Fix types in qv_partition_with_scores() and train_no_init() (#546)

Author: jparismorgan

src/include/detail/flat/qv.h

@@ -536,7 +536,7 @@ auto qv_partition_with_scores(
 
   // Just need a single vector
   std::vector<size_t> top_k(q.num_cols());
-  std::vector<size_t> top_k_scores(q.num_cols());
+  std::vector<float> top_k_scores(q.num_cols());
 
   auto par = stdx::execution::indexed_parallel_policy{(size_t)nthreads};
   stdx::range_for_each(

src/include/index/kmeans.h

@@ -254,7 +254,6 @@ void train_no_init(
   if (::num_vectors(training_set) == 0) {
     return;
   }
-  using feature_type = typename V::value_type;
   using centroid_feature_type = typename C::value_type;
   using index_type = size_t;
 
@@ -276,10 +275,9 @@ void train_no_init(
     // How many centroids should we try to fix up
     size_t heap_size =
         std::ceil(reassign_ratio_ * static_cast<float>(num_partitions_)) + 5;
-    auto high_scores = fixed_min_pair_heap<
-        feature_type,
-        index_type,
-        std::greater<feature_type>>(heap_size, std::greater<feature_type>());
+    auto high_scores =
+        fixed_min_pair_heap<float, index_type, std::greater<float>>(
+            heap_size, std::greater<float>());
     auto low_degrees = fixed_min_pair_heap<index_type, index_type>(heap_size);
 
     // @todo parallelize -- by partition
@@ -326,7 +324,7 @@ void train_no_init(
       std::sort_heap(begin(high_scores), end(high_scores), [](auto a, auto b) {
         return std::get<0>(a) > std::get<0>(b);
       });
-      for (size_t i = 0; i < size(low_degrees) &&
+      for (size_t i = 0; i < std::min(size(low_degrees), size(high_scores)) &&
                          std::get<0>(low_degrees[i]) <= lower_degree_bound;
            ++i) {
         // std::cout << "i: " << i << " low_degrees: ("
@@ -527,10 +525,9 @@ auto sub_kmeans(
 #ifdef REASSIGN
     // How many centroids should we try to fix up
     size_t heap_size = std::ceil(reassign_ratio * num_clusters) + 5;
-    auto high_scores = fixed_min_pair_heap<
-        feature_type,
-        index_type,
-        std::greater<feature_type>>(heap_size, std::greater<feature_type>());
+    auto high_scores =
+        fixed_min_pair_heap<float, index_type, std::greater<float>>(
+            heap_size, std::greater<float>());
     auto low_degrees = fixed_min_pair_heap<index_type, index_type>(heap_size);
 #endif
 

src/include/test/unit_kmeans.cc

@@ -288,3 +288,116 @@ TEST_CASE(
   // Verify results for kmeans_pp
   verify_centroids(centroids_pp);
 }
+
+TEST_CASE("test kmeans train_no_init random data", "[kmeans]") {
+  // Sample data: 6-dimensional vectors, 10 vectors total (column major)
+  std::vector<float> data = {
+      7, 6,   249, 3,   2,   2,   254, 249, 7,   0,  9,  3,   248, 255, 4,
+      0, 249, 0,   251, 249, 245, 3,   250, 252, 6,  7,  5,   252, 4,   5,
+      9, 9,   248, 254, 7,   1,   4,   1,   253, 5,  2,  255, 250, 6,   3,
+      0, 2,   249, 0,   250, 5,   4,   5,   2,   99, 30, 3,   1,   55,  88};
+
+  ColMajorMatrix<float> training_set(6, 10);  // 6 rows, 10 columns
+  std::copy(data.begin(), data.end(), training_set.data());
+
+  // Initial centroids: 6-dimensional vectors, 3 centroids total
+  std::vector<float> centroids_data = {
+      3,
+      5,
+      250,
+      245,
+      249,
+      0,
+      249,
+      248,
+      250,
+      0,
+      5,
+      251,
+      251,
+      249,
+      245,
+      3,
+      250,
+      252};
+
+  ColMajorMatrix<float> centroids(6, 3);
+  std::copy(centroids_data.begin(), centroids_data.end(), centroids.data());
+
+  size_t dimension_ = 6;
+  size_t num_partitions_ = 3;
+  uint32_t max_iterations = 2;
+  float tol_ = 2.5e-05;
+  size_t num_threads_ = 12;
+  float reassign_ratio_ = 0.075;
+
+  CHECK(centroids.num_rows() == dimension_);
+  CHECK(centroids.num_cols() == num_partitions_);
+
+  train_no_init(
+      training_set,
+      centroids,
+      dimension_,
+      num_partitions_,
+      max_iterations,
+      tol_,
+      num_threads_,
+      reassign_ratio_);
+
+  CHECK(centroids.num_rows() == dimension_);
+  CHECK(centroids.num_cols() == num_partitions_);
+
+  {
+    ColMajorMatrix<float> original_centroids(6, 3);
+    std::copy(
+        centroids_data.begin(),
+        centroids_data.end(),
+        original_centroids.data());
+    float max_diff = 0.0;
+    for (size_t i = 0; i < centroids.num_cols(); ++i) {
+      float diff =
+          sum_of_squares_distance{}(centroids[i], original_centroids[i]);
+      max_diff = std::max(max_diff, diff);
+    }
+    REQUIRE_THAT(max_diff, Catch::Matchers::WithinAbs(91858.75f, 1e-2));
+  }
+}
+
+TEST_CASE("test kmeans train_no_init training_set is empty", "[kmeans]") {
+  ColMajorMatrix<float> training_set(0, 0);  // Empty training set
+  ColMajorMatrix<float> centroids(0, 0);     // Empty centroids
+
+  train_no_init(training_set, centroids, 0, 0, 2, 0.00001, 12, 0.075);
+
+  CHECK(centroids.num_cols() == 0);  // Expect centroids to remain empty
+  CHECK(centroids.num_rows() == 0);
+}
+
+TEST_CASE(
+    "test kmeans train_no_init number of centroids exceeds data points",
+    "[kmeans]") {
+  std::vector<float> small_data = {1, 2, 3, 4, 5, 6};
+  ColMajorMatrix<float> small_training_set(6, 1);  // 6 rows, 1 column
+  std::copy(small_data.begin(), small_data.end(), small_training_set.data());
+
+  ColMajorMatrix<float> more_centroids(
+      6, 3);  // More centroids than data points
+
+  train_no_init(
+      small_training_set, more_centroids, 6, 3, 2, 0.00001, 12, 0.075);
+
+  CHECK(more_centroids.num_cols() == 3);  // Verify centroids were generated
+  for (size_t i = 0; i < more_centroids.num_cols(); ++i) {
+    // Ensure some centroids match the data point and the rest are zeros
+    bool is_zero = std::all_of(
+        more_centroids[i].begin(), more_centroids[i].end(), [](float val) {
+          return val == 0.0f;
+        });
+    if (!is_zero) {
+      CHECK(std::equal(
+          more_centroids[i].begin(),
+          more_centroids[i].end(),
+          small_data.begin()));
+    }
+  }
+}