Added debugging info to kbmira.

7 files changed, 52 insertions, 2 deletions

diff --git a/mert/HypPackEnumerator.cpp b/mert/HypPackEnumerator.cpp
index ffbf3cfb5..808ba8335 100644
--- a/mert/HypPackEnumerator.cpp
+++ b/mert/HypPackEnumerator.cpp
@@ -132,6 +132,10 @@ const ScoreDataItem& StreamingHypPackEnumerator::scoresAt(size_t index) {
   return m_scoreDataIters[pij.first]->operator[](pij.second);
 }
 
+size_t StreamingHypPackEnumerator::cur_id() {
+  return m_sentenceId;
+}
+
 /* --------- RandomAccessHypPackEnumerator ------------- */
 
 RandomAccessHypPackEnumerator::RandomAccessHypPackEnumerator(vector<string> const& featureFiles,
@@ -181,7 +185,9 @@ const ScoreDataItem& RandomAccessHypPackEnumerator::scoresAt(size_t i) {
   return m_scores[m_indexes[m_cur_index]][i];
 }
 
-  
+size_t RandomAccessHypPackEnumerator::cur_id() {
+  return m_indexes[m_cur_index];
+}  
 // --Emacs trickery--
 // Local Variables:
 // mode:c++
diff --git a/mert/HypPackEnumerator.h b/mert/HypPackEnumerator.h
index d878c2625..420cd740a 100644
--- a/mert/HypPackEnumerator.h
+++ b/mert/HypPackEnumerator.h
@@ -27,6 +27,7 @@ public:
   virtual bool finished() = 0;
   virtual void next() = 0;
 
+  virtual std::size_t cur_id() = 0;
   virtual std::size_t cur_size() = 0;
   virtual std::size_t num_dense() const = 0;
   virtual const FeatureDataItem& featuresAt(std::size_t i) = 0;
@@ -46,6 +47,7 @@ public:
   virtual bool finished();
   virtual void next();
 
+  virtual std::size_t cur_id();
   virtual std::size_t cur_size();
   virtual const FeatureDataItem& featuresAt(std::size_t i);
   virtual const ScoreDataItem& scoresAt(std::size_t i);
@@ -79,6 +81,7 @@ public:
   virtual bool finished();
   virtual void next();
 
+  virtual std::size_t cur_id();
   virtual std::size_t cur_size();
   virtual const FeatureDataItem& featuresAt(std::size_t i);
   virtual const ScoreDataItem& scoresAt(std::size_t i);
diff --git a/mert/MiraFeatureVector.cpp b/mert/MiraFeatureVector.cpp
index a1a5a5bdb..6233d0ba9 100644
--- a/mert/MiraFeatureVector.cpp
+++ b/mert/MiraFeatureVector.cpp
@@ -1,4 +1,5 @@
 #include <cmath>
+#include <iomanip>
 
 #include "MiraFeatureVector.h"
 
@@ -139,6 +140,15 @@ MiraFeatureVector operator-(const MiraFeatureVector& a, const MiraFeatureVector&
   return MiraFeatureVector(dense,sparseFeats,sparseVals);
 }
 
+ostream& operator<<(ostream& o, const MiraFeatureVector& e)
+{
+  for(size_t i=0;i<e.size();i++) {
+    if(i>0) o << " ";
+    o << e.feat(i) << ":" << e.val(i);
+  }
+  return o;
+}
+
 // --Emacs trickery--
 // Local Variables:
 // mode:c++
diff --git a/mert/MiraFeatureVector.h b/mert/MiraFeatureVector.h
index 31dd025c3..89c7fc126 100644
--- a/mert/MiraFeatureVector.h
+++ b/mert/MiraFeatureVector.h
@@ -13,6 +13,7 @@
 #define MERT_MIRA_FEATURE_VECTOR_H
 
 #include <vector>
+#include <iostream>
 
 #include "FeatureDataIterator.h"
 
@@ -34,6 +35,8 @@ public:
   friend MiraFeatureVector operator-(const MiraFeatureVector& a,
                                      const MiraFeatureVector& b);
 
