From b33b3a6732c6b6a66bd5c44c615be56d66f4ed67 Mon Sep 17 00:00:00 2001 From: Yimmon Zhuang Date: Wed, 14 Oct 2015 15:37:20 +0800 Subject: support kaldi decoder --- kaldi_decode/utils/int2sym.pl | 71 +++++ kaldi_decode/utils/parse_options.sh | 97 ++++++ kaldi_decode/utils/queue.pl | 580 ++++++++++++++++++++++++++++++++++++ kaldi_decode/utils/run.pl | 264 ++++++++++++++++ kaldi_decode/utils/split_data.sh | 135 +++++++++ 5 files changed, 1147 insertions(+) create mode 100755 kaldi_decode/utils/int2sym.pl create mode 100755 kaldi_decode/utils/parse_options.sh create mode 100755 kaldi_decode/utils/queue.pl create mode 100755 kaldi_decode/utils/run.pl create mode 100755 kaldi_decode/utils/split_data.sh (limited to 'kaldi_decode/utils') diff --git a/kaldi_decode/utils/int2sym.pl b/kaldi_decode/utils/int2sym.pl new file mode 100755 index 0000000..d618939 --- /dev/null +++ b/kaldi_decode/utils/int2sym.pl @@ -0,0 +1,71 @@ +#!/usr/bin/env perl +# Copyright 2010-2012 Microsoft Corporation Johns Hopkins University (Author: Daniel Povey) +# Apache 2.0. + +undef $field_begin; +undef $field_end; + + +if ($ARGV[0] eq "-f") { + shift @ARGV; + $field_spec = shift @ARGV; + if ($field_spec =~ m/^\d+$/) { + $field_begin = $field_spec - 1; $field_end = $field_spec - 1; + } + if ($field_spec =~ m/^(\d*)[-:](\d*)/) { # accept e.g. 1:10 as a courtesty (properly, 1-10) + if ($1 ne "") { + $field_begin = $1 - 1; # Change to zero-based indexing. + } + if ($2 ne "") { + $field_end = $2 - 1; # Change to zero-based indexing. + } + } + if (!defined $field_begin && !defined $field_end) { + die "Bad argument to -f option: $field_spec"; + } +} +$symtab = shift @ARGV; +if(!defined $symtab) { + print STDERR "Usage: sym2int.pl [options] symtab [input] > output\n" . + "options: [-f (|-)]\n" . + "e.g.: -f 2, or -f 3-4\n"; + exit(1); +} + +open(F, "<$symtab") || die "Error opening symbol table file $symtab"; +while() { + @A = split(" ", $_); + @A == 2 || die "bad line in symbol table file: $_"; + $int2sym{$A[1]} = $A[0]; +} + +sub int2sym { + my $a = shift @_; + my $pos = shift @_; + if($a !~ m:^\d+$:) { # not all digits.. + $pos1 = $pos+1; # make it one-based. + die "int2sym.pl: found noninteger token $a [in position $pos1]\n"; + } + $s = $int2sym{$a}; + if(!defined ($s)) { + die "int2sym.pl: integer $a not in symbol table $symtab."; + } + return $s; +} + +$error = 0; +while (<>) { + @A = split(" ", $_); + for ($pos = 0; $pos <= $#A; $pos++) { + $a = $A[$pos]; + if ( (!defined $field_begin || $pos >= $field_begin) + && (!defined $field_end || $pos <= $field_end)) { + $a = int2sym($a, $pos); + } + print $a . " "; + } + print "\n"; +} + + + diff --git a/kaldi_decode/utils/parse_options.sh b/kaldi_decode/utils/parse_options.sh new file mode 100755 index 0000000..fdc8a36 --- /dev/null +++ b/kaldi_decode/utils/parse_options.sh @@ -0,0 +1,97 @@ +#!/bin/bash + +# Copyright 2012 Johns Hopkins University (Author: Daniel Povey); +# Arnab Ghoshal, Karel Vesely + +# 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. + + +# Parse command-line options. +# To be sourced by another script (as in ". parse_options.sh"). +# Option format is: --option-name arg +# and shell variable "option_name" gets set to value "arg." +# The exception is --help, which takes no arguments, but prints the +# $help_message variable (if defined). + + +### +### The --config file options have lower priority to command line +### options, so we need to import them first... +### + +# Now import all the configs specified by command-line, in left-to-right order +for ((argpos=1; argpos<$#; argpos++)); do + if [ "${!argpos}" == "--config" ]; then + argpos_plus1=$((argpos+1)) + config=${!argpos_plus1} + [ ! -r $config ] && echo "$0: missing config '$config'" && exit 1 + . $config # source the config file. + fi +done + + +### +### No we process the command line options +### +while true; do + [ -z "${1:-}" ] && break; # break if there are no arguments + case "$1" in + # If the enclosing script is called with --help option, print the help + # message and exit. Scripts should put help messages in $help_message + --help|-h) if [ -z "$help_message" ]; then echo "No help found." 1>&2; + else printf "$help_message\n" 1>&2 ; fi; + exit 0 ;; + --*=*) echo "$0: options to scripts must be of the form --name value, got '$1'" + exit 1 ;; + # If the first command-line argument begins with "--" (e.g. --foo-bar), + # then work out the variable name as $name, which will equal "foo_bar". + --*) name=`echo "$1" | sed s/^--// | sed s/-/_/g`; + # Next we test whether the variable in question is undefned-- if so it's + # an invalid option and we die. Note: $0 evaluates to the name of the + # enclosing script. + # The test [ -z ${foo_bar+xxx} ] will return true if the variable foo_bar + # is undefined. We then have to wrap this test inside "eval" because + # foo_bar is itself inside a variable ($name). + eval '[ -z "${'$name'+xxx}" ]' && echo "$0: invalid option $1" 1>&2 && exit 