kaldi-developers — Kaldi Developers

Re: [Kaldi-developers] extract-segments

[Daniel P. <dp...@gm...>]

I think Nagendra may have been using this, he should chime in.
Dan



On Thu, Apr 18, 2013 at 1:30 PM, Arnab Ghoshal <ar...@gm...> wrote:

> Hi all,
>
> we just noticed that there is an (unmentioned) assumption in
> extract-segments.cc that an end time of (0, -1] in the segments file
> means "include till the end of the file". But there are additional
> logical bugs that causes an end time of 0 to have the same effect. I
> do not like having this special value of the end time and plan to
> remove it. But is there anybody who has a good reason to keep such a
> functionality?
>
> -Arnab
>
>
> ------------------------------------------------------------------------------
> Precog is a next-generation analytics platform capable of advanced
> analytics on semi-structured data. The platform includes APIs for building
> apps and a phenomenal toolset for data science. Developers can use
> our toolset for easy data analysis & visualization. Get a free account!
> http://www2.precog.com/precogplatform/slashdotnewsletter
> _______________________________________________
> Kaldi-developers mailing list
> Kal...@li...
> https://lists.sourceforge.net/lists/listinfo/kaldi-developers
>

[Kaldi-developers] extract-segments

[Arnab G. <ar...@gm...>]

Hi all,

we just noticed that there is an (unmentioned) assumption in
extract-segments.cc that an end time of (0, -1] in the segments file
means "include till the end of the file". But there are additional
logical bugs that causes an end time of 0 to have the same effect. I
do not like having this special value of the end time and plan to
remove it. But is there anybody who has a good reason to keep such a
functionality?

-Arnab

Re: [Kaldi-developers] obtaining 0'os when computing posteriorgrams

[Daniel P. <dp...@gm...>]

Thanks-- and thanks for reporting the issue.  I won't change this one
though, because in my opinion (and I know this is subjective) it was
behaving as desired.
Dan


On Mon, Apr 8, 2013 at 3:30 PM, Xavier Anguera <xan...@gm...> wrote:

> I know,
> the problem happened when not applying this standard method to adapt the
> code to perform MAP adaptation.
> In my opinion the posteriors should always be >0, even if their value is
> equal to the minimum possible float value available (like in my code).
> I do agree though that in standard ASR model training this is not
> relevant. I am finding this because I am using Kaldi for things other than
> ASR.
> BTW, thanks for helping put together such a nice tool!
>
> X.
>
>
> On Mon, Apr 8, 2013 at 9:23 PM, Daniel Povey <dp...@gm...> wrote:
>
>> It's the normal practice when dealing with Gaussians to get rid of small
>> counts.
>> Dan
>>
>>
>>
>> On Mon, Apr 8, 2013 at 3:23 PM, Xavier Anguera <xan...@gm...>wrote:
>>
>>> I agree my "hack" is not the solution.
>>> I see that when performing an EM training there is a check for very
>>> small occupancy or weight, and eliminates a Gaussian if it is so. I am
>>> though not happy with such an approach and had commented out that line some
>>> time ago (I am implementing a MAP adaptation function that needs to deal
>>> with these cases)
>>>
>>> X.
>>>
>>>
>>> On Mon, Apr 8, 2013 at 9:19 PM, Daniel Povey <dp...@gm...> wrote:
>>>
>>>> Hm, thanks, but I don't think this is the right way to fix the problem.
>>>>  Update code should always take into account the possibility that
>>>> occupancies will be zero.  It's expected that exp() on very negative values
>>>> will produce zero.
>>>> Dan
>>>>
>>>>
>>>> On Mon, Apr 8, 2013 at 3:15 PM, Xavier Anguera <xan...@gm...>wrote:
>>>>
>>>>> Hi Dan,
>>>>> the segmentation fault comes from a division by 0 when using the
>>>>> occupancy of the Gaussians, that has been computed by adding together all
>>>>> the posterior probabilities for each Gaussian and a set of features. When
>>>>> all posteriors for a given Gaussian and all features is 0, there is a
>>>>> division by 0.
>>>>> I am pasting the "hack" I wrote to prevent this. I believe though that
>>>>> maybe the exp() function should be revisited. Tell me what you think.
>>>>>
>>>>> Real VectorBase<Real>::ApplySoftMax() {
>>>>>   Real max = this->Max(), sum = 0.0;
>>>>>   for (MatrixIndexT i = 0; i < dim_; i++) {
>>>>>       data_[i] = exp(data_[i] - max);
>>>>>       if(data_[i] < FLT_MIN )
>>>>>           data_[i] = FLT_MIN; //very small value
>>>>>       sum += data_[i];
>>>>>   }
>>>>>
>>>>>   this->Scale(1.0 / sum);
>>>>>   return max + log(sum);
>>>>> }
>>>>>
>>>>>
>>>>> On Mon, Apr 8, 2013 at 8:58 PM, Daniel Povey <dp...@gm...> wrote:
>>>>>
>>>>>> Firstly, this should give you numerical problems but not a
>>>>>> segmentation fault.
>>>>>> You'll have to look in the code and see if it's behaving as expected.
>>>>>>  E.g. is it due to a number so small that it cannot be represented in
>>>>>> floating point, or is it larger than that and unexpectedly becoming zero?
>>>>>> It might be an issue with your algorithm design.
>>>>>> Let me know if that function needs to be fixed.
>>>>>>
>>>>>> Dan
>>>>>>
>>>>>>
>>>>>> On Mon, Apr 8, 2013 at 2:54 PM, Xavier Anguera <xan...@gm...>wrote:
>>>>>>
>>>>>>>  Hi,
>>>>>>> when using the function template<typename Real> Real
>>>>>>> VectorBase<Real>::ApplySoftMax() in kaldi-vector.cc file, I noticed that
>>>>>>> very small likelihoods are rounded to a posterior probability of 0.0
>>>>>>> Is this an expected behavior? I am trying to perform an EM training
>>>>>>> of a simple GMM and I keep bumping into segmentation fault due to this.
>>>>>>>
>>>>>>> Thanks
>>>>>>>
>>>>>>> Xavi Anguera
>>>>>>>
>>>>>>>
>>>>>>> ------------------------------------------------------------------------------
>>>>>>> Minimize network downtime and maximize team effectiveness.
>>>>>>> Reduce network management and security costs.Learn how to hire
>>>>>>> the most talented Cisco Certified professionals. Visit the
>>>>>>> Employer Resources Portal
>>>>>>> http://www.cisco.com/web/learning/employer_resources/index.html
>>>>>>> _______________________________________________
>>>>>>> Kaldi-developers mailing list
>>>>>>> Kal...@li...
>>>>>>> https://lists.sourceforge.net/lists/listinfo/kaldi-developers
>>>>>>>
>>>>>>>
>>>>>>
>>>>>
>>>>
>>>
>>
>

