зеркало из https://github.com/mozilla/kaldi.git
Some updates to scripts.
git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@65 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8
This commit is contained in:
Dan Povey
Родитель
ac48d4d172
Коммит
ce34fdb4f0
|
|
@ -74,7 +74,7 @@ for test in mar87 oct87 feb89 oct89 feb91 sep92; do
|
|||
fi
|
||||
|
||||
if [ ! -f $dir/test_${test}.pass2.ali ]; then
|
||||
sgmm-decode-faster $utt2spk_opt --spk-vecs=ark:$dir/test_${test}.vecs --beam=20.0 --acoustic-scale=0.1 --word-symbol-table=data/words.txt $model $graphdir/HCLG.fst "$feats" ark,t:$dir/test_${test}.pass2.tra ark,t:$dir/test_${test}.pass2.ali 2> $dir/pass2_${test}.log
|
||||
sgmm-decode-faster "$gselect_opt" $utt2spk_opt --spk-vecs=ark:$dir/test_${test}.vecs --beam=20.0 --acoustic-scale=0.1 --word-symbol-table=data/words.txt $model $graphdir/HCLG.fst "$feats" ark,t:$dir/test_${test}.pass2.tra ark,t:$dir/test_${test}.pass2.ali 2> $dir/pass2_${test}.log
|
||||
fi
|
||||
|
||||
if [ ! -f $dir/test_${test}.fmllr_xforms ]; then
|
||||
|
|
@ -86,7 +86,7 @@ for test in mar87 oct87 feb89 oct89 feb91 sep92; do
|
|||
fi
|
||||
|
||||
if [ ! -f $dir/test_${test}.tra ]; then
|
||||
sgmm-decode-faster-fmllr $utt2spk_opt --spk-vecs=ark:$dir/test_${test}.vecs --beam=20.0 --acoustic-scale=0.1 --word-symbol-table=data/words.txt $fmllr_model $graphdir/HCLG.fst "$feats" ark:$dir/test_${test}.fmllr_xforms ark,t:$dir/test_${test}.tra ark,t:$dir/test_${test}.ali 2> $dir/decode_${test}.log
|
||||
sgmm-decode-faster-fmllr "$gselect_opt" $utt2spk_opt --spk-vecs=ark:$dir/test_${test}.vecs --beam=20.0 --acoustic-scale=0.1 --word-symbol-table=data/words.txt $fmllr_model $graphdir/HCLG.fst "$feats" ark:$dir/test_${test}.fmllr_xforms ark,t:$dir/test_${test}.tra ark,t:$dir/test_${test}.ali 2> $dir/decode_${test}.log
|
||||
fi
|
||||
|
||||
# the ,p option lets it score partial output without dying..
|
||||
|
|
|
|||
|
|
@ -54,7 +54,7 @@ for test in mar87 oct87 feb89 oct89 feb91 sep92; do
|
|||
ark:$dir/test_${test}.vecs ) 2>$dir/vecs_${test}.log
|
||||
|
||||
|
||||
sgmm-decode-faster $utt2spk_opt --spk-vecs=ark:$dir/test_${test}.vecs --beam=20.0 --acoustic-scale=0.1 --word-symbol-table=data/words.txt $model $graphdir/HCLG.fst "$feats" ark,t:$dir/test_${test}.tra ark,t:$dir/test_${test}.ali 2> $dir/decode_${test}.log
|
||||
sgmm-decode-faster "$gselect_opt" $utt2spk_opt --spk-vecs=ark:$dir/test_${test}.vecs --beam=20.0 --acoustic-scale=0.1 --word-symbol-table=data/words.txt $model $graphdir/HCLG.fst "$feats" ark,t:$dir/test_${test}.tra ark,t:$dir/test_${test}.ali 2> $dir/decode_${test}.log
|
||||
|
||||
# the ,p option lets it score partial output without dying..
|
||||
scripts/sym2int.pl --ignore-first-field data/words.txt data_prep/test_${test}_trans.txt | \
|
||||
|
|
|
|||
|
|
@ -55,6 +55,11 @@ system:
|
|||
tri2h 13.4 20.2 | [ splice-9-frames + HLDA... worse than tri2a]
|
||||
tri2i 12.4 18.4 | [ triple-deltas + HLDA... same as tri2a]
|
||||
tri2j 12.8 18.3 | [ triple-deltas+LDA+MLLT... slightly worse than tri2a]
|
||||
tri2k 11.1 15.1 | [ splice-9-frames + LDA + ET ]
|
||||
[utt] 11.1 15.0 | [adaptation per utterance]
|
||||
[spk,+fmllr] 10.1 14.0 | [per speaker, plus fMLLR]
|
||||
tri2l 9.6 13.7 | train with SAT; test with fMLLR
|
||||
[utt] 12.0 16.8 | [adaptation per utterance]
