Changeset 559 for trunk/lib/classifier/SVM.cc

Timestamp:

Mar 11, 2006, 11:21:27 PM (17 years ago)

Author:

Peter

Message:

some changes in EB

File:

• trunk/lib/classifier/SVM.cc (modified) (11 diffs)

trunk/lib/classifier/SVM.cc

@@ -35 +35 @@
       tolerance_(0.00000001)
   {
+#ifndef NDEBUG
+    for (size_t i=0; i<alpha_.size(); i++) 
+      for (size_t j=0; j<alpha_.size(); j++) 
+        assert(kernel(i,j)==kernel(j,i));
+    for (size_t i=0; i<alpha_.size(); i++) 
+      for (size_t j=0; j<alpha_.size(); j++) 
+        assert((*kernel_)(i,j)==(*kernel_)(j,i));
+    for (size_t i = 0; i<kernel_->rows(); i++)
+      for (size_t j = 0; j<kernel_->columns(); j++)
+        if (std::isnan((*kernel_)(i,j)))
+          std::cerr << "SVM: Found nan in kernel: " << i << " " 
+                    << j << std::endl;
+#endif        
+
   }
 
@@ -84 +98 @@
                                              const Target& target) const
   {
+    assert(data.rows()==data.columns());
+    assert(data.columns()==target.size());
     // Peter, should check success of dynamic_cast
     const KernelLookup& tmp = dynamic_cast<const KernelLookup&>(data);
@@ -92 +108 @@
       sc = new SVM(tmp,target);
 
+
     //Copy those variables possible to modify from outside
     return sc;
@@ -99 +116 @@
   {
     // Peter, should check success of dynamic_cast
-    const KernelLookup* input_kernel= &dynamic_cast<const KernelLookup&>(input);
-
+    const KernelLookup input_kernel = dynamic_cast<const KernelLookup&>(input);
+
+    const KernelLookup* kernel_pointer;
     if (ranker_) {// feature selection
       std::vector<size_t> index;
@@ -106 +124 @@
       for (size_t i=0; i<nof_inputs_; i++) 
         index.push_back(ranker_->id(i));
-      input_kernel = input_kernel->selected(index);
-    }
+      kernel_pointer = input_kernel.selected(index);
+    }
+    else 
+      kernel_pointer = &input_kernel;
 
     assert(input.rows()==alpha_.size());
     prediction = gslapi::matrix(2,input.columns(),0);
     for (size_t i = 0; i<input.columns(); i++){
-      for (size_t j = 0; j<input.rows(); j++)
-        prediction(0,i) += target(j)*alpha_(i)*(*input_kernel)(j,i);
+      for (size_t j = 0; j<input.rows(); j++){
+        prediction(0,i) += target(j)*alpha_(j)*(*kernel_pointer)(j,i);
+        assert(target(j));
+      }
       prediction(0,i) += bias_;
     }
@@ -119 +141 @@
     for (size_t i = 0; i<prediction.columns(); i++)
       prediction(1,i) = -prediction(0,i);
+    
+    if (ranker_)
+      delete kernel_pointer;
+    assert(prediction(0,0));
   }
 
@@ -150 +176 @@
       for (size_t j=0; j<E.size(); j++) 
         E(i) += kernel_mod(i,j)*target(j)*alpha_(j);
-      E(i)=E(i)-target(i);
+      E(i)-=target(i);
     }
     assert(target_.size()==E.size());
@@ -170 +196 @@
       double alpha_old1=alpha_(sample_.value_first());
       double alpha_old2=alpha_(sample_.value_second());
-
       alpha_new2 = ( alpha_(sample_.value_second()) + 
                      target(sample_.value_second())*
@@ -179 +204 @@
       else if (alpha_new2<u)
         alpha_new2 = u;
-      
       
       // Updating the alphas
@@ -254 +278 @@
       assert(alpha_(sample_.value_second())>tolerance_);
 
-      
       if (E(sample_.value_second()) - E(sample_.value_first()) > 2*tolerance_){
         return true;
@@ -279 +302 @@
     // if no support vectors - special case
     else{
+      // to avoid getting stuck we shuffle
+      sample_.shuffle();
       for (size_t i=0; i<sample_.n(); i++) {
         if (target_.binary(sample_(i))){