Re: [Kaldi-developers] obtaining 0'os when computing posteriorgrams

[Xavier A. <xan...@gm...>]

I know,
the problem happened when not applying this standard method to adapt the
code to perform MAP adaptation.
In my opinion the posteriors should always be >0, even if their value is
equal to the minimum possible float value available (like in my code).
I do agree though that in standard ASR model training this is not relevant.
I am finding this because I am using Kaldi for things other than ASR.
BTW, thanks for helping put together such a nice tool!

X.


On Mon, Apr 8, 2013 at 9:23 PM, Daniel Povey <dp...@gm...> wrote:

> It's the normal practice when dealing with Gaussians to get rid of small
> counts.
> Dan
>
>
>
> On Mon, Apr 8, 2013 at 3:23 PM, Xavier Anguera <xan...@gm...> wrote:
>
>> I agree my "hack" is not the solution.
>> I see that when performing an EM training there is a check for very small
>> occupancy or weight, and eliminates a Gaussian if it is so. I am though not
>> happy with such an approach and had commented out that line some time ago
>> (I am implementing a MAP adaptation function that needs to deal with these
>> cases)
>>
>> X.
>>
>>
>> On Mon, Apr 8, 2013 at 9:19 PM, Daniel Povey <dp...@gm...> wrote:
>>
>>> Hm, thanks, but I don't think this is the right way to fix the problem.
>>>  Update code should always take into account the possibility that
>>> occupancies will be zero.  It's expected that exp() on very negative values
>>> will produce zero.
>>> Dan
>>>
>>>
>>> On Mon, Apr 8, 2013 at 3:15 PM, Xavier Anguera <xan...@gm...>wrote:
>>>
>>>> Hi Dan,
>>>> the segmentation fault comes from a division by 0 when using the
>>>> occupancy of the Gaussians, that has been computed by adding together all
>>>> the posterior probabilities for each Gaussian and a set of features. When
>>>> all posteriors for a given Gaussian and all features is 0, there is a
>>>> division by 0.
>>>> I am pasting the "hack" I wrote to prevent this. I believe though that
>>>> maybe the exp() function should be revisited. Tell me what you think.
>>>>
>>>> Real VectorBase<Real>::ApplySoftMax() {
>>>>   Real max = this->Max(), sum = 0.0;
>>>>   for (MatrixIndexT i = 0; i < dim_; i++) {
>>>>       data_[i] = exp(data_[i] - max);
>>>>       if(data_[i] < FLT_MIN )
>>>>           data_[i] = FLT_MIN; //very small value
>>>>       sum += data_[i];
>>>>   }
>>>>
>>>>   this->Scale(1.0 / sum);
>>>>   return max + log(sum);
>>>> }
>>>>
>>>>
>>>> On Mon, Apr 8, 2013 at 8:58 PM, Daniel Povey <dp...@gm...> wrote:
>>>>
>>>>> Firstly, this should give you numerical problems but not a
>>>>> segmentation fault.
>>>>> You'll have to look in the code and see if it's behaving as expected.
>>>>>  E.g. is it due to a number so small that it cannot be represented in
>>>>> floating point, or is it larger than that and unexpectedly becoming zero?
>>>>> It might be an issue with your algorithm design.
>>>>> Let me know if that function needs to be fixed.
>>>>>
>>>>> Dan
>>>>>
>>>>>
>>>>> On Mon, Apr 8, 2013 at 2:54 PM, Xavier Anguera <xan...@gm...>wrote:
>>>>>
>>>>>>  Hi,
>>>>>> when using the function template<typename Real> Real
>>>>>> VectorBase<Real>::ApplySoftMax() in kaldi-vector.cc file, I noticed that
>>>>>> very small likelihoods are rounded to a posterior probability of 0.0
>>>>>> Is this an expected behavior? I am trying to perform an EM training
>>>>>> of a simple GMM and I keep bumping into segmentation fault due to this.
>>>>>>
>>>>>> Thanks
>>>>>>
>>>>>> Xavi Anguera
>>>>>>
>>>>>>
>>>>>> ------------------------------------------------------------------------------
>>>>>> Minimize network downtime and maximize team effectiveness.
>>>>>> Reduce network management and security costs.Learn how to hire
>>>>>> the most talented Cisco Certified professionals. Visit the
>>>>>> Employer Resources Portal
>>>>>> http://www.cisco.com/web/learning/employer_resources/index.html
>>>>>> _______________________________________________
>>>>>> Kaldi-developers mailing list
>>>>>> Kal...@li...
>>>>>> https://lists.sourceforge.net/lists/listinfo/kaldi-developers
>>>>>>
>>>>>>
>>>>>
>>>>
>>>
>>
>

Re: [Kaldi-developers] obtaining 0'os when computing posteriorgrams

[Daniel P. <dp...@gm...>]

It's the normal practice when dealing with Gaussians to get rid of small
counts.
Dan



On Mon, Apr 8, 2013 at 3:23 PM, Xavier Anguera <xan...@gm...> wrote:

> I agree my "hack" is not the solution.
> I see that when performing an EM training there is a check for very small
> occupancy or weight, and eliminates a Gaussian if it is so. I am though not
> happy with such an approach and had commented out that line some time ago
> (I am implementing a MAP adaptation function that needs to deal with these
> cases)
>
> X.
>
>
> On Mon, Apr 8, 2013 at 9:19 PM, Daniel Povey <dp...@gm...> wrote:
>
>> Hm, thanks, but I don't think this is the right way to fix the problem.
>>  Update code should always take into account the possibility that
>> occupancies will be zero.  It's expected that exp() on very negative values
>> will produce zero.
>> Dan
>>
>>
>> On Mon, Apr 8, 2013 at 3:15 PM, Xavier Anguera <xan...@gm...>wrote:
>>
>>> Hi Dan,
>>> the segmentation fault comes from a division by 0 when using the
>>> occupancy of the Gaussians, that has been computed by adding together all
>>> the posterior probabilities for each Gaussian and a set of features. When
>>> all posteriors for a given Gaussian and all features is 0, there is a
>>> division by 0.
>>> I am pasting the "hack" I wrote to prevent this. I believe though that
>>> maybe the exp() function should be revisited. Tell me what you think.
>>>
>>> Real VectorBase<Real>::ApplySoftMax() {
>>>   Real max = this->Max(), sum = 0.0;
>>>   for (MatrixIndexT i = 0; i < dim_; i++) {
>>>       data_[i] = exp(data_[i] - max);
>>>       if(data_[i] < FLT_MIN )
>>>           data_[i] = FLT_MIN; //very small value
>>>       sum += data_[i];
>>>   }
>>>
>>>   this->Scale(1.0 / sum);
>>>   return max + log(sum);
>>> }
>>>
>>>
>>> On Mon, Apr 8, 2013 at 8:58 PM, Daniel Povey <dp...@gm...> wrote:
>>>
>>>> Firstly, this should give you numerical problems but not a segmentation
>>>> fault.
>>>> You'll have to look in the code and see if it's behaving as expected.
>>>>  E.g. is it due to a number so small that it cannot be represented in
>>>> floating point, or is it larger than that and unexpectedly becoming zero?
>>>> It might be an issue with your algorithm design.
>>>> Let me know if that function needs to be fixed.
>>>>
>>>> Dan
>>>>
>>>>
>>>> On Mon, Apr 8, 2013 at 2:54 PM, Xavier Anguera <xan...@gm...>wrote:
>>>>
>>>>>  Hi,
>>>>> when using the function template<typename Real> Real
>>>>> VectorBase<Real>::ApplySoftMax() in kaldi-vector.cc file, I noticed that
>>>>> very small likelihoods are rounded to a posterior probability of 0.0
>>>>> Is this an expected behavior? I am trying to perform an EM training of
>>>>> a simple GMM and I keep bumping into segmentation fault due to this.
>>>>>
>>>>> Thanks
>>>>>
>>>>> Xavi Anguera
>>>>>
>>>>>
>>>>> ------------------------------------------------------------------------------
>>>>> Minimize network downtime and maximize team effectiveness.
>>>>> Reduce network management and security costs.Learn how to hire
>>>>> the most talented Cisco Certified professionals. Visit the
>>>>> Employer Resources Portal
>>>>> http://www.cisco.com/web/learning/employer_resources/index.html
>>>>> _______________________________________________
>>>>> Kaldi-developers mailing list
>>>>> Kal...@li...
>>>>> https://lists.sourceforge.net/lists/listinfo/kaldi-developers
>>>>>
>>>>>
>>>>
>>>
>>
>

Re: [Kaldi-developers] obtaining 0'os when computing posteriorgrams

[Xavier A. <xan...@gm...>]

I agree my "hack" is not the solution.
I see that when performing an EM training there is a check for very small
occupancy or weight, and eliminates a Gaussian if it is so. I am though not
happy with such an approach and had commented out that line some time ago
(I am implementing a MAP adaptation function that needs to deal with these
cases)

X.


On Mon, Apr 8, 2013 at 9:19 PM, Daniel Povey <dp...@gm...> wrote:

> Hm, thanks, but I don't think this is the right way to fix the problem.
>  Update code should always take into account the possibility that
> occupancies will be zero.  It's expected that exp() on very negative values
> will produce zero.
> Dan
>
>
> On Mon, Apr 8, 2013 at 3:15 PM, Xavier Anguera <xan...@gm...> wrote:
>
>> Hi Dan,
>> the segmentation fault comes from a division by 0 when using the
>> occupancy of the Gaussians, that has been computed by adding together all
>> the posterior probabilities for each Gaussian and a set of features. When
>> all posteriors for a given Gaussian and all features is 0, there is a
>> division by 0.
>> I am pasting the "hack" I wrote to prevent this. I believe though that
>> maybe the exp() function should be revisited. Tell me what you think.
>>
>> Real VectorBase<Real>::ApplySoftMax() {
>>   Real max = this->Max(), sum = 0.0;
>>   for (MatrixIndexT i = 0; i < dim_; i++) {
>>       data_[i] = exp(data_[i] - max);
>>       if(data_[i] < FLT_MIN )
>>           data_[i] = FLT_MIN; //very small value
>>       sum += data_[i];
>>   }
>>
>>   this->Scale(1.0 / sum);
>>   return max + log(sum);
>> }
>>
>>
>> On Mon, Apr 8, 2013 at 8:58 PM, Daniel Povey <dp...@gm...> wrote:
>>
>>> Firstly, this should give you numerical problems but not a segmentation
>>> fault.
>>> You'll have to look in the code and see if it's behaving as expected.
>>>  E.g. is it due to a number so small that it cannot be represented in
>>> floating point, or is it larger than that and unexpectedly becoming zero?
>>> It might be an issue with your algorithm design.
>>> Let me know if that function needs to be fixed.
>>>
>>> Dan
>>>
>>>
>>> On Mon, Apr 8, 2013 at 2:54 PM, Xavier Anguera <xan...@gm...>wrote:
>>>
>>>>  Hi,
>>>> when using the function template<typename Real> Real
>>>> VectorBase<Real>::ApplySoftMax() in kaldi-vector.cc file, I noticed that
>>>> very small likelihoods are rounded to a posterior probability of 0.0
>>>> Is this an expected behavior? I am trying to perform an EM training of
>>>> a simple GMM and I keep bumping into segmentation fault due to this.
>>>>
>>>> Thanks
>>>>
>>>> Xavi Anguera
>>>>
>>>>
>>>> ------------------------------------------------------------------------------
>>>> Minimize network downtime and maximize team effectiveness.
>>>> Reduce network management and security costs.Learn how to hire
>>>> the most talented Cisco Certified professionals. Visit the
>>>> Employer Resources Portal
>>>> http://www.cisco.com/web/learning/employer_resources/index.html
>>>> _______________________________________________
>>>> Kaldi-developers mailing list
>>>> Kal...@li...
>>>> https://lists.sourceforge.net/lists/listinfo/kaldi-developers
>>>>
>>>>
>>>
>>
>

