Attempt at fixing sparse features for hgmira

mert/HopeFearDecoder.cpp

@@ -180,7 +180,7 @@ HypergraphHopeFearDecoder::HypergraphHopeFearDecoder
   references_.Load(referenceFiles, vocab_);
 
   SparseVector weights;
-  wv.ToSparse(&weights);
+  wv.ToSparse(&weights,num_dense_);
   scorer_ = scorer;
 
   static const string kWeights = "weights";
@@ -243,7 +243,7 @@ void HypergraphHopeFearDecoder::HopeFear(
 {
   size_t sentenceId = *sentenceIdIter_;
   SparseVector weights;
-  wv.ToSparse(&weights);
+  wv.ToSparse(&weights, num_dense_);
   const Graph& graph = *(graphs_[sentenceId]);
 
   // ValType hope_scale = 1.0;
@@ -338,7 +338,7 @@ void HypergraphHopeFearDecoder::MaxModel(const AvgWeightVector& wv, vector<ValTy
   HgHypothesis bestHypo;
   size_t sentenceId = *sentenceIdIter_;
   SparseVector weights;
-  wv.ToSparse(&weights);
+  wv.ToSparse(&weights, num_dense_);
   vector<ValType> bg(scorer_->NumberOfScores());
   //cerr << "Calculating bleu on " << sentenceId << endl;
   Viterbi(*(graphs_[sentenceId]), weights, 0, references_, sentenceId, bg, &bestHypo);

mert/Jamfile

@@ -77,6 +77,7 @@ unit-test feature_data_test : FeatureDataTest.cpp mert_lib ..//boost_unit_test_f
 unit-test data_test : DataTest.cpp mert_lib ..//boost_unit_test_framework ;
 unit-test forest_rescore_test : ForestRescoreTest.cpp mert_lib ..//boost_unit_test_framework ;
 unit-test hypergraph_test : HypergraphTest.cpp mert_lib ..//boost_unit_test_framework ;
+unit-test mira_feature_vector_test : MiraFeatureVectorTest.cpp mert_lib ..//boost_unit_test_framework ;
 unit-test ngram_test : NgramTest.cpp mert_lib ..//boost_unit_test_framework ;
 unit-test optimizer_factory_test : OptimizerFactoryTest.cpp mert_lib ..//boost_unit_test_framework ;
 unit-test point_test : PointTest.cpp mert_lib ..//boost_unit_test_framework ;

mert/MiraFeatureVectorTest.cpp

@@ -0,0 +1,49 @@
+#include "MiraFeatureVector.h"
+#include "MiraWeightVector.h"
+
+#define BOOST_TEST_MODULE MiraFeatureVector
+#include <boost/test/unit_test.hpp>
+
+using namespace MosesTuning;
+
+/* Note that the conversion to and from SparseVector needs to know
+how many of the features are really "dense". This is because in hg mira
+all features (sparse and dense) are to get rolled in to SparseVector
+*/
+
+BOOST_AUTO_TEST_CASE(from_sparse) {
+  SparseVector sp;
+  sp.set("dense0", 0.2);
+  sp.set("dense1", 0.3);
+  sp.set("sparse0", 0.7);
+  sp.set("sparse1", 0.9);
+  sp.set("sparse2", 0.1);
+
+  MiraFeatureVector mfv(sp,2);
+  BOOST_CHECK_EQUAL(mfv.size(),5);
+
+  BOOST_CHECK_EQUAL(mfv.feat(0),0);
+  BOOST_CHECK_EQUAL(mfv.feat(1),1);
+  BOOST_CHECK_EQUAL(mfv.feat(2),4);
+  BOOST_CHECK_EQUAL(mfv.feat(3),5);
+  BOOST_CHECK_EQUAL(mfv.feat(4),6);
+
+  BOOST_CHECK_CLOSE(mfv.val(0), 0.2,1e-5);
+  BOOST_CHECK_CLOSE(mfv.val(1), 0.3,1e-5);
+  BOOST_CHECK_CLOSE(mfv.val(2), 0.7,1e-5);
+  BOOST_CHECK_CLOSE(mfv.val(3), 0.9,1e-5);
+  BOOST_CHECK_CLOSE(mfv.val(4), 0.1,1e-5);
+
+  MiraWeightVector mwv;
+  mwv.update(mfv,1.0);
+  SparseVector sp2;
+  mwv.ToSparse(&sp2,2);
+
+  //check we get back what we started with
+  BOOST_CHECK_CLOSE(sp2.get("dense0"), 0.2,1e-5);
+  BOOST_CHECK_CLOSE(sp2.get("dense1"), 0.3,1e-5);
+  BOOST_CHECK_CLOSE(sp2.get("sparse0"), 0.7,1e-5);
+  BOOST_CHECK_CLOSE(sp2.get("sparse1"), 0.9,1e-5);
+  BOOST_CHECK_CLOSE(sp2.get("sparse2"), 0.1,1e-5);
+
+}

mert/MiraWeightVector.cpp

@@ -93,11 +93,17 @@ void MiraWeightVector::update(size_t index, ValType delta)
   m_lastUpdated[index] = m_numUpdates;
 }
 
-void MiraWeightVector::ToSparse(SparseVector* sparse) const
+void MiraWeightVector::ToSparse(SparseVector* sparse, size_t denseSize) const
 {
   for (size_t i = 0; i < m_weights.size(); ++i) {
     if(abs(m_weights[i])>1e-8) {
-      sparse->set(i,m_weights[i]);
+      if (i < denseSize) {
+        sparse->set(i,m_weights[i]);
+      } else {
+        //The ids in MiraFeatureVector/MiraWeightVector for sparse features
+        //need to be translated when converting back to SparseVector.
+        sparse->set(i-denseSize, m_weights[i]);
+      }
     }
   }
 }
@@ -172,12 +178,18 @@ size_t AvgWeightVector::size() const
   return m_wv.m_weights.size();
 }
 
-void AvgWeightVector::ToSparse(SparseVector* sparse) const
+void AvgWeightVector::ToSparse(SparseVector* sparse, size_t denseSize) const
 {
   for (size_t i = 0; i < size(); ++i) {
     ValType w = weight(i);
     if(abs(w)>1e-8) {
-      sparse->set(i,w);
+      if (i < denseSize) {
+        sparse->set(i,w);
+      } else {
+        //The ids in MiraFeatureVector/MiraWeightVector for sparse features
+        //need to be translated when converting back to SparseVector.
+        sparse->set(i-denseSize, w);
+      }
     }
   }
 }

mert/MiraWeightVector.h

@@ -64,9 +64,9 @@ public:
   AvgWeightVector avg();
 
   /**
-    * Convert to sparse vector, interpreting all features as sparse.
+    * Convert to sparse vector, interpreting all features as sparse. Only used by hgmira.
    **/
-  void ToSparse(SparseVector* sparse) const;
+  void ToSparse(SparseVector* sparse, size_t denseSize) const;
 
   friend class AvgWeightVector;
 
@@ -104,7 +104,7 @@ public:
   ValType score(const MiraFeatureVector& fv) const;
   ValType weight(std::size_t index) const;
   std::size_t size() const;
-  void ToSparse(SparseVector* sparse) const;
+  void ToSparse(SparseVector* sparse, size_t num_dense) const;
 private:
   const MiraWeightVector& m_wv;
 };