tmp update

src/lib/lbfgs.c

@@ -99,6 +99,7 @@ struct tag_callback_data {
     void *instance; // 用户给出的运行实例
     lbfgs_evaluate_t proc_evaluate; // 目标函数与模型梯度计算函数指针
     lbfgs_progress_t proc_progress; // 迭代过程监控函数指针
+    lbfgs_precondition_t proc_precondition; // 预优函数指针
 };
 typedef struct tag_callback_data callback_data_t;
 
@@ -250,7 +251,8 @@ int lbfgs(
     lbfgs_evaluate_t proc_evaluate,
     lbfgs_progress_t proc_progress,
     void *instance,
-    lbfgs_parameter_t *_param
+    lbfgs_parameter_t *_param,
+    lbfgs_precondition_t proc_precondition
     )
 {
     int ret;
@@ -264,7 +266,7 @@ int lbfgs(
     const int m = param.m;
 
     lbfgsfloatval_t *xp = NULL;
-    lbfgsfloatval_t *g = NULL, *gp = NULL, *pg = NULL;
+    lbfgsfloatval_t *g = NULL, *gp = NULL, *pg = NULL; // gp (p for previous) pg (p for pesudo)
     lbfgsfloatval_t *d = NULL, *w = NULL, *pf = NULL;
     iteration_data_t *lm = NULL, *it = NULL;
     lbfgsfloatval_t ys, yy;
@@ -280,6 +282,7 @@ int lbfgs(
     cd.instance = instance;
     cd.proc_evaluate = proc_evaluate;
     cd.proc_progress = proc_progress;
+    cd.proc_precondition = proc_precondition;
 
 #if     defined(USE_SSE) && (defined(__SSE__) || defined(__SSE2__))
     /* Round out the number of variables. */
@@ -379,6 +382,7 @@ int lbfgs(
         ret = LBFGSERR_OUTOFMEMORY;
         goto lbfgs_exit;
     }
+
     // 初始化计算L1模的数组
     if (param.orthantwise_c != 0.) {
         /* Allocate working space for OW-LQN. */
@@ -455,6 +459,7 @@ int lbfgs(
         vec2norm(&gnorm, pg, n);
     }
     // 为啥要保证xnorm大于等于1？不明白
+    // 答 参考头文件中的注释 如果模型参数的模长小于1则使用梯度的模长作为收敛性的测试指标
     if (xnorm < 1.0) xnorm = 1.0;
     // 如果输入x即为最优化的解 则退出
     if (gnorm / xnorm <= param.epsilon) {
@@ -465,14 +470,14 @@ int lbfgs(
     /* Compute the initial step:
         step = 1.0 / sqrt(vecdot(d, d, n))
      */
-    // 计算估算的初始步长
+    // 计算估算的初始步长 这一步在实际应用中重要 在相似代码编写中需参考
     vec2norminv(&step, d, n); // 计算数组L2模的倒数，与注释的内容等效
 
     k = 1;
     end = 0;
     for (;;) {
         /* Store the current position and gradient vectors. */
-        veccpy(xp, x, n);
+        veccpy(xp, x, n); // p for previous
         veccpy(gp, g, n);
 
         /* Search for an optimal step. */
@@ -485,7 +490,8 @@ int lbfgs(
                 param.orthantwise_c, param.orthantwise_start, param.orthantwise_end
                 );
         }
-        // 线性搜索错误 此时则退回到上一次迭代的位置并退出
+
+        // 返回值小于0则表示线性搜索错误 此时则退回到上一次迭代的位置并退出
         if (ls < 0) {
             /* Revert to the previous point. */
             veccpy(x, xp, n);
@@ -560,6 +566,7 @@ int lbfgs(
                 s_{k+1} = x_{k+1} - x_{k} = \step * d_{k}.
                 y_{k+1} = g_{k+1} - g_{k}.
          */
+        // 计算最新保存的s和y
         it = &lm[end];
         vecdiff(it->s, x, xp, n); // 计算两个数组的差 it->s = x - xp
         vecdiff(it->y, g, gp, n);
@@ -582,8 +589,12 @@ int lbfgs(
                 Mathematics of Computation, Vol. 35, No. 151,
                 pp. 773--782, 1980.
          */
+
+        // m小于迭代次数
         bound = (m <= k) ? m : k;
         ++k;
+
+        // end+1等于m则将其重置为0 循环保存
         end = (end + 1) % m;
 
         /* Compute the steepest direction. */
@@ -607,6 +618,11 @@ int lbfgs(
 
         vecscale(d, ys / yy, n); // 适当缩放d的大小
 
+        // 我们在这里提供一个预优函数的接口 返回 d = H0^-1 * d
+        if (cd.proc_precondition) {
+            cd.proc_precondition(cd.instance, x, d, n);
+        }
+
         for (i = 0;i < bound;++i) {
             it = &lm[j];
             /* \beta_{j} = \rho_{j} y^t_{j} \cdot \gamma_{i}. */

src/lib/lbfgs.h

@@ -463,6 +463,15 @@ typedef int (*lbfgs_progress_t)(
     int ls
     );
 
+
+
+typedef void (*lbfgs_precondition_t)(
+    void *instance,
+    const lbfgsfloatval_t *x,
+    lbfgsfloatval_t *d,
+    const int n
+    );
+
 /*
 A user must implement a function compatible with ::lbfgs_evaluate_t (evaluation
 callback) and pass the pointer to the callback function to lbfgs() arguments.
@@ -550,7 +559,8 @@ int lbfgs(
     lbfgs_evaluate_t proc_evaluate,
     lbfgs_progress_t proc_progress,
     void *instance,
-    lbfgs_parameter_t *param
+    lbfgs_parameter_t *param,
+    lbfgs_precondition_t proc_precondition
     );
 
 // 将一个参数类型内的全部变量值重置为默认值

src/sample/sample.c

@@ -71,7 +71,7 @@ int main(int argc, char *argv[])
         Start the L-BFGS optimization; this will invoke the callback functions
         evaluate() and progress() when necessary.
      */
-    ret = lbfgs(N, x, &fx, evaluate, progress, NULL, &param);
+    ret = lbfgs(N, x, &fx, evaluate, progress, NULL, &param, NULL);
 
     /* Report the result. */
     printf("L-BFGS optimization terminated with status code = %d\n", ret);

src/sample/sample2.cpp

@@ -39,7 +39,7 @@ public:
             Start the L-BFGS optimization; this will invoke the callback functions
             evaluate() and progress() when necessary.
          */
-        int ret = lbfgs(N, m_x, &fx, _evaluate, _progress, this, NULL);
+        int ret = lbfgs(N, m_x, &fx, _evaluate, _progress, this, NULL, NULL);
 
         /* Report the result. */
         printf("L-BFGS optimization terminated with status code = %d\n", ret);

src/sample/sample3.cpp

@@ -88,7 +88,7 @@ int TEST_FUNC::Routine()
 	//self_para.max_linesearch = 40;
 	//self_para.linesearch = LBFGS_LINESEARCH_BACKTRACKING_WOLFE;
 
-	int ret = lbfgs(3, m_x, &fx, _Func, _Progress, this, &self_para);
+	int ret = lbfgs(3, m_x, &fx, _Func, _Progress, this, &self_para, NULL);
 
 	clog << "L-BFGS optimization terminated with status: " << endl << lbfgs_strerror(ret) << endl;
 	clog << m_x[0] << " " << m_x[1] << " " << m_x[2] << endl;