diff options
Diffstat (limited to 'pro-train')
| -rw-r--r-- | pro-train/Makefile.am | 13 | ||||
| -rw-r--r-- | pro-train/README.shared-mem | 9 | ||||
| -rwxr-xr-x | pro-train/dist-pro.pl | 657 | ||||
| -rwxr-xr-x | pro-train/mr_pro_generate_mapper_input.pl | 18 | ||||
| -rw-r--r-- | pro-train/mr_pro_map.cc | 347 | ||||
| -rw-r--r-- | pro-train/mr_pro_reduce.cc | 279 |
6 files changed, 1323 insertions, 0 deletions
diff --git a/pro-train/Makefile.am b/pro-train/Makefile.am new file mode 100644 index 00000000..fdaf43e2 --- /dev/null +++ b/pro-train/Makefile.am @@ -0,0 +1,13 @@ +bin_PROGRAMS = \ + mr_pro_map \ + mr_pro_reduce + +TESTS = lo_test + +mr_pro_map_SOURCES = mr_pro_map.cc +mr_pro_map_LDADD = $(top_srcdir)/decoder/libcdec.a $(top_srcdir)/mteval/libmteval.a $(top_srcdir)/utils/libutils.a -lz + +mr_pro_reduce_SOURCES = mr_pro_reduce.cc +mr_pro_reduce_LDADD = $(top_srcdir)/decoder/libcdec.a $(top_srcdir)/training/optimize.o $(top_srcdir)/mteval/libmteval.a $(top_srcdir)/utils/libutils.a -lz + +AM_CPPFLAGS = -W -Wall -Wno-sign-compare $(GTEST_CPPFLAGS) -I$(top_srcdir)/utils -I$(top_srcdir)/decoder -I$(top_srcdir)/mteval -I$(top_srcdir)/training diff --git a/pro-train/README.shared-mem b/pro-train/README.shared-mem new file mode 100644 index 00000000..7728efc0 --- /dev/null +++ b/pro-train/README.shared-mem @@ -0,0 +1,9 @@ +If you want to run dist-vest.pl on a very large shared memory machine, do the +following: + + ./dist-vest.pl --use-make I --decode-nodes J --weights weights.init --source-file=dev.src --ref-files=dev.ref.* cdec.ini + +This will use I jobs for doing the line search and J jobs to run the decoder. Typically, since the +decoder must load grammars, language models, etc., J should be smaller than I, but this will depend +on the system you are running on and the complexity of the models used for decoding. + diff --git a/pro-train/dist-pro.pl b/pro-train/dist-pro.pl new file mode 100755 index 00000000..dbfa329a --- /dev/null +++ b/pro-train/dist-pro.pl @@ -0,0 +1,657 @@ +#!/usr/bin/env perl +use strict; +my @ORIG_ARGV=@ARGV; +use Cwd qw(getcwd); +my $SCRIPT_DIR; BEGIN { use Cwd qw/ abs_path /; use File::Basename; $SCRIPT_DIR = dirname(abs_path($0)); push @INC, $SCRIPT_DIR, "$SCRIPT_DIR/../environment"; } + +# Skip local config (used for distributing jobs) if we're running in local-only mode +use LocalConfig; +use Getopt::Long; +use IPC::Open2; +use POSIX ":sys_wait_h"; +my $QSUB_CMD = qsub_args(mert_memory()); + +my $VEST_DIR="$SCRIPT_DIR/../vest"; +require "$VEST_DIR/libcall.pl"; + +# Default settings +my $srcFile; +my $refFiles; +my $bin_dir = $SCRIPT_DIR; +die "Bin directory $bin_dir missing/inaccessible" unless -d $bin_dir; +my $FAST_SCORE="$bin_dir/../mteval/fast_score"; +die "Can't execute $FAST_SCORE" unless -x $FAST_SCORE; +my $MAPINPUT = "$bin_dir/mr_pro_generate_mapper_input.pl"; +my $MAPPER = "$bin_dir/mr_pro_map"; +my $REDUCER = "$bin_dir/mr_pro_reduce"; +my $parallelize = "$VEST_DIR/parallelize.pl"; +my $libcall = "$VEST_DIR/libcall.pl"; +my $sentserver = "$VEST_DIR/sentserver"; +my $sentclient = "$VEST_DIR/sentclient"; +my $LocalConfig = "$SCRIPT_DIR/../environment/LocalConfig.pm"; + +my $SCORER = $FAST_SCORE; +die "Can't find $MAPPER" unless -x $MAPPER; +my $cdec = "$bin_dir/../decoder/cdec"; +die "Can't find decoder in $cdec" unless -x $cdec; +die "Can't find $parallelize" unless -x $parallelize; +die "Can't find $libcall" unless -e $libcall; +my $decoder = $cdec; +my $lines_per_mapper = 30; +my $iteration = 1; +my $run_local = 0; +my $best_weights; +my $max_iterations = 30; +my $decode_nodes = 15; # number of decode nodes +my $pmem = "4g"; +my $disable_clean = 0; +my %seen_weights; +my $help = 0; +my $epsilon = 0.0001; +my $dryrun = 0; +my $last_score = -10000000; +my $metric = "ibm_bleu"; +my $dir; +my $iniFile; +my $weights; +my $use_make; # use make to parallelize +my $usefork; +my $initial_weights; +my $pass_suffix = ''; +my $cpbin=1; + +# regularization strength +my $tune_regularizer = 0; +my $reg = 1e-2; + +# Process command-line options +Getopt::Long::Configure("no_auto_abbrev"); +if (GetOptions( + "decode-nodes=i" => \$decode_nodes, + "dont-clean" => \$disable_clean, + "pass-suffix=s" => \$pass_suffix, + "use-fork" => \$usefork, + "dry-run" => \$dryrun, + "epsilon=s" => \$epsilon, + "help" => \$help, + "weights=s" => \$initial_weights, + "tune-regularizer" => \$tune_regularizer, + "reg=f" => \$reg, + "local" => \$run_local, + "use-make=i" => \$use_make, + "max-iterations=i" => \$max_iterations, + "pmem=s" => \$pmem, + "cpbin!" => \$cpbin, + "ref-files=s" => \$refFiles, + "metric=s" => \$metric, + "source-file=s" => \$srcFile, + "workdir=s" => \$dir, +) == 0 || @ARGV!=1 || $help) { + print_help(); + exit; +} + +if ($usefork) { $usefork = "--use-fork"; } else { $usefork = ''; } + +if ($metric =~ /^(combi|ter)$/i) { + $lines_per_mapper = 5; +} + +($iniFile) = @ARGV; + + +sub write_config; +sub enseg; +sub print_help; + +my $nodelist; +my $host =check_output("hostname"); chomp $host; +my $bleu; +my $interval_count = 0; +my $logfile; +my $projected_score; + +# used in sorting scores +my $DIR_FLAG = '-r'; +if ($metric =~ /^ter$|^aer$/i) { + $DIR_FLAG = ''; +} + +my $refs_comma_sep = get_comma_sep_refs('r',$refFiles); + +unless ($dir){ + $dir = "protrain"; +} +unless ($dir =~ /^\//){ # convert relative path to absolute path + my $basedir = check_output("pwd"); + chomp $basedir; + $dir = "$basedir/$dir"; +} + + +# Initializations and helper functions +srand; + +my @childpids = (); +my @cleanupcmds = (); + +sub cleanup { + print STDERR "Cleanup...