 25, 2019 at 8:34 AM <jeyhou...@gmail.com> wrote:

I'm using tedlium recipe to train chain/tdnn models and was wondering if there is a simple way of asking the training script (steps/nnet3/train_dnn.py) to print the training loss value for each iteration?

(also on some dev set as well)

--
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.
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/b037a933-4fdb-4317-904b-67900cbb37b4%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Jeyhoun Marks]

for tedlium, I can see run_tdnnf.sh in egs/tedlium/s5_r3/local/chain - is this the recommended one?

(training with around 700h of data)

You received this message because you are subscribed to a topic in the Google Groups "kaldi-help" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/kaldi-help/UAKh81Oapyw/unsubscribe.
To unsubscribe from this group and all its topics, send an email to kaldi-help+...@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/CAEWAuyTJtxv7mz%3DPPRT0vgQ%2Bmi11Z2z5CiGJk6UTDBK36Nm3CA%40mail.gmail.com.

[Daniel Povey]

It kind of works but it's a buggy configuration.  You would do better to copy the current TDNN-F script from tedlium s5_r2.  You could make a PR with that change.  You can call it run_tdnn.sh, but make it a soft link to tuning/run_tdnn_1a.sh.  
The reason I haven't done it is that it needs WERs, which I don't have time to get right now.

Dan

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/CALpbs1KJEi3T7pD5pkC8joQhc49Qvvhv5rKftT87y58j1nF55g%40mail.gmail.com.

[jeyhou...@gmail.com]

thanks, will use that one then
also created this PR: https://github.com/kaldi-asr/kaldi/pull/3056