1; + + oldval="`eval echo \\$$name`"; + # Work out whether we seem to be expecting a Boolean argument. + if [ "$oldval" == "true" ] || [ "$oldval" == "false" ]; then + was_bool=true; + else + was_bool=false; + fi + + # Set the variable to the right value-- the escaped quotes make it work if + # the option had spaces, like --cmd "queue.pl -sync y" + eval $name=\"$2\"; + + # Check that Boolean-valued arguments are really Boolean. + if $was_bool && [[ "$2" != "true" && "$2" != "false" ]]; then + echo "$0: expected \"true\" or \"false\": $1 $2" 1>&2 + exit 1; + fi + shift 2; + ;; + *) break; + esac +done + + +# Check for an empty argument to the --cmd option, which can easily occur as a +# result of scripting errors. +[ ! -z "${cmd+xxx}" ] && [ -z "$cmd" ] && echo "$0: empty argument to --cmd option" 1>&2 && exit 1; + + +true; # so this script returns exit code 0. diff --git a/kaldi_decode/utils/queue.pl b/kaldi_decode/utils/queue.pl new file mode 100755 index 0000000..1e36de6 --- /dev/null +++ b/kaldi_decode/utils/queue.pl @@ -0,0 +1,580 @@ +#!/usr/bin/env perl +use strict; +use warnings; + +# Copyright 2012 Johns Hopkins University (Author: Daniel Povey). +# 2014 Vimal Manohar (Johns Hopkins University) +# Apache 2.0. + +use File::Basename; +use Cwd; +use Getopt::Long; + +# queue.pl has the same functionality as run.pl, except that +# it runs the job in question on the queue (Sun GridEngine). +# This version of queue.pl uses the task array functionality +# of the grid engine. Note: it's different from the queue.pl +# in the s4 and earlier scripts. + +# The script now supports configuring the queue system using a config file +# (default in conf/queue.conf; but can be passed specified with --config option) +# and a set of command line options. +# The current script handles: +# 1) Normal configuration arguments +# For e.g. a command line option of "--gpu 1" could be converted into the option +# "-q g.q -l gpu=1" to qsub. How the CLI option is handled is determined by a +# line in the config file like +# gpu=* -q g.q -l gpu=$0 +# $0 here in the line is replaced with the argument read from the CLI and the +# resulting string is passed to qsub. +# 2) Special arguments to options such as +# gpu=0 +# If --gpu 0 is given in the command line, then no special "-q" is given. +# 3) Default argument +# default gpu=0 +# If --gpu option is not passed in the command line, then the script behaves as +# if --gpu 0 was passed since 0 is specified as the default argument for that +# option +# 4) Arbitrary options and arguments. +# Any command line option starting with '--' and its argument would be handled +# as long as its defined in the config file. +# 5) Default behavior +# If the config file that is passed using is not readable, then the script +# behaves as if the queue has the following config file: +# $ cat conf/queue.conf +# # Default configuration +# command qsub -v PATH -cwd -S /bin/bash -j y -l arch=*64* +# option mem=* -l mem_free=$0,ram_free=$0 +# option mem=0 # Do not add anything to qsub_opts +# option num_threads=* -pe smp $0 +# option num_threads=1 # Do not add anything to qsub_opts +# option max_jobs_run=* -tc $0 +# default gpu=0 +# option gpu=0 -q all.q +# option gpu=* -l gpu=$0 -q g.q + +my $qsub_opts = ""; +my $sync = 0; +my $num_threads = 1; +my $gpu = 0; + +my $config = "conf/queue.conf"; + +my %cli_options = (); + +my $jobname; +my $jobstart; +my $jobend; + +my $array_job = 0; + +sub print_usage() { + print STDERR + "Usage: queue.pl [options] [JOB=1:n] log-file command-line arguments...\n" . + "e.g.: queue.pl foo.log echo baz\n" . + " (which will echo \"baz\", with stdout and stderr directed to foo.log)\n" . + "or: queue.pl -q all.q\@xyz foo.log echo bar \| sed s/bar/baz/ \n" . + " (which is an example of using a pipe; you can provide other escaped bash constructs)\n" . + "or: queue.pl -q all.q\@qyz JOB=1:10 foo.JOB.log echo JOB \n" . + " (which illustrates the mechanism to submit parallel jobs; note, you can use \n" . + " another string other than JOB)\n" . + "Note: if you pass the \"-sync y\" option to qsub, this script will take note\n" . + "and change its behavior. Otherwise it uses qstat to work out when the job finished\n" . + "Options:\n" . + " --config (default: $config)\n" . + " --mem (e.g. --mem 2G, --mem 500M, \n" . + " also support K and numbers mean bytes)\n" . + " --num-threads (default: $num_threads)\n" . + " --max-jobs-run \n" . + " --gpu <0|1> (default: $gpu)\n"; + exit 1; +} + +if (@ARGV < 2) { + print_usage(); +} + +for (my $x = 1; $x <= 3; $x++) { # This for-loop is to + # allow the JOB=1:n option to be interleaved with the + # options to qsub. + while (@ARGV >= 2 && $ARGV[0] =~ m:^-:) { + my $switch = shift @ARGV; + + if ($switch eq "-V") { + $qsub_opts .= "-V "; + } else { + my $argument = shift @ARGV; + if ($argument =~ m/^--/) { + print STDERR "WARNING: suspicious argument '$argument' to $switch; starts with '-'\n"; + } + if ($switch eq "-sync" && $argument =~ m/^[yY]/) { + $sync = 1; + $qsub_opts .