Re: [Kaldi-developers] obtaining 0'os when computing posteriorgrams

[Daniel P. <dp...@gm...>]

Hm, thanks, but I don't think this is the right way to fix the problem.
 Update code should always take into account the possibility that
occupancies will be zero.  It's expected that exp() on very negative values
will produce zero.
Dan


On Mon, Apr 8, 2013 at 3:15 PM, Xavier Anguera <xan...@gm...> wrote:

> Hi Dan,
> the segmentation fault comes from a division by 0 when using the occupancy
> of the Gaussians, that has been computed by adding together all the
> posterior probabilities for each Gaussian and a set of features. When all
> posteriors for a given Gaussian and all features is 0, there is a division
> by 0.
> I am pasting the "hack" I wrote to prevent this. I believe though that
> maybe the exp() function should be revisited. Tell me what you think.
>
> Real VectorBase<Real>::ApplySoftMax() {
>   Real max = this->Max(), sum = 0.0;
>   for (MatrixIndexT i = 0; i < dim_; i++) {
>       data_[i] = exp(data_[i] - max);
>       if(data_[i] < FLT_MIN )
>           data_[i] = FLT_MIN; //very small value
>       sum += data_[i];
>   }
>
>   this->Scale(1.0 / sum);
>   return max + log(sum);
> }
>
>
> On Mon, Apr 8, 2013 at 8:58 PM, Daniel Povey <dp...@gm...> wrote:
>
>> Firstly, this should give you numerical problems but not a segmentation
>> fault.
>> You'll have to look in the code and see if it's behaving as expected.
>>  E.g. is it due to a number so small that it cannot be represented in
>> floating point, or is it larger than that and unexpectedly becoming zero?
>> It might be an issue with your algorithm design.
>> Let me know if that function needs to be fixed.
>>
>> Dan
>>
>>
>> On Mon, Apr 8, 2013 at 2:54 PM, Xavier Anguera <xan...@gm...>wrote:
>>
>>>  Hi,
>>> when using the function template<typename Real> Real
>>> VectorBase<Real>::ApplySoftMax() in kaldi-vector.cc file, I noticed that
>>> very small likelihoods are rounded to a posterior probability of 0.0
>>> Is this an expected behavior? I am trying to perform an EM training of a
>>> simple GMM and I keep bumping into segmentation fault due to this.
>>>
>>> Thanks
>>>
>>> Xavi Anguera
>>>
>>>
>>> ------------------------------------------------------------------------------
>>> Minimize network downtime and maximize team effectiveness.
>>> Reduce network management and security costs.Learn how to hire
>>> the most talented Cisco Certified professionals. Visit the
>>> Employer Resources Portal
>>> http://www.cisco.com/web/learning/employer_resources/index.html
>>> _______________________________________________
>>> Kaldi-developers mailing list
>>> Kal...@li...
>>> https://lists.sourceforge.net/lists/listinfo/kaldi-developers
>>>
>>>
>>
>

Re: [Kaldi-developers] obtaining 0'os when computing posteriorgrams

[Xavier A. <xan...@gm...>]

Hi Dan,
the segmentation fault comes from a division by 0 when using the occupancy
of the Gaussians, that has been computed by adding together all the
posterior probabilities for each Gaussian and a set of features. When all
posteriors for a given Gaussian and all features is 0, there is a division
by 0.
I am pasting the "hack" I wrote to prevent this. I believe though that
maybe the exp() function should be revisited. Tell me what you think.

Real VectorBase<Real>::ApplySoftMax() {
  Real max = this->Max(), sum = 0.0;
  for (MatrixIndexT i = 0; i < dim_; i++) {
      data_[i] = exp(data_[i] - max);
      if(data_[i] < FLT_MIN )
          data_[i] = FLT_MIN; //very small value
      sum += data_[i];
  }

  this->Scale(1.0 / sum);
  return max + log(sum);
}


On Mon, Apr 8, 2013 at 8:58 PM, Daniel Povey <dp...@gm...> wrote:

> Firstly, this should give you numerical problems but not a segmentation
> fault.
> You'll have to look in the code and see if it's behaving as expected.
>  E.g. is it due to a number so small that it cannot be represented in
> floating point, or is it larger than that and unexpectedly becoming zero?
> It might be an issue with your algorithm design.
> Let me know if that function needs to be fixed.
>
> Dan
>
>
> On Mon, Apr 8, 2013 at 2:54 PM, Xavier Anguera <xan...@gm...> wrote:
>
>> Hi,
>> when using the function template<typename Real> Real
>> VectorBase<Real>::ApplySoftMax() in kaldi-vector.cc file, I noticed that
>> very small likelihoods are rounded to a posterior probability of 0.0
>> Is this an expected behavior? I am trying to perform an EM training of a
>> simple GMM and I keep bumping into segmentation fault due to this.
>>
>> Thanks
>>
>> Xavi Anguera
>>
>>
>> ------------------------------------------------------------------------------
>> Minimize network downtime and maximize team effectiveness.
>> Reduce network management and security costs.Learn how to hire
>> the most talented Cisco Certified professionals. Visit the
>> Employer Resources Portal
>> http://www.cisco.com/web/learning/employer_resources/index.html
>> _______________________________________________
>> Kaldi-developers mailing list
>> Kal...@li...
>> https://lists.sourceforge.net/lists/listinfo/kaldi-developers
>>
>>
>

