Simple question about nnet3 config and indexes

[Titouan Parcollet]

Hello everyone,

I'm trying to dive a bit more in nnet3 and play with basic recipes. But I can't get what does input=Append(..,..,..) means ? Let's take the example of this configuration (For a very straightforward DNN) :

What I understand is that I should have three hidden layers of size 512 right ? But what does input=Append(-1,0,1) means ? I read the DataType informations on the kaldi documentation, and even if I'm familiar with Neural Networks, I do not understand.

Thanks you very much !

Titouan,

[Daniel Povey]

that's an abbreviation that's used in xconfig files.  It means that the input to that layer is the previous layer appended over frame offsets -1, 0, 1.  So the previous layer at frames t-1, t, and t+1.

--
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/67ca9d16-e5c0-482a-95d4-863c1d805c1a%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Titouan Parcollet]

Okay so it introduces a temporal dimension ? so if I wanted a very basic DNN I should set this indexes to only 0 ?

Thanks you very much,

Le vendredi 20 octobre 2017 18:38:37 UTC+2, Dan Povey a écrit :

that's an abbreviation that's used in xconfig files.  It means that the input to that layer is the previous layer appended over frame offsets -1, 0, 1.  So the previous layer at frames t-1, t, and t+1.

On Fri, Oct 20, 2017 at 5:38 AM, Titouan Parcollet <parcolle...@gmail.com> wrote:

Hello everyone,

I'm trying to dive a bit more in nnet3 and play with basic recipes. But I can't get what does input=Append(..,..,..) means ? Let's take the example of this configuration (For a very straightforward DNN) :

What I understand is that I should have three hidden layers of size 512 right ? But what does input=Append(-1,0,1) means ? I read the DataType informations on the kaldi documentation, and even if I'm familiar with Neural Networks, I do not understand.

Thanks you very much !

Titouan,

--
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.

[Daniel Povey]

Yes, you could do that.  But the performance won't be as good.

Actually it defaults to just taking the previous layer, un-spliced, as input, so you can just remove the 'input=Append(-1,0,1)' entirely if that's what you want.

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/9e5272f0-dd17-4fa3-854f-e5d865ea5e65%40googlegroups.com.

[Titouan Parcollet]

Okay, thanks you very much dan. By the way this the the configuration that gave me the large decoding time .