= "$switch $argument "; + } elsif ($switch eq "-pe") { # e.g. -pe smp 5 + my $argument2 = shift @ARGV; + $qsub_opts .= "$switch $argument $argument2 "; + $num_threads = $argument2; + } elsif ($switch =~ m/^--/) { # Config options + # Convert CLI option to variable name + # by removing '--' from the switch and replacing any + # '-' with a '_' + $switch =~ s/^--//; + $switch =~ s/-/_/g; + $cli_options{$switch} = $argument; + } else { # Other qsub options - passed as is + $qsub_opts .= "$switch $argument "; + } + } + } + if ($ARGV[0] =~ m/^([\w_][\w\d_]*)+=(\d+):(\d+)$/) { # e.g. JOB=1:20 + $array_job = 1; + $jobname = $1; + $jobstart = $2; + $jobend = $3; + shift; + if ($jobstart > $jobend) { + die "queue.pl: invalid job range $ARGV[0]"; + } + if ($jobstart <= 0) { + die "run.pl: invalid job range $ARGV[0], start must be strictly positive (this is a GridEngine limitation)."; + } + } elsif ($ARGV[0] =~ m/^([\w_][\w\d_]*)+=(\d+)$/) { # e.g. JOB=1. + $array_job = 1; + $jobname = $1; + $jobstart = $2; + $jobend = $2; + shift; + } elsif ($ARGV[0] =~ m/.+\=.*\:.*$/) { + print STDERR "Warning: suspicious first argument to queue.pl: $ARGV[0]\n"; + } +} + +if (@ARGV < 2) { + print_usage(); +} + +if (exists $cli_options{"config"}) { + $config = $cli_options{"config"}; +} + +my $default_config_file = <<'EOF'; +# Default configuration +command qsub -v PATH -cwd -S /bin/bash -j y -l arch=*64* +option mem=* -l mem_free=$0,ram_free=$0 +option mem=0 # Do not add anything to qsub_opts +option num_threads=* -pe smp $0 +option num_threads=1 # Do not add anything to qsub_opts +option max_jobs_run=* -tc $0 +default gpu=0 +option gpu=0 +option gpu=* -l gpu=$0 -q g.q +EOF + +# Here the configuration options specified by the user on the command line +# (e.g. --mem 2G) are converted to options to the qsub system as defined in +# the config file. (e.g. if the config file has the line +# "option mem=* -l ram_free=$0,mem_free=$0" +# and the user has specified '--mem 2G' on the command line, the options +# passed to queue system would be "-l ram_free=2G,mem_free=2G +# A more detailed description of the ways the options would be handled is at +# the top of this file. + +my $opened_config_file = 1; + +open CONFIG, "<$config" or $opened_config_file = 0; + +my %cli_config_options = (); +my %cli_default_options = (); + +if ($opened_config_file == 0 && exists($cli_options{"config"})) { + print STDERR "Could not open config file $config\n"; + exit(1); +} elsif ($opened_config_file == 0 && !exists($cli_options{"config"})) { + # Open the default config file instead + open (CONFIG, "echo '$default_config_file' |") or die "Unable to open pipe\n"; + $config = "Default config"; +} + +my $qsub_cmd = ""; +my $read_command = 0; + +while() { + chomp; + my $line = $_; + $_ =~ s/\s*#.*//g; + if ($_ eq "") { next; } + if ($_ =~ /^command (.+)/) { + $read_command = 1; + $qsub_cmd = $1 . " "; + } elsif ($_ =~ m/^option ([^=]+)=\* (.+)$/) { + # Config option that needs replacement with parameter value read from CLI + # e.g.: option mem=* -l mem_free=$0,ram_free=$0 + my $option = $1; # mem + my $arg= $2; # -l mem_free=$0,ram_free=$0 + if ($arg !~ m:\$0:) { + die "Unable to parse line '$line' in config file ($config)\n"; + } + if (exists $cli_options{$option}) { + # Replace $0 with the argument read from command line. + # e.g. "-l mem_free=$0,ram_free=$0" -> "-l mem_free=2G,ram_free=2G" + $arg =~ s/\$0/$cli_options{$option}/g; + $cli_config_options{$option} = $arg; + } + } elsif ($_ =~ m/^option ([^=]+)=(\S+)\s?(.*)$/) { + # Config option that does not need replacement + # e.g. option gpu=0 -q all.q + my $option = $1; # gpu + my $value = $2; # 0 + my $arg = $3; # -q all.q + if (exists $cli_options{$option}) { + $cli_default_options{($option,$value)} = $arg; + } + } elsif ($_ =~ m/^default (\S+)=(\S+)/) { + # Default options. Used for setting default values to options i.e. when + # the user does not specify the option on the command line + # e.g. default gpu=0 + my $option = $1; # gpu + my $value = $2; # 0 + if (!exists $cli_options{$option}) { + # If the user has specified this option on the command line, then we + # don't have to do anything + $cli_options{$option} = $value; + } + } else { + print STDERR "queue.pl: unable to parse line '$line' in config file ($config)\n"; + exit(1); + } +} + +close(CONFIG); + +if ($read_command != 1) { + print STDERR "queue.pl: config file ($config) does not contain the line \"command .*\"\n"; + exit(1); +} + +for my $option (keys %cli_options) { + if ($option eq "config") { next; } + if ($option eq "max_jobs_run" && $array_job != 1) { next; } + my $value = $cli_options{$option}; + + if (exists $cli_default_options{($option,$value)}) { + $qsub_opts .= "$cli_default_options{($option,$value)} "; + } elsif (exists $cli_config_options{$option}) { + $qsub_opts .