Merge branch 'master' of github.com:redpony/cdec

5 files changed, 255 insertions, 6 deletions

diff --git a/decoder/decoder.cc b/decoder/decoder.cc
index 0d7e84ad..d8dd0c61 100644
--- a/decoder/decoder.cc
+++ b/decoder/decoder.cc
@@ -176,8 +176,11 @@ struct DecoderImpl {
   bool Decode(const string& input, DecoderObserver*);
   void SetWeights(const vector<double>& weights) {
     init_weights = weights;
-    for (int i = 0; i < rescoring_passes.size(); ++i)
+    for (int i = 0; i < rescoring_passes.size(); ++i) {
+      if (rescoring_passes[i].models)
+        rescoring_passes[i].models->SetWeights(weights);
       rescoring_passes[i].weight_vector = weights;
+    }
   }
   void SetId(int next_sent_id) { sent_id = next_sent_id - 1; }
 
diff --git a/decoder/ff.h b/decoder/ff.h
index 89b8b067..c3aa9b6d 100644
--- a/decoder/ff.h
+++ b/decoder/ff.h
@@ -264,6 +264,9 @@ class ModelSet {
   ModelSet(const std::vector<double>& weights,
            const std::vector<const FeatureFunction*>& models);
 
+  // TODO stop stupid copy
+  void SetWeights(const std::vector<double>& w) { weights_ = w; }
+
   // sets edge->feature_values_ and edge->edge_prob_
   // NOTE: edge must not necessarily be in hg.edges_ but its TAIL nodes
   // must be.  edge features are supposed to be overwritten, not added to (possibly because rule features aren't in ModelSet so need to be left alone
diff --git a/mira/Makefile.am b/mira/Makefile.am
new file mode 100644
index 00000000..7b4a4e12
--- /dev/null
+++ b/mira/Makefile.am
@@ -0,0 +1,6 @@
+bin_PROGRAMS = kbest_mira
+
+kbest_mira_SOURCES = kbest_mira.cc
+kbest_mira_LDADD = $(top_srcdir)/decoder/libcdec.a $(top_srcdir)/mteval/libmteval.a $(top_srcdir)/utils/libutils.a ../klm/lm/libklm.a ../klm/util/libklm_util.a -lz
+
+AM_CPPFLAGS = -W -Wall -Wno-sign-compare -I$(top_srcdir)/utils -I$(top_srcdir)/decoder -I$(top_srcdir)/mteval
diff --git a/mira/kbest_mira.cc b/mira/kbest_mira.cc
new file mode 100644
index 00000000..7ff207a8
--- /dev/null
+++ b/mira/kbest_mira.cc
@@ -0,0 +1,237 @@
+#include <sstream>
+#include <iostream>
+#include <vector>
+#include <cassert>
+#include <cmath>
+
+#include "config.h"
+
+#include <boost/shared_ptr.hpp>
+#include <boost/program_options.hpp>
+#include <boost/program_options/variables_map.hpp>
+
+#include "sentence_metadata.h"
+#include "scorer.h"
+#include "verbose.h"
+#include "viterbi.h"
+#include "hg.h"
+#include "prob.h"
+#include "kbest.h"
+#include "ff_register.h"
+#include "decoder.h"
+#include "filelib.h"
+#include "fdict.h"
+#include "weights.h"
+#include "sparse_vector.h"
+
+using namespace std;
+using boost::shared_ptr;
+namespace po = boost::program_options;
+
+void SanityCheck(const vector<double>& w) {
+  for (int i = 0; i < w.size(); ++i) {
+    assert(!isnan(w[i]));
+    assert(!isinf(w[i]));
+  }
+}
+
+struct FComp {
+  const vector<double>& w_;
+  FComp(const vector<double>& w) : w_(w) {}
+  bool operator()(int a, int b) const {
+    return fabs(w_[a]) > fabs(w_[b]);
+  }
+};
+
+void ShowLargestFeatures(const vector<double>& w) {
+  vector<int> fnums(w.size());
+  for (int i = 0; i < w.size(); ++i)
+    fnums[i] = i;
+  vector<int>::iterator mid = fnums.begin();
+  mid += (w.size() > 10 ? 10 : w.size());
+  partial_sort(fnums.begin(), mid, fnums.end(), FComp(w));
+  cerr << "TOP FEATURES:";
+  for (vector<int>::iterator i = fnums.begin(); i != mid; ++i) {
+    cerr << ' ' << FD::Convert(*i) << '=' << w[*i];
+  }
+  cerr << endl;
+}
+
+bool InitCommandLine(int argc, char** argv, po::variables_map* conf) {
+  po::options_description opts("Configuration options");
+  opts.add_options()
+        ("input_weights,w",po::value<string>(),"Input feature weights file")
+        ("source,i",po::value<string>(),"Source file for development set")
