Changeset 514

Timestamp:

Feb 20, 2006, 10:45:34 AM (17 years ago)

Author:

Peter

Message:

generalised binary functionality in Target

Location:

trunk

Files:

• lib/classifier/CrossSplitter.cc (modified) (1 diff)
• lib/classifier/CrossSplitter.h (modified) (3 diffs)
• lib/classifier/SVM.cc (modified) (4 diffs)
• lib/classifier/SVM.h (modified) (3 diffs)
• lib/classifier/Target.cc (modified) (6 diffs)
• lib/classifier/Target.h (modified) (3 diffs)
• lib/statistics/Fisher.cc (modified) (2 diffs)
• lib/statistics/FoldChange.cc (modified) (2 diffs)
• lib/statistics/Pearson.cc (modified) (2 diffs)
• lib/statistics/ROC.cc (modified) (3 diffs)
• lib/statistics/tScore.cc (modified) (2 diffs)
• test/crossvalidation_test.cc (modified) (1 diff)
• test/score_test.cc (modified) (1 diff)
• test/svm_test.cc (modified) (2 diffs)
• test/target_test.cc (modified) (1 diff)

trunk/lib/classifier/CrossSplitter.cc

@@ -44 +44 @@
       
       for (size_t part=0; part<k && i<N; i++, part++) {
-        
         std::vector<size_t> training_index;
         std::vector<size_t> validation_index;
-
         for (size_t j=0; j<v.size(); j++) {
           if (j%k==part)

trunk/lib/classifier/CrossSplitter.h

@@ -21 +21 @@
   /// to the proportions in the whole dataset. In the first \a k
   /// rounds each sample is returned k-1 times, for next round the
-  /// samples are shuffled and...
+  /// samples are shuffled and... In total there are N partitions, in
+  /// other words, each sample is in validation roughly N/k
   ///   
 
@@ -29 +30 @@
   public:
     ///
-    /// Constructor taking \a target and \a k for k-fold cross
-    /// validation 
+    /// @brief Constructor 
+    ///  
+    /// @parameter Target targets
+    /// @parameter data data to split up in validation and training.
+    /// @parameter N total number of partitions
+    /// @parameter k for k-fold crossvalidation
     ///
-
-    CrossSplitter(const Target& target, const DataLookup2D&, 
-                   const size_t N, const size_t k);
+    CrossSplitter(const Target& target, const DataLookup2D& data, 
+                  const size_t N, const size_t k);
 
     ///
@@ -122 +126 @@
     
     std::vector<const DataLookup2D*> training_data_;
+    std::vector<std::vector<size_t> > training_index_;
     std::vector<Target> training_target_;
-    std::vector<std::vector<size_t> > training_index_;
 
     std::vector<const DataLookup2D*> validation_data_;
+    std::vector<std::vector<size_t> > validation_index_;
     std::vector<Target> validation_target_;
-    std::vector<std::vector<size_t> > validation_index_;     
   };
 

trunk/lib/classifier/SVM.cc

@@ -250 +250 @@
   bool SVM::choose(const theplu::gslapi::vector& E)
   {
-    //std::cout << "e choose\n"  ;
     // First check for violation among SVs
     // E should be the same for all SVs
@@ -257 +256 @@
     sample_.update_first(0);
     if (sample_.nof_sv()>1){
-      //std::cout << "there is SVs\n";
+
       double max = E(sample_(0));
       double min = max;
@@ -280 +279 @@
       
       // If no violation check among non-support vectors
-      //std::cout << "no violation SVs\n";
 
       sample_.shuffle();
-      //std::cout << "randomized\n";
       
       for (size_t i=sample_.nof_sv(); i<sample_.n();i++){
         //std::cout << "nr: " << i << std::endl;
-        if (target_.one(sample_(i))){
+        if (target_.binary(sample_(i))){
           if(E(sample_(i)) < E(sample_.value_first()) - 2*tolerance_){
             sample_.update_second(i);
@@ -305 +302 @@
     else{
       for (size_t i=0; i<sample_.n(); i++) {
-        if (target_.one(sample_(i))){
+        if (target_.binary(sample_(i))){
           for (size_t j=0; j<sample_.n(); j++) {
-            if ( !target_.one(sample_(j)) && 
+            if ( !target_.binary(sample_(j)) && 
                  E(sample_(i)) < E(sample_(j))+2*tolerance_ ){
               sample_.update_first(i);

trunk/lib/classifier/SVM.h

@@ -189 +189 @@
 
     ///
-    /// @todo Function calculating bias
+    /// @brief calculates the bias term
     ///
     /// @return true if successful
@@ -214 +214 @@
     