\n"; + for my $pid (@childpids){ unchecked_call("kill $pid"); } + for my $cmd (@cleanupcmds){ unchecked_call("$cmd"); } + exit 1; +}; +# Always call cleanup, no matter how we exit +*CORE::GLOBAL::exit = + sub{ cleanup(); }; +$SIG{INT} = "cleanup"; +$SIG{TERM} = "cleanup"; +$SIG{HUP} = "cleanup"; + +my $decoderBase = check_output("basename $decoder"); chomp $decoderBase; +my $newIniFile = "$dir/$decoderBase.ini"; +my $inputFileName = "$dir/input"; +my $user = $ENV{"USER"}; + + +# process ini file +-e $iniFile || die "Error: could not open $iniFile for reading\n"; +open(INI, $iniFile); + +use File::Basename qw(basename); +#pass bindir, refs to vars holding bin +sub modbin { + local $_; + my $bindir=shift; + check_call("mkdir -p $bindir"); + -d $bindir || die "couldn't make bindir $bindir"; + for (@_) { + my $src=$$_; + $$_="$bindir/".basename($src); + check_call("cp -p $src $$_"); + } +} +sub dirsize { + opendir ISEMPTY,$_[0]; + return scalar(readdir(ISEMPTY))-1; +} +my @allweights; +if ($dryrun){ + write_config(*STDERR); + exit 0; +} else { + if (-e $dir && dirsize($dir)>1 && -e "$dir/hgs" ){ # allow preexisting logfile, binaries, but not dist-pro.pl outputs + die "ERROR: working dir $dir already exists\n\n"; + } else { + -e $dir || mkdir $dir; + mkdir "$dir/hgs"; + modbin("$dir/bin",\$LocalConfig,\$cdec,\$SCORER,\$MAPINPUT,\$MAPPER,\$REDUCER,\$parallelize,\$sentserver,\$sentclient,\$libcall) if $cpbin; + mkdir "$dir/scripts"; + my $cmdfile="$dir/rerun-pro.sh"; + open CMD,'>',$cmdfile; + print CMD "cd ",&getcwd,"\n"; +# print CMD &escaped_cmdline,"\n"; #buggy - last arg is quoted. + my $cline=&cmdline."\n"; + print CMD $cline; + close CMD; + print STDERR $cline; + chmod(0755,$cmdfile); + check_call("cp $initial_weights $dir/weights.0"); + die "Can't find weights.0" unless (-e "$dir/weights.0"); + } + write_config(*STDERR); +} + + +# Generate initial files and values +check_call("cp $iniFile $newIniFile"); +$iniFile = $newIniFile; + +my $newsrc = "$dir/dev.input"; +enseg($srcFile, $newsrc); +$srcFile = $newsrc; +my $devSize = 0; +open F, "<$srcFile" or die "Can't read $srcFile: $!"; +while(<F>) { $devSize++; } +close F; + +unless($best_weights){ $best_weights = $weights; } +unless($projected_score){ $projected_score = 0.0; } +$seen_weights{$weights} = 1; + +my $random_seed = int(time / 1000); +my $lastWeightsFile; +my $lastPScore = 0; +# main optimization loop +while (1){ + print STDERR "\n\nITERATION $iteration\n==========\n"; + + if ($iteration > $max_iterations){ + print STDERR "\nREACHED STOPPING CRITERION: Maximum iterations\n"; + last; + } + # iteration-specific files + my $runFile="$dir/run.raw.$iteration"; + my $onebestFile="$dir/1best.$iteration"; + my $logdir="$dir/logs.$iteration"; + my $decoderLog="$logdir/decoder.sentserver.log.$iteration"; + my $scorerLog="$logdir/scorer.log.$iteration"; + check_call("mkdir -p $logdir"); + + + #decode + print STDERR "RUNNING DECODER AT "; + print STDERR unchecked_output("date"); + my $im1 = $iteration - 1; + my $weightsFile="$dir/weights.$im1"; + push @allweights, "-w $dir/weights.$im1"; + `rm -f $dir/hgs/*.gz`; + my $decoder_cmd = "$decoder -c $iniFile --weights$pass_suffix $weightsFile -O $dir/hgs"; + my $pcmd; + if ($run_local) { + $pcmd = "cat $srcFile |"; + } elsif ($use_make) { + # TODO: Throw error when decode_nodes is specified along with use_make + $pcmd = "cat $srcFile | $parallelize --use-fork -p $pmem -e $logdir -j $use_make --"; + } else { + $pcmd = "cat $srcFile | $parallelize $usefork -p $pmem -e $logdir -j $decode_nodes --"; + } + my $cmd = "$pcmd $decoder_cmd 2> $decoderLog 1> $runFile"; + print STDERR "COMMAND:\n$cmd\n"; + check_bash_call($cmd); + my $num_hgs; + my $num_topbest; + my $retries = 0; + while($retries < 5) { + $num_hgs = check_output("ls $dir/hgs/*.gz | wc -l"); + $num_topbest = check_output("wc -l < $runFile"); + print STDERR "NUMBER OF HGs: $num_hgs\n"; + print STDERR "NUMBER OF TOP-BEST HYPs: $num_topbest\n"; + if($devSize == $num_hgs && $devSize == $num_topbest) { + last; + } else { + print STDERR "Incorrect number of hypergraphs or topbest. Waiting for distributed filesystem and retrying...\n"; + sleep(3); + } + $retries++; + } + die "Dev set contains $devSize sentences, but we don't have topbest and hypergraphs for all these! Decoder failure? Check $decoderLog\n" if ($devSize != $num_hgs || $devSize != $num_topbest); + my $dec_score = check_output("cat $runFile | $SCORER $refs_comma_sep -l $metric"); + chomp $dec_score; + print STDERR "DECODER SCORE: $dec_score\n"; + + # save space + check_call("gzip -f $runFile"); + check_call("gzip -f $decoderLog"); + + # run optimizer + print STDERR "RUNNING OPTIMIZER AT "; + print STDERR unchecked_output("date"); + print STDERR " - GENERATE TRAINING EXEMPLARS\n"; + my $mergeLog="$logdir/prune-merge.log.