add support to rampion for accumulating k-best lists

7 files changed, 78 insertions, 52 deletions

diff --git a/pro-train/mr_pro_map.cc b/pro-train/mr_pro_map.cc
index bb13fdf4..eef40b8a 100644
--- a/pro-train/mr_pro_map.cc
+++ b/pro-train/mr_pro_map.cc
@@ -101,8 +101,8 @@ void Sample(const unsigned gamma,
     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].score_stats);
-    float gb = metric->ComputeScore(J_i[b].score_stats);
+    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);
diff --git a/rampion/Makefile.am b/rampion/Makefile.am
index 12df39c2..f4dbb7cc 100644
--- a/rampion/Makefile.am
+++ b/rampion/Makefile.am
@@ -1,6 +1,6 @@
 bin_PROGRAMS = rampion_cccp
 
 rampion_cccp_SOURCES = rampion_cccp.cc
-rampion_cccp_LDADD = $(top_srcdir)/decoder/libcdec.a $(top_srcdir)/mteval/libmteval.a $(top_srcdir)/utils/libutils.a -lz
+rampion_cccp_LDADD = $(top_srcdir)/training/libtraining.a $(top_srcdir)/decoder/libcdec.a $(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
+AM_CPPFLAGS = -W -Wall $(GTEST_CPPFLAGS) -I$(top_srcdir)/utils -I$(top_srcdir)/decoder -I$(top_srcdir)/mteval -I$(top_srcdir)/training
diff --git a/rampion/rampion.pl b/rampion/rampion.pl
index 9884f453..55f7b3f1 100755
--- a/rampion/rampion.pl
+++ b/rampion/rampion.pl
@@ -65,12 +65,14 @@ my $cpbin=1;
 my $tune_regularizer = 0;
 my $reg = 500;
 my $reg_previous = 5000;
+my $dont_accum = 0;
 
 # Process command-line options
 Getopt::Long::Configure("no_auto_abbrev");
 if (GetOptions(
 	"jobs=i" => \$jobs,
 	"dont-clean" => \$disable_clean,
+	"dont-accumulate" => \$dont_accum,
 	"pass-suffix=s" => \$pass_suffix,
         "qsub" => \$useqsub,
 	"dry-run" => \$dryrun,
@@ -163,8 +165,6 @@ 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, $iniFile);
@@ -229,6 +229,13 @@ close F;
 unless($best_weights){ $best_weights = $weights; }
 unless($projected_score){ $projected_score = 0.0; }
 $seen_weights{$weights} = 1;
+my $kbest = "$dir/kbest";
+if ($dont_accum) {
+  $kbest = '';
+} else {
+  check_call("mkdir -p $kbest");
+  $kbest = "--kbest_repository $kbest";
+}
 
 my $random_seed = int(time / 1000);
 my $lastWeightsFile;
@@ -305,7 +312,7 @@ while (1){
 	$cmd="$MAPINPUT $dir/hgs > $dir/agenda.$im1";
 	print STDERR "COMMAND:\n$cmd\n";
 	check_call($cmd);
-	$cmd="$MAPPER $refs_comma_sep -m $metric -i $dir/agenda.$im1 -w $inweights > $outweights";
+	$cmd="$MAPPER $refs_comma_sep -m $metric -i $dir/agenda.$im1 $kbest -w $inweights > $outweights";
 	check_call($cmd);
 	$lastWeightsFile = $outweights;
 	$iteration++;
@@ -445,6 +452,9 @@ General options:
 	--help
 		Print this message and exit.
 
+	--dont-accumulate
+		Don't accumulate k-best lists from multiple iterations.
+
 	--max-iterations <M>
 		Maximum number of iterations to run.  If not specified, defaults
 		to $default_max_iter.
diff --git a/rampion/rampion_cccp.cc b/rampion/rampion_cccp.cc
index 7a6f1f0c..1e36dc51 100644
--- a/rampion/rampion_cccp.cc
+++ b/rampion/rampion_cccp.cc
@@ -14,6 +14,7 @@
 #include "viterbi.h"
 #include "ns.h"
 #include "ns_docscorer.h"
+#include "candidate_set.h"
 
 using namespace std;
 namespace po = boost::program_options;
@@ -25,6 +26,7 @@ void InitCommandLine(int argc, char** argv, po::variables_map* conf) {
         ("weights,w",po::value<string>(), "[REQD] Weights files from current iterations")
         ("input,i",po::value<string>()->default_value("-"), "Input file to map (- is STDIN)")
         ("evaluation_metric,m",po::value<string>()->default_value("IBM_BLEU"), "Evaluation metric (ibm_bleu, koehn_bleu, nist_bleu, ter, meteor, etc.)")
+        ("kbest_repository,R",po::value<string>(), "Accumulate k-best lists from previous iterations (parameter is path to repository)")
         ("kbest_size,k",po::value<unsigned>()->default_value(500u), "Top k-hypotheses to extract")
         ("cccp_iterations,I", po::value<unsigned>()->default_value(10u), "CCCP iterations (T')")
         ("ssd_iterations,J", po::value<unsigned>()->default_value(5u), "Stochastic subgradient iterations (T'')")
@@ -50,38 +52,36 @@ void InitCommandLine(int argc, char** argv, po::variables_map* conf) {
   }
 }
 
