[Subject Prev][Subject Next][Thread Prev][Thread Next][Date Index][Thread Index]

[hts-users:02873] Re: Error in Adaptation demo in HTS-2.2 beta


Hello,
I have re-started the adapt script with a lot less context features, just 4 context features, and it finished without problem.
Before I was using all the features available in MARY (this was not a problem on the speaker independent demo) and I got the problem described below.

Now I still do not understand what the problem was when I used a lot more context features, if somebody have an idea, or can point me where to find more information, would be very much appreciated!

Now, my cmp/regTrees/dec.base contains:
~b "dec.base"
<MMFIDMASK> *
<PARAMETERS> MIXBASE
<STREAMINFO>  5 105 1 1 1 15
<NUMCLASSES> 33713

The number of models in full_all.list is 31224
and the number of models in full.list is 31348

So it seems that the base class is not actually a subset of full_all.list, right? then why is it call base class?

Regards,
Marcela.
 



On 05/24/2011 07:15 AM, 那兴宇 wrote:
The missing models maybe from the list-all, and the base class should be a subset of the full-all.list other than full.list. After makeing unseen models, this problem should be solved.

At 2011-05-24 00:00:53,"Marcela Charfuelan" <Marcela.Charfuelan@xxxxxxx> wrote:
Hello,
I would like to ask how this problem was solved, I can not figure out, I am training the HTS-demo_CMU-ARCTIC-ADAPT scripts of
HTS-2.2 beta and using MARY features.

In my case I have:
  ERROR [+999]  CheckBaseClass: Components missing from Base Class list (31628 42094)
  ERROR [+999]  BaseClass check failed

the cmp/regTrees/dec.base contains:

~b "dec.base"
<MMFIDMASK> *
<PARAMETERS> MIXBASE
<STREAMINFO>  5 105 1 1 1 15
<NUMCLASSES> 40994

(15 comes from voicing strengths I am using)

the number of models (full list) is:
HHEd
 50252/50252 Models Loaded [7 states max, 2 mixes max]

Some questions:
- the base class should be a sub-set of the full list of models right?
- the missing models from the base class list are models of the full list that are not found in the base class list?

Thanks in advance for any help....

Marcela.






On 08/30/2010 04:01 PM, 那兴宇 wrote:
Hi , Heiga

Trying to run the adaptation demo, encounter the error below:
  ERROR [+999]  CheckBaseClass: Components missing from Base Class list (15202 15252)
  ERROR [+999]  BaseClass check failed

Then I tracked into the training script and found 'reg.base' was used in this step, below is part of reg.base:
======================
~b "reg.base"
<MMFIDMASK> *
<PARAMETERS> MIXBASE
<STREAMINFO>  5 120 1 1 1 15
<NUMCLASSES> 130
<CLASS> 1 {..........
=======================

Then I used HError to output the total class number, and got this:
  ERROR [+999]  CheckBaseClass: Components missing from Base Class list (15202 15252 130)
  ERROR [+999]  BaseClass check failed

It seems that the list was complete, and the number of mix[*] in 'reg.base' was actually 15252.
Then I recheck the following code in HModel.c:
======================================================
  /* ensure that each component is assigned to a base class */
  for (b=1;b<=bclass->numClasses;b++) {
    for (i=bclass->ilist[b]; i!=NULL; i=i->next) {
      mp = ((MixtureElem *)i->item)->mpdf;
      if (mp->mIdx>0) {
        mp->mIdx = -mp->mIdx;
        ncomp++;
      } else { /* item appears in list multiple times */
        HRError(999,"CheckBaseClass: Component specified multiple times");
        return(FAIL);
      }
    }
  }
=============================================================
If the numClass euqals to the <NUMCLASSES> in 'reg.base', then the error wouldnt happens. Then how could this happen?
Thanks in advance!



您想拥有和网易免费邮箱一样强大的软件吗?


-- 
_______________________________________________________________
 Marcela Charfuelan, Researcher, DFKI GmbH
 Projektbuero Berlin, Alt-Moabit 91c, 10559 Berlin, Germany
 Phone: +49 (0)30 39491821
 URL  : http://www.dfki.de/~charfuel/
_______________________________________________________________
 Deutsches Forschungszentrum fuer Kuenstliche Intelligenz GmbH
 Firmensitz: Trippstadter Strasse 122, D-67663 Kaiserslautern
 Geschaeftsfuehrung:
 Prof. Dr. Dr. h.c. mult. Wolfgang Wahlster (Vorsitzender)
 Dr. Walter Olthoff 
 Vorsitzender des Aufsichtsrats:
 Prof. Dr. h.c. Hans A. Aukes
 Amtsgericht Kaiserslautern, HRB 2313
_______________________________________________________________




-- 
_______________________________________________________________
 Marcela Charfuelan, Researcher, DFKI GmbH
 Projektbuero Berlin, Alt-Moabit 91c, 10559 Berlin, Germany
 Phone: +49 (0)30 39491821
 URL  : http://www.dfki.de/~charfuel/
_______________________________________________________________
 Deutsches Forschungszentrum fuer Kuenstliche Intelligenz GmbH
 Firmensitz: Trippstadter Strasse 122, D-67663 Kaiserslautern
 Geschaeftsfuehrung:
 Prof. Dr. Dr. h.c. mult. Wolfgang Wahlster (Vorsitzender)
 Dr. Walter Olthoff 
 Vorsitzender des Aufsichtsrats:
 Prof. Dr. h.c. Hans A. Aukes
 Amtsgericht Kaiserslautern, HRB 2313
_______________________________________________________________

References
[hts-users:02859] Re: Error in Adaptation demo in HTS-2.2 beta, Marcela Charfuelan
[hts-users:02860] Re: Error in Adaptation demo in HTS-2.2 beta, 那兴宇