|
||||
|
||||
# TODO: tri2k, and tri3*?, and SGMM stuff.
|
||||
|
||||
|
|
@ -102,6 +107,16 @@ exp/decode_tri2i_tgpr_eval92/wer:%WER 12.39 [ 699 / 5641, 130 ins, 72 del, 497 s
|
|||
exp/decode_tri2i_tgpr_eval93/wer:%WER 18.35 [ 631 / 3439, 58 ins, 102 del, 471 sub ]
|
||||
exp/decode_tri2j_tgpr_eval92/wer:%WER 12.82 [ 723 / 5641, 127 ins, 70 del, 526 sub ]
|
||||
exp/decode_tri2j_tgpr_eval93/wer:%WER 18.26 [ 628 / 3439, 59 ins, 99 del, 470 sub ]
|
||||
exp/decode_tri2k_tgpr_eval92/wer:%WER 11.06 [ 624 / 5641, 133 ins, 42 del, 449 sub ]
|
||||
exp/decode_tri2k_tgpr_eval93/wer:%WER 15.09 [ 519 / 3439, 72 ins, 72 del, 375 sub ]
|
||||
exp/decode_tri2k_tgpr_fmllr_eval92/wer:%WER 10.07 [ 568 / 5641, 120 ins, 48 del, 400 sub ]
|
||||
exp/decode_tri2k_tgpr_fmllr_eval93/wer:%WER 14.02 [ 482 / 3439, 77 ins, 67 del, 338 sub ]
|
||||
exp/decode_tri2k_tgpr_utt_eval92/wer:%WER 11.13 [ 628 / 5641, 135 ins, 44 del, 449 sub ]
|
||||
exp/decode_tri2k_tgpr_utt_eval93/wer:%WER 14.95 [ 514 / 3439, 70 ins, 74 del, 370 sub ]
|
||||
exp/decode_tri2l_tgpr_eval92/wer:%WER 9.64 [ 544 / 5641, 121 ins, 44 del, 379 sub ]
|
||||
exp/decode_tri2l_tgpr_utt_eval92/wer:%WER 12.00 [ 677 / 5641, 141 ins, 60 del, 476 sub ]
|
||||
exp/decode_tri2l_tgpr_utt_eval93/wer:%WER 16.75 [ 576 / 3439, 59 ins, 93 del, 424 sub ]
|
||||
|
||||
exp/decode_tri3a_bg_eval92/wer:%WER 11.82 [ 667 / 5641, 132 ins, 63 del, 472 sub ]
|
||||
exp/decode_tri3a_bg_eval93/wer:%WER 15.00 [ 516 / 3439, 62 ins, 75 del, 379 sub ]
|
||||
exp/decode_tri3a_tgpr_dfmllr_eval92/wer:%WER 10.51 [ 593 / 5641, 111 ins, 51 del, 431 sub ]
|
||||
|
|
|
|||
|
|
@ -324,15 +324,20 @@ steps/train_tri2j.sh
|
|||
# LDA+ET
|
||||
steps/train_tri2k.sh
|
||||
(scripts/mkgraph.sh data/G_tg_pruned.fst exp/tri2k/tree exp/tri2k/final.mdl exp/graph_tri2k_tg_pruned || exit 1;
|
||||
scripts/decode.sh exp/decode_tri2k_tgpr_utt_eval92 exp/graph_tri2k_tg_pruned/HCLG.fst steps/decode_tri2k.sh data/eval_nov92.scp
|
||||
scripts/decode.sh --per-spk exp/decode_tri2k_tgpr_eval92 exp/graph_tri2k_tg_pruned/HCLG.fst steps/decode_tri2k.sh data/eval_nov92.scp
|
||||
for year in 92 93; do
|
||||
scripts/decode.sh exp/decode_tri2k_tgpr_utt_eval$year exp/graph_tri2k_tg_pruned/HCLG.fst steps/decode_tri2k.sh data/eval_nov$year.scp
|
||||
scripts/decode.sh --per-spk exp/decode_tri2k_tgpr_eval$year exp/graph_tri2k_tg_pruned/HCLG.fst steps/decode_tri2k.sh data/eval_nov$year.scp
|
||||
scripts/decode.sh --per-spk exp/decode_tri2k_tgpr_fmllr_eval$year exp/graph_tri2k_tg_pruned/HCLG.fst steps/decode_tri2k_fmllr.sh data/eval_nov$year.scp
|
||||
done
|
||||
)&
|
||||
|
||||
# LDA+MLLT+SAT
|
||||
steps/train_tri2l.sh
|
||||
(scripts/mkgraph.sh data/G_tg_pruned.fst exp/tri2l/tree exp/tri2l/final.mdl exp/graph_tri2l_tg_pruned || exit 1;
|
||||
scripts/decode.sh exp/decode_tri2l_tgpr_utt_eval92 exp/graph_tri2l_tg_pruned/HCLG.fst steps/decode_tri2l.sh data/eval_nov92.scp
|
||||
scripts/decode.sh exp/decode_tri2l_tgpr_utt_eval93 exp/graph_tri2l_tg_pruned/HCLG.fst steps/decode_tri2l.sh data/eval_nov93.scp
|
||||
scripts/decode.sh --per-spk exp/decode_tri2l_tgpr_eval92 exp/graph_tri2l_tg_pruned/HCLG.fst steps/decode_tri2l.sh data/eval_nov92.scp
|
||||
scripts/decode.sh --per-spk exp/decode_tri2l_tgpr_eval93 exp/graph_tri2l_tg_pruned/HCLG.fst steps/decode_tri2l.sh data/eval_nov93.scp
|
||||
)&
|
||||
|
||||
|
||||
|
|
|
|||
|
|
@ -71,11 +71,11 @@ echo running on `hostname` > $dir/predecode${job}.log
|
|||
gmm-decode-faster --beam=$prebeam --max-active=$max_active --acoustic-scale=$acwt --word-symbol-table=data/words.txt $alimodel $graph "$defaultfeats" ark,t:$dir/$job.pre_tra ark,t:$dir/$job.pre_ali 2>>$dir/predecode${job}.log
|
||||
|
||||
# Estimate transforms
|
||||
ali-to-post ark:$dir/$job.pre_ali ark:- | \
|
||||
(ali-to-post ark:$dir/$job.pre_ali ark:- | \
|
||||
weight-silence-post 0.0 $silphones $alimodel ark:- ark:- | \
|
||||
gmm-post-to-gpost $alimodel "$defaultfeats" ark,o:- ark:- | \
|
||||
gmm-est-et $spk2utt_opt $model $et "$sifeats" ark,o:- \
|
||||
ark:$dir/$job.trans ark,t:$dir/$job.warp 2>$dir/et${job}.log
|
||||
ark:$dir/$job.trans ark,t:$dir/$job.warp ) 2>$dir/et${job}.log
|
||||
|
||||
feats="ark:splice-feats --print-args=false scp:$scp ark:- | transform-feats $ldamat ark:- ark:- | transform-feats $utt2spk_opt ark:$dir/$job.trans ark:- ark:- |"
|
||||
|
||||
|
|
|
|||
|
|
@ -0,0 +1,97 @@
|
|||
#!/bin/bash
|
||||
# Copyright 2010-2011 Microsoft Corporation
|
||||
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
|
||||
# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
|
||||
# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
|
||||
# MERCHANTABLITY OR NON-INFRINGEMENT.
|
||||
# See the Apache 2 License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
|
||||
# This script does the decoding of a single batch of test data (on one core).
|
||||
# It requires arguments. It takes the graphdir and decoding directory, and the
|
||||
# job number. It expects a file $decode_dir/test${job_number}.scp to exist, and
|
||||
# puts its output in $decode_dir/${job_number}.tra
|
||||
#
|
||||
# If the files
|
||||
# $decode_dir/${job_number}.utt2spk and $decode_dir/${job_number}.spk2utt exist,
|
||||
# this script will assume you want to do per-speaker (not per-utterance) adaptation.
|
||||
|
||||
if [ $# != 3 ]; then
|
||||
echo "Usage: scripts/decode_tri2k_fmllr.sh <graph> <decode-dir> <job-number>"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
. path.sh || exit 1;
|
||||
|
||||
acwt=0.0625
|
||||
beam=13.0
|
||||
mincount=300 # for fMLLR
|
||||
prebeam=12.0 # first-pass decoding beam...
