From: Francois Fleuret <francois@fleuret.org>
Date: Wed, 27 Mar 2013 11:36:16 +0000 (+0100)
Subject: Cosmetics.
X-Git-Url: https://www.fleuret.org/cgi-bin/gitweb/gitweb.cgi?p=clueless-kmeans.git;a=commitdiff_plain;h=792d1a58d91f607d47ea5316ec680fb5e3454e5e

Cosmetics.
---

diff --git a/clueless-kmean.cc b/clueless-kmean.cc
index 1eea96f..16f7691 100644
--- a/clueless-kmean.cc
+++ b/clueless-kmean.cc
@@ -76,7 +76,9 @@ int main(int argc, char **argv) {
 
   glp_term_out(0);
 
-  clusterer.train(nb_clusters, sample_set.dim,
+  clusterer.train(Clusterer::UNINFORMATIVE_LP_ASSOCIATION,
+                  nb_clusters,
+                  sample_set.dim,
                   sample_set.nb_points, sample_set.points,
                   sample_set.nb_classes, sample_set.labels,
                   associated_clusters);
diff --git a/clusterer.cc b/clusterer.cc
index 42af81b..47d9ac3 100644
--- a/clusterer.cc
+++ b/clusterer.cc
@@ -177,7 +177,6 @@ scalar_t Clusterer::uninformative_lp_cluster_association(int nb_points, scalar_t
     for(int c = 1; c <= nb_classes; c++) {
       int row = nb_points + (k - 1) * nb_classes + c;
       scalar_t tau = nb_samples_per_class[c-1] / scalar_t(_nb_clusters);
-      // cout << "tau " << k << " " << c << " " << tau << endl;
       glp_set_row_bnds(lp, row, GLP_FX, tau, tau);
     }
   }
@@ -241,35 +240,6 @@ scalar_t Clusterer::uninformative_lp_cluster_association(int nb_points, scalar_t
     }
   }
 
-  // { // ******************************* START ***************************
-// #warning Test code added on 2013 Feb 07 20:32:05
-    // // for(int n = 0; n < nb_points; n++) {
-    // // scalar_t sum = 0;
-    // // for(int k = 0; k < _nb_clusters; k++) {
-    // // ASSERT(gamma[n][k] >= 0 && gamma[n][k] <= 1);
-    // // sum += gamma[n][k];
-    // // }
-    // // cout << sum << endl;
-    // // }
-
-    // scalar_t *sum_gamma = new scalar_t[nb_classes];
-
-    // for(int k = 0; k < _nb_clusters; k++) {
-      // for(int c = 0; c < nb_classes; c++) { sum_gamma[c] = 0.0; }
-      // for(int n = 0; n < nb_points; n++) {
-        // sum_gamma[labels[n]] += gamma[n][k];
-      // }
-      // cout << "CLUSTER" << k;
-      // for(int c = 0; c < nb_classes; c++) {
-        // cout << " " << sum_gamma[c];
-      // }
-      // cout << endl;
-    // }
-
-    // delete sum_gamma;
-
-  // } // ******************************** END ****************************
-
   delete[] nb_samples_per_class;
   delete[] ia;
   delete[] ja;
@@ -279,7 +249,7 @@ scalar_t Clusterer::uninformative_lp_cluster_association(int nb_points, scalar_t
   return total_dist;
 }
 
-void Clusterer::baseline_update_clusters(int nb_points, scalar_t **points, scalar_t **gamma)  {
+void Clusterer::update_clusters(int nb_points, scalar_t **points, scalar_t **gamma)  {
   for(int k = 0; k < _nb_clusters; k++) {
 
     for(int d = 0; d < _dim; d++) {
@@ -325,7 +295,8 @@ void Clusterer::initialize_clusters(int nb_points, scalar_t **points) {
   delete[] used;
 }
 
