[hts-users:04332] Re: A patch to HTS from Google

[Tóth Bálint Pál <toth.b@xxxxxxxxxxx>]

Dear Heiga,

Thank you for your answer. If I revert the patch than everything works fine.
Is that maybe possible that after the patch the Training.pl don’t use the new ‘-r’ parameter and in some if-else branch of the patched modules not the parallelized code is running (and that branch might have some mistakes)?

Thanks,
Balint

11/20/2015 14:58 keltezéssel, Heiga ZEN (Byung Ha CHUN) írta:

Interesting.

<STREAMINFO> 5 1 1 1 1 1
<MSDINFO> 5 0 0 0 0 0 0
<VECSIZE> 5 <NULLD><DIAGC><USER>

There are 6 zeros after 5 in <MSDINFO> line.

Heiga

On Fri, Nov 20, 2015 at 1:17 PM Tóth Bálint Pál <toth.b@xxxxxxxxxxx> wrote:

Dear Heiga,

Thank you very much for supporting the community. Should scripts\Training.pl be modified to make the patched modules run in parallel mode? (eg. adding –r parameter)
I applied the patch to HTS2.3beta and tried to train HTS-demo_CMU-ARCTIC-SLT. In the “making a monophone mmf” step the training exits with the following error (I replaced the full path with …):

HMM Def Error: GetToken: Symbol expected at line 4/col 0/char 51 in /.../HTS-demo_CMU-ARCTIC-SLT/models/qst001/ver1/dur/HRest/b

  ERROR [+7050]  HMError:

HMM Def Error: GetOption: GetToken failed at line 4/col 1/char 52 in /.../HTS-demo_CMU-ARCTIC-SLT/models/qst001/ver1/dur/HRest/b

  ERROR [+7050]  HMError:

HMM Def Error: GetOptions: GetOption failed at line 4/col 2/char 53 in /.../HTS-demo_CMU-ARCTIC-SLT/models/qst001/ver1/dur/HRest/b

  ERROR [+7050]  HMError:

HMM Def Error: LoadHMMSet: GetOptions failed at line 1/col 0/char -1 in /.../HTS-demo_CMU-ARCTIC-SLT/models/qst001/ver1/dur/HRest/b

  ERROR [+7050]  HMError:

  ERROR [+2628]  Initialise: LoadHMMSet failed

 FATAL ERROR - Terminating program /usr/local/HTS-2.3beta/bin/HHEd

Error in /usr/local/HTS-2.3beta/bin/HHEd      -A -B -C /.../HTS-demo_CMU-ARCTIC-SLT/configs/qst001/ver1/trn.cnf -D -T 1 -p -i  -d /.../HTS-demo_CMU-ARCTIC-SLT/models/qst001/ver1/dur/HRest -w /.../HTS-demo_CMU-ARCTIC-SLT/models/qst001/ver1/dur/monophone.mmf /.../HTS-demo_CMU-ARCTIC-SLT/edfiles/qst001/ver1/dur/lvf.hed /.../HTS-demo_CMU-ARCTIC-SLT/data/lists/mono.list

The content of models/qst001/ver1/dur/HRest/b is the following:

 ~o
<STREAMINFO> 5 1 1 1 1 1
<MSDINFO> 5 0 0 0 0 0 0
<VECSIZE> 5 <NULLD><DIAGC><USER>
~h "b"
<BEGINHMM>
<NUMSTATES> 3
<STATE> 2
<STREAM> 1
<MEAN> 1
 3.066904e+00
<VARIANCE> 1
 4.757239e+00
<STREAM> 2
<MEAN> 1
 4.053977e+00
<VARIANCE> 1
 8.995836e+00
<STREAM> 3
<MEAN> 1
 2.597377e+00
<VARIANCE> 1
 9.469593e+00
<STREAM> 4
<MEAN> 1
 1.809504e+00
<VARIANCE> 1
 4.504592e-01
<STREAM> 5
<MEAN> 1
 2.494573e+00
<VARIANCE> 1
 1.228845e+00
<TRANSP> 3
0 1 0
0 0 1
0 0 0
<ENDHMM>

Do you have an idea what can cause the problem?

Thanks,
Balint

11/9/2015 10:46 keltezéssel, Heiga ZEN (Byung Ha CHUN) írta:

Hi all,

At Google I've made minor changes to HTS to make it more scalable on large data and added some low-level tuning to make it faster.

Now I'm happy to announce that Google allows me to donate these changes to the HTS community.  Please find the attached patch code.  It is not big; it just adds the following functionalities.

• HERest-style parallel training to HCompV, HInit, and HRest.
• Reduce the size of dumped statistics from HERest.
• Low-level tunings in HModel, HTrain, HInit, and HRest.

The patch code should be applied to HTS-2.3beta (not HTS-2.2 stable version).  Please run the following command at your HTS root directory.

$ patch -p1 -d . < HTS-2.3beta_Google_patch.txt

This patch code was made manually from the internal version of HTS to make it independent from other internal libraries.  I did some tests after creating this patch, but it may not work on your environment.  Please feel free to post problems/questions about it to the hts-users ML (please don't send these questions to my e-mail address directly).

Cheers,

Heiga