[mlpack-svn] r13966 - mlpack/trunk/src/mlpack/methods/nca

[fastlab-svn at coffeetalk-1.cc.gatech.edu]

Author: rcurtin
Date: 2012-12-02 23:55:19 -0500 (Sun, 02 Dec 2012)
New Revision: 13966

Modified:
   mlpack/trunk/src/mlpack/methods/nca/nca.hpp
   mlpack/trunk/src/mlpack/methods/nca/nca_impl.hpp
   mlpack/trunk/src/mlpack/methods/nca/nca_main.cpp
Log:
Refactor NCA to take an optimizer as a template parameter, and default to SGD.


Modified: mlpack/trunk/src/mlpack/methods/nca/nca.hpp
===================================================================
--- mlpack/trunk/src/mlpack/methods/nca/nca.hpp	2012-12-01 22:59:57 UTC (rev 13965)
+++ mlpack/trunk/src/mlpack/methods/nca/nca.hpp	2012-12-03 04:55:19 UTC (rev 13966)
@@ -9,7 +9,10 @@
 
 #include <mlpack/core.hpp>
 #include <mlpack/core/metrics/lmetric.hpp>
+#include <mlpack/core/optimizers/sgd/sgd.hpp>
 
+#include "nca_softmax_error_function.hpp"
+
 namespace mlpack {
 namespace nca /** Neighborhood Components Analysis. */ {
 
@@ -36,7 +39,8 @@
  * }
  * @endcode
  */
-template<typename MetricType>
+template<typename MetricType = metric::SquaredEuclideanDistance,
+         template<typename> class OptimizerType = optimization::SGD>
 class NCA
 {
  public:
@@ -55,10 +59,6 @@
    */
   NCA(const arma::mat& dataset,
       const arma::uvec& labels,
-      const double stepSize = 0.01,
-      const size_t maxIterations = 500000,
-      const double tolerance = 1e-10,
-      const bool shuffle = true,
       MetricType metric = MetricType());
 
   /**
@@ -77,21 +77,12 @@
   //! Get the labels reference.
   const arma::uvec& Labels() const { return labels; }
 
-  //! Get the step size for stochastic gradient descent.
-  double StepSize() const { return stepSize; }
-  //! Modify the step size for stochastic gradient descent.
-  double& StepSize() { return stepSize; }
+  //! Get the optimizer.
+  const OptimizerType<SoftmaxErrorFunction<MetricType> >& Optimizer() const
+  { return optimizer; }
+  OptimizerType<SoftmaxErrorFunction<MetricType> >& Optimizer()
+  { return optimizer; }
 
-  //! Get the maximum number of iterations for stochastic gradient descent.
-  size_t MaxIterations() const { return maxIterations; }
-  //! Modify the maximum number of iterations for stochastic gradient descent.
-  size_t& MaxIterations() { return maxIterations; }
-
-  //! Get the tolerance for the termination of stochastic gradient descent.
-  double Tolerance() const { return tolerance; }
-  //! Modify the tolerance for the termination of stochastic gradient descent.
-  double& Tolerance() { return tolerance; }
-
  private:
   //! Dataset reference.
   const arma::mat& dataset;
@@ -101,14 +92,11 @@
   //! Metric to be used.
   MetricType metric;
 
-  //! Step size for stochastic gradient descent.
-  double stepSize;
-  //! Maximum iterations for stochastic gradient descent.
-  size_t maxIterations;
-  //! Tolerance for termination of stochastic gradient descent.
-  double tolerance;
-  //! Whether or not to shuffle the dataset for SGD.
-  bool shuffle;
+  //! The function to optimize.
+  SoftmaxErrorFunction<MetricType> errorFunction;
+
+  //! The optimizer to use.
+  OptimizerType<SoftmaxErrorFunction<MetricType> > optimizer;
 };
 
 }; // namespace nca

Modified: mlpack/trunk/src/mlpack/methods/nca/nca_impl.hpp
===================================================================
--- mlpack/trunk/src/mlpack/methods/nca/nca_impl.hpp	2012-12-01 22:59:57 UTC (rev 13965)
+++ mlpack/trunk/src/mlpack/methods/nca/nca_impl.hpp	2012-12-03 04:55:19 UTC (rev 13966)
@@ -10,48 +10,32 @@
 // In case it was not already included.
 #include "nca.hpp"
 
-#include <mlpack/core/optimizers/sgd/sgd.hpp>
-
-#include "nca_softmax_error_function.hpp"
-
 namespace mlpack {
 namespace nca {
 
 // Just set the internal matrix reference.
-template<typename MetricType>
-NCA<MetricType>::NCA(const arma::mat& dataset,
-                     const arma::uvec& labels,
-                     const double stepSize,
-                     const size_t maxIterations,
-                     const double tolerance,
-                     const bool shuffle,
-                     MetricType metric) :
+template<typename MetricType, template<typename> class OptimizerType>
+NCA<MetricType, OptimizerType>::NCA(const arma::mat& dataset,
+                                    const arma::uvec& labels,
+                                    MetricType metric) :
     dataset(dataset),
     labels(labels),
     metric(metric),
-    stepSize(stepSize),
-    maxIterations(maxIterations),
-    tolerance(tolerance),
-    shuffle(shuffle)
+    errorFunction(dataset, labels, metric),
+    optimizer(OptimizerType<SoftmaxErrorFunction<MetricType> >(errorFunction))
 { /* Nothing to do. */ }
 
-template<typename MetricType>
-void NCA<MetricType>::LearnDistance(arma::mat& outputMatrix)
+template<typename MetricType, template<typename> class OptimizerType>
+void NCA<MetricType, OptimizerType>::LearnDistance(arma::mat& outputMatrix)
 {
   // See if we were passed an initialized matrix.
   if ((outputMatrix.n_rows != dataset.n_rows) ||
       (outputMatrix.n_cols != dataset.n_rows))
     outputMatrix.eye(dataset.n_rows, dataset.n_rows);
 
-  SoftmaxErrorFunction<MetricType> errorFunc(dataset, labels, metric);
-
-  // We will use stochastic gradient descent to optimize the NCA error function.
-  optimization::SGD<SoftmaxErrorFunction<MetricType> > sgd(errorFunc, stepSize,
-      maxIterations, tolerance, shuffle);
-
   Timer::Start("nca_sgd_optimization");
 
-  sgd.Optimize(outputMatrix);
+  optimizer.Optimize(outputMatrix);
 
   Timer::Stop("nca_sgd_optimization");
 }

Modified: mlpack/trunk/src/mlpack/methods/nca/nca_main.cpp
===================================================================
--- mlpack/trunk/src/mlpack/methods/nca/nca_main.cpp	2012-12-01 22:59:57 UTC (rev 13965)
+++ mlpack/trunk/src/mlpack/methods/nca/nca_main.cpp	2012-12-03 04:55:19 UTC (rev 13966)
@@ -126,8 +126,11 @@
   }
 
   // Now create the NCA object and run the optimization.
-  NCA<LMetric<2> > nca(data, labels.unsafe_col(0), stepSize, maxIterations,
-      tolerance, shuffle);
+  NCA<LMetric<2> > nca(data, labels.unsafe_col(0));
+  nca.Optimizer().StepSize() = stepSize;
+  nca.Optimizer().MaxIterations() = maxIterations;
+  nca.Optimizer().Tolerance() = tolerance;
+  nca.Optimizer().Shuffle() = shuffle;
 
   nca.LearnDistance(distance);