Changeset 812 for trunk

Timestamp:

Mar 16, 2007, 2:02:07 AM (16 years ago)

Author:

Peter

Message:

implemented NBC predict. Have to check what happens in the weighted case though, Refs #57

Location:

trunk/yat/classifier

Files:

• NBC.cc (modified) (3 diffs)
• NBC.h (modified) (1 diff)

trunk/yat/classifier/NBC.cc

@@ -31 +31 @@
 
 #include <cassert>
+#include <cmath>
 #include <vector>
 
@@ -74 +75 @@
   bool NBC::train()
   {   
-    sigma2_=centroids_=utility::matrix(data_.rows(), target_.nof_classes());
+    sigma2_.resize(data_.rows(), target_.nof_classes());
+    centroids_.resize(data_.rows(), target_.nof_classes());
     utility::matrix nof_in_class(data_.rows(), target_.nof_classes());
-    
     
     for(size_t i=0; i<data_.rows(); ++i) {
@@ -99 +100 @@
 
 
-  void NBC::predict(const DataLookup2D& input,                    
+  void NBC::predict(const DataLookup2D& x,                    
                     utility::matrix& prediction) const
   {   
-    std::cerr << "NBC::predict not implemented\n";
-    exit(1);
-    assert(data_.rows()==input.rows());
+    assert(data_.rows()==x.rows());
 
-    //    utility
-    //for (size_t i=0; 
-
-
-    prediction = utility::matrix(centroids_.columns(),input.columns());
-    for (size_t c=0; c<centroids_.columns(); ++c) {
+    // each row in prediction corresponds to a sample label (class)
+    prediction.resize(centroids_.columns(), x.columns(), 0);
+    // first calculate -lnP = sum sigma_i + (x_i-m_i)^2/2sigma_i^2
+    for (size_t label=0; label<prediction.columns(); ++label) {
       double sum_ln_sigma=0;
-      for (size_t i=0; i<sigma2_.rows(); ++i) 
-        sum_ln_sigma += log(sigma2_(i,c));
-      sum_ln_sigma /= 2;
-
-      for (size_t s=0; s<input.columns(); ++s) {
-        // -lnp = sum{ln(sigma_i)} + sum{(x_i-m_i)^2/(2sigma_i)}
-        prediction(c,s) = sum_ln_sigma;
-        for (size_t i=0; i<input.columns(); ++i) {
-          prediction(c,s) += std::pow(input(i,s)-centroids_(i,c),2)/sigma2_(i,c);
+      for (size_t i=0; i<x.rows(); ++i) 
+        sum_ln_sigma += std::log(sigma2_(i, label));
+      sum_ln_sigma /= 2; // taking sum of log(sigma) not sigma2
+      for (size_t sample=0; sample<prediction.rows(); ++sample) {
+        for (size_t i=0; i<x.rows(); ++i) {
+          prediction(label, sample) += 
+            std::pow(x(i, label)-centroids_(i, label),2)/sigma2_(i, label);
         }
       }
     }
-    // exponentiate and normalize
+
+    // -lnP might be a large number, in order to avoid out of bound
+    // problems when calculating P = exp(- -lnP), we centralize matrix
+    // by adding a constant.
+    double m=0;
+    for (size_t i=0; i<prediction.rows(); ++i)
+      for (size_t j=0; j<prediction.columns(); ++j)
+        m+=prediction(i,j);
+    prediction -= m/prediction.rows()/prediction.columns();
+
+    // exponentiate
+    for (size_t i=0; i<prediction.rows(); ++i)
+      for (size_t j=0; j<prediction.columns(); ++j)
+        prediction(i,j) = std::exp(prediction(i,j));
+
+    // normalize each row (label) to sum up to unity (probability)
+    for (size_t i=0; i<prediction.rows(); ++i)
+      utility::vector(prediction,i) *= 
+        1.0/utility::sum(utility::vector(prediction,i));
+
   }
 

trunk/yat/classifier/NBC.h

@@ -90 +90 @@
 
   private:
-    double gaussian(double x, double m, double sigma) const;
-
     utility::matrix centroids_;
     utility::matrix sigma2_;