$iteration"; + + my $score = 0; + my $icc = 0; + my $inweights="$dir/weights.$im1"; + $cmd="$MAPINPUT $dir/hgs > $dir/agenda.$im1"; + print STDERR "COMMAND:\n$cmd\n"; + check_call($cmd); + check_call("mkdir -p $dir/splag.$im1"); + $cmd="split -a 3 -l $lines_per_mapper $dir/agenda.$im1 $dir/splag.$im1/mapinput."; + print STDERR "COMMAND:\n$cmd\n"; + check_call($cmd); + opendir(DIR, "$dir/splag.$im1") or die "Can't open directory: $!"; + my @shards = grep { /^mapinput\./ } readdir(DIR); + closedir DIR; + die "No shards!" unless scalar @shards > 0; + my $joblist = ""; + my $nmappers = 0; + @cleanupcmds = (); + my %o2i = (); + my $first_shard = 1; + my $mkfile; # only used with makefiles + my $mkfilename; + if ($use_make) { + $mkfilename = "$dir/splag.$im1/domap.mk"; + open $mkfile, ">$mkfilename" or die "Couldn't write $mkfilename: $!"; + print $mkfile "all: $dir/splag.$im1/map.done\n\n"; + } + my @mkouts = (); # only used with makefiles + my @mapoutputs = (); + for my $shard (@shards) { + my $mapoutput = $shard; + my $client_name = $shard; + $client_name =~ s/mapinput.//; + $client_name = "pro.$client_name"; + $mapoutput =~ s/mapinput/mapoutput/; + push @mapoutputs, "$dir/splag.$im1/$mapoutput"; + $o2i{"$dir/splag.$im1/$mapoutput"} = "$dir/splag.$im1/$shard"; + my $script = "$MAPPER -s $srcFile -l $metric $refs_comma_sep -w $inweights -K $dir/kbest < $dir/splag.$im1/$shard > $dir/splag.$im1/$mapoutput"; + if ($run_local) { + print STDERR "COMMAND:\n$script\n"; + check_bash_call($script); + } elsif ($use_make) { + my $script_file = "$dir/scripts/map.$shard"; + open F, ">$script_file" or die "Can't write $script_file: $!"; + print F "#!/bin/bash\n"; + print F "$script\n"; + close F; + my $output = "$dir/splag.$im1/$mapoutput"; + push @mkouts, $output; + chmod(0755, $script_file) or die "Can't chmod $script_file: $!"; + if ($first_shard) { print STDERR "$script\n"; $first_shard=0; } + print $mkfile "$output: $dir/splag.$im1/$shard\n\t$script_file\n\n"; + } else { + my $script_file = "$dir/scripts/map.$shard"; + open F, ">$script_file" or die "Can't write $script_file: $!"; + print F "$script\n"; + close F; + if ($first_shard) { print STDERR "$script\n"; $first_shard=0; } + + $nmappers++; + my $qcmd = "$QSUB_CMD -N $client_name -o /dev/null -e $logdir/$client_name.ER $script_file"; + my $jobid = check_output("$qcmd"); + chomp $jobid; + $jobid =~ s/^(\d+)(.*?)$/\1/g; + $jobid =~ s/^Your job (\d+) .*$/\1/; + push(@cleanupcmds, "qdel $jobid 2> /dev/null"); + print STDERR " $jobid"; + if ($joblist == "") { $joblist = $jobid; } + else {$joblist = $joblist . "\|" . $jobid; } + } + } + my @dev_outs = (); + my @devtest_outs = (); + if ($tune_regularizer) { + for (my $i = 0; $i < scalar @mapoutputs; $i++) { + if ($i % 3 == 1) { + push @devtest_outs, $mapoutputs[$i]; + } else { + push @dev_outs, $mapoutputs[$i]; + } + } + if (scalar @devtest_outs == 0) { + die "Not enough training instances for regularization tuning! Rerun without --tune-regularizer\n"; + } + } else { + @dev_outs = @mapoutputs; + } + if ($run_local) { + print STDERR "\nCompleted extraction of training exemplars.\n"; + } elsif ($use_make) { + print $mkfile "$dir/splag.$im1/map.done: @mkouts\n\ttouch $dir/splag.$im1/map.done\n\n"; + close $mkfile; + my $mcmd = "make -j $use_make -f $mkfilename"; + print STDERR "\nExecuting: $mcmd\n"; + check_call($mcmd); + } else { + print STDERR "\nLaunched $nmappers mappers.\n"; + sleep 8; + print STDERR "Waiting for mappers to complete...\n"; + while ($nmappers > 0) { + sleep 5; + my @livejobs = grep(/$joblist/, split(/\n/, unchecked_output("qstat | grep -v ' C '"))); + $nmappers = scalar @livejobs; + } + print STDERR "All mappers complete.\n"; + } + my $tol = 0; + my $til = 0; + my $dev_test_file = "$dir/splag.$im1/devtest.gz"; + if ($tune_regularizer) { + my $cmd = "cat @devtest_outs | gzip > $dev_test_file"; + check_bash_call($cmd); + die "Can't find file $dev_test_file" unless -f $dev_test_file; + } + #print STDERR "MO: @mapoutputs\n"; + for my $mo (@mapoutputs) { + #my $olines = get_lines($mo); + #my $ilines = get_lines($o2i{$mo}); + #die "$mo: no training instances generated!" if $olines == 0; + } + print STDERR "\nRUNNING CLASSIFIER (REDUCER)\n"; + print STDERR unchecked_output("date"); + $cmd="cat @dev_outs | $REDUCER -w $dir/weights.$im1 -s $reg"; + if ($tune_regularizer) { + $cmd .= " -T -t $dev_test_file"; + } + $cmd .= " > $dir/weights.$iteration"; + print STDERR "COMMAND:\n$cmd\n"; + check_bash_call($cmd); + $lastWeightsFile = "$dir/weights.$iteration"; + if ($tune_regularizer) { + open W, "<$lastWeightsFile" or die "Can't read $lastWeightsFile: $!"; + my $line = <W>; + close W; + my ($sharp, $label, $nreg) = split /\s|=/, $line; + print STDERR "REGULARIZATION STRENGTH ($label) IS $nreg\n"; + $reg = $nreg; + # only tune regularizer on first iteration? + $tune_regularizer = 0; + } + $lastPScore = $score; + $iteration++; + print STDERR "\n==========\n"; +} + +print STDERR "\nFINAL WEIGHTS: $lastWeightsFile\n(Use -w <this file> with the decoder)\n\n"; + +print STDOUT "$lastWeightsFile\n"; + +exit 0; + +sub get_lines { + my $fn = shift @_; + open FL, "<$fn" or die "Couldn't read $fn: $!"; + my $lc = 0; + while(<FL>) { $lc++; } + return $lc; +} + +sub get_comma_sep_refs { + my ($r,$p) = @_; + my $o = check_output("echo $p"); + chomp $o; + my @files = split /\s+/, $o; + return "-$r " . join(" -$r ", @files); +} + +sub read_weights_file { + my ($file) = @_; + open F, "<$file" or die "Couldn't read $file: $!"; + my @r = (); + my $pm = -1; + while(<F>) { + next if /^#/; + next if /^\s*$/; + chomp; + if (/^(.