Use early stop strategy in K-Means (#4608)

* Use early stop strategy

Signed-off-by: sahuang <xiaohaix@student.unimelb.edu.au>

core/src/index/archive/KnowhereResource.cpp

@@ -71,8 +71,8 @@ KnowhereResource::Initialize() {
     }
 
     // init faiss global variable
-    int64_t use_blas_threshold = config.engine.use_blas_threshold();
-    faiss::distance_compute_blas_threshold = use_blas_threshold;
+    faiss::distance_compute_blas_threshold = config.engine.use_blas_threshold();
+    faiss::early_stop_threshold = config.engine.early_stop_threshold();
 
     int64_t clustering_type = config.engine.clustering_type();
     switch (clustering_type) {

core/src/index/thirdparty/faiss/Clustering.cpp

@@ -262,6 +262,7 @@ int split_clusters (size_t d, size_t k, size_t n,
 };
 
 ClusteringType clustering_type = ClusteringType::K_MEANS;
+double early_stop_threshold = 0.0;
 
 void Clustering::kmeans_algorithm(std::vector<int>& centroids_index, int64_t random_seed,
                                   size_t n_input_centroids, size_t d, size_t k,
@@ -532,6 +533,7 @@ void Clustering::train_encoded (idx_t nx, const uint8_t *x_in,
         // k-means iterations
 
         float err = 0;
+        float prev_objective = 0;
         for (int i = 0; i < niter; i++) {
             double t0s = getmillisecs();
 
@@ -600,6 +602,14 @@ void Clustering::train_encoded (idx_t nx, const uint8_t *x_in,
             }
 
             index.add (k, centroids.data());
+
+            // Early stop strategy
+            float diff = (prev_objective == 0) ? std::numeric_limits<float>::max() : (prev_objective - stats.obj) / prev_objective;
+            prev_objective = stats.obj;
+            if (diff < early_stop_threshold / 100.) {
+                break;
+            }
+
             InterruptCallback::check ();
         }
 

core/src/index/thirdparty/faiss/Clustering.h

@@ -25,9 +25,12 @@ enum ClusteringType
     K_MEANS_TWO,
 };
 
-//The default algorithm use the K_MEANS
+// The default algorithm use the K_MEANS
 extern ClusteringType clustering_type;
 
+// K-Means Early Stop Threshold; defaults to 0.0
+extern double early_stop_threshold;
+
 
 /** Class for the clustering parameters. Can be passed to the
  * constructor of the Clustering object.

core/src/value/config/ConfigInit.cpp

@@ -191,6 +191,7 @@ InitConfig() {
         Enum(engine.clustering_type, &ClusteringMap, ClusteringType::K_MEANS),
         Enum(engine.simd_type, &SimdMap, SimdType::AUTO),
         Integer(engine.statistics_level, 0, 3, 1),
+        Floating(engine.early_stop_threshold, 0.0, 100.0, 0.0),
 
         Bool(system.lock.enable, true),
 

core/src/value/config/ConfigMgr.cpp

@@ -69,6 +69,7 @@ ConfigMgr::ConfigMgr() : ValueMgr(InitConfig()) {
         "engine.search_combine_nq",
         "engine.use_blas_threshold",
         "engine.omp_thread_num",
+        "engine.early_stop_threshold",
     };
 }
 

core/src/value/config/ServerConfig.h

@@ -110,6 +110,7 @@ struct ServerConfig {
         Integer clustering_type;
         Integer simd_type;
         Integer statistics_level;
+        Floating early_stop_threshold;
     } engine;
 
     struct GPU {