+  friend std::ostream& operator<<(std::ostream& o, const MiraFeatureVector& e);
+
 private:
   std::vector<ValType> m_dense;
   std::vector<std::size_t> m_sparseFeats;
diff --git a/mert/MiraWeightVector.cpp b/mert/MiraWeightVector.cpp
index 7e17a2714..092c5a44b 100644
--- a/mert/MiraWeightVector.cpp
+++ b/mert/MiraWeightVector.cpp
@@ -1,5 +1,7 @@
 #include "MiraWeightVector.h"
 
+#include <cmath>
+
 using namespace std;
 
 /**
@@ -113,6 +115,17 @@ AvgWeightVector::AvgWeightVector(const MiraWeightVector& wv)
   :m_wv(wv)
 {}
 
+ostream& operator<<(ostream& o, const MiraWeightVector& e) 
+{
+  for(size_t i=0;i<e.m_weights.size();i++) {
+    if(abs(e.m_weights[i])>1e-8) {
+      if(i>0) o << " ";
+      cerr << i << ":" << e.m_weights[i];
+    }
+  }
+  return o;
+}
+
 ValType AvgWeightVector::weight(size_t index) const
 {
   if(m_wv.m_numUpdates==0) return m_wv.weight(index);
diff --git a/mert/MiraWeightVector.h b/mert/MiraWeightVector.h
index 65b374625..20d7e0314 100644
--- a/mert/MiraWeightVector.h
+++ b/mert/MiraWeightVector.h
@@ -11,6 +11,7 @@
 #define MERT_MIRA_WEIGHT_VECTOR_H
 
 #include <vector>
+#include <iostream>
 
 #include "MiraFeatureVector.h"
 
@@ -59,6 +60,8 @@ public:
 
   friend class AvgWeightVector;
 
+  friend std::ostream& operator<<(std::ostream& o, const MiraWeightVector& e);
+
 private:
   /**
    * Updates a weight and lazily updates its total
diff --git a/mert/kbmira.cpp b/mert/kbmira.cpp
index 0723e9975..80a797860 100644
--- a/mert/kbmira.cpp
+++ b/mert/kbmira.cpp
@@ -84,6 +84,7 @@ int main(int argc, char** argv)
   bool streaming = false; // Stream all k-best lists?
   bool no_shuffle = false; // Don't shuffle, even for in memory version
   bool model_bg = false; // Use model for background corpus
+  bool verbose = false; // Verbose updates
   
   // Command-line processing follows pro.cpp
   po::options_description desc("Allowed options");
@@ -100,7 +101,8 @@ int main(int argc, char** argv)
       ("sparse-init,s", po::value<string>(&sparseInitFile), "Weight file for sparse features")
       ("streaming", po::value(&streaming)->zero_tokens()->default_value(false), "Stream n-best lists to save memory, implies --no-shuffle")
       ("no-shuffle", po::value(&no_shuffle)->zero_tokens()->default_value(false), "Don't shuffle hypotheses before each epoch")
-      ("model-bg", po::value(&model_bg)->zero_tokens()->default_value(false), "Use model instead of hope for BLEU background");
+      ("model-bg", po::value(&model_bg)->zero_tokens()->default_value(false), "Use model instead of hope for BLEU background")
+      ("verbose", po::value(&verbose)->zero_tokens()->default_value(false), "Verbose updates")
       ;
 
   po::options_description cmdline_options;
@@ -115,6 +117,8 @@ int main(int argc, char** argv)
       exit(0);
   }
 
+  cerr << "kbmira with c=" << c << " decay=" << decay << " no_shuffle=" << no_shuffle << endl;
+
   if (vm.count("random-seed")) {
     cerr << "Initialising random seed to " << seed << endl;
     srand(seed);
@@ -233,6 +237,14 @@ int main(int argc, char** argv)
         // Loss and update
         ValType diff_score = wv.score(diff);
         ValType loss = delta - diff_score;
+	if(verbose) {
+	  cerr << "Updating sent " << train->cur_id() << endl;
+	  cerr << "Wght: " << wv << endl;
+	  cerr << "Hope: " << hope << " => " << hopeBleu << " <> " << wv.score(hope) << endl;
+	  cerr << "Fear: " << fear << " => " << fearBleu << " <> " << wv.score(fear) << endl;
+	  cerr << "Diff: " << diff << " => " << delta << " <> " << diff_score << endl;
+	  cerr << endl;
+	}
         if(loss > 0) {
           ValType eta = min(c, loss / diff.sqrNorm());
           wv.update(diff,eta);