|
||||
max_active=7000
|
||||
alimodel=exp/tri2k/final.alimdl # first-pass model...
|
||||
model=exp/tri2k/final.mdl
|
||||
et=exp/tri2k/final.et
|
||||
defaultmat=exp/tri2k/default.mat
|
||||
ldamat=exp/tri2k/lda.mat
|
||||
silphones=`cat data/silphones.csl`
|
||||
graph=$1
|
||||
dir=$2
|
||||
job=$3
|
||||
scp=$dir/$job.scp
|
||||
sifeats="ark:splice-feats --print-args=false scp:$scp ark:- | transform-feats $ldamat ark:- ark:- |"
|
||||
defaultfeats="ark:splice-feats --print-args=false scp:$scp ark:- | transform-feats $defaultmat ark:- ark:- |"
|
||||
if [ -f $dir/$job.spk2utt ]; then
|
||||
if [ ! -f $dir/$job.utt2spk ]; then
|
||||
echo "spk2utt but not utt2spk file present!"
|
||||
exit 1
|
||||
fi
|
||||
spk2utt_opt=--spk2utt=ark:$dir/$job.spk2utt
|
||||
utt2spk_opt=--utt2spk=ark:$dir/$job.utt2spk
|
||||
fi
|
||||
|
||||
filenames="$scp $model $alimodel $et $graph data/words.txt"
|
||||
for file in $filenames; do
|
||||
if [ ! -f $file ] ; then
|
||||
echo "No such file $file";
|
||||
exit 1;
|
||||
fi
|
||||
done
|
||||
|
||||
echo running on `hostname` > $dir/predecode${job}.log
|
||||
|
||||
# First-pass decoding
|
||||
|
||||
gmm-decode-faster --beam=$prebeam --max-active=$max_active --acoustic-scale=$acwt --word-symbol-table=data/words.txt $alimodel $graph "$defaultfeats" ark,t:$dir/$job.pre_tra ark,t:$dir/$job.pre_ali 2>>$dir/predecode${job}.log
|
||||
|
||||
# Estimate transforms
|
||||
(ali-to-post ark:$dir/$job.pre_ali ark:- | \
|
||||
weight-silence-post 0.0 $silphones $alimodel ark:- ark:- | \
|
||||
gmm-post-to-gpost $alimodel "$defaultfeats" ark,o:- ark:- | \
|
||||
gmm-est-et $spk2utt_opt $model $et "$sifeats" ark,o:- \
|
||||
ark:$dir/$job.trans ark,t:$dir/$job.warp ) 2>$dir/et${job}.log
|
||||
|
||||
feats="ark:splice-feats --print-args=false scp:$scp ark:- | transform-feats $ldamat ark:- ark:- | transform-feats $utt2spk_opt ark:$dir/$job.trans ark:- ark:- |"
|
||||
|
||||
# Intermediate decoding
|
||||
echo running on `hostname` > $dir/pre2decode$job.log
|
||||
gmm-decode-faster --beam=$prebeam --max-active=$max_active --acoustic-scale=$acwt --word-symbol-table=data/words.txt $model $graph "$feats" ark,t:$dir/$job.pre2_tra ark,t:$dir/$job.pre2_ali 2>>$dir/pre2decode$job.log
|
||||
|
||||
# Estimate fMLLR transforms.
|
||||
(ali-to-post ark:$dir/$job.pre2_ali ark:- | \
|
||||
weight-silence-post 0.0 $silphones $model ark:- ark:- | \
|
||||
gmm-est-fmllr $spk2utt_opt --fmllr-min-count=$mincount $model "$feats" ark,o:- \
|
||||
ark:$dir/$job.fmllr ) 2>$dir/fmllr${job}.log
|
||||
|
||||
feats="ark:splice-feats --print-args=false scp:$scp ark:- | transform-feats $ldamat ark:- ark:- | transform-feats $utt2spk_opt ark:$dir/$job.trans ark:- ark:- | transform-feats $utt2spk_opt ark:$dir/$job.fmllr ark:- ark:- |"
|
||||
|
||||
# Final
|
||||
echo running final decoding pass on `hostname` > $dir/decode$job.log
|
||||
gmm-decode-faster --beam=$beam --max-active=$max_active --acoustic-scale=$acwt --word-symbol-table=data/words.txt $model $graph "$feats" ark,t:$dir/$job.tra ark,t:$dir/$job.ali 2>>$dir/decode$job.log
|
||||
|
|
@ -0,0 +1,280 @@
|
|||
#!/bin/bash
|
||||
# Copyright 2010-2011 Microsoft Corporation Arnab Ghoshal
|
||||
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
|
||||
# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
|
||||
# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
|
||||
# MERCHANTABLITY OR NON-INFRINGEMENT.
|
||||
# See the Apache 2 License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
|
||||
# tri2k is as tri2b ("exponential transform"), but using splice-9-frames + LDA
|
||||
# features instead.
|
||||
|
||||
if [ -f path.sh ]; then . path.sh; fi
|
||||
|
||||
dir=exp/tri2k
|
||||
srcdir=exp/tri1
|
||||
srcmodel=$srcdir/final.mdl
|
||||
scale_opts="--transition-scale=1.0 --acoustic-scale=0.1 --self-loop-scale=0.1"
|
||||
|
||||
# The 3 following settings relate to ET.
|
||||
dim=40 # the dim of our features.
|
||||
normtype=mean
|
||||
numiters_et=15 # Before this, update et.
|
||||
|
||||
numiters=35
|
||||
maxiterinc=20 # By this iter, we have all the Gaussians.
