debugged pro trainer

5 files changed, 253 insertions, 70 deletions

diff --git a/pro-train/dist-pro.pl b/pro-train/dist-pro.pl
index 55d7f1fa..c42e3876 100755
--- a/pro-train/dist-pro.pl
+++ b/pro-train/dist-pro.pl
@@ -66,6 +66,7 @@ my $bleu_weight=1;
 my $use_make;  # use make to parallelize line search
 my $dirargs='';
 my $usefork;
+my $initial_weights;
 my $pass_suffix = '';
 my $cpbin=1;
 # Process command-line options
@@ -79,6 +80,7 @@ if (GetOptions(
 	"dry-run" => \$dryrun,
 	"epsilon=s" => \$epsilon,
 	"help" => \$help,
+        "weights=s" => \$initial_weights,
 	"interval" => \$interval,
 	"iteration=i" => \$iteration,
 	"local" => \$run_local,
@@ -212,7 +214,7 @@ if ($dryrun){
         close CMD;
         print STDERR $cline;
         chmod(0755,$cmdfile);
-	check_call("touch $dir/weights.0");
+	check_call("cp $initial_weights $dir/weights.0");
 	die "Can't find weights.0" unless (-e "$dir/weights.0");
 	}
 	write_config(*STDERR);
@@ -239,7 +241,6 @@ my $random_seed = int(time / 1000);
 my $lastWeightsFile;
 my $lastPScore = 0;
 # main optimization loop
-my @mapoutputs = (); # aggregate map outputs over all iters
 while (1){
 	print STDERR "\n\nITERATION $iteration\n==========\n";
 
@@ -262,6 +263,7 @@ while (1){
 	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) {
@@ -333,6 +335,7 @@ while (1){
 		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;
@@ -341,7 +344,7 @@ while (1){
 		$mapoutput =~ s/mapinput/mapoutput/;
 		push @mapoutputs, "$dir/splag.$im1/$mapoutput";
 		$o2i{"$dir/splag.$im1/$mapoutput"} = "$dir/splag.$im1/$shard";
-		my $script = "$MAPPER -s $srcFile -l $metric $refs_comma_sep @allweights < $dir/splag.$im1/$shard > $dir/splag.$im1/$mapoutput";
+		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);
diff --git a/pro-train/mr_pro_map.cc b/pro-train/mr_pro_map.cc
index 128d93ce..4324e8de 100644
--- a/pro-train/mr_pro_map.cc
+++ b/pro-train/mr_pro_map.cc
@@ -2,7 +2,9 @@
 #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>
@@ -22,16 +24,63 @@
 using namespace std;
 namespace po = boost::program_options;
 
+struct ApproxVectorHasher {
+  static const size_t MASK = 0xFFFFFFFFull;
+  union UType {
+    double f;
+    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<double>& x) const {
+    size_t h = 0x573915839;
+    for (SparseVector<double>::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<double>& a, const SparseVector<double>& b) const {
+    SparseVector<double>::const_iterator bit = b.begin();
+    for (SparseVector<double>::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")
-        ("input,i",po::value<string>()->default_value("-"), "Input file to map (- is STDIN)")
-        ("weights,w",po::value<vector<string> >(), "[REQD] Weights files from previous and current iterations")
         ("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)")
@@ -46,7 +95,7 @@ void InitCommandLine(int argc, char** argv, po::variables_map* conf) {
     flag = true;
   }
   if (!conf->count("weights")) {
-    cerr << "Please specify one or more weights using -w <WEIGHTS.TXT>\n";
+    cerr << "Please specify weights using -w <WEIGHTS.TXT>\n";
     flag = true;
   }
   if (flag || conf->count("help")) {
@@ -56,6 +105,7 @@ void InitCommandLine(int argc, char** argv, po::variables_map* conf) {
 }
 
 struct HypInfo {
+  HypInfo() : g_(-100.0) {}
   HypInfo(const vector<WordID>& h, const SparseVector<double>& feats) : hyp(h), g_(-100.0), x(feats) {}
 
