sequenzo 0.1.21__cp312-cp312-macosx_11_0_arm64.whl

sequenzo/clustering/src/PAMonce.cpp

@@ -0,0 +1,234 @@
+#include <pybind11/pybind11.h>
+#include <pybind11/numpy.h>
+#include <vector>
+#include <iostream>
+#include <sstream>
+#include <algorithm>
+#define WEIGHTED_CLUST_TOL -1e-10
+#include <cfloat>
+#include <cmath>
+
+namespace py = pybind11;
+
+class PAMonce {
+public:
+    PAMonce(int nelement, py::array_t<double> diss, py::array_t<int> centroids, int npass, py::array_t<double> weights){
+        py::print("[>] Starting Partitioning Around Medoids with a Once-Only Swap Pass (PAMonce)...");
+        std::cout << std::flush;
+
+        try {
+            this->nelement = nelement;
+            this->diss = diss;
+            this->centroids = centroids;
+            this->npass = npass;
+            this->weights = weights;
+
+            clusterid = py::array_t<int>(nelement);
+            tclusterid.resize(nelement, -1);
+
+            maxdist = find_max_value(diss);
+            dysma.resize(nelement, maxdist);
+            dysmb.resize(nelement, maxdist);
+
+            fvect.resize(nelement, 0);
+            nclusters = centroids.size();
+        } catch (const std::exception& e){
+            py::print("Error in constructor: ", e.what());
+            throw;
+        }
+    }
+
+    // 小工具：把 vector 打成一行文本
+//    static void debug_print_vec(const char* name,
+//                                const std::vector<int>& v,
+//                                std::size_t maxn = 50) {
+//        std::ostringstream oss;
+//        oss << name << " (n=" << v.size() << ") [";
+//        std::size_t n = std::min<std::size_t>(v.size(), maxn);
+//        for (std::size_t i = 0; i < n; ++i) {
+//            if (i) oss << ", ";
+//            oss << v[i];
+//        }
+//        if (v.size() > n) oss << ", ...";
+//        oss << "]\n";
+//        // 用 cerr 更容易立刻看到，并避免与 Python 输出缓冲混在一起
+//        std::cerr << oss.str() << std::flush;
+//    }
+
+    double find_max_value(py::array_t<double> diss) {
+        auto buf_info = diss.shape();
+        auto ptr = diss.unchecked<2>();
+
+        int rows = buf_info[0];
+        int cols = buf_info[1];
+
+        double max_val = -std::numeric_limits<double>::infinity();
+
+        #pragma omp parallel
+        {
+            double thread_max = -std::numeric_limits<double>::infinity();
+            #pragma omp for nowait
+            for (int i = 0; i < rows; ++i) {
+                for (int j = 0; j < cols; ++j) {
+                    thread_max = std::max(thread_max, ptr(i, j));
+                }
+            }
+
+            #pragma omp critical
+            {
+                max_val = std::max(max_val, thread_max);
+            }
+        }
+
+        return max_val;
+    }
+
+    py::array_t<int> runclusterloop() {
+        auto ptr_diss = diss.unchecked<2>();
+        auto ptr_weights = weights.unchecked<1>();
+        auto ptr_centroids = centroids.mutable_data();
+        auto ptr_clusterid = clusterid.mutable_data();
+
+        for (int i = 0; i < nelement; i++) {
+            ptr_clusterid[i] = -1;
+        }
+
+        double dzsky = 1;
+        int hbest = -1, nbest = -1;
+        double total = -1;
+
+        do {
+            // 为每个点寻找距离它最近和次近的中心点
+            for (int i = 0; i < nelement; i++) {
+                dysma[i] = maxdist;
+                dysmb[i] = maxdist;
+                for (int k = 0; k < nclusters; k++) {
+                    int i_cluster = ptr_centroids[k];
+                    double dist = ptr_diss(i, i_cluster);
+
+                    if (dysma[i] >= dist) {
+                        // 原码是‘>’，现改为‘>=’。
+                        // 因为如果当前点 i 与所有 medoids 的距离都是 maxdist，那么将无法进入这个分支
+                        dysmb[i] = dysma[i];
+                        dysma[i] = dist;
+
+                        tclusterid[i] = k;  // tclusterid 是中间变量，如果没有进入这个分支，那么将采用初始值-1