     /// @return 1 if i belong to class one and -1 if i belong to rest
-    inline int target(size_t i) const { return target_.one(i) ? 1 : -1; }
+    inline int target(size_t i) const { return target_.binary(i) ? 1 : -1; }
 
     gslapi::vector alpha_;
@@ -233 +233 @@
   /// @todo The output operator for the SVM class.
   ///
-  std::ostream& operator<< (std::ostream& s, const SVM&);
+  //std::ostream& operator<< (std::ostream& s, const SVM&);
   
   

trunk/lib/classifier/Target.cc

@@ -17 +17 @@
 
   Target::Target(const std::vector<std::string>& label)
-    : one_(0)
   {
     init(label);
@@ -24 +23 @@
   Target::Target(const Target& t, 
                  const std::vector<size_t>& index)
-    : class_map_(t.class_map_), one_(t.one_)
+    : class_map_(t.class_map_)
   {
+    binary_.resize(index.size());
     classes_.resize(index.size());
     for (size_t i=0; i<index.size(); i++) {
       assert(index[i]<t.size());
       classes_[i]=t.classes_[index[i]];
+      binary_[i]=t.binary_[index[i]];
     }
     labels_ = t.labels_;
@@ -35 +36 @@
 
 
-  // Peter to Markus, align with gslapi
   Target::Target(std::istream& is, char sep) 
     throw (utility::IO_error,std::exception)
-    : one_(0)
   {
     std::vector<std::string> vec;
@@ -72 +71 @@
     class_map_.clear();
     labels_.clear();
-
     
     for (size_t i=0; i<label.size(); i++) {
-      std::map<std::string,size_t>::const_iterator iter = 
+      std::map<std::string,size_t>::iterator iter = 
         class_map_.lower_bound(label[i]);
 
@@ -87 +85 @@
       else{
         classes_.push_back(class_map_.size());
-        // Peter, should use iter to hint on position
-        class_map_.insert(std::make_pair(label[i],classes_[i])); 
+        class_map_.insert(iter, std::make_pair(label[i],classes_[i])); 
         labels_.push_back(label[i]);
       }
     }
+    // setting binary to false for every class
+    binary_=std::vector<short>(label.size(),0);
+    
+    set_binary(0,true);
 
   }
@@ -102 +103 @@
   //  return *this;
   //}
+
+  void Target::set_binary(const size_t target, const bool b)
+  {
+    for (size_t i=0; i<classes_.size(); i++)
+      if (classes_[i]==target)
+        binary_[i]=b;
+  }
+
 
   std::ostream& operator<<(std::ostream& s, const Target& a)

trunk/lib/classifier/Target.h

@@ -69 +69 @@
 
     ///
-    /// @return true if class of sample @a i is equal to variable one
+    /// Default binary is set to false for all classes except class 0
     ///
-    /// @see set_one(const size_t one)
+    /// @return true if class of sample @a i is set to true
     ///
-    inline bool one(const size_t i) const 
-    { assert(i<size()); return classes_[i]==one_; }
+    /// @see set_binary
+    ///
+    inline bool binary(const size_t i) const 
+    { assert(i<size()); return binary_[i]; }
     
     ///
-    /// Function to set variable one. This variable is used in 2-class
-    /// algorithm in rder to use the Targer object in a one-versus-all
-    /// manner.
+    /// Class @a target is set to @a b. Default is binary set to false
+    /// for each class.
     ///
-    /// @see one(const size_t i)
-    ///
-    inline void set_one(const size_t one) { assert(one<size()); one_=one; }
+    void set_binary(const size_t target, const bool b); 
 
     ///
@@ -95 +94 @@
     const size_t size(const std::string& label) const;
 
-    //
-    //@brief sort
-    //
-    //inline void sort(void) { std::sort(classes_.begin(), classes_.end()); }
-
     ///
     /// @return the class of @a sample
@@ -106 +100 @@
     { assert(sample<size()); return classes_[sample]; }
 
