summaryrefslogtreecommitdiff
path: root/training/pro
diff options
context:
space:
mode:
Diffstat (limited to 'training/pro')
-rw-r--r--training/pro/Makefile.am13
-rwxr-xr-xtraining/pro/mr_pro_generate_mapper_input.pl18
-rw-r--r--training/pro/mr_pro_map.cc201
-rw-r--r--training/pro/mr_pro_reduce.cc286
-rwxr-xr-xtraining/pro/pro.pl555
5 files changed, 1073 insertions, 0 deletions
diff --git a/training/pro/Makefile.am b/training/pro/Makefile.am
new file mode 100644
index 00000000..09364804
--- /dev/null
+++ b/training/pro/Makefile.am
@@ -0,0 +1,13 @@
+bin_PROGRAMS = \
+ mr_pro_map \
+ mr_pro_reduce
+
+mr_pro_map_SOURCES = mr_pro_map.cc
+mr_pro_map_LDADD = ../../training/utils/libtraining_utils.a ../../decoder/libcdec.a ../../mteval/libmteval.a ../../utils/libutils.a
+
+mr_pro_reduce_SOURCES = mr_pro_reduce.cc
+mr_pro_reduce_LDADD = ../../training/liblbfgs/liblbfgs.a ../../utils/libutils.a
+
+EXTRA_DIST = mr_pro_generate_mapper_input.pl pro.pl
+
+AM_CPPFLAGS = -W -Wall -Wno-sign-compare -I$(top_srcdir)/utils -I$(top_srcdir)/decoder -I$(top_srcdir)/mteval -I$(top_srcdir)/training/utils -I$(top_srcdir)/training
diff --git a/training/pro/mr_pro_generate_mapper_input.pl b/training/pro/mr_pro_generate_mapper_input.pl
new file mode 100755
index 00000000..b30fc4fd
--- /dev/null
+++ b/training/pro/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/training/pro/mr_pro_map.cc b/training/pro/mr_pro_map.cc
new file mode 100644
index 00000000..eef40b8a
--- /dev/null
+++ b/training/pro/mr_pro_map.cc
@@ -0,0 +1,201 @@
+#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 "candidate_set.h"
+#include "sampler.h"
+#include "filelib.h"
+#include "stringlib.h"
+#include "weights.h"
+#include "inside_outside.h"
+#include "hg_io.h"
+#include "ns.h"
+#include "ns_docscorer.h"
+
+// This is Figure 4 (Algorithm Sampler) from Hopkins&May (2011)
+
+using namespace std;
+namespace po = boost::program_options;
+
+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)")
+ ("evaluation_metric,m",po::value<string>()->default_value("IBM_BLEU"), "Evaluation metric (ibm_bleu, koehn_bleu, nist_bleu, ter, meteor, etc.)")
+ ("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 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 training::CandidateSet& J_i,
+ const EvaluationMetric* metric,
+ vector<TrainingInstance>* pv) {
+ const bool invert_score = metric->IsErrorMetric();
+ 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 = metric->ComputeScore(J_i[a].eval_feats);
+ float gb = metric->ComputeScore(J_i[b].eval_feats);
+ 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].fmap - J_i[b].fmap).erase_zeros();
+ if (xdiff.empty()) {
+ cerr << "Empty diff:\n " << TD::GetString(J_i[a].ewords) << endl << "x=" << J_i[a].fmap << endl;
+ cerr << " " << TD::GetString(J_i[b].ewords) << endl << "x=" << J_i[b].fmap << 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 evaluation_metric = conf["evaluation_metric"].as<string>();
+
+ EvaluationMetric* metric = EvaluationMetric::Instance(evaluation_metric);
+ DocumentScorer ds(metric, conf["reference"].as<vector<string> >());
+ cerr << "Loaded " << ds.size() << " references for scoring with " << evaluation_metric << 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;
+ training::CandidateSet J_i;
+ os << kbest_repo << "/kbest." << sent_id << ".txt.gz";
+ const string kbest_file = os.str();
+ if (FileExists(kbest_file))
+ J_i.ReadFromFile(kbest_file);
+ HypergraphIO::ReadFromJSON(rf.stream(), &hg);
+ hg.Reweight(weights);
+ J_i.AddKBestCandidates(hg, kbest_size, ds[sent_id]);
+ J_i.WriteToFile(kbest_file);
+
+ Sample(gamma, xi, J_i, metric, &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/training/pro/mr_pro_reduce.cc b/training/pro/mr_pro_reduce.cc
new file mode 100644
index 00000000..5ef9b470
--- /dev/null
+++ b/training/pro/mr_pro_reduce.cc
@@ -0,0 +1,286 @@
+#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"