Re: [Kaldi-developers] obtaining 0'os when computing posteriorgrams

[Daniel P. <dp...@gm...>]

Firstly, this should give you numerical problems but not a segmentation
fault.
You'll have to look in the code and see if it's behaving as expected.  E.g.
is it due to a number so small that it cannot be represented in floating
point, or is it larger than that and unexpectedly becoming zero?
It might be an issue with your algorithm design.
Let me know if that function needs to be fixed.

Dan


On Mon, Apr 8, 2013 at 2:54 PM, Xavier Anguera <xan...@gm...> wrote:

> Hi,
> when using the function template<typename Real> Real
> VectorBase<Real>::ApplySoftMax() in kaldi-vector.cc file, I noticed that
> very small likelihoods are rounded to a posterior probability of 0.0
> Is this an expected behavior? I am trying to perform an EM training of a
> simple GMM and I keep bumping into segmentation fault due to this.
>
> Thanks
>
> Xavi Anguera
>
>
> ------------------------------------------------------------------------------
> Minimize network downtime and maximize team effectiveness.
> Reduce network management and security costs.Learn how to hire
> the most talented Cisco Certified professionals. Visit the
> Employer Resources Portal
> http://www.cisco.com/web/learning/employer_resources/index.html
> _______________________________________________
> Kaldi-developers mailing list
> Kal...@li...
> https://lists.sourceforge.net/lists/listinfo/kaldi-developers
>
>

[Kaldi-developers] obtaining 0'os when computing posteriorgrams

[Xavier A. <xan...@gm...>]

Hi,
when using the function template<typename Real> Real
VectorBase<Real>::ApplySoftMax() in kaldi-vector.cc file, I noticed that
very small likelihoods are rounded to a posterior probability of 0.0
Is this an expected behavior? I am trying to perform an EM training of a
simple GMM and I keep bumping into segmentation fault due to this.

Thanks

Xavi Anguera

Re: [Kaldi-developers] bugfix

[Arnab G. <ar...@gm...>]

Done. Thanks.

On Fri, Apr 5, 2013 at 8:02 PM, Alexander Kain <lx...@gm...> wrote:
> a link (or copy) is required to enable the yesno example format_data.sh
> script to run
>
> devel@VB:~/asr/kaldi-stable/egs/yesno/s3$ ln -s ../../wsj/s3/scripts scripts
>
>
> ------------------------------------------------------------------------------
> Minimize network downtime and maximize team effectiveness.
> Reduce network management and security costs.Learn how to hire
> the most talented Cisco Certified professionals. Visit the
> Employer Resources Portal
> http://www.cisco.com/web/learning/employer_resources/index.html
> _______________________________________________
> Kaldi-developers mailing list
> Kal...@li...
> https://lists.sourceforge.net/lists/listinfo/kaldi-developers

[Kaldi-developers] also

[Alexander K. <lx...@gm...>]

devel@VB:~/asr/kaldi-stable/egs/yesno/s3$ ln -s ../../wsj/s3/steps/ steps

also seems necessary to continue run.sh

[Kaldi-developers] bugfix

[Alexander K. <lx...@gm...>]

a link (or copy) is required to enable the yesno example format_data.sh 
script to run

devel@VB:~/asr/kaldi-stable/egs/yesno/s3$ ln -s ../../wsj/s3/scripts scripts

Re: [Kaldi-developers] Hello

[Arnab G. <ar...@gm...>]

Already added him. -Arnab

On Thu, Mar 21, 2013 at 4:01 PM, Andrew Fandrianto <ft...@cs...> wrote:
> Hi -- I'd like to be added to the kaldi users list
>
> Thanks,
> Andrew
>
> ------------------------------------------------------------------------------
> Everyone hates slow websites. So do we.
> Make your web apps faster with AppDynamics
> Download AppDynamics Lite for free today:
> http://p.sf.net/sfu/appdyn_d2d_mar
> _______________________________________________
> Kaldi-developers mailing list
> Kal...@li...
> https://lists.sourceforge.net/lists/listinfo/kaldi-developers
>

[Kaldi-developers] Hello

[Andrew F. <ft...@cs...>]

Hi -- I'd like to be added to the kaldi users list

Thanks,
Andrew

Re: [Kaldi-developers] Moved Kaldi repository (action required to stay updated)

[Arnab G. <ar...@gm...>]

You may notice that the svn commands now require a password. You can
get password-less access by posting your ssh keys to the link in the
account services page:
https://sourceforge.net/account/services

Apparently the keys are synced every 10mins, but it may take longer.
So you won't get password-less access immediately after posting the
keys.