|
||||
realign_iters="10 20 30";
|
||||
numleaves=2000
|
||||
numgauss=2000 # initial num-gauss smallish so that transform-training
|
||||
# code (when we modify this script) is a bit faster.
|
||||
totgauss=10000 # Total num-gauss
|
||||
incgauss=$[($totgauss-$numgauss)/$maxiterinc] # per-iter increment for #Gauss
|
||||
|
||||
silphonelist=`cat data/silphones.csl`
|
||||
|
||||
mkdir -p $dir
|
||||
cp $srcdir/train.scp $dir
|
||||
cp $srcdir/train.tra $dir
|
||||
|
||||
scripts/filter_scp.pl $dir/train.scp data/train.utt2spk > $dir/train.utt2spk
|
||||
scripts/utt2spk_to_spk2utt.pl $dir/train.utt2spk > $dir/train.spk2utt
|
||||
|
||||
scripts/split_scp.pl --utt2spk=$dir/train.utt2spk $dir/train{,1,2,3}.scp
|
||||
scripts/split_scp.pl --utt2spk=$dir/train.utt2spk $dir/train{,1,2,3}.tra
|
||||
scripts/split_scp.pl --utt2spk=$dir/train.utt2spk $dir/train{,1,2,3}.utt2spk
|
||||
|
||||
for n in 1 2 3 ""; do # The "" handles the un-split one. Creating spk2utt files..
|
||||
scripts/utt2spk_to_spk2utt.pl $dir/train$n.utt2spk > $dir/train$n.spk2utt
|
||||
done
|
||||
|
||||
|
||||
# also see featspart below, used for sub-parts of the features;
|
||||
# try to keep them in sync.
|
||||
srcfeats="ark,s,cs:add-deltas scp:$dir/train.scp ark:- |"
|
||||
feats="ark,s,cs:splice-feats scp:$dir/train.scp ark:- | transform-feats $dir/lda.mat ark:- ark:- |"
|
||||
origfeats="ark,s,cs:splice-feats scp:$dir/train.scp ark:- | transform-feats $dir/lda.mat ark:- ark:- |"
|
||||
for n in 1 2 3; do
|
||||
srcfeatspart[$n]="ark,s,cs:add-deltas scp:$dir/train${n}.scp ark:- |"
|
||||
rawfeatspart[$n]="ark,s,cs:splice-feats scp:$dir/train${n}.scp ark:- |"
|
||||
featspart[$n]="ark,s,cs:splice-feats scp:$dir/train${n}.scp ark:- | transform-feats $dir/lda.mat ark:- ark:- |"
|
||||
origfeatspart[$n]="ark,s,cs:splice-feats scp:$dir/train${n}.scp ark:- | transform-feats $dir/lda.mat ark:- ark:- |"
|
||||
done
|
||||
|
||||
cp $srcdir/topo $dir
|
||||
|
||||
# Align all training data using old model (and old graphs, since we
|
||||
# use the same data-subset as last time).
|
||||
# Note: a few fail to get aligned here due to the difference between
|
||||
# per-speaker and per-utterance splitting, but this doesn't really matter.
|
||||
|
||||
echo "Aligning all training data"
|
||||
|
||||
rm -f $dir/.error
|
||||
for n in 1 2 3; do
|
||||
gmm-align-compiled $scale_opts --beam=8 --retry-beam=40 $srcmodel \
|
||||
"ark:gunzip -c $srcdir/graphs${n}.fsts.gz|" "${srcfeatspart[$n]}" \
|
||||
"ark:|gzip -c >$dir/0.${n}.ali.gz" \
|
||||
2> $dir/align.0.${n}.log || touch $dir/.error &
|
||||
done
|
||||
wait;
|
||||
[ -f $dir/.error ] && echo alignment error RE old system && exit 1
|
||||
|
||||
|
||||
# Now accumulate LDA stats.
|
||||
|
||||
for n in 1 2 3; do
|
||||
( ali-to-post "ark:gunzip -c $dir/0.$n.ali.gz|" ark:- | \
|
||||
weight-silence-post 0.0 $silphonelist $srcmodel ark:- ark:- | \
|
||||
acc-lda $srcmodel "${rawfeatspart[$n]}" \
|
||||
ark:- $dir/lda$n.acc ) 2>$dir/lda_acc.$n.log || touch $dir/.error &
|
||||
done
|
||||
wait
|
||||
|
||||
[ -f $dir/.error ] && echo lda-acc error && exit 1
|
||||
|
||||
est-lda --dim=$dim $dir/lda.mat $dir/lda?.acc 2>$dir/lda_est.log || exit 1
|
||||
|
||||
|
||||
acc-tree-stats --ci-phones=$silphonelist $srcmodel "$feats" "ark:gunzip -c $dir/0.?.ali.gz|" $dir/treeacc 2> $dir/acc.tree.log || exit 1;
|
||||
|
||||
|
||||
# The next few commands are involved with making the questions
|
||||
# for tree clustering. The extra complexity vs. the RM recipe has
|
||||
# to do with the desire to ask questions about the "real" phones
|
||||
# ignoring things like stress and position-in-word, and ask questions
|
||||
# separately about stress and position-in-word.
|
||||
|
||||
# Don't include silences as things to be clustered -> --nosil option.
|
||||
scripts/make_shared_phones.sh --nosil | scripts/sym2int.pl data/phones.txt > $dir/phone_sets.list
|
||||
cluster-phones $dir/treeacc $dir/phone_sets.list $dir/questions.txt 2> $dir/cluster_phones.log || exit 1;
|
||||
scripts/int2sym.pl data/phones.txt < $dir/questions.txt > $dir/questions_syms.txt
|
||||
scripts/make_extra_questions.sh | cat $dir/questions_syms.txt - > $dir/questions_syms_all.txt
|
||||
scripts/sym2int.pl data/phones.txt < $dir/questions_syms_all.txt > $dir/questions_all.txt
|
||||
|
||||
compile-questions $dir/topo $dir/questions_all.txt $dir/questions.qst 2>$dir/compile_questions.log || exit 1;
|
||||
|
||||
scripts/make_roots.sh > $dir/roots_syms.txt
|
||||
scripts/sym2int.pl --ignore-oov data/phones.txt < $dir/roots_syms.txt > $dir/roots.txt
|
||||
|
||||
|
||||
build-tree --verbose=1 --max-leaves=$numleaves \
|
||||
$dir/treeacc $dir/roots.txt \
|
||||
$dir/questions.qst $dir/topo $dir/tree 2> $dir/train_tree.log || exit 1;
|
||||
|
||||
gmm-init-model --write-occs=$dir/1.occs \
|
||||
$dir/tree $dir/treeacc $dir/topo $dir/1.mdl 2> $dir/init_model.log || exit 1;
|
||||
|
||||
gmm-mixup --mix-up=$numgauss $dir/1.mdl $dir/1.occs $dir/1.mdl \
|
||||
2>$dir/mixup.log || exit 1;
