Re: [Kaldi-developers] Errors when training mono-state NN in SWBD recipe

SourceForge Headquarters 1320 Columbia Street Suite 310 San Diego, CA 92101 +1 (858) 422-6466

Dear Dan,
thank you for your help.
Next are the tests you asked me to perform:

Running utils/validate_lang.pl data/lang_test_phn-mono/ gives:

Checking data/lang_test_phn-mono//phones/roots.{txt, int} ...
--> 30 entry/entries in data/lang_test_phn-mono//phones/roots.txt
--> data/lang_test_phn-mono//phones/roots.int corresponds to
data/lang_test_phn-mono//phones/roots.txt
--> data/lang_test_phn-mono//phones/roots.{txt, int} are OK

Checking data/lang_test_phn-mono//phones/sets.{txt, int} ...
--> 30 entry/entries in data/lang_test_phn-mono//phones/sets.txt
--> data/lang_test_phn-mono//phones/sets.int corresponds to
data/lang_test_phn-mono//phones/sets.txt
--> data/lang_test_phn-mono//phones/sets.{txt, int} are OK

Checking data/lang_test_phn-mono//phones/extra_questions.{txt, int} ...
--> 9 entry/entries in data/lang_test_phn-mono//phones/extra_questions.txt
--> data/lang_test_phn-mono//phones/extra_questions.int corresponds to
data/lang_test_phn-mono//phones/extra_questions.txt
--> data/lang_test_phn-mono//phones/extra_questions.{txt, int} are OK

Checking disjoint: silence.txt, nosilenct.txt, disambig.txt ...
--> silence.txt and nonsilence.txt are disjoint
--> silence.txt and disambig.txt are disjoint
--> disambig.txt and nonsilence.txt are disjoint
--> disjoint property is OK

Checking sumation: silence.txt, nonsilence.txt, disambig.txt ...
--> ERROR: data/lang_test_phn-mono//phones/disambig.txt is empty or not
exists

Checking optional_silence.txt ...
--> reading data/lang_test_phn-mono//phones/optional_silence.txt
--> data/lang_test_phn-mono//phones/optional_silence.txt is OK

Checking disambiguation symbols: #0 and #1
--> ERROR: data/lang_test_phn-mono//phones/disambig.txt is empty or not
exists
--> ERROR: data/lang_test_phn-mono//phones/disambig.txt doesn't have "#0"
or "#1"
Checking topo ...
--> data/lang_test_phn-mono//topo's nonsilence section is OK
--> data/lang_test_phn-mono//topo's silence section is OK
--> data/lang_test_phn-mono//topo is OK

Checking data/lang_test_phn-mono//oov.{txt, int} ...
--> ERROR: fail to open data/lang_test_phn-mono//oov.txt

--> ERROR

Apparently I do not have either oov.txt nore disambig.txt
Probably the test data I am using does not have any OOV in it. I can add it
artificially, but I guess this is not the main problem here... regarding
the disambig.txt file, what should it contain?

I did run gdb as you indicated (thank you for such detailed info) and gives
me:
(gdb) p subsequential_symbol
$1 = 97
(gdb) p  disambig_syms_.count(subsequential_symbol)
$2 = 0
(gdb) p phone_syms_.count(subsequential_symbol)
$3 = 1
(gdb) p phone_syms_.size()
$4 = 78
(gdb) p disambig_syms_.size()
$5 = 0

Finally, the contents of cat data/lang_test_phn-mono/phones/disambig.int is
also empty.

Thanks again for your help!

yours,

Xavier Anguera

On Fri, Dec 27, 2013 at 10:26 PM, Daniel Povey <dp...@gm...> wrote:

> Could you please do the following. [apologies if you already know gdb]
>
> First do utils/validate_lang.pl data/lang_test_phn-mono/
> and let me know if it fails.
> If it doesn't fail, do:
>
> gdb --args fstcomposecontext --context-size=1 --central-position=0
> --read-disambig-syms=data/lang_test_phn-mono/phones/disambig.int
> --write-disambig-syms=data/lang_test_
> phn-mono/tmp/disambig_ilabels_1_0.int data/lang_test_phn-mono/tmp/ilabels_1_0
> data/lang_test_phn-mono/tmp/LG.fst
>
> (gdb) r
> # wait till it crashes
> # go up the stack by typing "up" until you get to the right frame; type
> "down" if you go too far
>
> (gdb) p subsequential_symbol
> (gdb) p  disambig_syms_.count(subsequential_symbol)
> (gdb) p phone_syms_.count(subsequential_symbol)
> (gdb) p phone_syms_.size()
> (gdb) p disambig_syms_.size()
> (gdb) quit
>
> [I hope this works; sometimes it will fail because functions are inlined].
> Anyway, send the output, and also
> cat data/lang_test_phn-mono/phones/disambig.int
> and show me that output too.
>
> Dan
>
>
>
>
> On Fri, Dec 27, 2013 at 10:23 AM, Xavier Anguera <xan...@gm...>wrote:
>
>> Dear all,
>> I am encounering a problem when training mono-state NN using a recipe
>> adapted from the SWBD S5 recipe. I am able to train, decode and phone-align
>> a GMM system, but when I use these results to train the NN I get the
>> following error (see below). I have used this recipe in the past to
>> successfully train one ASR system and now the only difference is that I am
>> trying to train a similar system using graphemes are phonemes (for which I
>> have assigned the graphemes of the words as transcriptions to each word).
>> Any help is appreciated.
>>
>> This is the beginning of the file exp/mono/graph/mkgraph_phn.log:
>>
>> # utils/mkgraph.sh --mono data/lang_test_phn-mono exp/mono
>> exp/mono/graph_phn
>> # Started at Fri Dec 27 18:57:19 CET 2013
>> #
>> fsttablecompose data/lang_test_phn-mono/L_disambig.fst
>> data/lang_test_phn-mono/G.fst
>> fstdeterminizestar --use-log=true
>> fstminimizeencoded
>> fstisstochastic data/lang_test_phn-mono/tmp/LG.fst
>> 0.000358155 -0.000356635
>> fstcomposecontext --context-size=1 --central-position=0
>> --read-disambig-syms=data/lang_test_phn-mono/phones/disambig.int--write-disambig-syms=data/lang_test_
>> phn-mono/tmp/disambig_ilabels_1_0.intdata/lang_test_phn-mono/tmp/ilabels_1_0
>> fstcomposecontext: ../fstext/context-fst-inl.h:105:
>> fst::ContextFstImpl<Arc, LabelT>::ContextFstImpl(typename Arc::Label, const
>> std::vector<B, std::allocator<
>> _T2> >&, const std::vector<B, std::allocator<_T2> >&, int, int) [with Arc
>> = fst::ArcTpl<fst::TropicalWeightTpl<float> >, LabelT = int]: Assertion
>> `subsequenti
>> al_symbol != 0 && disambig_syms_.count(subsequential_symbol) == 0 &&
>> phone_syms_.count(subsequential_symbol) == 0' failed.
>> utils/mkgraph.sh: line 76:  6263 Aborted
>> fstcomposecontext --context-size=$N --central-position=$P
>> --read-disambig-syms=$lang/phones/disambig.
>> int --write-disambig-syms=$lang/tmp/disambig_ilabels_${N}_${P}.int
>> $lang/tmp/ilabels_${N}_${P} < $lang/tmp/LG.fst > $clg
>> fstisstochastic data/lang_test_phn-mono/tmp/CLG_1_0.fst
>> ERROR: FstHeader::Read: Bad FST header:
>> data/lang_test_phn-mono/tmp/CLG_1_0.fst
>> ERROR (fstisstochastic:ReadFstKaldi():fstext/fstext-utils-inl.h:1183)
>> Reading FST: error reading FST header from
>> data/lang_test_phn-mono/tmp/CLG_1_0.fst
>> ERROR (fstisstochastic:ReadFstKaldi():fstext/fstext-utils-inl.h:1183)
>> Reading FST: error reading FST header from
>> data/lang_test_phn-mono/tmp/CLG_1_0.fst
>>
>>
>>
>> ------------------------------------------------------------------------------
>> Rapidly troubleshoot problems before they affect your business. Most IT
>> organizations don't have a clear picture of how application performance
>> affects their revenue. With AppDynamics, you get 100% visibility into your
>> Java,.NET, & PHP application. Start your 15-day FREE TRIAL of AppDynamics
>> Pro!
>>
>> http://pubads.g.doubleclick.net/gampad/clk?id=84349831&iu=/4140/ostg.clktrk
>> _______________________________________________
>> Kaldi-developers mailing list
>> Kal...@li...
>> https://lists.sourceforge.net/lists/listinfo/kaldi-developers
>>
>>
>