+)\s+(.+)$/) { + my $m = $1; + my $w = $2; + die "Weights out of order: $m <= $pm" unless $m > $pm; + push @r, $w; + } else { + warn "Unexpected feature name in weight file: $_"; + } + } + close F; + return join ' ', @r; +} + +# subs +sub write_config { + my $fh = shift; + my $cleanup = "yes"; + if ($disable_clean) {$cleanup = "no";} + + print $fh "\n"; + print $fh "DECODER: $decoder\n"; + print $fh "INI FILE: $iniFile\n"; + print $fh "WORKING DIR: $dir\n"; + print $fh "SOURCE (DEV): $srcFile\n"; + print $fh "REFS (DEV): $refFiles\n"; + print $fh "EVAL METRIC: $metric\n"; + print $fh "MAX ITERATIONS: $max_iterations\n"; + print $fh "DECODE NODES: $decode_nodes\n"; + print $fh "HEAD NODE: $host\n"; + print $fh "PMEM (DECODING): $pmem\n"; + print $fh "CLEANUP: $cleanup\n"; +} + +sub update_weights_file { + my ($neww, $rfn, $rpts) = @_; + my @feats = @$rfn; + my @pts = @$rpts; + my $num_feats = scalar @feats; + my $num_pts = scalar @pts; + die "$num_feats (num_feats) != $num_pts (num_pts)" unless $num_feats == $num_pts; + open G, ">$neww" or die; + for (my $i = 0; $i < $num_feats; $i++) { + my $f = $feats[$i]; + my $lambda = $pts[$i]; + print G "$f $lambda\n"; + } + close G; +} + +sub enseg { + my $src = shift; + my $newsrc = shift; + open(SRC, $src); + open(NEWSRC, ">$newsrc"); + my $i=0; + while (my $line=<SRC>){ + chomp $line; + if ($line =~ /^\s*<seg/i) { + if($line =~ /id="[0-9]+"/) { + print NEWSRC "$line\n"; + } else { + die "When using segments with pre-generated <seg> tags, you must include a zero-based id attribute"; + } + } else { + print NEWSRC "<seg id=\"$i\">$line</seg>\n"; + } + $i++; + } + close SRC; + close NEWSRC; + die "Empty dev set!" if ($i == 0); +} + +sub print_help { + + my $executable = check_output("basename $0"); chomp $executable; + print << "Help"; + +Usage: $executable [options] <ini file> + + $executable [options] <ini file> + Runs a complete MERT optimization and test set decoding, using + the decoder configuration in ini file. Note that many of the + options have default values that are inferred automatically + based on certain conventions. For details, refer to descriptions + of the options --decoder, --weights, and --workdir. + +Required: + + --ref-files <files> + Dev set ref files. This option takes only a single string argument. + To use multiple files (including file globbing), this argument should + be quoted. + + --source-file <file> + Dev set source file. + + --weights <file> + Initial weights file (use empty file to start from 0) + +General options: + + --local + Run the decoder and optimizer locally with a single thread. + + --decode-nodes <I> + Number of decoder processes to run in parallel. [default=15] + + --help + Print this message and exit. + + --max-iterations <M> + Maximum number of iterations to run. If not specified, defaults + to 10. + + --metric <method> + Metric to optimize. + Example values: IBM_BLEU, NIST_BLEU, Koehn_BLEU, TER, Combi + + --pass-suffix <S> + If the decoder is doing multi-pass decoding, the pass suffix "2", + "3", etc., is used to control what iteration of weights is set. + + --pmem <N> + Amount of physical memory requested for parallel decoding jobs. + + --use-make <I> + Use make -j <I> to run the optimizer commands (useful on large + shared-memory machines where qsub is unavailable). + + --workdir <dir> + Directory for intermediate and output files. If not specified, the + name is derived from the ini filename. Assuming that the ini + filename begins with the decoder name and ends with ini, the default + name of the working directory is inferred from the middle part of + the filename. E.g. an ini file named decoder.foo.ini would have + a default working directory name foo. + +Regularization options: + + --tune-regularizer + Hold out one third of the tuning data and used this to tune the + regularization parameter. + + --reg <F> + +Help +} + +sub convert { + my ($str) = @_; + my @ps = split /;/, $str; + my %dict = (); + for my $p (@ps) { + my ($k, $v) = split /=/, $p; + $dict{$k} = $v; + } + return %dict; +} + + +sub cmdline { + return join ' ',($0,@ORIG_ARGV); +} + +#buggy: last arg gets quoted sometimes? +my $is_shell_special=qr{[ \t\n\\><|&;"'`~*?{}$!()]}; +my $shell_escape_in_quote=qr{[\\"\$`!]}; + +sub escape_shell { + my ($arg)=@_; + return undef unless defined $arg; + if ($arg =~ /$is_shell_special/) { + $arg =~ s/($shell_escape_in_quote)/\\$1/g; + return "\"$arg\""; + } + return $arg; +} + +sub escaped_shell_args { + return map {local $_=$_;chomp;escape_shell($_)} @_; +} + +sub escaped_shell_args_str { + return join ' ',&escaped_shell_args(@_); +} + +sub escaped_cmdline { + return "$0 ".&escaped_shell_args_str(@ORIG_ARGV); +} diff --git a/pro-train/mr_pro_generate_mapper_input.pl b/pro-train/mr_pro_generate_mapper_input.pl new file mode 100755 index 00000000..b30fc4fd --- /dev/null +++ b/pro-train/mr_pro_generate_mapper_input.pl @@ -0,0 +1,18 @@ +#!