-struct HypInfo {
-  HypInfo() : g(-100.0f) {}
-  HypInfo(const vector<WordID>& h,
-          const SparseVector<weight_t>& feats,
-          const SegmentEvaluator& scorer, const EvaluationMetric* metric) : hyp(h), x(feats) {
-    SufficientStats ss;
-    scorer.Evaluate(hyp, &ss);
-    g = metric->ComputeScore(ss);
+struct GainFunction {
+  explicit GainFunction(const EvaluationMetric* m) : metric(m) {}
+  float operator()(const SufficientStats& eval_feats) const {
+    float g = metric->ComputeScore(eval_feats);
     if (!metric->IsErrorMetric()) g = 1 - g;
+    return g;
   }
-
-  vector<WordID> hyp;
-  float g;
-  SparseVector<weight_t> x;
+  const EvaluationMetric* metric;
 };
 
-void CostAugmentedSearch(const vector<HypInfo>& kbest,
+template <typename GainFunc>
+void CostAugmentedSearch(const GainFunc& gain,
+                         const training::CandidateSet& cs,
                          const SparseVector<double>& w,
                          double alpha,
                          SparseVector<double>* fmap) {
   unsigned best_i = 0;
   double best = -numeric_limits<double>::infinity();
-  for (unsigned i = 0; i < kbest.size(); ++i) {
-    double s = kbest[i].x.dot(w) + alpha * kbest[i].g;
+  for (unsigned i = 0; i < cs.size(); ++i) {
+    double s = cs[i].fmap.dot(w) + alpha * gain(cs[i].eval_feats);
     if (s > best) {
       best = s;
       best_i = i;
     }
   }
-  *fmap = kbest[best_i].x;
+  *fmap = cs[best_i].fmap;
 }
 
+
+
 // runs lines 4--15 of rampion algorithm
 int main(int argc, char** argv) {
   po::variables_map conf;
@@ -97,6 +97,11 @@ int main(int argc, char** argv) {
   Hypergraph hg;
   string last_file;
   ReadFile in_read(conf["input"].as<string>());
+  string kbest_repo;
+  if (conf.count("kbest_repository")) {
+    kbest_repo = conf["kbest_repository"].as<string>();
+    MkDirP(kbest_repo);
+  }
   istream &in=*in_read.stream();
   const unsigned kbest_size = conf["kbest_size"].as<unsigned>();
   const unsigned tp = conf["cccp_iterations"].as<unsigned>();
@@ -112,40 +117,44 @@ int main(int argc, char** argv) {
     Weights::InitSparseVector(vweights, &weights);
   }
   string line, file;
-  vector<vector<HypInfo> > kis;
+  vector<training::CandidateSet> kis;
   cerr << "Loading hypergraphs...\n";
   while(getline(in, line)) {
     istringstream is(line);
     int sent_id;
     kis.resize(kis.size() + 1);
-    vector<HypInfo>& curkbest = kis.back();
+    training::CandidateSet& curkbest = kis.back();
+    string kbest_file;
+    if (kbest_repo.size()) {
+      ostringstream os;
+      os << kbest_repo << "/kbest." << sent_id << ".txt.gz";
+      kbest_file = os.str();
+      if (FileExists(kbest_file))
+        curkbest.ReadFromFile(kbest_file);
+    }
     is >> file >> sent_id;
     ReadFile rf(file);
     if (kis.size() % 5 == 0) { cerr << '.'; }
     if (kis.size() % 200 == 0) { cerr << " [" << kis.size() << "]\n"; }
     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;
-      curkbest.push_back(HypInfo(d->yield, d->feature_values, *ds[sent_id], metric));
-    }
+    curkbest.AddKBestCandidates(hg, kbest_size, ds[sent_id]);
+    if (kbest_file.size())
+      curkbest.WriteToFile(kbest_file);
   }
   cerr << "\nHypergraphs loaded.\n";
 
