Merge pull request mlpack#1310 from yamidark/fix_mean_shift_bug

Fix bug in mean shift when max_iterations is too low w.r.t size of data
vinods07 · Mar 26, 2018 · 789c194 · 789c194
2 parents ff6f638 + a77b0b8
commit 789c194
Show file tree

Hide file tree

Showing 4 changed files with 79 additions and 10 deletions.
diff --git a/src/mlpack/methods/mean_shift/mean_shift.hpp b/src/mlpack/methods/mean_shift/mean_shift.hpp
@@ -33,9 +33,11 @@ namespace meanshift /** Mean shift clustering. */ {
  * extern arma::mat data; // Dataset we want to run mean shift on.
  * arma::Row<size_t> assignments; // Cluster assignments.
  * arma::mat centroids; // Cluster centroids.
+ * bool forceConvergence = true; // Flag whether to force each centroid seed
+ * to converge regardless of maxIterations.
  *
  * MeanShift<> meanShift();
- * meanShift.Cluster(dataset, assignments, centroids);
+ * meanShift.Cluster(dataset, assignments, centroids, forceConvergence);
  * @endcode
  *
  * @tparam UseKernel Use kernel or mean to calculate new centroid.
@@ -80,10 +82,13 @@ class MeanShift
    * @param data Dataset to cluster.
    * @param assignments Vector to store cluster assignments in.
    * @param centroids Matrix in which centroids are stored.
+   * @param forceConvergence Flag whether to force each centroid seed to
+   * converge regardless of maxIterations.
    */
   void Cluster(const MatType& data,
                arma::Row<size_t>& assignments,
                arma::mat& centroids,
+               bool forceConvergence = true,
                bool useSeeds = true);
 
   //! Get the maximum number of iterations.

diff --git a/src/mlpack/methods/mean_shift/mean_shift_impl.hpp b/src/mlpack/methods/mean_shift/mean_shift_impl.hpp
@@ -185,6 +185,7 @@ inline void MeanShift<UseKernel, KernelType, MatType>::Cluster(
     const MatType& data,
     arma::Row<size_t>& assignments,
     arma::mat& centroids,
+    bool forceConvergence,
     bool useSeeds)
 {
   if (radius <= 0)
@@ -216,8 +217,8 @@ inline void MeanShift<UseKernel, KernelType, MatType>::Cluster(
   {
     // Initial centroid is the seed itself.
     allCentroids.col(i) = pSeeds->unsafe_col(i);
-    for (size_t completedIterations = 0; completedIterations < maxIterations;
-         completedIterations++)
+    for (size_t completedIterations = 0; completedIterations < maxIterations
+      || forceConvergence; completedIterations++)
     {
       // Store new centroid in this.
       arma::colvec newCentroid = arma::zeros<arma::colvec>(pSeeds->n_rows);
@@ -260,12 +261,36 @@ inline void MeanShift<UseKernel, KernelType, MatType>::Cluster(
     }
   }
 
-  // Assign centroids to each point.
-  neighbor::KNN neighborSearcher(centroids);
-  arma::mat neighborDistances;
-  arma::Mat<size_t> resultingNeighbors;
-  neighborSearcher.Search(data, 1, resultingNeighbors, neighborDistances);
-  assignments = resultingNeighbors;
+  // If no centroid has converged due to too little iterations and without
+  // forcing convergence, take 1 random centroid calculated.
+  if (centroids.empty())
+  {
+    Log::Warn << "No clusters converge, setting 1 random centroid calculated. "
+    "Try a larger max_iterations or pass force_convergence flag." << std::endl;
+
+    if (maxIterations == 0)
+    {
+      centroids.insert_cols(centroids.n_cols, data.col(0));
+    }
+    else
+    {
+      centroids.insert_cols(centroids.n_cols, allCentroids.col(0));
+    }
+    assignments.zeros();
+  }
+  else if (centroids.n_cols == 1)
+  {
+    assignments.zeros();
+  }
+  else
+  {
+    // Assign centroids to each point.
+    neighbor::KNN neighborSearcher(centroids);
+    arma::mat neighborDistances;
+    arma::Mat<size_t> resultingNeighbors;
+    neighborSearcher.Search(data, 1, resultingNeighbors, neighborDistances);
+    assignments = resultingNeighbors;
+  }
 }
 
 } // namespace meanshift

diff --git a/src/mlpack/methods/mean_shift/mean_shift_main.cpp b/src/mlpack/methods/mean_shift/mean_shift_main.cpp
@@ -53,6 +53,9 @@ PARAM_FLAG("in_place", "If specified, a column containing the learned cluster "
     "--output_file is overridden.  (Do not use with Python.)", "P");
 PARAM_FLAG("labels_only", "If specified, only the output labels will be "
     "written to the file specified by --output_file.", "l");
+PARAM_FLAG("force_convergence", "If specified, the mean shift algorithm will "
+  "continue running regardless of max_iterations until the clusters converge."
+  , "f");
 PARAM_MATRIX_OUT("output", "Matrix to write output labels or labeled data to.",
     "o");
 PARAM_MATRIX_OUT("centroid", "If specified, the centroids of each cluster will "
@@ -89,7 +92,8 @@ static void mlpackMain()
 
   Timer::Start("clustering");
   Log::Info << "Performing mean shift clustering..." << endl;
-  meanShift.Cluster(dataset, assignments, centroids);
+  meanShift.Cluster(dataset, assignments, centroids,
+    CLI::HasParam("force_convergence"));
   Timer::Stop("clustering");
 
   Log::Info << "Found " << centroids.n_cols << " centroids." << endl;

diff --git a/src/mlpack/tests/main_tests/mean_shift_test.cpp b/src/mlpack/tests/main_tests/mean_shift_test.cpp
@@ -118,6 +118,41 @@ BOOST_AUTO_TEST_CASE(MeanShiftInPlaceTest)
   BOOST_REQUIRE_EQUAL(CLI::GetParam<arma::mat>("output").n_cols, numCols);
 }
 
+/**
+ * Ensure that force_convergence is used by testing that the
+ * force_convergence flag makes a difference in the program.
+ */
+BOOST_AUTO_TEST_CASE(MeanShiftForceConvergenceTest)
+{
+  arma::mat x;
+  if (!data::Load("iris_test.csv", x))
+    BOOST_FAIL("Cannot load test dataset iris_test.csv!");
+
+  // Input random data points.
+  SetInputParam("input", x);
+  // Set a very small max_iterations.
+  SetInputParam("max_iterations", (int) 1);
+
+  mlpackMain();
+
+  const int numCentroids1 = CLI::GetParam<arma::mat>("centroid").n_cols;
+
+  ResetSettings();
+
+  // Input same random data points.
+  SetInputParam("input", std::move(x));
+  // Set the same small max_iterations.
+  SetInputParam("max_iterations", (int) 1);
+  // Set the force_convergence flag on.
+  SetInputParam("force_convergence", true);
+
+  mlpackMain();
+
+  const int numCentroids2 = CLI::GetParam<arma::mat>("centroid").n_cols;
+  // Resulting number of centroids should be different.
+  BOOST_REQUIRE_NE(numCentroids1, numCentroids2);
+}
+
 /**
  * Ensure that radius is used by testing that the radius
  * makes a difference in the program.