Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[WIP] Lattice-faster-decoder-combine #3061

Open
wants to merge 29 commits into
base: master
Choose a base branch
from
Open
Show file tree
Hide file tree
Changes from 25 commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
128 changes: 128 additions & 0 deletions egs/wsj/s5/steps/decode_combine_test.sh
Original file line number Diff line number Diff line change
@@ -0,0 +1,128 @@
#!/bin/bash

# Copyright 2012 Johns Hopkins University (Author: Daniel Povey)
# Apache 2.0

# Begin configuration.
nj=4
cmd=run.pl
maxactive=7000
beam=15.0
lattice_beam=8.0
expand_beam=30.0
acwt=1.0
skip_scoring=false
combine_version=false

stage=0
online_ivector_dir=
post_decode_acwt=10.0
extra_left_context=0
extra_right_context=0
extra_left_context_initial=0
extra_right_context_final=0
chunk_width=140,100,160
use_gpu=no
# End configuration.

echo "$0 $@" # Print the command line for logging

[ -f ./path.sh ] && . ./path.sh; # source the path.
. parse_options.sh || exit 1;

if [ $# != 3 ]; then
echo "Usage: steps/decode_combine_test.sh [options] <graph-dir> <data-dir> <decode-dir>"
echo "... where <decode-dir> is assumed to be a sub-directory of the directory"
echo " where the model is."
echo "e.g.: steps/decode_combine_test.sh exp/mono/graph_tgpar data/test_dev93 exp/mono/decode_dev93_tgpr"
echo ""
echo "This script works on CMN + (delta+delta-delta | LDA+MLLT) features; it works out"
echo "what type of features you used (assuming it's one of these two)"
echo ""
echo "main options (for others, see top of script file)"
echo " --config <config-file> # config containing options"
echo " --nj <nj> # number of parallel jobs"
echo " --cmd (utils/run.pl|utils/queue.pl <queue opts>) # how to run jobs."
exit 1;
fi


graphdir=$1
data=$2
dir=$3

srcdir=`dirname $dir`; # The model directory is one level up from decoding directory.
sdata=$data/split$nj;
splice_opts=`cat $srcdir/splice_opts 2>/dev/null`
cmvn_opts=`cat $srcdir/cmvn_opts 2>/dev/null`
delta_opts=`cat $srcdir/delta_opts 2>/dev/null`

mkdir -p $dir/log
[[ -d $sdata && $data/feats.scp -ot $sdata ]] || split_data.sh $data $nj || exit 1;
echo $nj > $dir/num_jobs


for f in $sdata/1/feats.scp $sdata/1/cmvn.scp $srcdir/final.mdl $graphdir/HCLG.fst; do
[ ! -f $f ] && echo "decode_combine_test.sh: no such file $f" && exit 1;
done


if [ -f $srcdir/final.mat ]; then feat_type=lda; else feat_type=delta; fi
echo "decode_combine_test.sh: feature type is $feat_type"

feats="ark,s,cs:apply-cmvn $cmvn_opts --utt2spk=ark:$sdata/JOB/utt2spk scp:$sdata/JOB/cmvn.scp scp:$sdata/JOB/feats.scp ark:- |"

posteriors="ark,scp:$sdata/JOB/posterior.ark,$sdata/JOB/posterior.scp"
posteriors_scp="scp:$sdata/JOB/posterior.scp"

if [ ! -z "$online_ivector_dir" ]; then
ivector_period=$(cat $online_ivector_dir/ivector_period) || exit 1;
ivector_opts="--online-ivectors=scp:$online_ivector_dir/ivector_online.scp --online-ivector-period=$ivector_period"
fi

if [ "$post_decode_acwt" == 1.0 ]; then
lat_wspecifier="ark:|gzip -c >$dir/lat.JOB.gz"
else
lat_wspecifier="ark:|lattice-scale --acoustic-scale=$post_decode_acwt ark:- ark:- | gzip -c >$dir/lat.JOB.gz"
fi

frame_subsampling_opt=
if [ -f $srcdir/frame_subsampling_factor ]; then
# e.g. for 'chain' systems
frame_subsampling_opt="--frame-subsampling-factor=$(cat $srcdir/frame_subsampling_factor)"
fi

