Extract landmarks code

include/tapkee/methods/landmark_isomap.hpp

@@ -8,6 +8,7 @@
 #include <tapkee/methods/base.hpp>
 #include <tapkee/routines/isomap.hpp>
 #include <tapkee/routines/multidimensional_scaling.hpp>
+#include <tapkee/routines/landmarks.hpp>
 /* End of Tapkee includes */
 
 namespace tapkee

include/tapkee/methods/landmark_multidimensional_scaling.hpp

@@ -7,6 +7,7 @@
 /* Tapkee includes */
 #include <tapkee/methods/base.hpp>
 #include <tapkee/routines/multidimensional_scaling.hpp>
+#include <tapkee/routines/landmarks.hpp>
 /* End of Tapkee includes */
 
 namespace tapkee

include/tapkee/routines/landmarks.hpp

@@ -0,0 +1,78 @@
+/* This software is distributed under BSD 3-clause license (see LICENSE file).
+ *
+ * Copyright (c) 2012-2024 Sergey Lisitsyn
+ */
+#pragma once
+
+/* Tapkee includes */
+#include <tapkee/defines.hpp>
+#include <tapkee/utils/time.hpp>
+/* End of Tapkee includes */
+
+namespace tapkee
+{
+namespace tapkee_internal
+{
+
+template <class RandomAccessIterator>
+Landmarks select_landmarks_random(RandomAccessIterator begin, RandomAccessIterator end, ScalarType ratio)
+{
+    Landmarks landmarks;
+    landmarks.reserve(end - begin);
+    for (RandomAccessIterator iter = begin; iter != end; ++iter)
+        landmarks.push_back(iter - begin);
+    tapkee::random_shuffle(landmarks.begin(), landmarks.end());
+    landmarks.erase(landmarks.begin() + static_cast<IndexType>(landmarks.size() * ratio), landmarks.end());
+    return landmarks;
+}
+
+template <class RandomAccessIterator, class PairwiseCallback>
+DenseMatrix triangulate(RandomAccessIterator begin, RandomAccessIterator end, PairwiseCallback distance_callback,
+                        Landmarks& landmarks, DenseVector& landmark_distances_squared,
+                        EigendecompositionResult& landmarks_embedding, IndexType target_dimension)
+{
+    timed_context context("Landmark triangulation");
+
+    const IndexType n_vectors = end - begin;
+    const IndexType n_landmarks = landmarks.size();
+
+    std::vector<bool> to_process(n_vectors, true);
+
+    DenseMatrix embedding(n_vectors, target_dimension);
+
+    for (IndexType index_iter = 0; index_iter < n_landmarks; ++index_iter)
+    {
+        to_process[landmarks[index_iter]] = false;
+        embedding.row(landmarks[index_iter]).noalias() = landmarks_embedding.first.row(index_iter);
+    }
+
+    for (IndexType i = 0; i < target_dimension; ++i)
+        landmarks_embedding.first.col(i).array() /= landmarks_embedding.second(i);
+
+#pragma omp parallel
+    {
+        DenseVector distances_to_landmarks(n_landmarks);
+        IndexType index_iter;
+#pragma omp for nowait
+        for (index_iter = 0; index_iter < n_vectors; ++index_iter)
+        {
+            if (!to_process[index_iter])
+                continue;
+
+            for (IndexType i = 0; i < n_landmarks; ++i)
+            {
+                ScalarType d = distance_callback.distance(begin[index_iter], begin[landmarks[i]]);
+                distances_to_landmarks(i) = d * d;
+            }
+            // distances_to_landmarks.array().square();
+
+            distances_to_landmarks -= landmark_distances_squared;
+            embedding.row(index_iter).noalias() = -0.5 * landmarks_embedding.first.transpose() * distances_to_landmarks;
+        }
+    }
+
+    return embedding;
+}
+
+} // End of namespace tapkee_internal
+} // End of namespace tapkee

include/tapkee/routines/multidimensional_scaling.hpp

@@ -14,18 +14,6 @@ namespace tapkee
 namespace tapkee_internal
 {
 
-template <class RandomAccessIterator>
-Landmarks select_landmarks_random(RandomAccessIterator begin, RandomAccessIterator end, ScalarType ratio)
-{
-    Landmarks landmarks;
-    landmarks.reserve(end - begin);
-    for (RandomAccessIterator iter = begin; iter != end; ++iter)
-        landmarks.push_back(iter - begin);
-    tapkee::random_shuffle(landmarks.begin(), landmarks.end());
-    landmarks.erase(landmarks.begin() + static_cast<IndexType>(landmarks.size() * ratio), landmarks.end());
-    return landmarks;
-}
-
 template <class RandomAccessIterator, class PairwiseCallback>
 DenseSymmetricMatrix compute_distance_matrix(RandomAccessIterator begin, RandomAccessIterator /*end*/,
                                              Landmarks& landmarks, PairwiseCallback callback)
@@ -53,57 +41,6 @@ DenseSymmetricMatrix compute_distance_matrix(RandomAccessIterator begin, RandomA
     return distance_matrix;
 }
 
-template <class RandomAccessIterator, class PairwiseCallback>
-DenseMatrix triangulate(RandomAccessIterator begin, RandomAccessIterator end, PairwiseCallback distance_callback,
-                        Landmarks& landmarks, DenseVector& landmark_distances_squared,
-                        EigendecompositionResult& landmarks_embedding, IndexType target_dimension)
-{
-    timed_context context("Landmark triangulation");
-
-    const IndexType n_vectors = end - begin;
-    const IndexType n_landmarks = landmarks.size();
-
-    bool* to_process = new bool[n_vectors];
-    std::fill(to_process, to_process + n_vectors, true);
-
-    DenseMatrix embedding(n_vectors, target_dimension);
-
-    for (IndexType index_iter = 0; index_iter < n_landmarks; ++index_iter)
-    {
-        to_process[landmarks[index_iter]] = false;
-        embedding.row(landmarks[index_iter]).noalias() = landmarks_embedding.first.row(index_iter);
-    }
-
-    for (IndexType i = 0; i < target_dimension; ++i)
-        landmarks_embedding.first.col(i).array() /= landmarks_embedding.second(i);
-
-#pragma omp parallel
-    {
-        DenseVector distances_to_landmarks(n_landmarks);
-        IndexType index_iter;
-#pragma omp for nowait
-        for (index_iter = 0; index_iter < n_vectors; ++index_iter)
-        {
-            if (!to_process[index_iter])
-                continue;
-
-            for (IndexType i = 0; i < n_landmarks; ++i)
-            {
-                ScalarType d = distance_callback.distance(begin[index_iter], begin[landmarks[i]]);
-                distances_to_landmarks(i) = d * d;
-            }
-            // distances_to_landmarks.array().square();
-
-            distances_to_landmarks -= landmark_distances_squared;
-            embedding.row(index_iter).noalias() = -0.5 * landmarks_embedding.first.transpose() * distances_to_landmarks;
-        }
-    }
-
-    delete[] to_process;
-
-    return embedding;
-}
-
 template <class RandomAccessIterator, class PairwiseCallback>
 DenseSymmetricMatrix compute_distance_matrix(RandomAccessIterator begin, RandomAccessIterator end,
                                              PairwiseCallback callback)