[mlpack] 200/324: Changes to implementation of adaboost. Implemented adaboost.m1

[Barak A. Pearlmutter]

This is an automated email from the git hooks/post-receive script.

bap pushed a commit to branch svn-trunk
in repository mlpack.

commit 03c58b15b568eb49b1b16217a206a996429dacb1
Author: saxena.udit <saxena.udit at 9d5b8971-822b-0410-80eb-d18c1038ef23>
Date:   Sun Jul 20 18:02:20 2014 +0000

    Changes to implementation of adaboost. Implemented adaboost.m1
    
    git-svn-id: http://svn.cc.gatech.edu/fastlab/mlpack/trunk@16842 9d5b8971-822b-0410-80eb-d18c1038ef23
---
 src/mlpack/methods/CMakeLists.txt             |  2 +-
 src/mlpack/methods/adaboost/adaboost_impl.hpp | 76 +++++++++++++++++++++------
 2 files changed, 60 insertions(+), 18 deletions(-)

diff --git a/src/mlpack/methods/CMakeLists.txt b/src/mlpack/methods/CMakeLists.txt
index 041e475..8ea92d2 100644
--- a/src/mlpack/methods/CMakeLists.txt
+++ b/src/mlpack/methods/CMakeLists.txt
@@ -1,6 +1,6 @@
 # Recurse into each method mlpack provides.
 set(DIRS
-#  adaboost 
+  adaboost 
   amf
   cf
   decision_stump
diff --git a/src/mlpack/methods/adaboost/adaboost_impl.hpp b/src/mlpack/methods/adaboost/adaboost_impl.hpp
index 8b335be..c49fc8a 100644
--- a/src/mlpack/methods/adaboost/adaboost_impl.hpp
+++ b/src/mlpack/methods/adaboost/adaboost_impl.hpp
@@ -14,6 +14,12 @@
 namespace mlpack {
 namespace adaboost {
 
+//
+// Currently this is an implementation of adaboost.m1
+// which will be templatized later and adaboost.mh and
+// adaboost.samme will be added.
+// 
+
 template<typename MatType, typename WeakLearner>
 Adaboost<MatType, WeakLearner>::Adaboost(const MatType& data, const arma::Row<size_t>& labels,
          int iterations, size_t classes, const WeakLearner& other)
@@ -26,15 +32,25 @@ Adaboost<MatType, WeakLearner>::Adaboost(const MatType& data, const arma::Row<si
   // load the initial weights
   
   const double initWeight = 1 / (data.n_cols * classes);
-  arma::Row<double> D(data.n_cols);
+  arma::rowvec D(data.n_cols);
   D.fill(initWeight);
 
-  double rt, alphat = 0.0, zt;
+  size_t countMP; // for counting mispredictions.
+  double rt, alphat = 0.0, zt, et;
   arma::Row<size_t> predictedLabels(labels.n_cols);
   MatType tempData(data);
+  
+  // This behaves as ht(x)
+  arma::rowvec mispredict(predictedLabels.n_cols);
+
+  arma::mat sumFinalH(data.n_cols, classes);
+  sumFinalH.fill(0.0);
+
+  arma::rowvec finalH(labels.n_cols);
   // now start the boosting rounds
   for (i = 0; i < iterations; i++)
   {
+    countMP = 0;
     rt = 0.0;
     zt = 0.0;
     
@@ -47,37 +63,63 @@ Adaboost<MatType, WeakLearner>::Adaboost(const MatType& data, const arma::Row<si
     // building a helper rowvector, mispredict to help in calculations.
     // this stores the value of Yi(l)*ht(xi,l)
     
-    arma::Row<double> mispredict(predictedLabels.n_cols);
-    
+    // first calculate error: 
     for(j = 0;j < predictedLabels.n_cols; j++)
     {
       if (predictedLabels(j) != labels(j))
+      {  
         mispredict(j) = -predictedLabels(j);
+        countMP++;
+      }
       else
         mispredict(j) = predictedLabels(j);
     }
+    et = ((double) countMP / predictedLabels.n_cols);
 
-    // begin calculation of rt
+    if (et < 0.5)
+    {
+      // begin calculation of rt
 
-    for (j = 0;j < predictedLabels.n_cols; j++)
-      rt +=(D(j) * mispredict(j));
+      // for (j = 0;j < predictedLabels.n_cols; j++)
+      //   rt +=(D(j) * mispredict(j));
 
-    // end calculation of rt
+      // end calculation of rt
 
-    alphat = 0.5 * log((1 + rt) / (1 - rt));
+      // alphat = 0.5 * log((1 + rt) / (1 - rt));
 
-    // end calculation of alphat
-    
-    for (j = 0;j < mispredict.n_cols; j++)
-    {
-      zt += D(i) * exp(-1 * alphat * mispredict(i));
-      D(i) = D(i) * exp(-1 * alphat * mispredict(i));
-    }
+      alphat = 0.5 * log((1 - et) / et);  
+
+      // end calculation of alphat
+      
+      // now start modifying weights
 
-    D = D / zt;
+      for (j = 0;j < mispredict.n_cols; j++)
+      {
+        // we calculate zt, the normalization constant
+        zt += D(j) * exp(-1 * alphat * (mispredict(j) / predictedLabels(j)));
+        D(j) = D(j) * exp(-1 * alphat * (mispredict(j) / predictedLabels(j)));
 
+        // adding to the matrix of FinalHypothesis 
+        if (mispredict(j) == predictedLabels(j)) // if correct prediction
+          sumFinalH(j, mispredict(j)) += alphat;
+      }
+      // normalization of D
+
+      D = D / zt;
+    }
   }
 
+  // build a strong hypothesis from a weighted combination of these weak hypotheses.
+  
+  // This step of storing it in a temporary row vector can be improved upon.
+  arma::rowvec tempSumFinalH;
+
+  for (i = 0;i < sumFinalH.n_rows; i++)
+  {
+    tempSumFinalH = sumFinalH.row(i);
+    tempSumFinalH.max(max_index);
+    finalH(i) = max_index;
+  }
 }
 
 } // namespace adaboost

-- 
Alioth's /usr/local/bin/git-commit-notice on /srv/git.debian.org/git/debian-science/packages/mlpack.git