Blame view

Scripts/steps/train_lda_mllt.sh 7.84 KB
ec85f8892   bigot benjamin   first commit
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
  #!/bin/bash
  
  # Copyright 2012  Johns Hopkins University (Author: Daniel Povey)
  # Apache 2.0.
  
  # Begin configuration.
  cmd=run.pl
  config=
  stage=-5
  scale_opts="--transition-scale=1.0 --acoustic-scale=0.1 --self-loop-scale=0.1"
  realign_iters="10 20 30";
  mllt_iters="2 4 6 12";
  num_iters=35    # Number of iterations of training
  max_iter_inc=25  # Last iter to increase #Gauss on.
  dim=40
  beam=10
  retry_beam=40
  boost_silence=1.0 # Factor by which to boost silence likelihoods in alignment
  power=0.25 # Exponent for number of gaussians according to occurrence counts
  randprune=4.0 # This is approximately the ratio by which we will speed up the
                # LDA and MLLT calculations via randomized pruning.
  splice_opts=
  cluster_thresh=-1  # for build-tree control final bottom-up clustering of leaves
  # End configuration.
  train_tree=true  # if false, don't actually train the tree.
  use_lda_mat=  # If supplied, use this LDA[+MLLT] matrix.
  
  echo "$0 $@"  # Print the command line for logging
  
  [ -f path.sh ] && . ./path.sh
  . parse_options.sh || exit 1;
  
  if [ $# != 6 ]; then
    echo "Usage: steps/train_lda_mllt.sh [options] <#leaves> <#gauss> <data> <lang> <alignments> <dir>"
    echo " e.g.: steps/train_lda_mllt.sh 2500 15000 data/train_si84 data/lang exp/tri1_ali_si84 exp/tri2b"
    echo "Main options (for others, see top of script file)"
    echo "  --cmd (utils/run.pl|utils/queue.pl <queue opts>) # how to run jobs."
    echo "  --config <config-file>                           # config containing options"
    echo "  --stage <stage>                                  # stage to do partial re-run from."
    exit 1;
  fi
  
  numleaves=$1
  totgauss=$2
  data=$3
  lang=$4
  alidir=$5
  dir=$6
  
  for f in $alidir/final.mdl $alidir/ali.1.gz $data/feats.scp $lang/phones.txt; do
    [ ! -f $f ] && echo "train_lda_mllt.sh: no such file $f" && exit 1;
  done
  
  numgauss=$numleaves
  incgauss=$[($totgauss-$numgauss)/$max_iter_inc] # per-iter #gauss increment
  oov=`cat $lang/oov.int` || exit 1;
  nj=`cat $alidir/num_jobs` || exit 1;
  silphonelist=`cat $lang/phones/silence.csl` || exit 1;
  ciphonelist=`cat $lang/phones/context_indep.csl` || exit 1;
  
  mkdir -p $dir/log
  echo $nj >$dir/num_jobs
  echo "$splice_opts" >$dir/splice_opts # keep track of frame-splicing options
             # so that later stages of system building can know what they were.
  
  sdata=$data/split$nj;
  split_data.sh $data $nj || exit 1;
  
  
  splicedfeats="ark,s,cs:apply-cmvn --norm-vars=false --utt2spk=ark:$sdata/JOB/utt2spk scp:$sdata/JOB/cmvn.scp scp:$sdata/JOB/feats.scp ark:- | splice-feats $splice_opts ark:- ark:- |"
  # Note: $feats gets overwritten later in the script.
  feats="$splicedfeats transform-feats $dir/0.mat ark:- ark:- |"
  
  
  
  if [ $stage -le -5 ]; then
    if [ -z "$use_lda_mat" ]; then
      echo "Accumulating LDA statistics."
      $cmd JOB=1:$nj $dir/log/lda_acc.JOB.log \
      ali-to-post "ark:gunzip -c $alidir/ali.JOB.gz|" ark:- \| \
        weight-silence-post 0.0 $silphonelist $alidir/final.mdl ark:- ark:- \| \
        acc-lda --rand-prune=$randprune $alidir/final.mdl "$splicedfeats" ark,s,cs:- \
        $dir/lda.JOB.acc || exit 1;
      est-lda --write-full-matrix=$dir/full.mat --dim=$dim $dir/0.mat $dir/lda.*.acc \
        2>$dir/log/lda_est.log || exit 1;
      rm $dir/lda.*.acc
    else
      echo "Using supplied LDA matrix $use_lda_mat"
      cp $use_lda_mat $dir/0.mat || exit 1;
      [ ! -z "$mllt_iters" ] && \
        echo "Warning: using supplied LDA matrix $use_lda_mat but we will do MLLT," && \
        echo "which you might not want; to disable MLLT, specify --mllt-iters ''" && \
        sleep 5
    fi
  fi
  
  cur_lda_iter=0
  
  if [ $stage -le -4 ] && $train_tree; then
    echo "Accumulating tree stats"
    $cmd JOB=1:$nj $dir/log/acc_tree.JOB.log \
     acc-tree-stats  --ci-phones=$ciphonelist $alidir/final.mdl "$feats" \
       "ark:gunzip -c $alidir/ali.JOB.gz|" $dir/JOB.treeacc || exit 1;
    [ `ls $dir/*.treeacc | wc -w` -ne "$nj" ] && echo "Wrong #tree-accs" && exit 1;
    $cmd $dir/log/sum_tree_acc.log \
      sum-tree-stats $dir/treeacc $dir/*.treeacc || exit 1;
    rm $dir/*.treeacc
  fi
  