+#include "liblbfgs/lbfgs++.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")
+ ("regularization_strength,C",po::value<double>()->default_value(500.0), "l2 regularization strength")
+ ("l1",po::value<double>()->default_value(0.0), "l1 regularization strength")
+ ("regularize_to_weights,y",po::value<double>()->default_value(5000.0), "Differences in learned weights to previous weights are penalized with an l2 penalty with this strength; 0.0 = no effect")
+ ("memory_buffers,m",po::value<unsigned>()->default_value(100), "Number of memory buffers (LBFGS)")
+ ("min_reg,r",po::value<double>()->default_value(0.01), "When tuning (-T) regularization strength, minimum regularization strenght")
+ ("max_reg,R",po::value<double>()->default_value(1e6), "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")
+ ("interpolate_with_weights,p",po::value<double>()->default_value(1.0), "[deprecated] Output weights are p*w + (1-p)*w_prev; 1.0 = no effect")
+ ("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, weight_t* acc) {
+ for (SparseVector<weight_t>::const_iterator it = v.begin();
+ it != v.end(); ++it) {
+ acc[it->first] += it->second * scale;
+ }
+}
+
+double ApplyRegularizationTerms(const double C,
+ const double T,
+ const vector<weight_t>& weights,
+ const vector<weight_t>& prev_weights,
+ weight_t* g) {
+ double reg = 0;
+ for (size_t i = 0; i < weights.size(); ++i) {
+ const double prev_w_i = (i < prev_weights.size() ? prev_weights[i] : 0.0);
+ const double& w_i = weights[i];
+ reg += C * w_i * w_i;
+ g[i] += 2 * C * w_i;
+
+ const double diff_i = w_i - prev_w_i;
+ reg += T * diff_i * diff_i;
+ g[i] += 2 * T * diff_i;
+ }
+ return reg;
+}
+
+double TrainingInference(const vector<weight_t>& x,
+ const vector<pair<bool, SparseVector<weight_t> > >& corpus,
+ 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;
+}
+
+struct ProLoss {
+ ProLoss(const vector<pair<bool, SparseVector<weight_t> > >& tr,
+ const vector<pair<bool, SparseVector<weight_t> > >& te,
+ const double c,
+ const double t,
+ const vector<weight_t>& px) : training(tr), testing(te), C(c), T(t), prev_x(px){}
+ double operator()(const vector<double>& x, double* g) const {
+ fill(g, g + x.size(), 0.0);
+ double cll = TrainingInference(x, training, g);
+ tppl = 0;
+ if (testing.size())
+ tppl = pow(2.0, TrainingInference(x, testing, g) / (log(2) * testing.size()));
+ double ppl = cll / log(2);
+ ppl /= training.size();
+ ppl = pow(2.0, ppl);
+ double reg = ApplyRegularizationTerms(C, T, x, prev_x, g);
+ return cll + reg;
+ }
+ const vector<pair<bool, SparseVector<weight_t> > >& training, testing;
+ const double C, T;
+ const vector<double>& prev_x;
+ mutable double tppl;
+};
+
+// 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 C,
+ const double C1,
+ const double T,
+ const unsigned memory_buffers,
+ const vector<weight_t>& prev_x,
+ vector<weight_t>* px) {
+ assert(px->size() == prev_x.size());
+ ProLoss loss(training, testing, C, T, prev_x);
+ LBFGS<ProLoss> lbfgs(px, loss, memory_buffers, C1);
+ lbfgs.MinimizeFunction();
+ return loss.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 C = conf["regularization_strength"].as<double>(); // will be overridden if parameter is tuned
+ double C1 = conf["l1"].as<double>(); // will be overridden if parameter is tuned
+ const double T = conf["regularize_to_weights"].as<double>();
+ assert(C >= 0.0);
+ assert(min_reg >= 0.0);
+ assert(max_reg >= 0.0);
+ assert(max_reg > min_reg);
+ const double psi = conf["interpolate_with_weights"].as<double>();
+ if (psi < 0.0 || psi > 1.0) { cerr << "Invalid interpolation weight: " << psi << endl; return 1; }
+ 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) {
+ C = min_reg;
+ const double steps = 18;
+ double sweep_factor = exp((log(max_reg) - log(min_reg)) / steps);
+ cerr << "SWEEP FACTOR: " << sweep_factor << endl;
+ while(C < max_reg) {