+                    } else if (dysmb[i] > dist) {
+                        dysmb[i] = dist;
+                    }
+                }
+            }
+
+            if (total < 0) {
+                total = 0;
+                #pragma omp parallel for reduction(+:total) schedule(static)
+                for (int i = 0; i < nelement; i++) {
+                    total += ptr_weights(i) * dysma[i];
+                }
+            }
+
+            dzsky = 1;
+            hbest = -1;
+            nbest = -1;
+
+            // 遍历每个聚类中心 i，寻找替换中心 h 的可能性
+            for (int k = 0; k < nclusters; k++) {
+                int i = ptr_centroids[k];
+                double removeCost = 0;
+
+                // 计算移除该 medoid 的成本
+                #pragma omp parallel for reduction(+:removeCost) schedule(static)
+                for (int j = 0; j < nelement; j++) {
+                    if (tclusterid[j] == k) {
+                        removeCost += ptr_weights(j) * (dysmb[j] - dysma[j]);
+                        fvect[j] = dysmb[j];
+                    } else {
+                        fvect[j] = dysma[j];
+                    }
+                }
+
+                // 查找最优的新 medoid h
+                #pragma omp parallel
+                {
+                    double local_dzsky = 1;
+                    int local_hbest = -1, local_nbest = -1;
+
+                    #pragma omp for schedule(static)
+                    for (int h = 0; h < nelement; h++) {
+                        if (ptr_diss(h, i) > 0) {
+                            double addGain = removeCost;
+                            for (int j = 0; j < nelement; j++) {
+                                if (ptr_diss(h, j) < fvect[j]) {
+                                    addGain += ptr_weights(j) * (ptr_diss(h, j) - fvect[j]);
+                                }
+                            }
+
+                            if (local_dzsky > addGain) {
+                                local_dzsky = addGain;
+                                local_hbest = h;
+                                local_nbest = i;
+                            }
+                        }
+                    }
+
+                    // 合并线程局部结果
+                    #pragma omp critical
+                    {
+                        if (dzsky > local_dzsky) {
+                            dzsky = local_dzsky;
+                            hbest = local_hbest;
+                            nbest = local_nbest;
+                        }
+                    }
+                }
+            }
+
+            // 更新 medoids
+            if (dzsky < WEIGHTED_CLUST_TOL) {
+                for (int k = 0; k < nclusters; k++) {
+                    if (ptr_centroids[k] == nbest) {
+                        ptr_centroids[k] = hbest;
+                    }
+                }
+                total += dzsky;
+            }
+        } while (dzsky < WEIGHTED_CLUST_TOL);
+
+        // ---- 安全打印（无 pybind11，线程/进程均可）----
+//        {
+//            // 复制指针区的数据，形成可打印的 vector
+//            std::vector<int> centroids_vec(ptr_centroids, ptr_centroids + nclusters);
+//
+//            // 注意：如果你用了 OpenMP，可选地只让一个线程打印，避免多线程交叉输出
+//            // #pragma omp single
+//            debug_print_vec("tclusterid", tclusterid);
+//            debug_print_vec("ptr_centroids", centroids_vec);
+//        }
+
+        // 更新最终聚类分配
+        #pragma omp parallel for schedule(static)
+        for (int j = 0; j < nelement; j++) {
+            ptr_clusterid[j] = ptr_centroids[tclusterid[j]];
+        }
+
+        return clusterid;
+    }
+
+
+private:
+    int nelement;
+    py::array_t<double> diss;
+    py::array_t<int> centroids;
+    int npass;
+    py::array_t<double> weights;
+
+    py::array_t<int> clusterid;
+    std::vector<int> tclusterid;
+
+    double maxdist;
+    std::vector<double> dysma;
+    std::vector<double> dysmb;
+
+    std::vector<double> fvect;
+    int nclusters;
+};