+        ("reference,r",po::value<vector<string> >(), "[REQD] Reference translation(s) (tokenized text file)")
+        ("mt_metric,m",po::value<string>()->default_value("ter"), "Scoring metric (ibm_bleu, nist_bleu, koehn_bleu, ter, combi)")
+        ("max_step_size,C", po::value<double>()->default_value(0.0001), "maximum step size (C)")
+        ("mt_metric_scale,s", po::value<double>()->default_value(1.0), "Amount to scale MT loss function by")
+        ("decoder_config,c",po::value<string>(),"Decoder configuration file");
+  po::options_description clo("Command line options");
+  clo.add_options()
+        ("config", po::value<string>(), "Configuration file")
+        ("help,h", "Print this help message and exit");
+  po::options_description dconfig_options, dcmdline_options;
+  dconfig_options.add(opts);
+  dcmdline_options.add(opts).add(clo);
+  
+  po::store(parse_command_line(argc, argv, dcmdline_options), *conf);
+  if (conf->count("config")) {
+    ifstream config((*conf)["config"].as<string>().c_str());
+    po::store(po::parse_config_file(config, dconfig_options), *conf);
+  }
+  po::notify(*conf);
+
+  if (conf->count("help") || !conf->count("input_weights") || !conf->count("source") || !conf->count("decoder_config") || !conf->count("reference")) {
+    cerr << dcmdline_options << endl;
+    return false;
+  }
+  return true;
+}
+
+static const double kMINUS_EPSILON = -1e-6;
+
+struct HypothesisInfo {
+  SparseVector<double> features;
+  double mt_metric;
+};
+
+struct GoodBadOracle {
+  shared_ptr<HypothesisInfo> good;
+  shared_ptr<HypothesisInfo> bad;
+};
+
+struct TrainingObserver : public DecoderObserver {
+  TrainingObserver(const DocScorer& d, vector<GoodBadOracle>* o) : ds(d), oracles(*o) {}
+  const DocScorer& ds;
+  vector<GoodBadOracle>& oracles;
+  shared_ptr<HypothesisInfo> cur_best;
+
+  const HypothesisInfo& GetCurrentBestHypothesis() const {
+    return *cur_best;
+  }
+
+  virtual void NotifyTranslationForest(const SentenceMetadata& smeta, Hypergraph* hg) {
+    UpdateOracles(smeta.GetSentenceID(), *hg);
+  }
+
+  shared_ptr<HypothesisInfo> MakeHypothesisInfo(const SparseVector<double>& feats, const double score) {
+    shared_ptr<HypothesisInfo> h(new HypothesisInfo);
+    h->features = feats;
+    h->mt_metric = score;
+    return h;
+  }
+
+  void UpdateOracles(int sent_id, const Hypergraph& forest) {
+    int kbest_size = 330;
+    shared_ptr<HypothesisInfo>& cur_good = oracles[sent_id].good;
+    shared_ptr<HypothesisInfo>& cur_bad = oracles[sent_id].bad;
+    cur_bad.reset();  // TODO get rid of??
+    KBest::KBestDerivations<vector<WordID>, ESentenceTraversal> kbest(forest, kbest_size);
+    for (int i = 0; i < kbest_size; ++i) {
+      const KBest::KBestDerivations<vector<WordID>, ESentenceTraversal>::Derivation* d =
+        kbest.LazyKthBest(forest.nodes_.size() - 1, i);
+      if (!d) break;
+      float sentscore = ds[sent_id]->ScoreCandidate(d->yield)->ComputeScore();
+//      cerr << TD::GetString(d->yield) << " ||| " << d->score << " ||| " << sentscore << endl;
+      if (i == 0)
+        cur_best = MakeHypothesisInfo(d->feature_values, sentscore);
+      if (!cur_good || sentscore < cur_good->mt_metric)
+        cur_good = MakeHypothesisInfo(d->feature_values, sentscore);
+      if (!cur_bad || sentscore > cur_bad->mt_metric)
+        cur_bad = MakeHypothesisInfo(d->feature_values, sentscore);
+    }
+    cerr << "GOOD: " << cur_good->mt_metric << endl;
+    cerr << " BAD: " << cur_bad->mt_metric << endl;
+    cerr << "  #1: " << cur_best->mt_metric << endl;
+  }
+};
+
+void ReadTrainingCorpus(const string& fname, vector<string>* c) {
+  ReadFile rf(fname);
+  istream& in = *rf.stream();
+  string line;
+  while(in) {
+    getline(in, line);
+    if (!in) break;
+    c->push_back(line);
+  }
+}
+
