bug fixes

git-svn-id: https://mosesdecoder.svn.sourceforge.net/svnroot/mosesdecoder/trunk@1719 1f5c12ca-751b-0410-a591-d2e778427230

3 files changed, 48 insertions, 13 deletions

diff --git a/mert/Optimizer.cpp b/mert/Optimizer.cpp
index 118d285fe..655c46322 100644
--- a/mert/Optimizer.cpp
+++ b/mert/Optimizer.cpp
@@ -90,9 +90,11 @@ statscore_t Optimizer::LineOptimize(const Point& origin,const Point& direction,P
   thresholdlist.push_back(threshold(MIN_FLOAT,vector<diff>()));
   vector<unsigned> first1best;//the vector of nbrests for x=-inf
   for(int S=0;S<size();S++){
+    cerr<<"SENTENCE"<<S;
     //first we determine the translation with the best feature score for each sentence and each value of x
     multimap<float,unsigned> gradient;
     vector<float> f0;
+    f0.resize(size());
     for(unsigned j=0;j<FData->get(S).size();j++){
       gradient.insert(pair<float,unsigned>(direction*(FData->get(S,j)),j));//gradient of the feature function for this particular target sentence
       f0[j]=origin*FData->get(S,j);//compute the feature function at the origin point
@@ -111,19 +113,24 @@ statscore_t Optimizer::LineOptimize(const Point& origin,const Point& direction,P
     while(it!=gradient.end()&&it->first==smallest){
       if(f0[it->second]>f0[index])
 	index=it->second;//the highest line is the one with he highest f0
-      }
+      it++;
+    }
     --it;//we went one step too far in the while loop
     onebest.push_back(pair<float,unsigned>(MIN_FLOAT,index));//first 1best is the lowest gradient. 
     //now we look for the intersections points indicating a change of 1 best
     //we use the fact that the function is convex, which means that the gradient can only go up   
+    int c=0;
     while(it!=gradient.end()){
+      cerr<<c++<<endl;
       map<float,unsigned>::iterator leftmost=it;
       float leftmostx=onebest.back().first;
       float m=it->first;
       float b=f0[it->second];
       multimap<float,unsigned>::iterator it2=it;
       it2++;
+      int d=0;
       for(;it2!=gradient.end();it2++){
+	//cerr<<"--"<<d++<<' '<<it2->first<<' '<<it2->second<<endl;
 	//look for all candidate with a gradient bigger than the current one and fond the one with the leftmost intersection
 	float curintersect=intersect(m,b,it2->first,f0[it2->second]);
 	if(curintersect<leftmostx){
@@ -147,7 +154,11 @@ statscore_t Optimizer::LineOptimize(const Point& origin,const Point& direction,P
 	onebest.back()=pair<float,unsigned>(leftmostx,leftmost->second);//leftmost->first is the gradient, we are interested in the value of the intersection
       else //normal case: we add a new threshold
 	onebest.push_back(pair<float,unsigned>(leftmostx,leftmost->second));
-      it=leftmost;
+      if(it==leftmost)
+	it=gradient.end();
+      else
+	cerr<<"ok"<<endl;
+	it=leftmost;
     }
     //we have the onebest list and the threshold for the current sentence.
     //now we update the thresholdlist: we add the new threshold and the  value of the onebest.
@@ -253,6 +264,10 @@ statscore_t Optimizer::Run(Point& P)const{
     cerr<<"error trying to optimize without a Scorer loaded"<<endl;
     exit(2);
   }
+  if (scorer->getReferenceSize()!=FData->size()){
+    cerr<<"errror size mismatch between FeatureData and Scorer"<<endl;
+    exit(2);
+  }
   statscore_t s=TrueRun(P);
   P.score=s;//just in case its not done in TrueRun
   return s;
diff --git a/mert/ScoreData.cpp b/mert/ScoreData.cpp
index 6b5e7975f..fb1d4b9fb 100644
--- a/mert/ScoreData.cpp
+++ b/mert/ScoreData.cpp
@@ -17,7 +17,7 @@ theScorer(&ptr)
 {
 	score_type = theScorer->getName();
 	TRACE_ERR("score_type:" << score_type << std::endl);
-	
+	theScorer->setScoreData(this);//this is not dangerous: we dont use the this pointer in SetScoreData 	
 };
 
