Menu

kmeans- aborting

Help
Anonymous
2002-05-24
2012-09-22
  • Anonymous

    Anonymous - 2002-05-24

    Hi,
    This is the log I got after running kmeans

    INFO: kmeans.c(159): km iter [65] 1.414042e-03 ...
    INFO: kmeans.c(159): km iter [66] 1.195984e-03 ...
    INFO: kmeans.c(159): km iter [67] 1.026603e-03 ...
    INFO: kmeans.c(159): km iter [68] 1.061375e-03 ...
    INFO: kmeans.c(172): kmtrineq n_iter 69 sqerr 1.297945e+04 conv_ratio 8.719380e-04
    INFO: main.c(589):     best-so-far sqerr = 1.297945e+04
    INFO: main.c(240): alloc'ing 4Mb obs buf
    INFO: main.c(544): Initializing means using random k-means
    INFO: main.c(547): Trial 0: 256 means
    INFO: kmeans.c(159): km iter [0] 1.000000e+00 ...
    WARNING: "kmeans.c", line 437: Empty cluster 182
    INFO: main.c(580):     -> Aborting k-means, bad initialization
    INFO: kmeans.c(159): km iter [0] 1.000000e+00 ...
    INFO: kmeans.c(159): km iter [1] 1.165755e-01 ...
    INFO: kmeans.c(159): km iter [2] 3.344532e-02 ...

    The parameters I used were

    /Speech/speech/SphinxTrain/robot/bin/kmeans_init \ -gthobj single \ -stride 1 \ -ntrial 1 \ -minratio 0.001 \ -ndensity 256 \ -meanfn /Speech/speech/SphinxTrain/robot/model_parameters/robot.ci_semi_flatinitial/means \ -varfn /Speech/speech/SphinxTrain/robot/model_parameters/robot.ci_semi_flatinitial/variances \ -reest no \ -segdmpdirs /Speech/speech/SphinxTrain/robot/bwaccumdir/robot_buff_1 \ -segdmpfn /Speech/speech/SphinxTrain/robot/bwaccumdir/robot_buff_1/robot.dmp \ -ceplen 13 \ -feat c/1..L-1/,d/1..L-1/,c/0/d/0/dd/0/,dd/1..L-1/ \ -agc none \ -cmn current

    From the faq I came to know that this may happen for the following reasons

    1. features are byte-swapped : Since I recorded & trained on a linux
        machine with intel this cannot be the problem

    2. feature file is small : The smallest feature file I have is 7 KB and the
        corresponding raw audio file is 44kB. ( I am using 16kHz , 16 bit data)
        
    3. audio fiiles may have headers : I used sox to convert wav to raw; &
        none of the audio files have the wav header.

    4. presence of digital zeros in the wav data: In the few files I saw There were none.

    Is it posible to get more info from kmeans.c/main.c as to in which feature/raw file did
    disaster take place?
       
    The faq says that there are executables cepview & seecep for viewing the features.
    But I couldn't find them. Can you tell me where to find them?

    Thanks in advance for your help
    Edison

     
    • Anonymous

      Anonymous - 2002-06-24

      Hi,

      I'm also encountering similar problem.  If you receive any respond regarding this problem,
      please post them.

      Here's  part of the log file:
      INFO: main.c(486): No mdef files.  Assuming 1-class init
      INFO: main.c(1257): 1-class dump file
      INFO: main.c(1295): Corpus 0: sz==1276 frames
      INFO: main.c(1304): Convergence ratios are abs(cur - prior) / abs(prior)
      INFO: main.c(240): alloc'ing 0Mb obs buf
      INFO: main.c(544): Initializing means using random k-means
      INFO: main.c(547): Trial 0: 256 means
      INFO: kmeans.c(159): km iter [0] 1.000000e+00 ...
      WARNING: "kmeans.c", line 437: Empty cluster 59
      WARNING: "kmeans.c", line 437: Empty cluster 80
      WARNING: "kmeans.c", line 437: Empty cluster 90
      WARNING: "kmeans.c", line 437: Empty cluster 92
      . . .

      I only train 4 wav files with 3 seconds duration for each.  Could a small data set
      have cause this error?

      Also, I set the following arguments when using wave2feat to create the feature files:

          %wave2feat -c control.ctl -raw -di in -ei wav -do out -eo feat

      I'm using '-raw' as input format, because i couldn't use '-wav'.  Could this also have cause
      the error?

      Thanks,

      wilson

       

Log in to post a comment.

Want the latest updates on software, tech news, and AI?
Get latest updates about software, tech news, and AI from SourceForge directly in your inbox once a month.