= "$cli_config_options{$option} "; + } else { + if ($opened_config_file == 0) { $config = "default config file"; } + die "queue.pl: Command line option $option not described in $config (or value '$value' not allowed)\n"; + } +} + +my $cwd = getcwd(); +my $logfile = shift @ARGV; + +if ($array_job == 1 && $logfile !~ m/$jobname/ + && $jobend > $jobstart) { + print STDERR "queue.pl: you are trying to run a parallel job but " + . "you are putting the output into just one log file ($logfile)\n"; + exit(1); +} + +# +# Work out the command; quote escaping is done here. +# Note: the rules for escaping stuff are worked out pretty +# arbitrarily, based on what we want it to do. Some things that +# we pass as arguments to queue.pl, such as "|", we want to be +# interpreted by bash, so we don't escape them. Other things, +# such as archive specifiers like 'ark:gunzip -c foo.gz|', we want +# to be passed, in quotes, to the Kaldi program. Our heuristic +# is that stuff with spaces in should be quoted. This doesn't +# always work. +# +my $cmd = ""; + +foreach my $x (@ARGV) { + if ($x =~ m/^\S+$/) { $cmd .= $x . " "; } # If string contains no spaces, take + # as-is. + elsif ($x =~ m:\":) { $cmd .= "'$x' "; } # else if no dbl-quotes, use single + else { $cmd .= "\"$x\" "; } # else use double. +} + +# +# Work out the location of the script file, and open it for writing. +# +my $dir = dirname($logfile); +my $base = basename($logfile); +my $qdir = "$dir/q"; +$qdir =~ s:/(log|LOG)/*q:/q:; # If qdir ends in .../log/q, make it just .../q. +my $queue_logfile = "$qdir/$base"; + +if (!-d $dir) { system "mkdir -p $dir 2>/dev/null"; } # another job may be doing this... +if (!-d $dir) { die "Cannot make the directory $dir\n"; } +# make a directory called "q", +# where we will put the log created by qsub... normally this doesn't contain +# anything interesting, evertyhing goes to $logfile. +if (! -d "$qdir") { + system "mkdir $qdir 2>/dev/null"; + sleep(5); ## This is to fix an issue we encountered in denominator lattice creation, + ## where if e.g. the exp/tri2b_denlats/log/15/q directory had just been + ## created and the job immediately ran, it would die with an error because nfs + ## had not yet synced. I'm also decreasing the acdirmin and acdirmax in our + ## NFS settings to something like 5 seconds. +} + +my $queue_array_opt = ""; +if ($array_job == 1) { # It's an array job. + $queue_array_opt = "-t $jobstart:$jobend"; + $logfile =~ s/$jobname/\$SGE_TASK_ID/g; # This variable will get + # replaced by qsub, in each job, with the job-id. + $cmd =~ s/$jobname/\$\{SGE_TASK_ID\}/g; # same for the command... + $queue_logfile =~ s/\.?$jobname//; # the log file in the q/ subdirectory + # is for the queue to put its log, and this doesn't need the task array subscript + # so we remove it. +} + +# queue_scriptfile is as $queue_logfile [e.g. dir/q/foo.log] but +# with the suffix .sh. +my $queue_scriptfile = $queue_logfile; +($queue_scriptfile =~ s/\.[a-zA-Z]{1,5}$/.sh/) || ($queue_scriptfile .= ".sh"); +if ($queue_scriptfile !~ m:^/:) { + $queue_scriptfile = $cwd . "/" . $queue_scriptfile; # just in case. +} + +# We'll write to the standard input of "qsub" (the file-handle Q), +# the job that we want it to execute. +# Also keep our current PATH around, just in case there was something +# in it that we need (although we also source ./path.sh) + +my $syncfile = "$qdir/done.$$"; + +system("rm $queue_logfile $syncfile 2>/dev/null"); +# +# Write to the script file, and then close it. +# +open(Q, ">$queue_scriptfile") || die "Failed to write to $queue_scriptfile"; + +print Q "#!/bin/bash\n"; +print Q "cd $cwd\n"; +print Q ". ./path.sh\n"; +print Q "( echo '#' Running on \`hostname\`\n"; +print Q " echo '#' Started at \`date\`\n"; +print Q " echo -n '# '; cat <$logfile\n"; +print Q "time1=\`date +\"%s\"\`\n"; +print Q " ( $cmd ) 2>>$logfile >>$logfile\n"; +print Q "ret=\$?\n"; +print Q "time2=\`date +\"%s\"\`\n"; +print Q "echo '#' Accounting: time=\$((\$time2-\$time1)) threads=$num_threads >>$logfile\n"; +print Q "echo '#' Finished at \`date\` with status \$ret >>$logfile\n"; +print Q "[ \$ret -eq 137 ] && exit 100;\n"; # If process was killed (e.g. oom) it will exit with status 137; + # let the script return with status 100 which will put it to E state; more easily rerunnable. +if ($array_job == 0) { # not an array job + print Q "touch $syncfile\n"; # so we know it's done. +} else { + print Q "touch $syncfile.\$SGE_TASK_ID\n"; # touch a bunch of sync-files. +} +print Q "exit \$[\$ret ? 1 : 0]\n"; # avoid status 100 which grid-engine +print Q "## submitted with:\n"; # treats specially. +$qsub_cmd .