-    ///
-    /// @brief assignment operator
-    ///
-    //const Target& operator=(const Target& other);
 
   private:
+    std::vector<short> binary_; // avoid using vector<bool>
     std::vector<size_t> classes_; // class of sample i
     std::map<std::string,size_t> class_map_;  
     std::vector<std::string> labels_; // label of class i
-    size_t one_;
     
     void init(const std::vector<std::string>&);

trunk/lib/statistics/Fisher.cc

@@ -116 +116 @@
     a_=b_=c_=d_=0;
     for (size_t i=0; i<target.size(); i++)
-      if (target.one(i))
+      if (target.binary(i))
         if (value(i)>value_cutoff_)
           a_++;
@@ -143 +143 @@
     a_=b_=c_=d_=0;
     for (size_t i=0; i<target.size(); i++)
-      if (target.one(i))
+      if (target.binary(i))
         if (value(i)>value_cutoff_)
           a_+=weight(i);

trunk/lib/statistics/FoldChange.cc

@@ -39 +39 @@
 
     for (size_t i=0; i<value.size(); i++) 
-      if (target.one(i))
+      if (target.binary(i))
         pos.add(value(i));
       else
@@ -58 +58 @@
 
     for (size_t i=0; i<value.size(); i++) 
-      if (target.one(i))
+      if (target.binary(i))
         pos.add(value(i),weight(i));
       else

trunk/lib/statistics/Pearson.cc

@@ -46 +46 @@
     AveragerPair ap;
     for (size_t i=0; i<target.size(); i++){
-      if (target.one(i))
+      if (target.binary(i))
         ap.add(1, value(i));
       else
@@ -66 +66 @@
     AveragerPairWeighted ap;
     for (size_t i=0; i<target.size(); i++){
-      if (target.one(i))
+      if (target.binary(i))
         ap.add(1, value(i),1,weight(i));
       else

trunk/lib/statistics/ROC.cc

@@ -75 +75 @@
     vec_pair_.reserve(target.size());
     for (size_t i=0; i<target.size(); i++)
-      vec_pair_.push_back(std::make_pair(target.one(i),value(i)));
+      vec_pair_.push_back(std::make_pair(target.binary(i),value(i)));
 
     std::sort(vec_pair_.begin(),vec_pair_.end(),
@@ -110 +110 @@
     for (unsigned int i=0; i<target.size(); i++)
       if (weight(i))
-        vec_pair_.push_back(std::make_pair(target.one(i),value(i)));
+        vec_pair_.push_back(std::make_pair(target.binary(i),value(i)));
 
     std::sort(vec_pair_.begin(),vec_pair_.end(),
@@ -120 +120 @@
 
     for (size_t i=0; i<n(); i++)
-      if (target.one(i))
+      if (target.binary(i))
         for (size_t j=0; j<n(); j++)
-          if (!target.one(j)){
+          if (!target.binary(j)){
             if (value(i)>value(j))
               area_+=weight(i)*weight(j);

trunk/lib/statistics/tScore.cc

@@ -27 +27 @@
     dof_=target.size()-2;
     for(size_t i=0; i<target.size(); i++){
-      if (target.one(i))
+      if (target.binary(i))
         positive.add(value(i));
       else
@@ -52 +52 @@
     dof_=target.size()-2;
     for(size_t i=0; i<target.size(); i++){
-      if (target.one(i))
+      if (target.binary(i))
         positive.add(value(i),weight(i));
       else

trunk/test/crossvalidation_test.cc

@@ -27 +27 @@
   bool ok = true;
 
-  
   std::vector<std::string> label(10,"default");
   label[2]=label[7]="white";

trunk/test/score_test.cc

@@ -45 +45 @@
     ok = false;
   }
-  target.set_one(1);
+  target.set_binary(0,false);
+  target.set_binary(1,true);
   area = roc.score(target, value);
   if (area!=1.0){

trunk/test/svm_test.cc

@@ -59 +59 @@
   double tmp=0;
   for (size_t i=0; i<target2.size(); i++) 
-    if (target2.one(i))
+    if (target2.binary(i))
       tmp += classifier2.alpha()(i);
     else
@@ -121 +121 @@
   for (unsigned int i=0; i<target.size(); i++){
     if (output(i)*target(i) < 1){
-      if (target.one(i))
+      if (target.binary(i))
         slack += 1 - output(i);
       else

trunk/test/target_test.cc

@@ -60 +60 @@
     *error << "Error: target(20)!=2" << std::endl;
   }
-  if (!target.one(0)){
+  if (!target.binary(0)){
     ok=false;
-    *error << "Error: target.one(0) not true" << std::endl;
+    *error << "Error: target.binary(0) not true" << std::endl;
   }   
-  if (target.one(20)){
+  if (target.binary(20)){
     ok=false;
-    *error << "Error: target.one(20) not false" << std::endl;
+    *error << "Error: target.binary(20) not false" << std::endl;
   }   
-  target.set_one(2);
-  if (!target.one(20)){
+  target.set_binary(2,true);
+  if (!target.binary(20)){
     ok=false;
-    *error << "Error: target.one(20) not true" << std::endl;
+    *error << "Error: target.binary(20) not true" << std::endl;
   }   
   std::ifstream is("data/rank_target.txt");