If what I just said made no sense, you can look at the following on
how to generate ssh keys, etc.
https://sourceforge.net/apps/trac/sourceforge/wiki/SSH%20keys

On Thu, Mar 21, 2013 at 12:51 AM, Daniel Povey <dp...@gm...> wrote:
>
>
>
> Hi all,
>
> This evening something broke on Sourceforge's svn server (error message:
> svn: Can't find a temporary directory: Internal error) and I decided that
> now would be a good time to move Kaldi over to the "new" version of
> Sourceforge's software.  I have now done this, and it works, but because of
> the move, the old repositories are no longer valid.  (I don't know if you
> will still be be able to access them after they fix the server issue, but
> regardless, they would be out of date).
>
> In order to switch your repositories over to the "new" location, please do
> as follows.
>
> If you DO NOT need to commit changes to the repository, change directory to
> the top level of where you checked Kaldi out, and do:
>
>   svn switch --relocate  https://kaldi.svn.sourceforge.net/svnroot/kaldi
> svn://svn.code.sf.net/p/kaldi/code
>
> If you DO need to commit changes at some point, then do as follows, with
> USERNAME replaced with your Sourceforge username:
>
>   svn switch --relocate  https://kaldi.svn.sourceforge.net/svnroot/kaldi
> svn+ssh://USE...@sv.../p/kaldi/code/
>
> Note: if at some point in the future you need to change your repository from
> a non-commit to commit-is-possible kind of repository, you can always do as
> follows:
>   svn switch --relocate  svn://svn.code.sf.net/p/kaldi/code
> svn+ssh://USE...@sv.../p/kaldi/code/
>
> BTW, I will hopefully have time to give a proper update in a few days.   I
> have quite a lot to report but no time right now to prepare something.
>
> Dan
>
>
>
>
> ------------------------------------------------------------------------------
> Everyone hates slow websites. So do we.
> Make your web apps faster with AppDynamics
> Download AppDynamics Lite for free today:
> http://p.sf.net/sfu/appdyn_d2d_mar
> _______________________________________________
> Kaldi-developers mailing list
> Kal...@li...
> https://lists.sourceforge.net/lists/listinfo/kaldi-developers
>

Re: [Kaldi-developers] Word Arc Posterior Probabilities

[Daniel P. <dp...@gm...>]

There is not exactly such a tool because there is no very natural way to
produce the output of this in WFST form, and preserve the normal WFST
semantics.   There are some related tools, for instance, lattice-mbr-decode
does a Minimum Bayes Risk decoding and as a byproduct can create some stats
similar to a confusion networks (or "sausage string") with the probability
of different words at different positions, and associated timing
information.
If you want to work more directly with the posteriors of arcs in the
lattice, you would have to work at the C++ level, and possibly write the
forward-backward routine yourself working with the CompactLattice format--
the existing forward backward routines such as LatticeForwardBackward in
lattice-functions.{h,cc} seem to use the Lattice format which is not
suitable if you want to work at the whole-word level.
Dan

On Thu, Mar 21, 2013 at 12:13 AM, Kartik Audhkhasi <aud...@us...> wrote:

> Hi Dan,
>
> I can compute the transition-id posterior distribution for each arc in a
> Kaldi lattice using lattice-to-post. Is there a similar tool to compute the
> posterior probability of each arc in a word lattice using the
> forward-backward algorithm?
>
> Thanks,
> Kartik
>
> --
> Kartik Audhkhasi
>
>
> ------------------------------------------------------------------------------
> Everyone hates slow websites. So do we.
> Make your web apps faster with AppDynamics
> Download AppDynamics Lite for free today:
> http://p.sf.net/sfu/appdyn_d2d_mar
> _______________________________________________
> Kaldi-developers mailing list
> Kal...@li...
> https://lists.sourceforge.net/lists/listinfo/kaldi-developers
>
>

[Kaldi-developers] Word Arc Posterior Probabilities

[Kartik A. <aud...@us...>]

Hi Dan,

I can compute the transition-id posterior distribution for each arc in a
Kaldi lattice using lattice-to-post. Is there a similar tool to compute the
posterior probability of each arc in a word lattice using the
forward-backward algorithm?

Thanks,
Kartik

-- 
Kartik Audhkhasi

[Kaldi-developers] Moved Kaldi repository (action required to stay updated)

[Daniel P. <dp...@gm...>]

Hi all,

This evening something broke on Sourceforge's svn server (error message:
svn: Can't find a temporary directory: Internal error) and I decided that
now would be a good time to move Kaldi over to the "new" version of
Sourceforge's software.  I have now done this, and it works, but because of
the move, the old repositories are no longer valid.  (I don't know if you
will still be be able to access them after they fix the server issue, but
regardless, they would be out of date).

In order to switch your repositories over to the "new" location, please do
as follows.

If you DO NOT need to commit changes to the repository, change directory to
the top level of where you checked Kaldi out, and do:

  svn switch --relocate  https://kaldi.svn.sourceforge.net/svnroot/kaldisvn://
svn.code.sf.net/p/kaldi/code

If you DO need to commit changes at some point, then do as follows, with
USERNAME replaced with your Sourceforge username:

  svn switch --relocate
https://kaldi.svn.sourceforge.net/svnroot/kaldisvn+ssh://
USE...@sv.../p/kaldi/code/

Note: if at some point in the future you need to change your repository
from a non-commit to commit-is-possible kind of repository, you can always
do as follows:
  svn switch --relocate  svn://svn.code.sf.net/p/kaldi/code svn+ssh://
USE...@sv.../p/kaldi/code/

BTW, I will hopefully have time to give a proper update in a few days.   I
have quite a lot to report but no time right now to prepare something.

Dan

Re: [Kaldi-developers] ask for help about compiling the kaldi_lm package in the cygwin

[Daniel P. <dp...@gm...>]

It is probably some kind of windows-related bug.
I recently clarified in the Windows README and in the online documentation
that Kaldi does not really work on Windows, i.e. it has not really been
properly tested.  Sorry.  If you can figure out the bug and fix it that
would be great.  Probably something to do with newline translation.
Dan

On Wed, Mar 20, 2013 at 12:44 PM, 牛铜 <niu...@gm...> wrote:

> this is the error when i run the timit_format_data.sh.
>
>
> 2013/3/21 牛铜 <niu...@gm...>
>
>> Thank you for your attention. I have solved the problem by just change
>> the default g++ compiler version in the cygwin. It does work with the g++
>> 4,5,3.
>> Thank you all the same.
>> But when i run the example in the eggs/timit/s3, I have met a new
>> problem. It is excute the /timit_format_data.sh file, it has the
>> information below the picture. Will this affect the following steps?
>> Niu Tong
>>
>>
>>
>> 2013/3/20 Daniel Povey <dp...@gm...>
>>
>>> Replied separately-- but try g++ version 4.2 or so, those are either too
>>> old or too new.
>>> Dan
>>>
>>> On Wed, Mar 20, 2013 at 5:02 AM, 牛铜 <niu...@gm...> wrote:
>>>
>>>> Hi, recently I download the kaldi package and compiled it with vs2010.
>>>> When I run the examples in the eggs/timit/s3 floader, i met one problem. As
>>>> I don't have the kaldi_lm package, I download it from the web url
>>>> http://merlin.fit.vutbr.cz/kaldi/kaldi_lm.tar.gz . But when I compiled
>>>> it in the cygwin, I can't make success. Fisrt, I use the default g++ which
>>>> is versioned 3.4, it has 'can't open the file tr1/unordered_map'. So I
>>>> asigned the g++ version with 4.5.3, make it again, still has other problem
>>>> such as ' in the function _Z12P\process_linePci, get_raw_ngrams.cc:33:
>>>> undefine __cxa_guard_acquire'.
>>>> Is there any compiled exe in the windows version or any solution to
>>>> build it in the cygwin?
>>>> Thanks for check this email.
>>>>
>>>>
>>>> ------------------------------------------------------------------------------
>>>> Everyone hates slow websites. So do we.
>>>> Make your web apps faster with AppDynamics
>>>> Download AppDynamics Lite for free today:
>>>> http://p.sf.net/sfu/appdyn_d2d_mar
>>>> _______________________________________________
>>>> Kaldi-developers mailing list
>>>> Kal...@li...
>>>> https://lists.sourceforge.net/lists/listinfo/kaldi-developers
>>>>
>>>>
>>>
>>
>

Re: [Kaldi-developers] ask for help about compiling the kaldi_lm package in the cygwin

[Daniel P. <dp...@gm...>]

Replied separately-- but try g++ version 4.2 or so, those are either too
old or too new.
Dan

On Wed, Mar 20, 2013 at 5:02 AM, 牛铜 <niu...@gm...> wrote:

> Hi, recently I download the kaldi package and compiled it with vs2010.
> When I run the examples in the eggs/timit/s3 floader, i met one problem. As
> I don't have the kaldi_lm package, I download it from the web url
> http://merlin.fit.vutbr.cz/kaldi/kaldi_lm.tar.gz . But when I compiled it
> in the cygwin, I can't make success. Fisrt, I use the default g++ which is
> versioned 3.4, it has 'can't open the file tr1/unordered_map'. So I asigned
> the g++ version with 4.5.3, make it again, still has other problem such as
> ' in the function _Z12P\process_linePci, get_raw_ngrams.cc:33: undefine
> __cxa_guard_acquire'.
> Is there any compiled exe in the windows version or any solution to build
> it in the cygwin?
> Thanks for check this email.
>
>
> ------------------------------------------------------------------------------
> Everyone hates slow websites. So do we.
> Make your web apps faster with AppDynamics
> Download AppDynamics Lite for free today:
> http://p.sf.net/sfu/appdyn_d2d_mar
> _______________________________________________
> Kaldi-developers mailing list
> Kal...@li...
> https://lists.sourceforge.net/lists/listinfo/kaldi-developers
>
>

Re: [Kaldi-developers] ask for help about compiling the kaldi_lm package in the cygwin

[Daniel P. <dp...@gm...>]

Try an intermediate g++ such as 4.2.
Dan

On Wed, Mar 20, 2013 at 4:55 AM, jerry_newton <jer...@16...> wrote:

> Hi, recently I download the kaldi package and compiled it with vs2010.
> When I run the examples in the eggs/timit/s3 floader, i met one problem. As
> I don't have the kaldi_lm package, I download it from the web url
> http://merlin.fit.vutbr.cz/kaldi/kaldi_lm.tar.gz . But when I compiled it
> in the cygwin, I can't make success. Fisrt, I use the default g++ which is
> versioned 3.4, it has 'can't open the file tr1/unordered_map'. So I asigned
> the g++ version with 4.5.3, make it again, still has other problem such as
> ' in the function _Z12P\process_linePci, get_raw_ngrams.cc:33: undefine
> __cxa_guard_acquire'.
> Is there any compiled exe in the windows version or any solution to build
> it in the cygwin?
> Thanks for check this email.
>
>
>
>
> ------------------------------------------------------------------------------
> Everyone hates slow websites. So do we.
> Make your web apps faster with AppDynamics
> Download AppDynamics Lite for free today:
> http://p.sf.net/sfu/appdyn_d2d_mar
> _______________________________________________
> Kaldi-developers mailing list
> Kal...@li...
> https://lists.sourceforge.net/lists/listinfo/kaldi-developers
>
>

[Kaldi-developers] ask for help about compiling the kaldi_lm package in the cygwin

[牛铜 <niu...@gm...>]

Hi, recently I download the kaldi package and compiled it with vs2010. When
I run the examples in the eggs/timit/s3 floader, i met one problem. As I
don't have the kaldi_lm package, I download it from the web url
http://merlin.fit.vutbr.cz/kaldi/kaldi_lm.tar.gz . But when I compiled it
in the cygwin, I can't make success. Fisrt, I use the default g++ which is
versioned 3.4, it has 'can't open the file tr1/unordered_map'. So I asigned
the g++ version with 4.5.3, make it again, still has other problem such as
' in the function _Z12P\process_linePci, get_raw_ngrams.cc:33: undefine
__cxa_guard_acquire'.
Is there any compiled exe in the windows version or any solution to build
it in the cygwin?
Thanks for check this email.