 void ScoreData::save(std::ofstream& outFile, bool bin)
diff --git a/mert/mert.cpp b/mert/mert.cpp
index 02834888f..590c74f91 100755
--- a/mert/mert.cpp
+++ b/mert/mert.cpp
@@ -24,20 +24,25 @@ float min_interval = 1e-3;
 using namespace std;
 
 void usage(void) {
-  cerr<<"usage: mert -d <dimensions>"<<endl;
-  cerr<<"[-n retry ntimes]"<<endl;
-  cerr<<"[-o indexes to optimize]"<<endl;
-  cerr<<"[-t optimizer]"<<endl;
-  cerr<<"[-s scorer]"<<endl;
+  cerr<<"usage: mert -d <dimensions> (mandatory )"<<endl;
+  cerr<<"[-n retry ntimes (default 1)]"<<endl;
+  cerr<<"[-o\tthe indexes to optimize(default all)]"<<endl;
+  cerr<<"[-t\tthe optimizer(default Powell)]"<<endl;
+  cerr<<"[-sctype] the scorer type (default BLEU)"<<endl;
+  cerr<<"[-scfile] the scorer data file (default score.data)"<<endl;
+  cerr<<"[-ffile] the feature data file data file (default feature.data)"<<endl;
   exit(1);
 }
 
 static struct option long_options[] =
   {
     {"dim", 1, 0, 'd'},
-    {"only",1,0,'o'},
+    {"ntry",1,0,'n'},
+    {"optimize",1,0,'o'},
     {"type",1,0,'t'},
-    {"scorer",1,0,'s'},
+    {"sctype",1,0,'s'},
+    {"scfile",1,0,'S'},
+    {"ffile",1,0,'F'},
     {0, 0, 0, 0}
   };
 int option_index;
@@ -48,22 +53,30 @@ int main (int argc, char **argv) {
   int ntry=1;
   string type("powell");
   string scorertype("BLEU");
+  string scorerfile("statscore.data");
+  string featurefile("features.data");
   vector<unsigned> tooptimize;
   vector<parameter_t> start;
-  while (getopt_long (argc, argv, "d:n:o:t:s:", long_options, &option_index) != -1) {
+  while ((c=getopt_long (argc, argv, "d:n:t:s:S:F:", long_options, &option_index)) != -1) {
     switch (c) {
     case 'd':
       dim = strtol(optarg, NULL, 10);
+      cerr<<dim;
       break;
     case 'n':
       ntry=strtol(optarg, NULL, 10);
       break;
-    case 'o':
-      //TODO
     case 't':
       type=string(optarg);
+      break;
       case's':
 	scorertype=string(optarg);
+      break;
+    case 'S':
+      scorerfile=string(optarg);
+    case 'F':
+      featurefile=string(optarg);
+      break;
     default:
       usage();
     }
@@ -73,6 +86,7 @@ int main (int argc, char **argv) {
     for(i=0;i<dim;i++)
       tooptimize[i]=i;
   }
+  ScorerFactory SF;
   Optimizer *O;
   Scorer *TheScorer=NULL;;
   FeatureData *FD=NULL;
@@ -86,6 +100,12 @@ int main (int argc, char **argv) {
   Point bestP;
   //it make sense to know what parameter set where used to generate the nbest
   O=BuildOptimizer(dim,tooptimize,start,"powell");
+  
+  TheScorer=SF.getScorer(scorertype);
+  ScoreData *SD=new ScoreData(*TheScorer);
+  FD=new FeatureData();
+  FD->load(featurefile);
+  SD->load(scorerfile);
   O->SetScorer(TheScorer);
   O->SetFData(FD);
   Point min;//to: initialize