lmnn-function.patch

diff --git a/src/mlpack/methods/lmnn/lmnn_function_impl.hpp b/src/mlpack/methods/lmnn/lmnn_function_impl.hpp
index b9fb88ad6..5de8e8279 100644
--- a/src/mlpack/methods/lmnn/lmnn_function_impl.hpp
+++ b/src/mlpack/methods/lmnn/lmnn_function_impl.hpp
@@ -193,6 +193,8 @@ double LMNNFunction<MetricType>::Evaluate(const arma::mat& transformation,
         begin, batchSize);
   }
 
+  size_t prunes = 0;
+  size_t checks = 0;
   for (size_t i = begin; i < begin + batchSize; i++)
   {
     for (size_t j = 0; j < k ; j++)
@@ -211,6 +213,7 @@ double LMNNFunction<MetricType>::Evaluate(const arma::mat& transformation,
         // Calculate cost due to {data point, target neighbors, impostors}
         // triplets.
         double eval = 0;
+        ++checks;
         if (transformationOld.n_elem != 0)
         {
           eval = evalOld[i] + transformationDiff * (norm(targetNeighbors(j, i)) +
@@ -232,6 +235,10 @@ double LMNNFunction<MetricType>::Evaluate(const arma::mat& transformation,
                        transformedDataset.col(impostors(l, i)));
             }
           }
+          else
+          {
+            ++prunes;
+          }
         }
         else
         {
@@ -256,6 +263,9 @@ double LMNNFunction<MetricType>::Evaluate(const arma::mat& transformation,
     }
   }
 
+  std::cout << "Evaluate(): " << prunes << " pruned of " << checks << ", "
+      << "transformation diff " << transformationDiff << ".\n";
+
   // Update cache transformation matrix.
   transformationOld = transformation;
 
@@ -373,6 +383,7 @@ double LMNNFunction<MetricType>::EvaluateWithGradient(
                                    const arma::mat& transformation,
                                    GradType& gradient)
 {
+  Timer::Start("lmnn_function_evaluate_with_gradient");
   double cost = 0;
 
   // Apply metric over dataset.
@@ -399,6 +410,12 @@ double LMNNFunction<MetricType>::EvaluateWithGradient(
   // Calculate gradient due to impostors.
   arma::mat cil = arma::zeros(dataset.n_rows, dataset.n_rows);
 
+  size_t prunes = 0;
+  size_t checks = 0;
+  size_t active = 0;
+  size_t close = 0;
+  size_t inactive = 0;
+  Timer::Start("lmnn_function_compute_evals");
   for (size_t i = 0; i < dataset.n_cols; i++)
   {
     for (size_t j = 0; j < k ; j++)
@@ -417,10 +434,15 @@ double LMNNFunction<MetricType>::EvaluateWithGradient(
         // Calculate cost due to {data point, target neighbors, impostors}
         // triplets.
         double eval = 0;
+        ++checks;
         if (transformationOld.n_elem != 0)
         {
           eval = evalOld[i] + transformationDiff * (norm(targetNeighbors(j, i)) +
               norm(impostors(l, i)) + 2 * norm(i));
+//          std::cout << "eval bound of point " << i << " with l " << l << ": " <<
+//eval << "; tdiff " << transformationDiff << " norm target " <<
+//norm(targetNeighbors(j, i)) << ", norm impostors " << norm(impostors(l, i)) <<
+//", norm i " << norm(i) << "; evalOld " << evalOld[i] << "\n";
           if (eval > -1)
           {
             // Calculate exact eval.
@@ -438,6 +460,10 @@ double LMNNFunction<MetricType>::EvaluateWithGradient(
                        transformedDataset.col(impostors(l, i)));
             }
           }
+          else
+          {
+            ++prunes;
+          }
         }
         else
         {
@@ -446,16 +472,21 @@ double LMNNFunction<MetricType>::EvaluateWithGradient(
                  distance(l, i);
         }
 
+        // Update cache eval value.
+        evalOld[i] = eval;
+
         // Check bounding condition.
         if (eval <= -1)
         {
+          if (eval < -2)
+            ++inactive;
+          else
+            ++close;
           // update bound.
           bp = l;
           break;
         }
-
-        // Update cache eval value.
-        evalOld[i] = eval;
+        ++active;
 
         cost += regularization * (1 + eval);
 
@@ -468,12 +499,18 @@ double LMNNFunction<MetricType>::EvaluateWithGradient(
       }
     }
   }
+  Timer::Stop("lmnn_function_compute_evals");
 
   gradient = 2 * transformation * ((1 - regularization) * cij +
       regularization * cil);
 
+  std::cout << "Evaluate(): " << prunes << " pruned of " << checks << ", "
+      << "transformation diff " << transformationDiff << ".  Active: " << active
+      << ", close " << close << ", inactive " << inactive << ".\n";
+
   // Update cache transformation matrix.
   transformationOld = transformation;
+  Timer::Stop("lmnn_function_evaluate_with_gradient");
 
   return cost;
 }