= "-o $queue_logfile $qsub_opts $queue_array_opt $queue_scriptfile >>$queue_logfile 2>&1"; +print Q "# $qsub_cmd\n"; +if (!close(Q)) { # close was not successful... || die "Could not close script file $shfile"; + die "Failed to close the script file (full disk?)"; +} + +my $ret = system ($qsub_cmd); +if ($ret != 0) { + if ($sync && $ret == 256) { # this is the exit status when a job failed (bad exit status) + if (defined $jobname) { $logfile =~ s/\$SGE_TASK_ID/*/g; } + print STDERR "queue.pl: job writing to $logfile failed\n"; + } else { + print STDERR "queue.pl: error submitting jobs to queue (return status was $ret)\n"; + print STDERR "queue log file is $queue_logfile, command was $qsub_cmd\n"; + print STDERR `tail $queue_logfile`; + } + exit(1); +} + +my $sge_job_id; +if (! $sync) { # We're not submitting with -sync y, so we + # need to wait for the jobs to finish. We wait for the + # sync-files we "touched" in the script to exist. + my @syncfiles = (); + if (!defined $jobname) { # not an array job. + push @syncfiles, $syncfile; + } else { + for (my $jobid = $jobstart; $jobid <= $jobend; $jobid++) { + push @syncfiles, "$syncfile.$jobid"; + } + } + # We will need the sge_job_id, to check that job still exists + { # Get the SGE job-id from the log file in q/ + open(L, "<$queue_logfile") || die "Error opening log file $queue_logfile"; + undef $sge_job_id; + while () { + if (m/Your job\S* (\d+)[. ].+ has been submitted/) { + if (defined $sge_job_id) { + die "Error: your job was submitted more than once (see $queue_logfile)"; + } else { + $sge_job_id = $1; + } + } + } + close(L); + if (!defined $sge_job_id) { + die "Error: log file $queue_logfile does not specify the SGE job-id."; + } + } + my $check_sge_job_ctr=1; + # + my $wait = 0.1; + my $counter = 0; + foreach my $f (@syncfiles) { + # wait for them to finish one by one. + while (! -f $f) { + sleep($wait); + $wait *= 1.2; + if ($wait > 3.0) { + $wait = 3.0; # never wait more than 3 seconds. + # the following (.kick) commands are basically workarounds for NFS bugs. + if (rand() < 0.25) { # don't do this every time... + if (rand() > 0.5) { + system("touch $qdir/.kick"); + } else { + system("rm $qdir/.kick 2>/dev/null"); + } + } + if ($counter++ % 10 == 0) { + # This seems to kick NFS in the teeth to cause it to refresh the + # directory. I've seen cases where it would indefinitely fail to get + # updated, even though the file exists on the server. + # Only do this every 10 waits (every 30 seconds) though, or if there + # are many jobs waiting they can overwhelm the file server. + system("ls $qdir >/dev/null"); + } + } + + # Check that the job exists in SGE. Job can be killed if duration + # exceeds some hard limit, or in case of a machine shutdown. + if (($check_sge_job_ctr++ % 10) == 0) { # Don't run qstat too often, avoid stress on SGE. + if ( -f $f ) { next; }; #syncfile appeared: OK. + $ret = system("qstat -j $sge_job_id >/dev/null 2>/dev/null"); + # system(...) : To get the actual exit value, shift $ret right by eight bits. + if ($ret>>8 == 1) { # Job does not seem to exist + # Don't consider immediately missing job as error, first wait some + # time to make sure it is not just delayed creation of the syncfile. + + sleep(3); + # Sometimes NFS gets confused and thinks it's transmitted the directory + # but it hasn't, due to timestamp issues. Changing something in the + # directory will usually fix that. + system("touch $qdir/.kick"); + system("rm $qdir/.kick 2>/dev/null"); + if ( -f $f ) { next; } #syncfile appeared, ok + sleep(7); + system("touch $qdir/.kick"); + sleep(1); + system("rm $qdir/.kick 2>/dev/null"); + if ( -f $f ) { next; } #syncfile appeared, ok + sleep(60); + system("touch $qdir/.kick"); + sleep(1); + system("rm $qdir/.kick 2>/dev/null"); + if ( -f $f ) { next; } #syncfile appeared, ok + $f =~ m/\.(\d+)$/ || die "Bad sync-file name $f"; + my $job_id = $1; + if (defined $jobname) { + $logfile =~ s/\$SGE_TASK_ID/$job_id/g; + } + my $last_line = `tail -n 1 $logfile`; + if ($last_line =~ m/status 0$/ && (-M $logfile) < 0) { + # if the last line of $logfile ended with "status 0" and + # $logfile is newer than this program [(-M $logfile) gives the + # time elapsed between file modification and the start of this + # program], then we assume the program really finished OK, + # and maybe something is up with the file system. + print STDERR "**queue.pl: syncfile $f was not created but job seems\n" . + "**to have finished OK. Probably your file-system has problems.\n" . + "**This is just a warning.\n"; + last; + } else { + chop $last_line; + print STDERR "queue.pl: Error, unfinished job no " . + "longer exists, log is in $logfile, last line is '$last_line', " . + "syncfile is $f, return status of qstat was $ret\n" . + "Possible reasons: a) Exceeded time limit? -> Use more jobs!" . + " b) Shutdown/Frozen machine? -> Run again!\n"; + exit(1); + } + } elsif ($ret != 0) { + print STDERR "queue.pl: Warning: qstat command returned status $ret (qstat -j $sge_job_id,$!)