   // lazy evaluation
@@ -66,10 +116,92 @@ struct HypInfo {
   }
   vector<WordID> hyp;
   mutable double g_;
- public:
   SparseVector<double> 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<double>* out) {
+  SparseVector<double>& 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 {
@@ -81,6 +213,7 @@ struct ThresholdAlpha {
 struct TrainingInstance {
   TrainingInstance(const SparseVector<double>& feats, bool positive, double diff) : x(feats), y(positive), gdiff(diff) {}
   SparseVector<double> x;
+#undef DEBUGGING_PRO
 #ifdef DEBUGGING_PRO
   vector<WordID> a;
   vector<WordID> b;
@@ -88,6 +221,11 @@ struct TrainingInstance {
   bool y;
   double 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 {
@@ -95,36 +233,51 @@ struct DiffOrder {
   }
 };
 
-template<typename Alpha>
-void Sample(const unsigned gamma, const unsigned xi, const vector<HypInfo>& J_i, const SentenceScorer& scorer, const Alpha& alpha_i, bool invert_score, vector<TrainingInstance>* pv) {
-  vector<TrainingInstance> v;
+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;
+  double avg_diff = 0;
   for (unsigned i = 0; i < gamma; ++i) {
-    size_t a = rng->inclusive(0, J_i.size() - 1)();
-    size_t b = rng->inclusive(0, J_i.size() - 1)();
+    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;
     double ga = J_i[a].g(scorer);
     double gb = J_i[b].g(scorer);
-    bool positive = ga < gb;
+    bool positive = gb < ga;
     if (invert_score) positive = !positive;
-    double gdiff = fabs(ga - gb);
+    const double gdiff = fabs(ga - gb);
     if (!gdiff) continue;
-    if (rng->next() < alpha_i(gdiff)) {
-      v.push_back(TrainingInstance((J_i[a].x - J_i[b].x).erase_zeros(), positive, gdiff));
+    avg_diff += gdiff;
+    SparseVector<double> 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
-      v.back().a = J_i[a].hyp;
-      v.back().b = J_i[b].hyp;
+    v1.back().a = J_i[a].hyp;
+    v1.back().b = J_i[b].hyp;
+    cerr << "N: " << v1.back() << endl;
 #endif
-    }
   }
-  vector<TrainingInstance>::iterator mid = v.begin() + xi;
-  if (xi > v.size()) mid = v.end();
-  partial_sort(v.begin(), mid, v.end(), DiffOrder());
-  copy(v.begin(), mid, back_inserter(*pv));
+  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 (v.size() >= 5)
-    for (int i =0; i < 5; ++i) {
-      cerr << v[i].gdiff << " y=" << v[i].y << "\tA:" << TD::GetString(v[i].a) << "\n\tB: " << TD::GetString(v[i].b) << endl;
+  if (v2.size() >= 5) {
+    for (int i =0; i < (mid - v2.begin()); ++i) {
+      cerr << v2[i] << endl;
     }
+    cerr << pv->back() << endl;
+  }
 #endif
 }
 
@@ -136,6 +289,7 @@ int main(int argc, char** argv) {
   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;
@@ -146,13 +300,15 @@ int main(int argc, char** argv) {
   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>();
-  vector<string> weights_files = conf["weights"].as<vector<string> >();
-  vector<vector<double> > weights(weights_files.size());
-  for (int i = 0; i < weights.size(); ++i) {
+  string weightsf = conf["weights"].as<string>();
+  vector<double> weights;
+  {
     Weights w;
-    w.InitFromFile(weights_files[i]);
-    w.InitVector(&weights[i]);
+    w.InitFromFile(weightsf);
+    w.InitVector(&weights);
   }
+  string kbest_repo = conf["kbest_repository"].as<string>();
+  MkDirP(kbest_repo);
   while(in) {
     vector<TrainingInstance> v;
     string line;
@@ -164,24 +320,26 @@ int main(int argc, char** argv) {
     // path-to-file (JSON) sent_id
     is >> file >> sent_id;
     ReadFile rf(file);
-    HypergraphIO::ReadFromJSON(rf.stream(), &hg);
+    ostringstream os;
     vector<HypInfo> J_i;
-    int start = weights.size();
-    start -= 4;
-    if (start < 0) start = 0;
-    for (int i = start; i < weights.size(); ++i) {
-      hg.Reweight(weights[i]);
-      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));
-      }
+    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], ThresholdAlpha(0.05), (type == TER), &v);
+    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;
diff --git a/pro-train/mr_pro_reduce.cc b/pro-train/mr_pro_reduce.cc
index 2b9c5ce7..e1a7db8a 100644
--- a/pro-train/mr_pro_reduce.cc
+++ b/pro-train/mr_pro_reduce.cc
@@ -24,7 +24,7 @@ void InitCommandLine(int argc, char** argv, po::variables_map* conf) {
         ("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(0.5), "Sigma squared for Gaussian prior")
+        ("sigma_squared,s",po::value<double>()->default_value(1.0), "Sigma squared for Gaussian prior")
         ("help,h", "Help");
   po::options_description dcmdline_options;
   dcmdline_options.add(opts);
@@ -35,6 +35,31 @@ void InitCommandLine(int argc, char** argv, po::variables_map* conf) {
   }
 }
 