|
||||
|
||||
|
||||
rm $dir/treeacc $dir/1.occs
|
||||
|
||||
|
||||
# Convert alignments generated from previous model, to use as initial alignments.
|
||||
|
||||
for n in 1 2 3; do
|
||||
convert-ali $srcmodel $dir/1.mdl $dir/tree \
|
||||
"ark:gunzip -c $dir/0.$n.ali.gz|" \
|
||||
"ark:|gzip -c > $dir/cur$n.ali.gz" 2>$dir/convert.$n.log
|
||||
done
|
||||
rm $dir/0.?.ali.gz
|
||||
|
||||
# Make training graphs
|
||||
echo "Compiling training graphs"
|
||||
|
||||
rm -f $dir/.error
|
||||
for n in 1 2 3; do
|
||||
compile-train-graphs $dir/tree $dir/1.mdl data/L.fst ark:$dir/train${n}.tra \
|
||||
"ark:|gzip -c > $dir/graphs${n}.fsts.gz" \
|
||||
2>$dir/compile_graphs.${n}.log || touch $dir/.error &
|
||||
done
|
||||
|
||||
wait
|
||||
[ -f $dir/.error ] && echo compile-graphs error && exit 1
|
||||
|
||||
gmm-init-et --normalize-type=$normtype --binary=false --dim=$dim $dir/1.et 2>$dir/init_et.log || exit 1
|
||||
|
||||
x=1
|
||||
while [ $x -lt $numiters ]; do
|
||||
echo "Pass $x"
|
||||
if echo $realign_iters | grep -w $x >/dev/null; then
|
||||
echo "Aligning data"
|
||||
rm -f $dir/.error
|
||||
for n in 1 2 3; do
|
||||
gmm-align-compiled $scale_opts --beam=8 --retry-beam=40 $dir/$x.mdl \
|
||||
"ark:gunzip -c $dir/graphs${n}.fsts.gz|" "${featspart[$n]}" \
|
||||
"ark:|gzip -c >$dir/cur${n}.ali.gz" 2> $dir/align.$x.$n.log \
|
||||
|| touch $dir/.error &
|
||||
done
|
||||
wait
|
||||
[ -f $dir/.error ] && echo error aligning data && exit 1
|
||||
fi
|
||||
if [ $x -lt $numiters_et ]; then
|
||||
# Work out current transforms (in parallel).
|
||||
echo "Computing ET transforms"
|
||||
rm -f $dir/.error
|
||||
for n in 1 2 3; do
|
||||
( ali-to-post "ark:gunzip -c $dir/cur${n}.ali.gz|" ark:- | \
|
||||
weight-silence-post 0.0 $silphonelist $dir/$x.mdl ark:- ark:- | \
|
||||
gmm-post-to-gpost $dir/$x.mdl "${featspart[$n]}" ark,o:- ark:- | \
|
||||
gmm-est-et --spk2utt=ark:$dir/train$n.spk2utt --verbose=1 $dir/$x.mdl $dir/$x.et \
|
||||
"${origfeatspart[$n]}" ark,s,cs:- ark:$dir/$x.$n.trans ark,t:$dir/$x.$n.warp ) \
|
||||
2> $dir/trans.$x.$n.log || touch $dir/.error &
|
||||
done
|
||||
wait
|
||||
[ -f $dir/.error ] && echo error aligning data && exit 1
|
||||
|
||||
# Remove previous transforms, if present.
|
||||
if [ $x -gt 1 ]; then rm $dir/$[$x-1].?.trans; fi
|
||||
# Now change $feats and $featspart to correspond to the transformed features.
|
||||
feats="ark,s,cs:splice-feats scp:$dir/train.scp ark:- | transform-feats $dir/lda.mat ark:- ark:- | transform-feats --utt2spk=ark:$dir/train.utt2spk \"ark,s,cs:cat $dir/$x.?.trans|\" ark:- ark:- |"
|
||||
for n in 1 2 3; do
|
||||
featspart[$n]="ark,s,cs:splice-feats scp:$dir/train${n}.scp ark:- | transform-feats $dir/lda.mat ark:- ark:- | transform-feats --utt2spk=ark:$dir/train$n.utt2spk ark:$dir/$x.$n.trans ark:- ark:- |"
|
||||
done
|
||||
fi
|
||||
gmm-acc-stats-ali --binary=false $dir/$x.mdl "$feats" \
|
||||
"ark,s,cs:gunzip -c $dir/cur?.ali.gz|" $dir/$x.acc 2> $dir/acc.$x.log || exit 1;
|
||||
gmm-est --write-occs=$dir/$[$x+1].occs --mix-up=$numgauss $dir/$x.mdl $dir/$x.acc $dir/$[$x+1].mdl 2> $dir/update.$x.log || exit 1;