frames_per_chunk=$(echo $chunk_width | cut -d, -f1)
# generate log-likelihood
if [ $stage -le 1 ]; then
$cmd JOB=1:$nj $dir/log/nnet_compute.JOB.log \
nnet3-compute $ivector_opts $frame_subsampling_opt \
--acoustic-scale=$acwt \
--extra-left-context=$extra_left_context \
--extra-right-context=$extra_right_context \
--extra-left-context-initial=$extra_left_context_initial \
--extra-right-context-final=$extra_right_context_final \
--frames-per-chunk=$frames_per_chunk \
--use-gpu=$use_gpu --use-priors=true \
$srcdir/final.mdl "$feats" "$posteriors"
fi

if [ $stage -le 2 ]; then
suffix=
if $combine_version ; then
suffix="-combine"
fi
$cmd JOB=1:$nj $dir/log/decode.JOB.log \
latgen-faster-mapped$suffix --max-active=$maxactive --beam=$beam --lattice-beam=$lattice_beam \
--acoustic-scale=$acwt --allow-partial=true --word-symbol-table=$graphdir/words.txt \
$srcdir/final.mdl $graphdir/HCLG.fst "$posteriors_scp" "$lat_wspecifier" || exit 1;
fi

if ! $skip_scoring ; then
[ ! -x local/score.sh ] && \
echo "Not scoring because local/score.sh does not exist or not executable." && exit 1;
local/score.sh --cmd "$cmd" $data $graphdir $dir ||
{ echo "$0: Scoring failed. (ignore by '--skip-scoring true')"; exit 1; }
fi

exit 0;
2 changes: 1 addition & 1 deletion src/bin/Makefile
Original file line number Diff line number Diff line change
Expand Up @@ -22,7 +22,7 @@ BINFILES = align-equal align-equal-compiled acc-tree-stats \
matrix-sum build-pfile-from-ali get-post-on-ali tree-info am-info \
vector-sum matrix-sum-rows est-pca sum-lda-accs sum-mllt-accs \
transform-vec align-text matrix-dim post-to-smat compile-graph \
compare-int-vector
compare-int-vector latgen-faster-mapped-combine


OBJFILES =
Expand Down
179 changes: 179 additions & 0 deletions src/bin/latgen-faster-mapped-combine.cc
Original file line number Diff line number Diff line change
@@ -0,0 +1,179 @@
// bin/latgen-faster-mapped.cc

// Copyright 2009-2012 Microsoft Corporation, Karel Vesely
// 2013 Johns Hopkins University (author: Daniel Povey)
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is OK for testing but eventually this should just be a change to lattice-faster-decoder.cc. It's just a small optimization.

// 2014 Guoguo Chen

// See ../../COPYING for clarification regarding multiple authors
//
// 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.


#include "base/kaldi-common.h"
#include "util/common-utils.h"
#include "tree/context-dep.h"
#include "hmm/transition-model.h"
#include "fstext/fstext-lib.h"
#include "decoder/decoder-wrappers.h"
#include "decoder/decodable-matrix.h"
#include "base/timer.h"


int main(int argc, char *argv[]) {
try {
using namespace kaldi;
typedef kaldi::int32 int32;
using fst::SymbolTable;
using fst::Fst;
using fst::StdArc;

const char *usage =
"Generate lattices, reading log-likelihoods as matrices\n"
" (model is needed only for the integer mappings in its transition-model)\n"
"Usage: latgen-faster-mapped [options] trans-model-in (fst-in|fsts-rspecifier) loglikes-rspecifier"
" lattice-wspecifier [ words-wspecifier [alignments-wspecifier] ]\n";
ParseOptions po(usage);
Timer timer;
bool allow_partial = false;
BaseFloat acoustic_scale = 0.1;
LatticeFasterDecoderCombineConfig config;

std::string word_syms_filename;
config.Register(&po);
po.Register("acoustic-scale", &acoustic_scale, "Scaling factor for acoustic likelihoods");

po.Register("word-symbol-table", &word_syms_filename, "Symbol table for words [for debug output]");
po.Register("allow-partial", &allow_partial, "If true, produce output even if end state was not reached.");

po.Read(argc, argv);

if (po.NumArgs() < 4 || po.NumArgs() > 6) {
po.PrintUsage();
exit(1);
}

std::string model_in_filename = po.GetArg(1),
fst_in_str = po.GetArg(2),
feature_rspecifier = po.GetArg(3),
lattice_wspecifier = po.GetArg(4),
words_wspecifier = po.GetOptArg(5),
alignment_wspecifier = po.GetOptArg(6);

TransitionModel trans_model;
ReadKaldiObject(model_in_filename, &trans_model);

bool determinize = config.determinize_lattice;
CompactLatticeWriter compact_lattice_writer;
LatticeWriter lattice_writer;
if (! (determinize ? compact_lattice_writer.Open(lattice_wspecifier)
: lattice_writer.Open(lattice_wspecifier)))
KALDI_ERR << "Could not open table for writing lattices: "
<< lattice_wspecifier;