+ cerr << "C=" << C << "\tT=" <<T << endl;
+ tppl = LearnParameters(training, testing, C, C1, T, conf["memory_buffers"].as<unsigned>(), prev_x, &x);
+ sp.push_back(make_pair(C, tppl));
+ C *= 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;
+ }
+ }
+ C = sp[best_i].first;
+ } // tune regularizer
+ tppl = LearnParameters(training, testing, C, C1, T, conf["memory_buffers"].as<unsigned>(), prev_x, &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 << "# C=" << C << "\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;
+}
diff --git a/training/pro/pro.pl b/training/pro/pro.pl
new file mode 100755
index 00000000..3b30c379
--- /dev/null
+++ b/training/pro/pro.pl
@@ -0,0 +1,555 @@
+#!/usr/bin/env perl
+use strict;
+use File::Basename qw(basename);
+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", "$SCRIPT_DIR/../utils"; }
+
+# 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 $default_jobs = env_default_jobs();
+
+my $UTILS_DIR="$SCRIPT_DIR/../utils";
+require "$UTILS_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 $parallelize = "$UTILS_DIR/parallelize.pl";
+my $libcall = "$UTILS_DIR/libcall.pl";
+my $sentserver = "$UTILS_DIR/sentserver";
+my $sentclient = "$UTILS_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 $best_weights;
+my $psi = 1;
+my $default_max_iter = 30;
+my $max_iterations = $default_max_iter;
+my $jobs = $default_jobs; # 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 = 1; # use make to parallelize
+my $useqsub = 0;
+my $initial_weights;
+my $pass_suffix = '';
+my $devset;
+
+# regularization strength
+my $reg = 500;
+my $reg_previous = 5000;
+
+# Process command-line options
+if (GetOptions(
+ "config=s" => \$iniFile,
+ "weights=s" => \$initial_weights,
+ "devset=s" => \$devset,
+ "jobs=i" => \$jobs,
+ "metric=s" => \$metric,
+ "pass-suffix=s" => \$pass_suffix,
+ "qsub" => \$useqsub,
+ "help" => \$help,
+ "reg=f" => \$reg,
+ "reg-previous=f" => \$reg_previous,
+ "output-dir=s" => \$dir,
+) == 0 || @ARGV!=0 || $help) {
+ print_help();
+ exit;
+}
+
+if ($useqsub) {
+ $use_make = 0;
+ die "LocalEnvironment.pm does not have qsub configuration for this host. Cannot run with --qsub!\n" unless has_qsub();
+}
+
+my @missing_args = ();
+if (!defined $iniFile) { push @missing_args, "--config"; }
+if (!defined $devset) { push @missing_args, "--devset"; }
+if (!defined $initial_weights) { push @missing_args, "--weights"; }
+die "Please specify missing arguments: " . join (', ', @missing_args) . "\n" if (@missing_args);
+
+if ($metric =~ /^(combi|ter)$/i) {
+ $lines_per_mapper = 5;
+}
+
+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) {
+ $DIR_FLAG = '';
+}
+
+unless ($dir){
+ $dir = 'pro';
+}
+unless ($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);
+
+if (-e $dir) {
+ die "ERROR: working dir $dir already exists\n\n";
+} else {
+ mkdir "$dir" or die "Can't mkdir $dir: $!";
+ mkdir "$dir/hgs" or die;
+ mkdir "$dir/scripts" or die;
+ print STDERR <<EOT;
+ DECODER: $decoder
+ INI FILE: $iniFile
+ WORKING DIR: $dir
+ DEVSET: $devset
+ EVAL METRIC: $metric
+ MAX ITERATIONS: $max_iterations
+ PARALLEL JOBS: $jobs
+ HEAD NODE: $host
+ PMEM (DECODING): $pmem
+ INITIAL WEIGHTS: $initial_weights
+EOT
+}
+
+# Generate initial files and values
+check_call("cp $iniFile $newIniFile");
+check_call("cp $initial_weights $dir/weights.0");
+$iniFile = $newIniFile;
+
+my $refs = "$dir/dev.refs";
+split_devset($devset, "$dir/dev.input.raw", $refs);
+my $newsrc = "$dir/dev.input";
+enseg("$dir/dev.input.raw", $newsrc);
+$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
+my @allweights;
+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 $pcmd;
+ if ($use_make) {
+ $pcmd = "cat $srcFile | $parallelize --use-fork -p $pmem -e $logdir -j $jobs --";
+ } else {
+ $pcmd = "cat $srcFile | $parallelize -p $pmem -e $logdir -j $jobs --";
+ }