\n"; + } + } + } + } + my $all_syncfiles = join(" ", @syncfiles); + system("rm $all_syncfiles 2>/dev/null"); +} + +# OK, at this point we are synced; we know the job is done. +# But we don't know about its exit status. We'll look at $logfile for this. +# First work out an array @logfiles of file-locations we need to +# read (just one, unless it's an array job). +my @logfiles = (); +if (!defined $jobname) { # not an array job. + push @logfiles, $logfile; +} else { + for (my $jobid = $jobstart; $jobid <= $jobend; $jobid++) { + my $l = $logfile; + $l =~ s/\$SGE_TASK_ID/$jobid/g; + push @logfiles, $l; + } +} + +my $num_failed = 0; +my $status = 1; +foreach my $l (@logfiles) { + my @wait_times = (0.1, 0.2, 0.2, 0.3, 0.5, 0.5, 1.0, 2.0, 5.0, 5.0, 5.0, 10.0, 25.0); + for (my $iter = 0; $iter <= @wait_times; $iter++) { + my $line = `tail -10 $l 2>/dev/null`; # Note: although this line should be the last + # line of the file, I've seen cases where it was not quite the last line because + # of delayed output by the process that was running, or processes it had called. + # so tail -10 gives it a little leeway. + if ($line =~ m/with status (\d+)/) { + $status = $1; + last; + } else { + if ($iter < @wait_times) { + sleep($wait_times[$iter]); + } else { + if (! -f $l) { + print STDERR "Log-file $l does not exist.\n"; + } else { + print STDERR "The last line of log-file $l does not seem to indicate the " + . "return status as expected\n"; + } + exit(1); # Something went wrong with the queue, or the + # machine it was running on, probably. + } + } + } + # OK, now we have $status, which is the return-status of + # the command in the job. + if ($status != 0) { $num_failed++; } +} +if ($num_failed == 0) { exit(0); } +else { # we failed. + if (@logfiles == 1) { + if (defined $jobname) { $logfile =~ s/\$SGE_TASK_ID/$jobstart/g; } + print STDERR "queue.pl: job failed with status $status, log is in $logfile\n"; + if ($logfile =~ m/JOB/) { + print STDERR "queue.pl: probably you forgot to put JOB=1:\$nj in your script.\n"; + } + } else { + if (defined $jobname) { $logfile =~ s/\$SGE_TASK_ID/*/g; } + my $numjobs = 1 + $jobend - $jobstart; + print STDERR "queue.pl: $num_failed / $numjobs failed, log is in $logfile\n"; + } + exit(1); +} diff --git a/kaldi_decode/utils/run.pl b/kaldi_decode/utils/run.pl new file mode 100755 index 0000000..6145a7a --- /dev/null +++ b/kaldi_decode/utils/run.pl @@ -0,0 +1,264 @@ +#!/usr/bin/env perl +use warnings; #sed replacement for -w perl parameter + +# In general, doing +# run.pl some.log a b c is like running the command a b c in +# the bash shell, and putting the standard error and output into some.log. +# To run parallel jobs (backgrounded on the host machine), you can do (e.g.) +# run.pl JOB=1:4 some.JOB.log a b c JOB is like running the command a b c JOB +# and putting it in some.JOB.log, for each one. [Note: JOB can be any identifier]. +# If any of the jobs fails, this script will fail. + +# A typical example is: +# run.pl some.log my-prog "--opt=foo bar" foo \| other-prog baz +# and run.pl will run something like: +# ( my-prog '--opt=foo bar' foo | other-prog baz ) >& some.log +# +# Basically it takes the command-line arguments, quotes them +# as necessary to preserve spaces, and evaluates them with bash. +# In addition it puts the command line at the top of the log, and +# the start and end times of the command at the beginning and end. +# The reason why this is useful is so that we can create a different +# version of this program that uses a queueing system instead. + +# use Data::Dumper; + +@ARGV < 2 && die "usage: run.pl log-file command-line arguments..."; + + +$max_jobs_run = -1; +$jobstart = 1; +$jobend = 1; +$ignored_opts = ""; # These will be ignored. + +# First parse an option like JOB=1:4, and any +# options that would normally be given to +# queue.pl, which we will just discard. + +if (@ARGV > 0) { + while (@ARGV >= 2 && $ARGV[0] =~ m:^-:) { # parse any options + # that would normally go to qsub, but which will be ignored here. + $switch = shift @ARGV; + if ($switch eq "-V") { + $ignored_opts .= "-V "; + } elsif ($switch eq "--max-jobs-run" || $switch eq "-tc") { + # we do support the option --max-jobs-run n, and its GridEngine form -tc n. + $max_jobs_run = shift @ARGV; + if (! ($max_jobs_run > 0)) { + die "run.pl: invalid option --max-jobs-run $max_jobs_run"; + } + } else { + $option = shift @ARGV; + if ($switch eq "-sync" && $option =~ m/^[yY]/) { + $ignored_opts .= "-sync "; # Note: in the + # corresponding code in queue.pl it says instead, just "$sync = 1;". + } + $ignored_opts .= "$switch $option "; + if ($switch eq "-pe") { # e.g. -pe smp 5 + $option2 = shift @ARGV; + $ignored_opts .= "$option2 "; + } + } + } + if ($ARGV[0] =~ m/^([\w_][\w\d_]*)+=(\d+):(\d+)$/) { # e.g. JOB=1:10 + $jobname = $1; + $jobstart = $2; + $jobend = $3; + shift; + if ($jobstart > $jobend) { + die "run.pl: invalid job range $ARGV[0]"; + } + if ($jobstart <= 0) { + die "run.pl: invalid job range $ARGV[0], start must be strictly positive (this is required for GridEngine compatibility)."; + } + } elsif ($ARGV[0] =~ m/^([\w_][\w\d_]*)+=(\d+)$/) { # e.g. JOB=1. + $jobname = $1; + $jobstart = $2; + $jobend = $2; + shift; + } elsif ($ARGV[0] =~ m/.+\=.*\:.*$/) { + print STDERR "run.pl: Warning: suspicious first argument to run.pl: $ARGV[0]\n"; + } +} + +# Users found this message confusing so we are removing it. +# if ($ignored_opts ne "") { +# print STDERR "run.pl: Warning: ignoring options \"$ignored_opts\"\n"; +# } + +if ($max_jobs_run == -1) { # If --max-jobs-run option not set, + # then work out the number of processors if possible, + # and set it based on that. + $max_jobs_run = 0; + if (open(P, ") { if (m/^processor/) { $max_jobs_run++; } } + if ($max_jobs_run == 0) { + print STDERR "run.pl: Warning: failed to detect any processors from /proc/cpuinfo\n"; + $max_jobs_run = 10; # reasonable default. + } + close(P); + } elsif (open(P, "sysctl -a |")) { # BSD/Darwin + while (