Int32VectorWriter words_writer(words_wspecifier);

Int32VectorWriter alignment_writer(alignment_wspecifier);

fst::SymbolTable *word_syms = NULL;
if (word_syms_filename != "")
if (!(word_syms = fst::SymbolTable::ReadText(word_syms_filename)))
KALDI_ERR << "Could not read symbol table from file "
<< word_syms_filename;

double tot_like = 0.0;
kaldi::int64 frame_count = 0;
int num_success = 0, num_fail = 0;

if (ClassifyRspecifier(fst_in_str, NULL, NULL) == kNoRspecifier) {
SequentialBaseFloatMatrixReader loglike_reader(feature_rspecifier);
// Input FST is just one FST, not a table of FSTs.
Fst<StdArc> *decode_fst = fst::ReadFstKaldiGeneric(fst_in_str);
timer.Reset();

{
LatticeFasterDecoderCombine decoder(*decode_fst, config);

for (; !loglike_reader.Done(); loglike_reader.Next()) {
std::string utt = loglike_reader.Key();
Matrix<BaseFloat> loglikes (loglike_reader.Value());
loglike_reader.FreeCurrent();
if (loglikes.NumRows() == 0) {
KALDI_WARN << "Zero-length utterance: " << utt;
num_fail++;
continue;
}

DecodableMatrixScaledMapped decodable(trans_model, loglikes, acoustic_scale);

double like;
if (DecodeUtteranceLatticeFasterCombine(
decoder, decodable, trans_model, word_syms, utt,
acoustic_scale, determinize, allow_partial, &alignment_writer,
&words_writer, &compact_lattice_writer, &lattice_writer,
&like)) {
tot_like += like;
frame_count += loglikes.NumRows();
num_success++;
} else num_fail++;
}
}
delete decode_fst; // delete this only after decoder goes out of scope.
} else { // We have different FSTs for different utterances.
SequentialTableReader<fst::VectorFstHolder> fst_reader(fst_in_str);
RandomAccessBaseFloatMatrixReader loglike_reader(feature_rspecifier);
for (; !fst_reader.Done(); fst_reader.Next()) {
std::string utt = fst_reader.Key();
if (!loglike_reader.HasKey(utt)) {
KALDI_WARN << "Not decoding utterance " << utt
<< " because no loglikes available.";
num_fail++;
continue;
}
const Matrix<BaseFloat> &loglikes = loglike_reader.Value(utt);
if (loglikes.NumRows() == 0) {
KALDI_WARN << "Zero-length utterance: " << utt;
num_fail++;
continue;
}
LatticeFasterDecoderCombine decoder(fst_reader.Value(), config);
DecodableMatrixScaledMapped decodable(trans_model, loglikes, acoustic_scale);
double like;
if (DecodeUtteranceLatticeFasterCombine(
decoder, decodable, trans_model, word_syms, utt, acoustic_scale,
determinize, allow_partial, &alignment_writer, &words_writer,
&compact_lattice_writer, &lattice_writer, &like)) {
tot_like += like;
frame_count += loglikes.NumRows();
num_success++;
} else num_fail++;
}
}

double elapsed = timer.Elapsed();
KALDI_LOG << "Time taken "<< elapsed
<< "s: real-time factor assuming 100 frames/sec is "
<< (elapsed*100.0/frame_count);
KALDI_LOG << "Done " << num_success << " utterances, failed for "
<< num_fail;
KALDI_LOG << "Overall log-likelihood per frame is " << (tot_like/frame_count) << " over "
<< frame_count<<" frames.";

delete word_syms;
if (num_success != 0) return 0;
else return 1;
} catch(const std::exception &e) {
std::cerr << e.what();
return -1;
}
}
2 changes: 1 addition & 1 deletion src/decoder/Makefile
Original file line number Diff line number Diff line change
Expand Up @@ -7,7 +7,7 @@ TESTFILES =

OBJFILES = training-graph-compiler.o lattice-simple-decoder.o lattice-faster-decoder.o \
lattice-faster-online-decoder.o simple-decoder.o faster-decoder.o \
decoder-wrappers.o grammar-fst.o decodable-matrix.o
decoder-wrappers.o grammar-fst.o decodable-matrix.o lattice-faster-decoder-combine.o

LIBNAME = kaldi-decoder

Expand Down
Loading