TDNN relu-renorm-layer

[Delia Bullock]

I am trying to understand how the TDNN implemented in this file

          https://github.com/kaldi-asr/kaldi/blob/master/egs/tedlium/s5_r2/local/chain/tuning/run_tdnn_lstm_1a.sh

works and how it can be manipulated. I want to play around with it to understand it better. 

The first thing I wanted to is expand the relu-renorm-layer's splicing window on line 179. I changed the line from:

          relu-renorm-layer name=tdnn6 dim=512 input=Append(-3,0,3)

to

          relu-renorm-layer name=tdnn6 dim=512 input=Append(-4,0,3)

However, this caused an error while training on iteration 1. I attached the output of the error. It doesn't happen if I don't make the change. I must be misunderstanding what the input field is meant to do. Is there any documentation I could look at to better understand or is there anyone who could shed some light on this for me?

Thank you so much for your help!

Best,

Delia

[Daniel Povey]

It's running out of memory.

In common configurations, all the later layers have splicing indexes where the differences between all the indexes are multiples of 3; and the chain models only request the output at every 3 indexes.  This means that all the intermediate layers need to be computed only for every 3rd frame.  If you change that -3 to -4, it stops that, and for all the internal activations it now needs 3 times more memory (and 3 times more time).  You could fix the memory issue by reducing the minibatch size, but it would still be slower.

Dan

--
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.
For more options, visit https://groups.google.com/d/optout.