[Kaldi-developers] ask for help about compiling the kaldi_lm package in the cygwin

[jerry_newton <jer...@16...>]

Hi, recently I download the kaldi package and compiled it with vs2010. When I run the examples in the eggs/timit/s3 floader, i met one problem. As I don't have the kaldi_lm package, I download it from the web url http://merlin.fit.vutbr.cz/kaldi/kaldi_lm.tar.gz . But when I compiled it in the cygwin, I can't make success. Fisrt, I use the default g++ which is versioned 3.4, it has 'can't open the file tr1/unordered_map'. So I asigned the g++ version with 4.5.3, make it again, still has other problem such as ' in the function _Z12P\process_linePci, get_raw_ngrams.cc:33: undefine __cxa_guard_acquire'.
Is there any compiled exe in the windows version or any solution to build it in the cygwin?
Thanks for check this email.

Re: [Kaldi-developers] Times in Kaldi Lattices

[Daniel P. <dp...@gm...>]

No, all it does is it ensures that on each arc, the sequence of
transition-ids are the ones that "belong" to the word that is on that arc.
 By default they are asynchronized.


On Mon, Mar 11, 2013 at 12:02 PM, Kartik Audhkhasi <aud...@us...> wrote:

> Thanks. I understand the OpenFST format but was under the impression that
> lattice-align-words will replace FST state IDs with the corresponding frame
> numbers.
>
>
> On Mon, Mar 11, 2013 at 8:59 AM, Daniel Povey <dp...@gm...> wrote:
>
>> cc-ing kaldi-developers so it's archived, in case anyone has similar
>> questions.
>>
>> You need to understand the OpenFst format.  This is the "acceptor"
>> version of the format, which is (start-state end-state symbol weight), and
>> the weight is (graph cost, acoustic cost, sequence of transition-ids).
>>  [the "normal", non-acceptor format is start-state end-state input-symbol
>> output-symbol weight].  The time can be obtained by summing up the number
>> of transition-ids starting from the beginning of the lattice; in the code
>> it's LatticeStateTimes.
>>
>> Note that the state numbers are arbitrary in a sense, they contain no
>> real information.  See www.openfst.org for more info on WFSTs.
>>
>>
>> On Mon, Mar 11, 2013 at 11:55 AM, Kartik Audhkhasi <aud...@us...>wrote:
>>
>>> Thanks Dan. I have started using the new scripts. However the timing
>>> issue still remains. I used lattice-align-words to get times on the lattice
>>> nodes in the same way as is demonstrated in the run.sh script. I think I am
>>> not interpreting the times correctly. Do the IDs on both start and end
>>> nodes represent frame numbers? E.g. the first line of my lattice is:
>>>
>>> 0       2337    44870
>>> 17.0063,2553.48,9468_9482_9492_9491_9491_9491_9491_9910_9909_9909_9916_9924_9194_9210_9242_9241_9241_9241
>>>
>>> Does this says that word 44870 goes from frame 0 to frame 2237? The
>>> transition ID sequence however shows only 18 frames.
>>>
>>>
>>> On Sun, Mar 10, 2013 at 9:26 AM, Daniel Povey <dp...@gm...> wrote:
>>>
>>>> And RE how to debug it-- before and after lattice-align-words, you
>>>> could run something like lattice-to-post; this program will crash if there
>>>> are inconsistent times in the lattice, i.e. the lengths of input-symbol
>>>> sequences are not all the same.  I suspect you actually mixed something up.
>>>> Dan
>>>>
>>>>
>>>> On Sun, Mar 10, 2013 at 12:24 PM, Daniel Povey <dp...@gm...>wrote:
>>>>
>>>>> It's a shame that you're using the older versions of the script.
>>>>>  Currently the "s5" scripts are the canonical ones.  Your issue with times
>>>>> greater than the length of the file is very unexpected.  This is not the
>>>>> kind of error I would expect to ever arise.
>>>>> RE getting the N-best or 1-best sequences-- the programs lattice-nbest
>>>>> and lattice-1best are relevant here; they output stuff in the regular
>>>>> lattice format, and you can then put them through lattice-word-align (old
>>>>> scripts) or lattice-align-words (new scripts), and convert the output to,
>>>>> say, ctm format-- you can check the scripts for how to convert to ctm
>>>>> format, it's something like lattice-to-ctm, but there are scripts such as
>>>>> get_ctm.sh and get_train_ctm.sh in s5.
>>>>> Dan
>>>>>
>>>>>
>>>>>
>>>>> On Sat, Mar 9, 2013 at 8:41 PM, Kartik Audhkhasi <aud...@us...>wrote:
>>>>>
>>>>>> Hello,
>>>>>>
>>>>>> I am working with Kaldi lattices and used the walign_lats.sh script
>>>>>> to get times on nodes. I am using position dependent phones and believe
>>>>>> that all required files are in place. However, I see that some frame
>>>>>> numbers (in units of 10ms) exceed the total length of the file. Do you have
>>>>>> any suggestions as to what could be going wrong?
>>>>>>
>>>>>> Also: Is there an easy way to get the N-best or 1-best sequences with
>>>>>> word boundaries?
>>>>>>
>>>>>> Thanks,
>>>>>> Kartik
>>>>>>
>>>>>> --
>>>>>> Kartik Audhkhasi
>>>>>>
>>>>>>
>>>>>> ------------------------------------------------------------------------------
>>>>>> Symantec Endpoint Protection 12 positioned as A LEADER in The
>>>>>> Forrester
>>>>>> Wave(TM): Endpoint Security, Q1 2013 and "remains a good choice" in
>>>>>> the
>>>>>> endpoint security space. For insight on selecting the right partner to
>>>>>> tackle endpoint security challenges, access the full report.
>>>>>> http://p.sf.net/sfu/symantec-dev2dev
>>>>>> _______________________________________________
>>>>>> Kaldi-developers mailing list
>>>>>> Kal...@li...
>>>>>> https://lists.sourceforge.net/lists/listinfo/kaldi-developers
>>>>>>
>>>>>>
>>>>>
>>>>
>>>
>>>
>>> --
>>> Kartik Audhkhasi
>>>
>>
>>
>
>
> --
> Kartik Audhkhasi
>