  
  if [ $stage -le -3 ] && $train_tree; then
    echo "Getting questions for tree clustering."
    # preparing questions, roots file...
    cluster-phones $dir/treeacc $lang/phones/sets.int $dir/questions.int 2> $dir/log/questions.log || exit 1;
    cat $lang/phones/extra_questions.int >> $dir/questions.int
    compile-questions $lang/topo $dir/questions.int $dir/questions.qst 2>$dir/log/compile_questions.log || exit 1;
  
    echo "Building the tree"
    $cmd $dir/log/build_tree.log \
      build-tree --verbose=1 --max-leaves=$numleaves \
      --cluster-thresh=$cluster_thresh $dir/treeacc $lang/phones/roots.int \
      $dir/questions.qst $lang/topo $dir/tree || exit 1;
  fi
  
  if [ $stage -le -2 ]; then
    echo "$0: Initializing the model"
    if $train_tree; then
      gmm-init-model  --write-occs=$dir/1.occs  \
        $dir/tree $dir/treeacc $lang/topo $dir/1.mdl 2> $dir/log/init_model.log || exit 1;
      grep 'no stats' $dir/log/init_model.log && echo "This is a bad warning.";
      rm $dir/treeacc
    else
      cp $alidir/tree $dir/ || exit 1;
      $cmd JOB=1 $dir/log/init_model.log \
        gmm-init-model-flat $dir/tree $lang/topo $dir/1.mdl \
          "$feats subset-feats ark:- ark:-|" || exit 1;
    fi
  fi
  
  
  if [ $stage -le -1 ]; then
    # Convert the alignments.
    echo "Converting alignments from $alidir to use current tree"
    $cmd JOB=1:$nj $dir/log/convert.JOB.log \
      convert-ali $alidir/final.mdl $dir/1.mdl $dir/tree \
       "ark:gunzip -c $alidir/ali.JOB.gz|" "ark:|gzip -c >$dir/ali.JOB.gz" || exit 1;
  fi
  
  if [ $stage -le 0 ] && [ "$realign_iters" != "" ]; then
    echo "Compiling graphs of transcripts"
    $cmd JOB=1:$nj $dir/log/compile_graphs.JOB.log \
      compile-train-graphs $dir/tree $dir/1.mdl  $lang/L.fst  \
       "ark:utils/sym2int.pl --map-oov $oov -f 2- $lang/words.txt < $data/split$nj/JOB/text |" \
        "ark:|gzip -c >$dir/fsts.JOB.gz" || exit 1;
  fi
  
  
  x=1
  while [ $x -lt $num_iters ]; do
    echo Training pass $x
    if echo $realign_iters | grep -w $x >/dev/null && [ $stage -le $x ]; then
      echo Aligning data
      mdl="gmm-boost-silence --boost=$boost_silence `cat $lang/phones/optional_silence.csl` $dir/$x.mdl - |"
      $cmd JOB=1:$nj $dir/log/align.$x.JOB.log \
        gmm-align-compiled $scale_opts --beam=$beam --retry-beam=$retry_beam "$mdl" \
        "ark:gunzip -c $dir/fsts.JOB.gz|" "$feats" \
        "ark:|gzip -c >$dir/ali.JOB.gz" || exit 1;
    fi
    if echo $mllt_iters | grep -w $x >/dev/null; then
      if [ $stage -le $x ]; then
        echo "Estimating MLLT"
        $cmd JOB=1:$nj $dir/log/macc.$x.JOB.log \
          ali-to-post "ark:gunzip -c $dir/ali.JOB.gz|" ark:- \| \
          weight-silence-post 0.0 $silphonelist $dir/$x.mdl ark:- ark:- \| \
          gmm-acc-mllt --rand-prune=$randprune  $dir/$x.mdl "$feats" ark:- $dir/$x.JOB.macc \
          || exit 1;
        est-mllt $dir/$x.mat.new $dir/$x.*.macc 2> $dir/log/mupdate.$x.log || exit 1;
        gmm-transform-means  $dir/$x.mat.new $dir/$x.mdl $dir/$x.mdl \
          2> $dir/log/transform_means.$x.log || exit 1;
        compose-transforms --print-args=false $dir/$x.mat.new $dir/$cur_lda_iter.mat $dir/$x.mat || exit 1;
        rm $dir/$x.*.macc
      fi
      feats="$splicedfeats transform-feats $dir/$x.mat ark:- ark:- |"
      cur_lda_iter=$x
    fi
  
    if [ $stage -le $x ]; then
      $cmd JOB=1:$nj $dir/log/acc.$x.JOB.log \
        gmm-acc-stats-ali  $dir/$x.mdl "$feats" \
        "ark,s,cs:gunzip -c $dir/ali.JOB.gz|" $dir/$x.JOB.acc || exit 1;
      $cmd $dir/log/update.$x.log \
        gmm-est --write-occs=$dir/$[$x+1].occs --mix-up=$numgauss --power=$power \
          $dir/$x.mdl "gmm-sum-accs - $dir/$x.*.acc |" $dir/$[$x+1].mdl || exit 1;
      rm $dir/$x.mdl $dir/$x.*.acc $dir/$x.occs 
    fi
    [ $x -le $max_iter_inc ] && numgauss=$[$numgauss+$incgauss];
    x=$[$x+1];
  done
  
  rm $dir/final.{mdl,mat,occs} 2>/dev/null
  ln -s $x.mdl $dir/final.mdl
  ln -s $x.occs $dir/final.occs
  ln -s $cur_lda_iter.mat $dir/final.mat
  
  # Summarize warning messages...
  
  utils/summarize_warnings.pl $dir/log
  
  echo Done training system with LDA+MLLT features in $dir