+bool ApproxEqual(double a, double b) {
+  if (a == b) return true;
+  return (fabs(a-b)/fabs(b)) < 0.000001;
+}
+
+int main(int argc, char** argv) {
+  register_feature_functions();
+  //SetSilent(true);  // turn off verbose decoder output
+
+  po::variables_map conf;
+  if (!InitCommandLine(argc, argv, &conf)) return 1;
+
+  vector<string> corpus;
+  ReadTrainingCorpus(conf["source"].as<string>(), &corpus);
+  const string metric_name = conf["mt_metric"].as<string>();
+  ScoreType type = ScoreTypeFromString(metric_name);
+  DocScorer ds(type, conf["reference"].as<vector<string> >(), "");
+  cerr << "Loaded " << ds.size() << " references for scoring with " << metric_name << endl;
+  if (ds.size() != corpus.size()) {
+    cerr << "Mismatched number of references (" << ds.size() << ") and sources (" << corpus.size() << ")\n";
+    return 1;
+  }
+  // load initial weights
+  Weights weights;
+  weights.InitFromFile(conf["input_weights"].as<string>());
+  SparseVector<double> lambdas;
+  weights.InitSparseVector(&lambdas);
+
+  // freeze feature set (should be optional?)
+  const bool freeze_feature_set = true;
+  if (freeze_feature_set) FD::Freeze();
+
+  ReadFile ini_rf(conf["decoder_config"].as<string>());
+  Decoder decoder(ini_rf.stream());
+  const double max_step_size = conf["max_step_size"].as<double>();
+  const double mt_metric_scale = conf["mt_metric_scale"].as<double>();
+
+  assert(corpus.size() > 0);
+  vector<GoodBadOracle> oracles(corpus.size());
+
+  TrainingObserver observer(ds, &oracles);
+  int cur_sent = 0;
+  bool converged = false;
+  vector<double> dense_weights;
+  while (!converged) {
+    dense_weights.clear();
+    weights.InitFromVector(lambdas);
+    weights.InitVector(&dense_weights);
+    decoder.SetWeights(dense_weights);
+    if (corpus.size() == cur_sent) cur_sent = 0;
+    decoder.SetId(cur_sent);
+    decoder.Decode(corpus[cur_sent], &observer);  // update oracles
+    const HypothesisInfo& cur_hyp = observer.GetCurrentBestHypothesis();
+    const HypothesisInfo& cur_good = *oracles[cur_sent].good;
+    const HypothesisInfo& cur_bad = *oracles[cur_sent].bad;
+    if (!ApproxEqual(cur_hyp.mt_metric, cur_good.mt_metric)) {
+      const double loss = cur_bad.features.dot(dense_weights) - cur_good.features.dot(dense_weights) +
+          mt_metric_scale * (cur_good.mt_metric - cur_bad.mt_metric);
+      cerr << "LOSS: " << loss << endl;
+      if (loss > 0.0) {
+        SparseVector<double> diff = cur_good.features;
+        diff -= cur_bad.features;
+        double step_size = loss / diff.l2norm_sq();
+        //cerr << loss << " " << step_size << " " << diff << endl;
+        if (step_size > max_step_size) step_size = max_step_size;
+        lambdas += (cur_good.features * step_size);
+        lambdas -= (cur_bad.features * step_size);
+        //cerr << "L: " << lambdas << endl;
+      }
+    }
+    ++cur_sent;
+    static int cc = 0; ++cc; if (cc==250) converged = true;
+  }
+  weights.WriteToFile("-");
+  return 0;
+}
+
diff --git a/utils/sparse_vector.h b/utils/sparse_vector.h
index 274220ef..e3721b50 100644
--- a/utils/sparse_vector.h
+++ b/utils/sparse_vector.h
@@ -54,16 +54,16 @@ SparseVector<T> operator*(const SparseVector<T>& a, const T& b) {
   return result *= b;
 }
 
-template <class T>
-SparseVector<T> operator/(const SparseVector<T>& a, const double& b) {
+template <class T, typename S>
+SparseVector<T> operator/(const SparseVector<T>& a, const S& b) {
   SparseVector<T> result = a;
-  return result *= b;
+  return result /= b;
 }
 
 template <class T>
-SparseVector<T> operator/(const SparseVector<T>& a, const T& b) {
+SparseVector<T> operator/(const SparseVector<T>& a, const double& b) {
   SparseVector<T> result = a;
-  return result *= b;
+  return result /= b;
 }
 
 #include "fdict.h"