   vector<SparseVector<weight_t> > goals(kis.size());  // f(x_i,y+,h+)
   SparseVector<weight_t> fear;  // f(x,y-,h-)
+  const GainFunction gain(metric);
   for (unsigned iterp = 1; iterp <= tp; ++iterp) {
     cerr << "CCCP Iteration " << iterp << endl;
-    for (int i = 0; i < goals.size(); ++i)
-      CostAugmentedSearch(kis[i], weights, goodsign * alpha, &goals[i]);
+    for (unsigned i = 0; i < goals.size(); ++i)
+      CostAugmentedSearch(gain, kis[i], weights, goodsign * alpha, &goals[i]);
     for (unsigned iterpp = 1; iterpp <= tpp; ++iterpp) {
       cerr << "  SSD Iteration " << iterpp << endl;
-      for (int i = 0; i < goals.size(); ++i) {
-        CostAugmentedSearch(kis[i], weights, badsign * alpha, &fear);
+      for (unsigned i = 0; i < goals.size(); ++i) {
+        CostAugmentedSearch(gain, kis[i], weights, badsign * alpha, &fear);
         weights -= weights * (eta * reg / goals.size());
         weights += (goals[i] - fear) * eta;
       }
diff --git a/training/candidate_set.cc b/training/candidate_set.cc
index e2ca9ad2..8c086ece 100644
--- a/training/candidate_set.cc
+++ b/training/candidate_set.cc
@@ -112,7 +112,7 @@ void CandidateSet::WriteToFile(const string& file) const {
   for (unsigned i = 0; i < cs.size(); ++i) {
     out << TD::GetString(cs[i].ewords) << endl;
     out << cs[i].fmap << endl;
-    cs[i].score_stats.Encode(&ss);
+    cs[i].eval_feats.Encode(&ss);
     out << ss << endl;
   }
 }
@@ -131,7 +131,7 @@ void CandidateSet::ReadFromFile(const string& file) {
     cs.push_back(Candidate());
     TD::ConvertSentence(cand, &cs.back().ewords);
     ParseSparseVector(feats, 0, &cs.back().fmap);
-    cs.back().score_stats = SufficientStats(ss);
+    cs.back().eval_feats = SufficientStats(ss);
   }
   cerr << "  read " << cs.size() << " candidates\n";
 }
@@ -160,7 +160,7 @@ void CandidateSet::AddKBestCandidates(const Hypergraph& hg, size_t kbest_size, c
     if (!d) break;
     cs.push_back(Candidate(d->yield, d->feature_values));
     if (scorer)
-      scorer->Evaluate(d->yield, &cs.back().score_stats);
+      scorer->Evaluate(d->yield, &cs.back().eval_feats);
   }
   Dedup();
 }
diff --git a/training/candidate_set.h b/training/candidate_set.h
index 824a4de2..9d326ed0 100644
--- a/training/candidate_set.h
+++ b/training/candidate_set.h
@@ -15,16 +15,25 @@ namespace training {
 struct Candidate {
   Candidate() {}
   Candidate(const std::vector<WordID>& e, const SparseVector<double>& fm) :
-    ewords(e),
-    fmap(fm) {}
-  std::vector<WordID> ewords;
-  SparseVector<double> fmap;
-  SufficientStats score_stats;
+      ewords(e),
+      fmap(fm) {}
+  Candidate(const std::vector<WordID>& e,
+            const SparseVector<double>& fm,
+            const SegmentEvaluator& se) :
+      ewords(e),
+      fmap(fm) {
+    se.Evaluate(ewords, &eval_feats);
+  }
+
   void swap(Candidate& other) {
-    score_stats.swap(other.score_stats);
+    eval_feats.swap(other.eval_feats);
     ewords.swap(other.ewords);
     fmap.swap(other.fmap);
   }
+
+  std::vector<WordID> ewords;
+  SparseVector<double> fmap;
+  SufficientStats eval_feats;
 };
 
 // represents some kind of collection of translation candidates, e.g.
diff --git a/training/mpi_flex_optimize.cc b/training/mpi_flex_optimize.cc
index a9197208..a9ead018 100644
--- a/training/mpi_flex_optimize.cc
+++ b/training/mpi_flex_optimize.cc
@@ -179,18 +179,16 @@ double ApplyRegularizationTerms(const double C,
                                 const double T,
                                 const vector<double>& weights,
                                 const vector<double>& prev_weights,
-                                vector<double>* g) {
-  assert(weights.size() == g->size());
+                                double* 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];
-    double& g_i = (*g)[i];
     reg += C * w_i * w_i;
-    g_i += 2 * C * w_i;
+    g[i] += 2 * C * w_i;
 
     reg += T * (w_i - prev_w_i) * (w_i - prev_w_i);
-    g_i += 2 * T * (w_i - prev_w_i);
+    g[i] += 2 * T * (w_i - prev_w_i);
   }
   return reg;
 }
@@ -365,7 +363,7 @@ int main(int argc, char** argv) {
                                 time_series_strength, // * (iter == 0 ? 0.0 : 1.0),
                                 cur_weights,
                                 prev_weights,
-                                &gg);
+                                &gg[0]);
           obj += r;
           if (mi == 0 || mi == (minibatch_iterations - 1)) {
             if (!mi) cerr << iter << ' '; else cerr << ' ';