|
||||
rm $dir/$x.mdl $dir/$x.acc $dir/$x.occs 2>/dev/null
|
||||
|
||||
if [ $x -lt $numiters_et ]; then
|
||||
# Alternately estimate either A or B.
|
||||
|
||||
x1=$[$x+1]
|
||||
if [ $[$x%2] == 0 ]; then # Estimate A:
|
||||
for n in 1 2 3; do
|
||||
( ali-to-post "ark:gunzip -c $dir/cur${n}.ali.gz|" ark:- | \
|
||||
weight-silence-post 0.0 $silphonelist $dir/$x1.mdl ark:- ark:- | \
|
||||
gmm-post-to-gpost $dir/$x1.mdl "${featspart[$n]}" ark:- ark:- | \
|
||||
gmm-et-acc-a --spk2utt=ark:$dir/train$n.spk2utt --verbose=1 $dir/$x1.mdl \
|
||||
$dir/$x.et "${origfeatspart[$n]}" ark,s,cs:- $dir/$x.$n.et_acc_a ) \
|
||||
2> $dir/acc_a.$x.$n.log || touch $dir/.error &
|
||||
done
|
||||
wait
|
||||
[ -f $dir/.error ] && echo error computing stats to accumulate A && exit 1
|
||||
gmm-et-est-a --verbose=1 $dir/$x.et $dir/$x1.et $dir/$x.?.et_acc_a 2> $dir/update_a.$x.log || exit 1;
|
||||
rm $dir/$x.?.et_acc_a
|
||||
else
|
||||
for n in 1 2 3; do
|
||||
( ali-to-post "ark:gunzip -c $dir/cur${n}.ali.gz|" ark:- | \
|
||||
weight-silence-post 0.0 $silphonelist $dir/$x1.mdl ark:- ark:- | \
|
||||
gmm-post-to-gpost $dir/$x1.mdl "${featspart[$n]}" ark:- ark:- | \
|
||||
gmm-et-acc-b --spk2utt=ark:$dir/train$n.spk2utt --verbose=1 $dir/$x1.mdl $dir/$x.et \
|
||||
"${origfeatspart[$n]}" ark,s,cs:- ark:$dir/$x.$n.trans ark:$dir/$x.$n.warp $dir/$x.$n.et_acc_b ) \
|
||||
2> $dir/acc_b.$x.$n.log || touch $dir/.error &
|
||||
done
|
||||
wait
|
||||
[ -f $dir/.error ] && echo error computing stats to accumulate B && exit 1
|
||||
gmm-et-est-b --verbose=1 $dir/$x.et $dir/$x1.et $dir/$x.mat $dir/$x.?.et_acc_b \
|
||||
2> $dir/update_b.$x.log || exit 1;
|
||||
rm $dir/$x.?.et_acc_b
|
||||
# Careful!: gmm-transform-means here changes $x1.mdl in-place.
|
||||
# It's important that the very next thing we do after this is to re-estimate
|
||||
# the ET transforms: any existing transforms would be invalid.
|
||||
gmm-transform-means $dir/$x.mat $dir/$x1.mdl $dir/$x1.mdl 2> $dir/transform_means.$x.log
|
||||
fi
|
||||
fi
|
||||
|
||||
|
||||
if [ $x -le $maxiterinc ]; then
|
||||
numgauss=$[$numgauss+$incgauss];
|
||||
fi
|
||||
x=$[$x+1];
|
||||
done
|
||||
|
||||
# Accumulate stats for "alignment model" which is as the model but with
|
||||
# the baseline features (shares Gaussian-level alignments).
|
||||
|
||||
gmm-et-get-b $dir/$numiters_et.et $dir/B.mat 2>$dir/get_b.log || exit 1
|
||||
compose-transforms $dir/B.mat $dir/lda.mat $dir/default.mat 2>>$dir/get_b.log || exit 1
|
||||
|
||||
defaultfeats="ark,s,cs:splice-feats scp:$dir/train.scp ark:- | transform-feats $dir/default.mat ark:- ark:- |"
|
||||
|
||||
( ali-to-post "ark:gunzip -c $dir/cur?.ali.gz|" ark:- | \
|
||||
gmm-acc-stats-twofeats $dir/$x.mdl "$feats" "$defaultfeats" ark:- $dir/$x.acc2 ) 2>$dir/acc_alimdl.log || exit 1;
|
||||
# Update model.
|
||||
gmm-est --remove-low-count-gaussians=false $dir/$x.mdl $dir/$x.acc2 $dir/$x.alimdl \
|
||||
2>$dir/est_alimdl.log || exit 1;
|
||||
rm $dir/$x.acc2
|
||||
|
||||
# The following files may be be useful for display purposes.
|
||||
y=1
|
||||
while [ $y -lt $numiters_et ]; do
|
||||
cat $dir/$y.?.warp | scripts/process_warps.pl data/spk2gender.map > $dir/warps.$y
|
||||
y=$[$y+1]
|
||||
done
|
||||
|
||||
|
||||
( cd $dir; rm final.{mdl,alimdl,et} 2>/dev/null;
|
||||
ln -s $x.mdl final.mdl; ln -s $x.alimdl final.alimdl;
|
||||
ln -s $numiters_et.et final.et )
|
||||
|
|
@ -227,18 +227,19 @@ done
|
|||
ln -s $x.mdl final.mdl;
|
||||
ln -s `basename $cur_lda` final.mat )
|
||||
|
||||
defaultfeats="ark:splice-feats scp:$dir/train.scp ark:- | transform-feats $cur_lda ark:- ark:-|"
|
||||
defaultfeats="ark:splice-feats scp:$dir/train.scp ark:- | transform-feats $dir/final.mat ark:- ark:-|"
|
||||
feats="ark:splice-feats scp:$dir/train.scp ark:- | transform-feats $dir/final.mat ark:- ark:- | transform-feats --utt2spk=ark:$dir/train.utt2spk \"ark:cat $dir/cur?.trans|\" ark:- ark:- |"
|
||||
|
||||
# Accumulate stats for "alignment model" which is as the model but with
|
||||
# the unadapted, default features (shares Gaussian-level alignments).
|
||||
( ali-to-post "ark:cat $dir/cur?.ali.gz|" ark:- | \
|
||||
gmm-acc-stats-twofeats $dir/$x.mdl "$feats" "$defaultfeats" ark:- $dir/$x.acc2 ) 2>$dir/acc_alimdl.log || exit 1;
|
||||
( ali-to-post "ark:gunzip -c $dir/cur?.ali.gz|" ark:- | \
|
||||
gmm-acc-stats-twofeats $dir/final.mdl "$feats" "$defaultfeats" ark:- $dir/$x.acc2 ) 2>$dir/acc_alimdl.log || exit 1;
|
||||
# Update model.
|
||||
gmm-est --write-occs=$dir/final.occs --remove-low-count-gaussians=false \
|
||||
$dir/$x.mdl $dir/$x.acc2 $dir/$x.alimdl \
|
||||
2>$dir/est_alimdl.log || exit 1;
|
||||
rm $dir/$x.acc2
|
||||
|
||||
( cd $dir; rm final.{mdl,alimdl,mat} 2>/dev/null;
|
||||
ln -s $x.mdl final.mdl; ln -s $x.alimdl final.alimdl
|
||||
ln -s `basename $cur_lda` final.mat )
|
||||
( cd $dir; rm final.alimdl 2>/dev/null;
|
||||
ln -s $x.alimdl final.alimdl )
|
||||
|
||||
|
|
|
|||
|
|
@ -0,0 +1,205 @@
|
|||
#!/bin/bash
|
||||
# Copyright 2010-2011 Microsoft Corporation Arnab Ghoshal
|
||||
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
|
||||
# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
|
||||
# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
|
||||
# MERCHANTABLITY OR NON-INFRINGEMENT.
|
||||
# See the Apache 2 License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
|
||||
# tri3k is triphone training with splice-9-frames + LDA + ET (exponential
|
||||
# transform), using the entire si-284 training set, starting from the
|
||||
# model in tri2k. Initializing the states from the previous model's states
|
||||
# for faster training. We use the previous model to get the speaker transforms,
|
||||
# and do this only once (we use the alignment model).
|
||||
# This script uses (about) 3 CPUs.
|
||||
|
||||
if [ -f path.sh ]; then . path.sh; fi
|
||||
|
||||
dir=exp/tri3k
|
||||
srcdir=exp/tri2k
|
||||
srcmodel=$srcdir/final.mdl
|
||||
et=$srcdir/final.et
|
||||
ldamat=$srcdir/lda.mat
|
||||
defaultmat=$srcdir/default.mat # with the "default" exponential transform, used by alignment model.
