Merge pull request #3641 from MikeLing/clean_up_KNN_3

add tests for KNN and fix an error in KDTree solver

src/shogun/multiclass/KDTreeKNNsolver.cpp

@@ -40,7 +40,7 @@ CMulticlassLabels* CKDTREEKNNSolver::classify_objects(CDistance* distance, const
 		output->set_label(i, out_idx + m_min_label);
 	}
 	SG_UNREF(query);
-
+	SG_UNREF(kd_tree);
 	return output;
 }
 
@@ -49,7 +49,7 @@ int32_t* CKDTREEKNNSolver::classify_objects_k(CDistance* distance, const int32_t
 	int32_t* output=SG_MALLOC(int32_t, m_k*num_lab);
 
 	//allocation for distances to nearest neighbors
-	float64_t* dists=SG_MALLOC(float64_t, m_k);
+	SGVector<float64_t> dists(m_k);
 
 	CFeatures* lhs = distance->get_lhs();
 	CKDTree* kd_tree = new CKDTree(m_leaf_size);
@@ -65,12 +65,15 @@ int32_t* CKDTREEKNNSolver::classify_objects_k(CDistance* distance, const int32_t
 		for (index_t j=0; j<m_k; j++)
 		{
 			train_lab[j] = m_train_labels[ NN(j,i) ];
-			dists[j] = distance->distance(i, NN(j,i));
+			dists[j] = distance->distance(NN(j,i), i);
 		}
-		CMath::qsort_index(dists, train_lab, m_k);
+		CMath::qsort_index(dists.vector, train_lab, m_k);
 
 		choose_class_for_multiple_k(output+i, classes, train_lab, num_lab);
 	}
 
+	SG_UNREF(data);
+	SG_UNREF(kd_tree);
+
 	return output;
 }

tests/unit/multiclass/KNN_unittest.cc

@@ -91,7 +91,7 @@ TEST(KNN, kdtree_solver)
 
 	int32_t k=4;
 	CEuclideanDistance* distance = new CEuclideanDistance();	
-	CKNN* knn=new CKNN (k, distance, labels, KNN_BRUTE);
+	CKNN* knn=new CKNN (k, distance, labels, KNN_KDTREE);
 	SG_REF(knn);
 
 	features->add_subset(train);
@@ -159,6 +159,53 @@ TEST(KNN, lsh_solver)
 	SG_UNREF(knn);
 }
 
+TEST(KNN, classify_multiple_brute)
+{
+	int32_t num = 50;
+	int32_t feats = 2;
+	int32_t classes = 3;
+
+	SGVector< float64_t > lab(classes*num);
+	SGMatrix< float64_t > feat(feats, classes*num);
+
+	generate_knn_data(feat, lab, num, classes, feats);
+	SGVector<index_t> train (int32_t(num*classes*0.75));
+	SGVector<index_t> test (int32_t(num*classes*0.25));
+	train.random(0, classes*num-1);
+	test.random(0, classes*num-1);
+
+	CMulticlassLabels* labels = new CMulticlassLabels(lab);
+	CDenseFeatures< float64_t >* features = new CDenseFeatures< float64_t >(feat);
+	CFeatures* features_test = (CFeatures*) features->clone();  
+	CLabels* labels_test = (CLabels*) labels->clone();
+
+	int32_t k=4;
+	CEuclideanDistance* distance = new CEuclideanDistance();                
+	CKNN* knn=new CKNN (k, distance, labels, KNN_BRUTE);
+	SG_REF(knn);
+
+	features->add_subset(train);
+	labels->add_subset(train);
+	knn->train(features);
+
+	// classify for multiple k
+	features_test->add_subset(test);
+	labels_test->add_subset(test);
+
+	CEuclideanDistance* dist = new CEuclideanDistance(features, ((CDotFeatures*)features_test));  
+	knn->set_distance(dist);
+	SGMatrix<int32_t> out_mat =knn->classify_for_multiple_k();
+	features_test->remove_subset();
+
+	for ( index_t i = 0; i < labels_test->get_num_labels(); ++i )
+		for ( index_t j = 0; j < k; ++j )
+			EXPECT_EQ(out_mat(i, j), ((CMulticlassLabels*)labels_test)->get_label(i));
+
+	SG_UNREF(knn);
+	SG_UNREF(features_test);
+	SG_UNREF(labels_test);
+}
+
 
 TEST(KNN, classify_multiple_kdtree)
 {
@@ -169,7 +216,7 @@ TEST(KNN, classify_multiple_kdtree)
 
 	SGVector< float64_t > lab(classes*num);
 	SGMatrix< float64_t > feat(feats, classes*num);
-	
+
 	generate_knn_data(feat, lab, num, classes, feats);
 	SGVector<index_t> train (int32_t(num*classes*0.75));
 	SGVector<index_t> test (int32_t(num*classes*0.25));
@@ -183,7 +230,7 @@ TEST(KNN, classify_multiple_kdtree)
 
 	int32_t k=4;
 	CEuclideanDistance* distance = new CEuclideanDistance();	
-	CKNN* knn=new CKNN (k, distance, labels, KNN_BRUTE);
+	CKNN* knn=new CKNN (k, distance, labels, KNN_KDTREE);
 	SG_REF(knn);
 
 	features->add_subset(train);
@@ -193,7 +240,7 @@ TEST(KNN, classify_multiple_kdtree)
 	// classify for multiple k
 	features_test->add_subset(test);
 	labels_test->add_subset(test);
-	CEuclideanDistance* dist = new CEuclideanDistance(features, ((CDotFeatures*)features_test));	
+	CEuclideanDistance* dist = new CEuclideanDistance(features, ((CDotFeatures*)features_test));
 	knn->set_distance(dist);
 	SGMatrix<int32_t> out_mat =knn->classify_for_multiple_k();
 	features_test->remove_subset();