-void Clusterer::train(int nb_clusters, int dim,
+void Clusterer::train(int mode,
+                      int nb_clusters, int dim,
                       int nb_points, scalar_t **points,
                       int nb_classes, int *labels,
                       int *cluster_associations) {
@@ -347,48 +318,41 @@ void Clusterer::train(int nb_clusters, int dim,
 
   do {
     pred_total_distance = total_distance;
+
+    switch(mode) {
+
+    case STANDARD_ASSOCIATION:
+    total_distance =
+      baseline_cluster_association(nb_points, points, nb_classes, labels, gammas);
+      break;
+
+    case STANDARD_LP_ASSOCIATION:
+    total_distance =
+      baseline_lp_cluster_association(nb_points, points, nb_classes, labels, gammas);
+      break;
+
+    case UNINFORMATIVE_LP_ASSOCIATION:
     total_distance =
-      // baseline_cluster_association(nb_points, points, nb_classes, labels, gammas);
-      // baseline_lp_cluster_association(nb_points, points, nb_classes, labels, gammas);
       uninformative_lp_cluster_association(nb_points, points, nb_classes, labels, gammas);
+      break;
+
+    default:
+      cerr << "Unknown sample-cluster association mode." << endl;
+      abort();
+    }
+
     cout << "TRAIN " << nb_rounds << " " << total_distance << endl;
-    baseline_update_clusters(nb_points, points, gammas);
+    update_clusters(nb_points, points, gammas);
     nb_rounds++;
   } while(total_distance < min_iteration_improvement * pred_total_distance &&
           nb_rounds < max_nb_iterations);
 
-  {
-    cout << "TOTAL_NB_SAMPLES";
-    for(int c = 0; c < nb_classes; c++) {
-      int nb_samples = 0;
-      for(int n = 0; n < nb_points; n++) {
-        if(labels[n] == c) {
-          nb_samples++;
-        }
-      }
-      cout << " " << nb_samples;
-    }
-    cout << endl;
-
-    for(int k = 0; k < _nb_clusters; k++) {
-      cout << "CLUSTER_GAMMA_SUM " << k << " :";
-      for(int c = 0; c < nb_classes; c++) {
-        scalar_t sum = 0.0;
-        for(int n = 0; n < nb_points; n++) {
-          if(labels[n] == c) {
-            sum += gammas[n][k];
-          }
+  if(cluster_associations) {
+    for(int n = 0; n < nb_points; n++) {
+      for(int k = 0; k < _nb_clusters; k++) {
+        if(k == 0 || gammas[n][k] > gammas[n][cluster_associations[n]]) {
+          cluster_associations[n] = k;
         }
-        cout << " " << sum;
-      }
-      cout << endl;
-    }
-  }
-
-  for(int n = 0; n < nb_points; n++) {
-    for(int k = 0; k < _nb_clusters; k++) {
-      if(k == 0 || gammas[n][k] > gammas[n][cluster_associations[n]]) {
-        cluster_associations[n] = k;
       }
     }
   }
diff --git a/clusterer.h b/clusterer.h
index 88c168a..b45f8c3 100644
--- a/clusterer.h
+++ b/clusterer.h
@@ -29,6 +29,9 @@
 
 class Clusterer {
 public:
+
+  enum { STANDARD_ASSOCIATION, STANDARD_LP_ASSOCIATION, UNINFORMATIVE_LP_ASSOCIATION };
+
   const static int max_nb_iterations = 10;
   const static scalar_t min_iteration_improvement = 0.999;
 
@@ -38,26 +41,38 @@ public:
 
   void initialize_clusters(int nb_points, scalar_t **points);
 
+  // Does the standard hard k-mean association
+
   scalar_t baseline_cluster_association(int nb_points, scalar_t **points,
                                         int nb_classes, int *labels,
                                         scalar_t **gamma);
 
+  // Does the same with an LP formulation, as a sanity check
+
   scalar_t baseline_lp_cluster_association(int nb_points, scalar_t **points,
                                            int nb_classes, int *labels,
                                            scalar_t **gamma);
 
+  // Does the association under constraints that each cluster gets
+  // associated clusters with the same class proportion as the overall
+  // training set
+
   scalar_t uninformative_lp_cluster_association(int nb_points, scalar_t **points,
                                                 int nb_classes, int *labels,
                                                 scalar_t **gamma);
 
-  void baseline_update_clusters(int nb_points, scalar_t **points, scalar_t **gamma);
+  void update_clusters(int nb_points, scalar_t **points, scalar_t **gamma);
 
 public:
   Clusterer();
   ~Clusterer();
-  void train(int nb_clusters, int dim,
+
+  void train(int mode,
+             int nb_clusters, int dim,
              int nb_points, scalar_t **points,
              int nb_classes, int *labels,
+             // This last array returns for each sample to what
+             // cluster it was associated. It can be null.
              int *cluster_associations);
 
   int cluster(scalar_t *point);