) { + if (m/hw\.ncpu\s*[:=]\s*(\d+)/) { # hw.ncpu = 4, or hw.ncpu: 4 + $max_jobs_run = $1; + last; + } + } + close(P); + if ($max_jobs_run == 0) { + print STDERR "run.pl: Warning: failed to detect any processors from sysctl -a\n"; + $max_jobs_run = 10; # reasonable default. + } + } else { + # allow at most 32 jobs at once, on non-UNIX systems; change this code + # if you need to change this default. + $max_jobs_run = 32; + } + # The just-computed value of $max_jobs_run is just the number of processors + # (or our best guess); and if it happens that the number of jobs we need to + # run is just slightly above $max_jobs_run, it will make sense to increase + # $max_jobs_run to equal the number of jobs, so we don't have a small number + # of leftover jobs. + $num_jobs = $jobend - $jobstart + 1; + if ($num_jobs > $max_jobs_run && $num_jobs < 1.4 * $max_jobs_run) { + $max_jobs_run = $num_jobs; + } +} + +$logfile = shift @ARGV; + +if (defined $jobname && $logfile !~ m/$jobname/ && + $jobend > $jobstart) { + print STDERR "run.pl: you are trying to run a parallel job but " + . "you are putting the output into just one log file ($logfile)\n"; + exit(1); +} + +$cmd = ""; + +foreach $x (@ARGV) { + if ($x =~ m/^\S+$/) { $cmd .= $x . " "; } + elsif ($x =~ m:\":) { $cmd .= "'$x' "; } + else { $cmd .= "\"$x\" "; } +} + +#$Data::Dumper::Indent=0; +$ret = 0; +$numfail = 0; +%active_pids=(); + +use POSIX ":sys_wait_h"; +for ($jobid = $jobstart; $jobid <= $jobend; $jobid++) { + if (scalar(keys %active_pids) >= $max_jobs_run) { + + # Lets wait for a change in any child's status + # Then we have to work out which child finished + $r = waitpid(-1, 0); + $code = $?; + if ($r < 0 ) { die "run.pl: Error waiting for child process"; } # should never happen. + if ( defined $active_pids{$r} ) { + $jid=$active_pids{$r}; + $fail[$jid]=$code; + if ($code !=0) { $numfail++;} + delete $active_pids{$r}; + # print STDERR "Finished: $r/$jid " . Dumper(\%active_pids) . "\n"; + } else { + die "run.pl: Cannot find the PID of the chold process that just finished."; + } + + # In theory we could do a non-blocking waitpid over all jobs running just + # to find out if only one or more jobs finished during the previous waitpid() + # However, we just omit this and will reap the next one in the next pass + # through the for(;;) cycle + } + $childpid = fork(); + if (!defined $childpid) { die "run.pl: Error forking in run.pl (writing to $logfile)"; } + if ($childpid == 0) { # We're in the child... this branch + # executes the job and returns (possibly with an error status). + if (defined $jobname) { + $cmd =~ s/$jobname/$jobid/g; + $logfile =~ s/$jobname/$jobid/g; + } + system("mkdir -p `dirname $logfile` 2>/dev/null"); + open(F, ">$logfile") || die "run.pl: Error opening log file $logfile"; + print F "# " . $cmd . "\n"; + print F "# Started at " . `date`; + $starttime = `date +'%s'`; + print F "#\n"; + close(F); + + # Pipe into bash.. make sure we're not using any other shell. + open(B, "|bash") || die "run.pl: Error opening shell command"; + print B "( " . $cmd . ") 2>>$logfile >> $logfile"; + close(B); # If there was an error, exit status is in $? + $ret = $?; + + $lowbits = $ret & 127; + $highbits = $ret >> 8; + if ($lowbits != 0) { $return_str = "code $highbits; signal $lowbits" } + else { $return_str = "code $highbits"; } + + $endtime = `date +'%s'`; + open(F, ">>$logfile") || die "run.pl: Error opening log file $logfile (again)"; + $enddate = `date`; + chop $enddate; + print F "# Accounting: time=" . ($endtime - $starttime) . " threads=1\n"; + print F "# Ended ($return_str) at " . $enddate . ", elapsed time " . ($endtime-$starttime) . " seconds\n"; + close(F); + exit($ret == 0 ? 0 : 1); + } else { + $pid[$jobid] = $childpid; + $active_pids{$childpid} = $jobid; + # print STDERR "Queued: " . Dumper(\%active_pids) . "\n"; + } +} + +# Now we have submitted all the jobs, lets wait until all the jobs finish +foreach $child (keys %active_pids) { + $jobid=$active_pids{$child}; + $r = waitpid($pid[$jobid], 0); + $code = $?; + if ($r == -1) { die "run.pl: Error waiting for child process"; } # should never happen. + if ($r != 0) { $fail[$jobid]=$code; $numfail++ if $code!