/usr/bin/perl -w +use strict; + +die "Usage: $0 HG_DIR\n" unless scalar @ARGV == 1; +my $d = shift @ARGV; +die "Can't find directory $d" unless -d $d; + +opendir(DIR, $d) or die "Can't read $d: $!"; +my @hgs = grep { /\.gz$/ } readdir(DIR); +closedir DIR; + +for my $hg (@hgs) { + my $file = $hg; + my $id = $hg; + $id =~ s/(\.json)?\.gz//; + print "$d/$file $id\n"; +} + diff --git a/pro-train/mr_pro_map.cc b/pro-train/mr_pro_map.cc new file mode 100644 index 00000000..0a9b75d7 --- /dev/null +++ b/pro-train/mr_pro_map.cc @@ -0,0 +1,347 @@ +#include <sstream> +#include <iostream> +#include <fstream> +#include <vector> +#include <tr1/unordered_map> + +#include <boost/functional/hash.hpp> +#include <boost/shared_ptr.hpp> +#include <boost/program_options.hpp> +#include <boost/program_options/variables_map.hpp> + +#include "sampler.h" +#include "filelib.h" +#include "stringlib.h" +#include "weights.h" +#include "scorer.h" +#include "inside_outside.h" +#include "hg_io.h" +#include "kbest.h" +#include "viterbi.h" + +// This is Figure 4 (Algorithm Sampler) from Hopkins&May (2011) + +using namespace std; +namespace po = boost::program_options; + +struct ApproxVectorHasher { + static const size_t MASK = 0xFFFFFFFFull; + union UType { + double f; // leave as double + size_t i; + }; + static inline double round(const double x) { + UType t; + t.f = x; + size_t r = t.i & MASK; + if ((r << 1) > MASK) + t.i += MASK - r + 1; + else + t.i &= (1ull - MASK); + return t.f; + } + size_t operator()(const SparseVector<weight_t>& x) const { + size_t h = 0x573915839; + for (SparseVector<weight_t>::const_iterator it = x.begin(); it != x.end(); ++it) { + UType t; + t.f = it->second; + if (t.f) { + size_t z = (t.i >> 32); + boost::hash_combine(h, it->first); + boost::hash_combine(h, z); + } + } + return h; + } +}; + +struct ApproxVectorEquals { + bool operator()(const SparseVector<weight_t>& a, const SparseVector<weight_t>& b) const { + SparseVector<weight_t>::const_iterator bit = b.begin(); + for (SparseVector<weight_t>::const_iterator ait = a.begin(); ait != a.end(); ++ait) { + if (bit == b.end() || + ait->first != bit->first || + ApproxVectorHasher::round(ait->second) != ApproxVectorHasher::round(bit->second)) + return false; + ++bit; + } + if (bit != b.end()) return false; + return true; + } +}; + +boost::shared_ptr<MT19937> rng; + +void InitCommandLine(int argc, char** argv, po::variables_map* conf) { + po::options_description opts("Configuration options"); + opts.add_options() + ("reference,r",po::value<vector<string> >(), "[REQD] Reference translation (tokenized text)") + ("weights,w",po::value<string>(), "[REQD] Weights files from current iterations") + ("kbest_repository,K",po::value<string>()->default_value("./kbest"),"K-best list repository (directory)") + ("input,i",po::value<string>()->default_value("-"), "Input file to map (- is STDIN)") + ("source,s",po::value<string>()->default_value(""), "Source file (ignored, except for AER)") + ("loss_function,l",po::value<string>()->default_value("ibm_bleu"), "Loss function being optimized") + ("kbest_size,k",po::value<unsigned>()->default_value(1500u), "Top k-hypotheses to extract") + ("candidate_pairs,G", po::value<unsigned>()->default_value(5000u), "Number of pairs to sample per hypothesis (Gamma)") + ("best_pairs,X", po::value<unsigned>()->default_value(50u), "Number of pairs, ranked by magnitude of objective delta, to retain (Xi)") + ("random_seed,S", po::value<uint32_t>(), "Random seed (if not specified, /dev/random will be used)") + ("help,h", "Help"); + po::options_description dcmdline_options; + dcmdline_options.add(opts); + po::store(parse_command_line(argc, argv, dcmdline_options), *conf); + bool flag = false; + if (!conf->count("reference")) { + cerr << "Please specify one or more references using -r <REF.TXT>\n"; + flag = true; + } + if (!conf->count("weights")) { + cerr << "Please specify weights using -w <WEIGHTS.TXT>\n"; + flag = true; + } + if (flag || conf->count("help")) { + cerr << dcmdline_options << endl; + exit(1); + } +} + +struct HypInfo { + HypInfo() : g_(-100.0f) {} + HypInfo(const vector<WordID>& h, const SparseVector<weight_t>& feats) : hyp(h), g_(-100.0f), x(feats) {} + + // lazy evaluation + double g(const SentenceScorer& scorer) const { + if (g_ == -100.0f) + g_ = scorer.ScoreCandidate(hyp)->ComputeScore(); + return g_; + } + vector<WordID> hyp; + mutable float g_; + SparseVector<weight_t> x; +}; + +struct HypInfoCompare { + bool operator()(const HypInfo& a, const HypInfo& b) const { + ApproxVectorEquals comp; + return (a.hyp == b.hyp && comp(a.x,b.x)); + } +}; + +struct HypInfoHasher { + size_t operator()(const HypInfo& x) const { + boost::hash<vector<WordID> > hhasher; + ApproxVectorHasher vhasher; + size_t ha = hhasher(x.hyp); + boost::hash_combine(ha, vhasher(x.x)); + return ha; + } +}; + +void WriteKBest(const string& file, const vector<HypInfo>& kbest) { + WriteFile wf(file); + ostream& out = *wf.stream(); + out.precision(10); + for (int i = 0; i < kbest.size(); ++i) { + out << TD::GetString(kbest[i].hyp) << endl; + out << kbest[i].x << endl; + } +} + +void ParseSparseVector(string& line, size_t cur, SparseVector<weight_t>* out) { + SparseVector<weight_t>& x = *out; + size_t last_start = cur; + size_t last_comma = string::npos; + while(cur <= line.size()) { + if (line[cur] == ' ' || cur == line.size()) { + if (!