+ 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 -r $refs -m $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 -m $metric -r $refs -w $inweights -K $dir/kbest < $dir/splag.$im1/$shard > $dir/splag.$im1/$mapoutput";
+ if ($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 . "\|" . $jobid; }
+ }
+ }
+ my @dev_outs = ();
+ my @devtest_outs = ();
+ @dev_outs = @mapoutputs;
+ if ($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 $jobs -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";
+ print STDERR "\nRUNNING CLASSIFIER (REDUCER)\n";
+ print STDERR unchecked_output("date");
+ $cmd="cat @dev_outs | $REDUCER -w $dir/weights.$im1 -C $reg -y $reg_previous --interpolate_with_weights $psi";
+ $cmd .= " > $dir/weights.$iteration";
+ print STDERR "COMMAND:\n$cmd\n";
+ check_bash_call($cmd);
+ $lastWeightsFile = "$dir/weights.$iteration";
+ $lastPScore = $score;
+ $iteration++;
+ print STDERR "\n==========\n";
+}
+
+
+check_call("cp $lastWeightsFile $dir/weights.final");
+print STDERR "\nFINAL WEIGHTS: $dir/weights.final\n(Use -w <this file> with the decoder)\n\n";
+print STDOUT "$dir/weights.final\n";
+
+exit 0;
+
+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;
+}
+
+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 = basename($0); chomp $executable;
+ print << "Help";
+
+Usage: $executable [options]
+
+ $executable [options]
+ Runs a complete PRO optimization using the ini file specified.
+
+Required:
+
+ --config <cdec.ini>
+ Decoder configuration file.
+
+ --devset <files>
+ Dev set source and reference data.
+
+ --weights <file>
+ Initial weights file (use empty file to start from 0)
+
+General options:
+
+ --help
+ Print this message and exit.
+
+ --max-iterations <M>
+ Maximum number of iterations to run. If not specified, defaults
+ to $default_max_iter.
+
+ --metric <method>
+ Metric to optimize.
+ Example values: IBM_BLEU, NIST_BLEU, Koehn_BLEU, TER, Combi
+
+ --pass-suffix <S>
+ If the decoder is doing multi-pass decoding, the pass suffix "2",
+ "3", etc., is used to control what iteration of weights is set.
+
+ --workdir <dir>
+ Directory for intermediate and output files. If not specified, the
+ name is derived from the ini filename. Assuming that the ini
+ filename begins with the decoder name and ends with ini, the default
+ name of the working directory is inferred from the middle part of
+ the filename. E.g. an ini file named decoder.foo.ini would have
+ a default working directory name foo.
+
+Regularization options:
+
+ --reg <F>
+ l2 regularization strength [default=500]. The greater this value,
+ the closer to zero the weights will be.
+
+ --reg-previous <F>
+ l2 penalty for moving away from the weights from the previous
+ iteration. [default=5000]. The greater this value, the closer
+ to the previous iteration's weights the next iteration's weights
+ will be.
+
+Job control options:
+
+ --jobs <I>
+ Number of decoder processes to run in parallel. [default=$default_jobs]
+
+ --qsub
+ Use qsub to run jobs in parallel (qsub must be configured in
+ environment/LocalEnvironment.pm)
+
+ --pmem <N>
+ Amount of physical memory requested for parallel decoding jobs
+ (used with qsub requests only)
+
+Deprecated options:
+
+ --interpolate-with-weights <F>
+ [deprecated] At each iteration the resulting weights are
+ interpolated with the weights from the previous iteration, with
+ this factor. [default=1.0, i.e., no effect]
+
+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 "$0 ".&escaped_shell_args_str(@ORIG_ARGV);
+}
+
+sub split_devset {
+ my ($infile, $outsrc, $outref) = @_;
+ open F, "<$infile" or die "Can't read $infile: $!";
+ open S, ">$outsrc" or die "Can't write $outsrc: $!";
+ open R, ">$outref" or die "Can't write $outref: $!";
+ while(<F>) {
+ chomp;
+ my ($src, @refs) = split /\s*\|\|\|\s*/;
+ die "Malformed devset line: $_\n" unless scalar @refs > 0;
+ print S "$src\n";
+ print R join(' ||| ', @refs) . "\n";
+ }
+ close R;
+ close S;
+ close F;
+}
+