+void ParseSparseVector(string& line, size_t cur, SparseVector<double>* out) {
+  SparseVector<double>& 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;
+  }
+}
+
 int main(int argc, char** argv) {
   po::variables_map conf;
   InitCommandLine(argc, argv, &conf);
@@ -60,28 +85,7 @@ int main(int argc, char** argv) {
     assert(ks == 1);
     const bool y = line[0] == '1';
     SparseVector<double> x;
-    size_t last_start = ks + 1;
-    size_t last_comma = string::npos;
-    size_t cur = last_start;
-    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;
-    }
+    ParseSparseVector(line, ks + 1, &x);
     training.push_back(make_pair(y, x));
   }
   if (flag) cerr << endl;
@@ -95,6 +99,7 @@ int main(int argc, char** argv) {
   SparseVector<double> g;
   bool converged = false;
   LBFGSOptimizer opt(FD::NumFeats(), conf["memory_buffers"].as<unsigned>());
+  double ppl = 0;
   while(!converged) {
     double cll = 0;
     double dbias = 0;
@@ -114,14 +119,18 @@ int main(int argc, char** argv) {
       lp_false*=-1;
       if (training[i].first) {  // true label
         cll -= lp_true;
+        ppl += lp_true / log(2);
         g -= training[i].second * exp(lp_false);
         dbias -= exp(lp_false);
       } else {                  // false label
         cll -= lp_false;
+        ppl += lp_false / log(2);
         g += training[i].second * exp(lp_true);
         dbias += exp(lp_true);
       }
     }
+    ppl /= training.size();
+    ppl = pow(2.0, - ppl);
     vg.clear();
     g.init_vector(&vg);
     vg[0] = dbias;
@@ -139,7 +148,7 @@ int main(int argc, char** argv) {
     double reg = 0;
 #endif
     cll += reg;
-    cerr << cll << " (REG=" << reg << ")\t";
+    cerr << cll << " (REG=" << reg << ")\tPPL=" << ppl << "\t";
     bool failed = false;
     try {
       opt.Optimize(cll, vg, &x);
diff --git a/utils/filelib.cc b/utils/filelib.cc
index 79ad2847..a0969b1a 100644
--- a/utils/filelib.cc
+++ b/utils/filelib.cc
@@ -20,3 +20,15 @@ bool DirectoryExists(const string& dir) {
   return false;
 }
 
+void MkDirP(const string& dir) {
+  if (DirectoryExists(dir)) return;
+  if (mkdir(dir.c_str(), 0777)) {
+    perror(dir.c_str());
+    abort();
+  }
+  if (chmod(dir.c_str(), 07777)) {
+    perror(dir.c_str());
+    abort();
+  }
+}
+
diff --git a/utils/filelib.h b/utils/filelib.h
index dda98671..a8622246 100644
--- a/utils/filelib.h
+++ b/utils/filelib.h
@@ -12,6 +12,7 @@
 
 bool FileExists(const std::string& file_name);
 bool DirectoryExists(const std::string& dir_name);
+void MkDirP(const std::string& dir_name);
 
 // reads from standard in if filename is -
 // uncompresses if file ends with .gz