(cur > last_start && last_comma != string::npos && cur > last_comma)) { + cerr << "[ERROR] " << line << endl << " position = " << cur << endl; + exit(1); + } + const int fid = FD::Convert(line.substr(last_start, last_comma - last_start)); + if (cur < line.size()) line[cur] = 0; + const double val = strtod(&line[last_comma + 1], NULL); + x.set_value(fid, val); + + last_comma = string::npos; + last_start = cur+1; + } else { + if (line[cur] == '=') + last_comma = cur; + } + ++cur; + } +} + +void ReadKBest(const string& file, vector<HypInfo>* kbest) { + cerr << "Reading from " << file << endl; + ReadFile rf(file); + istream& in = *rf.stream(); + string cand; + string feats; + while(getline(in, cand)) { + getline(in, feats); + assert(in); + kbest->push_back(HypInfo()); + TD::ConvertSentence(cand, &kbest->back().hyp); + ParseSparseVector(feats, 0, &kbest->back().x); + } + cerr << " read " << kbest->size() << " hypotheses\n"; +} + +void Dedup(vector<HypInfo>* h) { + cerr << "Dedup in=" << h->size(); + tr1::unordered_set<HypInfo, HypInfoHasher, HypInfoCompare> u; + while(h->size() > 0) { + u.insert(h->back()); + h->pop_back(); + } + tr1::unordered_set<HypInfo, HypInfoHasher, HypInfoCompare>::iterator it = u.begin(); + while (it != u.end()) { + h->push_back(*it); + it = u.erase(it); + } + cerr << " out=" << h->size() << endl; +} + +struct ThresholdAlpha { + explicit ThresholdAlpha(double t = 0.05) : threshold(t) {} + double operator()(double mag) const { + if (mag < threshold) return 0.0; else return 1.0; + } + const double threshold; +}; + +struct TrainingInstance { + TrainingInstance(const SparseVector<weight_t>& feats, bool positive, float diff) : x(feats), y(positive), gdiff(diff) {} + SparseVector<weight_t> x; +#undef DEBUGGING_PRO +#ifdef DEBUGGING_PRO + vector<WordID> a; + vector<WordID> b; +#endif + bool y; + float gdiff; +}; +#ifdef DEBUGGING_PRO +ostream& operator<<(ostream& os, const TrainingInstance& d) { + return os << d.gdiff << " y=" << d.y << "\tA:" << TD::GetString(d.a) << "\n\tB: " << TD::GetString(d.b) << "\n\tX: " << d.x; +} +#endif + +struct DiffOrder { + bool operator()(const TrainingInstance& a, const TrainingInstance& b) const { + return a.gdiff > b.gdiff; + } +}; + +void Sample(const unsigned gamma, const unsigned xi, const vector<HypInfo>& J_i, const SentenceScorer& scorer, const bool invert_score, vector<TrainingInstance>* pv) { + vector<TrainingInstance> v1, v2; + float avg_diff = 0; + for (unsigned i = 0; i < gamma; ++i) { + const size_t a = rng->inclusive(0, J_i.size() - 1)(); + const size_t b = rng->inclusive(0, J_i.size() - 1)(); + if (a == b) continue; + float ga = J_i[a].g(scorer); + float gb = J_i[b].g(scorer); + bool positive = gb < ga; + if (invert_score) positive = !positive; + const float gdiff = fabs(ga - gb); + if (!gdiff) continue; + avg_diff += gdiff; + SparseVector<weight_t> xdiff = (J_i[a].x - J_i[b].x).erase_zeros(); + if (xdiff.empty()) { + cerr << "Empty diff:\n " << TD::GetString(J_i[a].hyp) << endl << "x=" << J_i[a].x << endl; + cerr << " " << TD::GetString(J_i[b].hyp) << endl << "x=" << J_i[b].x << endl; + continue; + } + v1.push_back(TrainingInstance(xdiff, positive, gdiff)); +#ifdef DEBUGGING_PRO + v1.back().a = J_i[a].hyp; + v1.back().b = J_i[b].hyp; + cerr << "N: " << v1.back() << endl; +#endif + } + avg_diff /= v1.size(); + + for (unsigned i = 0; i < v1.size(); ++i) { + double p = 1.0 / (1.0 + exp(-avg_diff - v1[i].gdiff)); + // cerr << "avg_diff=" << avg_diff << " gdiff=" << v1[i].gdiff << " p=" << p << endl; + if (rng->next() < p) v2.push_back(v1[i]); + } + vector<TrainingInstance>::iterator mid = v2.begin() + xi; + if (xi > v2.size()) mid = v2.end(); + partial_sort(v2.begin(), mid, v2.end(), DiffOrder()); + copy(v2.begin(), mid, back_inserter(*pv)); +#ifdef DEBUGGING_PRO + if (v2.size() >= 5) { + for (int i =0; i < (mid - v2.begin()); ++i) { + cerr << v2[i] << endl; + } + cerr << pv->back() << endl; + } +#endif +} + +int main(int argc, char** argv) { + po::variables_map conf; + InitCommandLine(argc, argv, &conf); + if (conf.count("random_seed")) + rng.reset(new MT19937(conf["random_seed"].as<uint32_t>())); + else + rng.reset(new MT19937); + const string loss_function = conf["loss_function"].as<string>(); + + ScoreType type = ScoreTypeFromString(loss_function); + DocScorer ds(type, conf["reference"].as<vector<string> >(), conf["source"].as<string>()); + cerr << "Loaded " << ds.size() << " references for scoring with " << loss_function << endl; + Hypergraph hg; + string last_file; + ReadFile in_read(conf["input"].as<string>()); + istream &in=*in_read.stream(); + const unsigned kbest_size = conf["kbest_size"].as<unsigned>(); + const unsigned gamma = conf["candidate_pairs"].as<unsigned>(); + const unsigned xi = conf["best_pairs"].as<unsigned>(); + string weightsf = conf["weights"].as<string>(); + vector<weight_t> weights; + Weights::InitFromFile(weightsf, &weights); + string kbest_repo = conf["kbest_repository"].as<string>(); + MkDirP(kbest_repo); + while(in) { + vector<TrainingInstance> v; + string line; + getline(in, line); + if (line.empty()) continue; + istringstream is(line); + int sent_id; + string file; + // path-to-file (JSON) sent_id + is >> file >> sent_id; + ReadFile rf(file); + ostringstream os; + vector<HypInfo> J_i; + os << kbest_repo << "/kbest." << sent_id << ".txt.gz"; + const string kbest_file = os.str(); + if (FileExists(kbest_file)) + ReadKBest(kbest_file, &J_i); + HypergraphIO::ReadFromJSON(rf.stream(), &hg); + hg.Reweight(weights); + KBest::KBestDerivations<vector<WordID>, ESentenceTraversal> kbest(hg, kbest_size); + + for (int i = 0; i < kbest_size; ++i) { + const KBest::KBestDerivations<vector<WordID>, ESentenceTraversal>::Derivation* d = + kbest.LazyKthBest(hg.nodes_.size() - 1, i); + if (!d) break; + J_i.push_back(HypInfo(d->yield, d->feature_values)); + } + Dedup(&J_i); + WriteKBest(kbest_file, J_i); + + Sample(gamma, xi, J_i, *ds[sent_id], (type == TER), &v); + for (unsigned i = 0; i < v.size(); ++i) { + const TrainingInstance& vi = v[i]; + cout << vi.y << "\t" << vi.x << endl; + cout << (!vi.y) << "\t" << (vi.x * -1.0) << endl; + } + } + return 0; +} + diff --git a/pro-train/mr_pro_reduce.cc b/pro-train/mr_pro_reduce.cc new file mode 100644 index 00000000..aff410a0 --- /dev/null +++ b/pro-train/mr_pro_reduce.cc @@ -0,0 +1,279 @@ +#include <cstdlib> +#include <sstream> +#include <iostream> +#include <fstream> +#include <vector> + +#include <boost/program_options.hpp> +#include <boost/program_options/variables_map.hpp> + +#include "filelib.h" +#include "weights.h" +#include "sparse_vector.h" +#include "optimize.h" + +using namespace std; +namespace po = boost::program_options; + +// since this is a ranking model, there should be equal numbers of +// positive and negative examples, so the bias should be 0 +static const double MAX_BIAS = 1e-10; + +void InitCommandLine(int argc, char** argv, po::variables_map* conf) { + po::options_description opts("Configuration options"); + opts.add_options() + ("weights,w", po::value<string>(), "Weights from previous iteration (used as initialization and interpolation") + ("interpolation,p",po::value<double>()->default_value(0.9), "Output weights are p*w + (1-p)*w_prev") + ("memory_buffers,m",po::value<unsigned>()->default_value(200), "Number of memory buffers (LBFGS)") + ("sigma_squared,s",po::value<double>()->default_value(0.1), "Sigma squared for Gaussian prior") + ("min_reg,r",po::value<double>()->default_value(1e-8), "When tuning (-T) regularization strength, minimum regularization strenght") + ("max_reg,R",po::value<double>()->default_value(10.0), "When tuning (-T) regularization strength, maximum regularization strenght") + ("testset,t",po::value<string>(), "Optional held-out test set") + ("tune_regularizer,T", "Use the held out test set (-t) to tune the regularization strength") + ("help,h", "Help"); + po::options_description dcmdline_options; + dcmdline_options.add(opts); + po::store(parse_command_line(argc, argv, dcmdline_options), *conf); + if (conf->count("help")) { + cerr << dcmdline_options << endl; + exit(1); + } +} + +void ParseSparseVector(string& line, size_t cur, SparseVector<weight_t>* out) { + SparseVector<weight_t>& x = *out; + size_t last_start = cur; + size_t last_comma = string::npos; + while(cur <= line.size()) { + if (line[cur] == ' ' || cur == line.size()) { + if (!(cur > last_start && last_comma != string::npos && cur > last_comma)) { + cerr << "[ERROR] " << line << endl << " position = " << cur << endl; + exit(1); + } + const int fid = FD::Convert(line.substr(last_start, last_comma - last_start)); + if (cur < line.size()) line[cur] = 0; + const weight_t val = strtod(&line[last_comma + 1], NULL); + x.set_value(fid, val); + + last_comma = string::npos; + last_start = cur+1; + } else { + if (line[cur] == '=') + last_comma = cur; + } + ++cur; + } +} + +void ReadCorpus(istream* pin, vector<pair<bool, SparseVector<weight_t> > >* corpus) { + istream& in = *pin; + corpus->clear(); + bool flag = false; + int lc = 0; + string line; + SparseVector<weight_t> x; + while(getline(in, line)) { + ++lc; + if (lc % 1000 == 0) { cerr << '.'; flag = true; } + if (lc % 40000 == 0) { cerr << " [" << lc << "]\n"; flag = false; } + if (line.empty()) continue; + const size_t ks = line.find("\t"); + assert(string::npos != ks); + assert(ks == 1); + const bool y = line[0] == '1'; + x.clear(); + ParseSparseVector(line, ks + 1, &x); + corpus->push_back(make_pair(y, x)); + } + if (flag) cerr << endl; +} + +void GradAdd(const SparseVector<weight_t>& v, const double scale, vector<weight_t>* acc) { + for (SparseVector<weight_t>::const_iterator it = v.begin(); + it != v.end(); ++it) { + (*acc)[it->first] += it->second * scale; + } +} + +double TrainingInference(const vector<weight_t>& x, + const vector<pair<bool, SparseVector<weight_t> > >& corpus, + vector<weight_t>* g = NULL) { + double cll = 0; + for (int i = 0; i < corpus.size(); ++i) { + const double dotprod = corpus[i].second.dot(x) + (x.size() ? x[0] : weight_t()); // x[0] is bias + double lp_false = dotprod; + double lp_true = -dotprod; + if (0 < lp_true) { + lp_true += log1p(exp(-lp_true)); + lp_false = log1p(exp(lp_false)); + } else { + lp_true = log1p(exp(lp_true)); + lp_false += log1p(exp(-lp_false)); + } + lp_true*=-1; + lp_false*=-1; + if (corpus[i].first) { // true label + cll -= lp_true; + if (g) { + // g -= corpus[i].second * exp(lp_false); + GradAdd(corpus[i].second, -exp(lp_false), g); + (*g)[0] -= exp(lp_false); // bias + } + } else { // false label + cll -= lp_false; + if (g) { + // g += corpus[i].second * exp(lp_true); + GradAdd(corpus[i].second, exp(lp_true), g); + (*g)[0] += exp(lp_true); // bias + } + } + } + return cll; +} + +// return held-out log likelihood +double LearnParameters(const vector<pair<bool, SparseVector<weight_t> > >& training, + const vector<pair<bool, SparseVector<weight_t> > >& testing, + const double sigsq, + const unsigned memory_buffers, + vector<weight_t>* px) { + vector<weight_t>& x = *px; + vector<weight_t> vg(FD::NumFeats(), 0.0); + bool converged = false; + LBFGSOptimizer opt(FD::NumFeats(), memory_buffers); + double tppl = 0.0; + while(!converged) { + fill(vg.begin(), vg.end(), 0.0); + double cll = TrainingInference(x, training, &vg); + double ppl = cll / log(2); + ppl /= training.size(); + ppl = pow(2.0, ppl); + + // evaluate optional held-out test set + if (testing.size()) { + tppl = TrainingInference(x, testing) / log(2); + tppl /= testing.size(); + tppl = pow(2.0, tppl); + } + + // handle regularizer +#if 1 + double norm = 0; + for (int i = 1; i < x.size(); ++i) { + const double mean_i = 0.0; + const double param = (x[i] - mean_i); + norm += param * param; + vg[i] += param / sigsq; + } + const double reg = norm / (2.0 * sigsq); +#else + double reg = 0; +#endif + cll += reg; + cerr << cll << " (REG=" << reg << ")\tPPL=" << ppl << "\t TEST_PPL=" << tppl << "\t"; + try { + vector<weight_t> old_x = x; + do { + opt.Optimize(cll, vg, &x); + converged = opt.HasConverged(); + } while (!converged && x == old_x); + } catch (...) { + cerr << "Exception caught, assuming convergence is close enough...\n"; + converged = true; + } + if (fabs(x[0]) > MAX_BIAS) { + cerr << "Biased model learned. Are your training instances wrong?\n"; + cerr << " BIAS: " << x[0] << endl; + } + } + return tppl; +} + +int main(int argc, char** argv) { + po::variables_map conf; + InitCommandLine(argc, argv, &conf); + string line; + vector<pair<bool, SparseVector<weight_t> > > training, testing; + const bool tune_regularizer = conf.count("tune_regularizer"); + if (tune_regularizer && !conf.count("testset")) { + cerr << "--tune_regularizer requires --testset to be set\n"; + return 1; + } + const double min_reg = conf["min_reg"].as<double>(); + const double max_reg = conf["max_reg"].as<double>(); + double sigsq = conf["sigma_squared"].as<double>(); // will be overridden if parameter is tuned + assert(sigsq > 0.0); + assert(min_reg > 0.0); + assert(max_reg > 0.0); + assert(max_reg > min_reg); + const double psi = conf["interpolation"].as<double>(); + if (psi < 0.0 || psi > 1.0) { cerr << "Invalid interpolation weight: " << psi << endl; } + ReadCorpus(&cin, &training); + if (conf.count("testset")) { + ReadFile rf(conf["testset"].as<string>()); + ReadCorpus(rf.stream(), &testing); + } + cerr << "Number of features: " << FD::NumFeats() << endl; + + vector<weight_t> x, prev_x; // x[0] is bias + if (conf.count("weights")) { + Weights::InitFromFile(conf["weights"].as<string>(), &x); + x.resize(FD::NumFeats()); + prev_x = x; + } else { + x.resize(FD::NumFeats()); + prev_x = x; + } + cerr << " Number of features: " << x.size() << endl; + cerr << "Number of training examples: " << training.size() << endl; + cerr << "Number of testing examples: " << testing.size() << endl; + double tppl = 0.0; + vector<pair<double,double> > sp; + vector<double> smoothed; + if (tune_regularizer) { + sigsq = min_reg; + const double steps = 18; + double sweep_factor = exp((log(max_reg) - log(min_reg)) / steps); + cerr << "SWEEP FACTOR: " << sweep_factor << endl; + while(sigsq < max_reg) { + tppl = LearnParameters(training, testing, sigsq, conf["memory_buffers"].as<unsigned>(), &x); + sp.push_back(make_pair(sigsq, tppl)); + sigsq *= sweep_factor; + } + smoothed.resize(sp.size(), 0); + smoothed[0] = sp[0].second; + smoothed.back() = sp.back().second; + for (int i = 1; i < sp.size()-1; ++i) { + double prev = sp[i-1].second; + double next = sp[i+1].second; + double cur = sp[i].second; + smoothed[i] = (prev*0.2) + cur * 0.6 + (0.2*next); + } + double best_ppl = 9999999; + unsigned best_i = 0; + for (unsigned i = 0; i < sp.size(); ++i) { + if (smoothed[i] < best_ppl) { + best_ppl = smoothed[i]; + best_i = i; + } + } + sigsq = sp[best_i].first; + } // tune regularizer + tppl = LearnParameters(training, testing, sigsq, conf["memory_buffers"].as<unsigned>(), &x); + if (conf.count("weights")) { + for (int i = 1; i < x.size(); ++i) { + x[i] = (x[i] * psi) + prev_x[i] * (1.0 - psi); + } + } + cout.precision(15); + cout << "# sigma^2=" << sigsq << "\theld out perplexity="; + if (tppl) { cout << tppl << endl; } else { cout << "N/A\n"; } + if (sp.size()) { + cout << "# Parameter sweep:\n"; + for (int i = 0; i < sp.size(); ++i) { + cout << "# " << sp[i].first << "\t" << sp[i].second << "\t" << smoothed[i] << endl; + } + } + Weights::WriteToFile("-", x); + return 0; +} |