|
||||
srcalimodel=$srcdir/final.alimdl
|
||||
scale_opts="--transition-scale=1.0 --acoustic-scale=0.1 --self-loop-scale=0.1"
|
||||
|
||||
numiters=20
|
||||
maxiterinc=10 # By this iter, we have all the Gaussians.
|
||||
realign_iters="10 15";
|
||||
|
||||
numleaves=4200
|
||||
numgauss=20000 # Initial num-gauss. Initializing states using tri2c model,
|
||||
# so can have a reasonably large number.
|
||||
totgauss=40000 # Total num-gauss
|
||||
incgauss=$[($totgauss-$numgauss)/$maxiterinc] # per-iter increment for #Gauss
|
||||
|
||||
silphonelist=`cat data/silphones.csl`
|
||||
|
||||
mkdir -p $dir
|
||||
|
||||
# Use all the SI-284 data.
|
||||
|
||||
cp data/train.{scp,tra,utt2spk} $dir
|
||||
|
||||
|
||||
# Split up the scp and related files to 3 parts; create spk2utt files.
|
||||
scripts/split_scp.pl --utt2spk=$dir/train.utt2spk $dir/train.scp $dir/train{1,2,3}.scp
|
||||
scripts/split_scp.pl --utt2spk=$dir/train.utt2spk $dir/train.tra $dir/train{1,2,3}.tra
|
||||
scripts/split_scp.pl --utt2spk=$dir/train.utt2spk $dir/train{,1,2,3}.utt2spk
|
||||
for n in 1 2 3 ""; do # The "" handles the un-split one. Creating spk2utt files..
|
||||
scripts/utt2spk_to_spk2utt.pl $dir/train$n.utt2spk > $dir/train$n.spk2utt
|
||||
done
|
||||
|
||||
# also see featspart below, used for sub-parts of the features.
|
||||
|
||||
# origfeats is feats with just LDA, no exponential trnsform.
|
||||
# defaultfeats is for the "default" speaker-- used with alignment model.
|
||||
origfeats="ark:splice-feats scp:$dir/train.scp ark:- | transform-feats $ldamat ark:- ark:- |"
|
||||
defaultfeats="ark:splice-feats scp:$dir/train.scp ark:- | transform-feats $defaultmat ark:- ark:- |"
|
||||
feats="ark:splice-feats scp:$dir/train.scp ark:- | transform-feats $ldamat ark:- ark:- | transform-feats --utt2spk=ark:$dir/train.utt2spk \"ark:cat $dir/?.trans|\" ark:- ark:- |"
|
||||
for n in 1 2 3; do
|
||||
featspart[$n]="ark,s,cs:splice-feats scp:$dir/train${n}.scp ark:- | transform-feats $ldamat ark:- ark:- | transform-feats --utt2spk=ark:$dir/train$n.utt2spk ark:$dir/$n.trans ark:- ark:- |"
|
||||
origfeatspart[$n]="ark,s,cs:splice-feats scp:$dir/train${n}.scp ark:- | transform-feats $ldamat ark:- ark:- |"
|
||||
defaultfeatspart[$n]="ark,s,cs:splice-feats scp:$dir/train${n}.scp ark:- | transform-feats $defaultmat ark:- ark:- |"
|
||||
done
|
||||
|
||||
cp $srcdir/topo $dir
|
||||
|
||||
|
||||
echo "Aligning all training data with alignment model"
|
||||
|
||||
rm -f $dir/.error
|
||||
for n in 1 2 3; do
|
||||
compile-train-graphs $srcdir/tree $srcalimodel data/L.fst ark:$dir/train${n}.tra ark:- 2>$dir/graphsold.${n}.log | \
|
||||
gmm-align-compiled $scale_opts --beam=8 --retry-beam=40 $srcalimodel ark:- "${defaultfeatspart[$n]}" "ark:|gzip -c > $dir/0.${n}.ali.gz" 2> $dir/align.0.${n}.log || touch $dir/.error &
|
||||
done
|
||||
wait;
|
||||
[ -f $dir/.error ] && echo compile-graphs error RE old system && exit 1
|
||||
|
||||
echo "Computing ET transforms."
|
||||
for n in 1 2 3; do
|
||||
( ali-to-post "ark:gunzip -c $dir/0.${n}.ali.gz|" ark:- | \
|
||||
weight-silence-post 0.0 $silphonelist $srcalimodel ark:- ark:- | \
|
||||
gmm-post-to-gpost $srcalimodel "${defaultfeatspart[$n]}" ark,o:- ark:- | \
|
||||
gmm-est-et --spk2utt=ark:$dir/train$n.spk2utt $srcmodel $et "${origfeatspart[$n]}" ark,o:- \
|
||||
ark:$dir/$n.trans ark,t:$dir/$n.warp ) 2>$dir/est_et.$n.log || touch $dir/.error &
|
||||
done
|
||||
wait;
|
||||
[ -f $dir/.error ] && echo Error computing ET transfirms && exit 1
|
||||
|
||||
# debug info for warping factors.
|
||||
cat $dir/?.warp | scripts/process_warps.pl data/spk2gender.map > $dir/warps
|
||||
|
||||
acc-tree-stats --ci-phones=$silphonelist $srcmodel "$feats" "ark:gunzip -c $dir/0.?.ali.gz|" $dir/treeacc 2> $dir/acc.tree.log || exit 1;