=0; } # Completed successfully +} + +# Some sanity checks: +# The $fail array should not contain undefined codes +# The number of non-zeros in that array should be equal to $numfail +# We cannot do foreach() here, as the JOB ids do not necessarily start by zero +$failed_jids=0; +for ($jobid = $jobstart; $jobid <= $jobend; $jobid++) { + $job_return = $fail[$jobid]; + if (not defined $job_return ) { + # print Dumper(\@fail); + + die "run.pl: Sanity check failed: we have indication that some jobs are running " . + "even after we waited for all jobs to finish" ; + } + if ($job_return != 0 ){ $failed_jids++;} +} +if ($failed_jids != $numfail) { + die "run.pl: Sanity check failed: cannot find out how many jobs failed ($failed_jids x $numfail)." +} +if ($numfail > 0) { $ret = 1; } + +if ($ret != 0) { + $njobs = $jobend - $jobstart + 1; + if ($njobs == 1) { + if (defined $jobname) { + $logfile =~ s/$jobname/$jobstart/; # only one numbered job, so replace name with + # that job. + } + print STDERR "run.pl: job failed, log is in $logfile\n"; + if ($logfile =~ m/JOB/) { + print STDERR "run.pl: probably you forgot to put JOB=1:\$nj in your script."; + } + } + else { + $logfile =~ s/$jobname/*/g; + print STDERR "run.pl: $numfail / $njobs failed, log is in $logfile\n"; + } +} + + +exit ($ret); diff --git a/kaldi_decode/utils/split_data.sh b/kaldi_decode/utils/split_data.sh new file mode 100755 index 0000000..941890c --- /dev/null +++ b/kaldi_decode/utils/split_data.sh @@ -0,0 +1,135 @@ +#!/bin/bash +# Copyright 2010-2013 Microsoft Corporation +# Johns Hopkins University (Author: Daniel Povey) + +# 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. + +split_per_spk=true +if [ "$1" == "--per-utt" ]; then + split_per_spk=false + shift +fi + +if [ $# != 2 ]; then + echo "Usage: split_data.sh [--per-utt] " + echo "This script will not split the data-dir if it detects that the output is newer than the input." + echo "By default it splits per speaker (so each speaker is in only one split dir)," + echo "but with the --per-utt option it will ignore the speaker information while splitting." + exit 1 +fi + +data=$1 +numsplit=$2 + +if [ $numsplit -le 0 ]; then + echo "Invalid num-split argument $numsplit"; + exit 1; +fi + +n=0; +feats="" +wavs="" +utt2spks="" +texts="" + +nu=`cat $data/utt2spk | wc -l` +nf=`cat $data/feats.scp 2>/dev/null | wc -l` +nt=`cat $data/text 2>/dev/null | wc -l` # take it as zero if no such file +if [ -f $data/feats.scp ] && [ $nu -ne $nf ]; then + echo "** split_data.sh: warning, #lines is (utt2spk,feats.scp) is ($nu,$nf); you can " + echo "** use utils/fix_data_dir.sh $data to fix this." +fi +if [ -f $data/text ] && [ $nu -ne $nt ]; then + echo "** split_data.sh: warning, #lines is (utt2spk,text) is ($nu,$nt); you can " + echo "** use utils/fix_data_dir.sh to fix this." +fi + +s1=$data/split$numsplit/1 +if [ ! -d $s1 ]; then + need_to_split=true +else + need_to_split=false + for f in utt2spk spk2utt spk2warp feats.scp text wav.scp cmvn.scp spk2gender \ + vad.scp segments reco2file_and_channel utt2lang; do + if [[ -f $data/$f && ( ! -f $s1/$f || $s1/$f -ot $data/$f ) ]]; then + need_to_split=true + fi + done +fi + +if ! $need_to_split; then + exit 0; +fi + +for n in `seq $numsplit`; do + mkdir -p $data/split$numsplit/$n + utt2spks="$utt2spks $data/split$numsplit/$n/utt2spk" +done + +if $split_per_spk; then + utt2spk_opt="--utt2spk=$data/utt2spk" +else + utt2spk_opt= +fi + +# If lockfile is not installed, just don't lock it. It's not a big deal. +which lockfile >&/dev/null && lockfile -l 60 $data/.split_lock + +utils/split_scp.pl $utt2spk_opt $data/utt2spk $utt2spks || exit 1 + +for n in `seq $numsplit`; do + dsn=$data/split$numsplit/$n + utils/utt2spk_to_spk2utt.pl $dsn/utt2spk > $dsn/spk2utt || exit 1; +done + +maybe_wav_scp= +if [ ! -f $data/segments ]; then + maybe_wav_scp=wav.scp # If there is no segments file, then wav file is + # indexed per utt. +fi + +# split some things that are indexed by utterance. +for f in feats.scp text vad.scp utt2lang $maybe_wav_scp; do + if [ -f $data/$f ]; then + utils/filter_scps.pl JOB=1:$numsplit \ + $data/split$numsplit/JOB/utt2spk $data/$f $data/split$numsplit/JOB/$f || exit 1; + fi +done + +# split some things that are indexed by speaker +for f in spk2gender spk2warp cmvn.scp; do + if [ -f $data/$f ]; then + utils/filter_scps.pl JOB=1:$numsplit \ + $data/split$numsplit/JOB/spk2utt $data/$f $data/split$numsplit/JOB/$f || exit 1; + fi +done + +for n in `seq $numsplit`; do + dsn=$data/split$numsplit/$n + if [ -f $data/segments ]; then + utils/filter_scp.pl $dsn/utt2spk $data/segments > $dsn/segments + awk '{print $2;}' $dsn/segments | sort | uniq > $data/tmp.reco # recording-ids. + if [ -f $data/reco2file_and_channel ]; then + utils/filter_scp.pl $data/tmp.reco $data/reco2file_and_channel > $dsn/reco2file_and_channel + fi + if [ -f $data/wav.scp ]; then + utils/filter_scp.pl $data/tmp.reco $data/wav.scp >$dsn/wav.scp + fi + rm $data/tmp.reco + fi # else it would have been handled above, see maybe_wav. +done + +rm -f $data/.split_lock + +exit 0 -- cgit v1.2.3-70-g09d2