|
||||
|
||||
# The next few commands are involved with making the questions
|
||||
# for tree clustering. The extra complexity vs. the RM recipe has
|
||||
# to do with the desire to ask questions about the "real" phones
|
||||
# ignoring things like stress and position-in-word, and ask questions
|
||||
# separately about stress and position-in-word.
|
||||
|
||||
# Don't include silences as things to be clustered -> --nosil option.
|
||||
scripts/make_shared_phones.sh --nosil | scripts/sym2int.pl data/phones.txt > $dir/phone_sets.list
|
||||
cluster-phones $dir/treeacc $dir/phone_sets.list $dir/questions.txt 2> $dir/cluster_phones.log || exit 1;
|
||||
scripts/int2sym.pl data/phones.txt < $dir/questions.txt > $dir/questions_syms.txt
|
||||
scripts/make_extra_questions.sh | cat $dir/questions_syms.txt - > $dir/questions_syms_all.txt
|
||||
scripts/sym2int.pl data/phones.txt < $dir/questions_syms_all.txt > $dir/questions_all.txt
|
||||
|
||||
compile-questions $dir/topo $dir/questions_all.txt $dir/questions.qst 2>$dir/compile_questions.log || exit 1;
|
||||
|
||||
scripts/make_roots.sh > $dir/roots_syms.txt
|
||||
scripts/sym2int.pl --ignore-oov data/phones.txt < $dir/roots_syms.txt > $dir/roots.txt
|
||||
|
||||
|
||||
build-tree --verbose=1 --max-leaves=$numleaves \
|
||||
$dir/treeacc $dir/roots.txt \
|
||||
$dir/questions.qst $dir/topo $dir/tree 2> $dir/train_tree.log || exit 1;
|
||||
|
||||
gmm-init-model --write-occs=$dir/1.occs \
|
||||
$dir/tree $dir/treeacc $dir/topo $dir/1.mdl \
|
||||
$srcdir/tree $srcmodel 2> $dir/init_model.log || exit 1;
|
||||
|
||||
# Mix down and mix up to get exactly the targets #Gauss
|
||||
# (note: the tool does mix-down first regardless of option order.)
|
||||
gmm-mixup --mix-down=$numgauss --mix-up=$numgauss $dir/1.mdl $dir/1.occs $dir/1.mdl \
|
||||
2>$dir/mixup.log || exit 1;
|
||||
|
||||
rm $dir/treeacc $dir/1.occs
|
||||
|
||||
|
||||
# Convert alignments generated from monophone model, to use as initial alignments.
|
||||
|
||||
for n in 1 2 3; do
|
||||
convert-ali $srcmodel $dir/1.mdl $dir/tree \
|
||||
"ark:gunzip -c $dir/0.$n.ali.gz|" \
|
||||
"ark:|gzip -c > $dir/cur$n.ali.gz" 2>$dir/convert.$n.log
|
||||
|
||||
done
|
||||
|
||||
rm $dir/0.?.ali.gz
|
||||
|
||||
# Make training graphs
|
||||
echo "Compiling training graphs"
|
||||
|
||||
rm -f $dir/.error
|
||||
for n in 1 2 3; do
|
||||
compile-train-graphs $dir/tree $dir/1.mdl data/L.fst ark:$dir/train${n}.tra \
|
||||
"ark:|gzip -c > $dir/graphs${n}.fsts.gz" \
|
||||
2>$dir/compile_graphs.${n}.log || touch $dir/.error &
|
||||
done
|
||||
|
||||
wait
|
||||
[ -f $dir/.error ] && echo compile-graphs error && exit 1
|
||||
|
||||
x=1
|
||||
while [ $x -lt $numiters ]; do
|
||||
echo "Pass $x"
|
||||
if echo $realign_iters | grep -w $x >/dev/null; then
|
||||
echo "Aligning data"
|
||||
rm -f $dir/.error
|
||||
for n in 1 2 3; do
|
||||
gmm-align-compiled $scale_opts --beam=8 --retry-beam=40 $dir/$x.mdl \
|
||||
"ark:gunzip -c $dir/graphs${n}.fsts.gz|" "${featspart[$n]}" \
|
||||
"ark:|gzip -c >$dir/cur${n}.ali.gz" 2> $dir/align.$x.$n.log \
|
||||
|| touch $dir/.error &
|
||||
done
|
||||
wait
|
||||
[ -f $dir/.error ] && echo compile-graphs error && exit 1
|
||||
fi
|
||||
for n in 1 2 3; do
|
||||
gmm-acc-stats-ali --binary=false $dir/$x.mdl "${featspart[$n]}" \
|
||||
"ark:gunzip -c $dir/cur${n}.ali.gz|" $dir/$x.$n.acc 2> $dir/acc.$x.$n.log || touch $dir/.error &
|
||||
done
|
||||
wait
|
||||
[ -f $dir/.error ] && echo accumulation error && exit 1
|
||||
gmm-sum-accs $dir/$x.acc $dir/$x.?.acc || exit 1;
|
||||
rm $dir/$x.?.acc
|
||||
gmm-est --write-occs=$dir/$[$x+1].occs --mix-up=$numgauss $dir/$x.mdl $dir/$x.acc $dir/$[$x+1].mdl 2> $dir/update.$x.log || exit 1;
|
||||
rm $dir/$x.mdl $dir/$x.acc $dir/$x.occs 2>/dev/null
|
||||
if [ $x -le $maxiterinc ]; then
|
||||
numgauss=$[$numgauss+$incgauss];
|
||||
fi
|
||||
x=$[$x+1];
|
||||
done
|
||||
|
||||
|
||||
( ali-to-post "ark:gunzip -c $dir/cur?.ali.gz|" ark:- | \
|
||||
gmm-acc-stats-twofeats $dir/$x.mdl "$feats" "$defaultfeats" ark:- $dir/$x.acc2 ) 2>$dir/acc_alimdl.log || exit 1;
|
||||
# Update model.
|
||||
gmm-est --remove-low-count-gaussians=false $dir/$x.mdl $dir/$x.acc2 $dir/$x.alimdl \
|
||||
2>$dir/est_alimdl.log || exit 1;
|
||||
rm $dir/$x.acc2
|
||||
|
||||
( cd $dir; rm final.{mdl,alimdl,et} 2>/dev/null;
|
||||
ln -s $x.mdl final.mdl; ln -s $x.alimdl final.alimdl;
|
||||
ln -s `pwd`/$et $dir/final.et )
|
||||
Загрузка…
Ссылка в новой задаче