由于DSP片上存储空间有限，只有2M，放了一个Haar特征检测模型之后，再没有多少存储空间了。用libsvm训练出来一个11M的模型没法移植上去，于是想到了试试liblinear。结果发现liblinear训练好的模型小了很多，精度和libsvm差不多。 在vs2012建立一个工程，把liblinear里的blas.h 、blasp.h、linear.h、tron.h、daxpy.cpp、ddot.cpp、dnrm2.cpp、dscal.cpp、linear.cpp、tron.cpp导入你的项目中。 CxLibLinear.h #ifndef _CXLIBLINEAR_H_H_ #define _CXLIBLINEAR_H_H_ #include #include #include #include "linear.h" using namespace std; //内存分配 #define Malloc(type,n) (type *)malloc((n)*sizeof(type)) class CxLibLinear { public: struct parameter param; private: struct feature_node *x_space; struct problem prob; struct model* model_; public: CxLibLinear(); ~CxLibLinear(); void init_linear_param(struct parameter& param); void train_model(const vector>& x, const vector& y, const struct parameter& param); int do_predict(const vector& x,double& prob_est); void do_cross_validation(const vector>& x, const vector& y, const struct parameter& param, const int & nr_fold); int load_linear_model(string model_path); int save_linear_model(string model_path); void free_model(); }; #endif CxLibLinear.cpp #include "stdafx.h" #include "CxLibLinear.h" CxLibLinear::CxLibLinear() { model_ = NULL; } CxLibLinear::~CxLibLinear() { free_model(); } void CxLibLinear::init_linear_param(struct parameter& param) { //参数初始化，参数调整部分在这里修改即可 // 默认参数 param.solver_type = L2R_L2LOSS_SVC; param.eps = 0.001; param.C = 2; param.nr_weight = 0; param.weight_label = NULL; param.weight = NULL; param.p = 0.001; param.init_sol = NULL; } void CxLibLinear::train_model(const vector>& x, const vector& y, const struct parameter& param) { if (x.size() == 0) { return; } //释放先前的模型 free_model(); /*初始化*/ long len = x.size(); long dim = x[0].size(); long elements = len * dim; //转换数据为liblinear格式 prob.l = len; prob.n = dim; prob.bias = -1.0; prob.y = Malloc(double, prob.l); prob.x = Malloc(struct feature_node *, prob.l); x_space = Malloc(struct feature_node, elements + len); int j = 0; for (int l = 0; l < len; l++) { prob.x[l] = &x_space[j]; for (int d = 0; d < dim; d++) { x_space[j].index = d+1; x_space[j].value = x[l][d]; j++; } x_space[j++].index = -1; prob.y[l] = y[l]; } /*训练*/ model_ = train(&prob, ¶m); } int CxLibLinear::do_predict(const vector& x,double& prob_est) { //int nr_class=get_nr_class(model_); //double *prob_estimates=NULL; //int n; //int nr_feature=get_nr_feature(model_); //if(model_->bias>=0) // n=nr_feature+1; //else // n=nr_feature; //double predict_label; //feature_node* x_test = Malloc(struct feature_node, x.size()+1); //for (unsigned int i = 0; i < x.size(); i++) //{ // if(model_->bias>=0) // { // x_test[i].index = n; // x_test[i].value = model_->bias; // } // else // { // x_test[i].index = i + 1; // x_test[i].value = x[i]; // } // //} //x_test[x.size()].index = -1; // //predict_label = predict(model_,x_test); // //return (int)predict_label; //数据转换 feature_node* x_test = Malloc(struct feature_node, x.size()+1); for (unsigned int i = 0; i < x.size(); i++) { x_test[i].index = i + 1; x_test[i].value = x[i]; } x_test[x.size()].index = -1; double *probs = new double[model_->nr_class];//存储了所有类别的概率 //预测类别和概率 int value = (int)predict_values(model_, x_test, probs); for (int k = 0; k < model_->nr_class; k++) {//查找类别相对应的概率 if (model_->label[k] == value) { prob_est = probs[k]; break; } } delete[] probs; return value; } void CxLibLinear::do_cross_validation(const vector>& x, const vector& y, const struct parameter& param, const int & nr_fold) { if (x.size() == 0) return; /*初始化*/ long len = x.size(); long dim = x[0].size(); long elements = len*dim; //转换数据为liblinear格式 prob.l = len; prob.n = dim; prob.y = Malloc(double, prob.l); prob.x = Malloc(struct feature_node *, prob.l); x_space = Malloc(struct feature_node, elements + len); int j = 0; for (int l = 0; l < len; l++) { prob.x[l] = &x_space[j]; for (int d = 0; d < dim; d++) { x_space[j].index = d + 1; x_space[j].value = x[l][d]; j++; } x_space[j++].index = -1; prob.y[l] = y[l]; } int i; int total_correct = 0; double total_error = 0; double sumv = 0, sumy = 0, sumvv = 0, sumyy = 0, sumvy = 0; double *target = Malloc(double, prob.l); cross_validation(&prob, ¶m, nr_fold, target); if(param.solver_type == L2R_L2LOSS_SVR || param.solver_type == L2R_L1LOSS_SVR_DUAL || param.solver_type == L2R_L2LOSS_SVR_DUAL) { for (i = 0; i < prob.l; i++) { double y = prob.y[i]; double v = target[i]; total_error += (v - y)*(v - y); sumv += v; sumy += y; sumvv += v*v; sumyy += y*y; sumvy += v*y; } printf("Cross Validation Mean squared error = %g ", total_error / prob.l); printf("Cross Validation Squared correlation coefficient = %g ", ((prob.l*sumvy - sumv*sumy)*(prob.l*sumvy - sumv*sumy)) / ((prob.l*sumvv - sumv*sumv)*(prob.l*sumyy - sumy*sumy)) ); } else { for (i = 0; i < prob.l; i++) if (target[i] == prob.y[i]) ++total_correct; printf("Cross Validation Accuracy = %g%% ", 100.0*total_correct / prob.l); } free(target); } int CxLibLinear::load_linear_model(string model_path) { //释放原来的模型 free_model(); //导入模型 model_ = load_model(model_path.c_str()); if (model_ == NULL) return -1; return 0; } int CxLibLinear::save_linear_model(string model_path) { int flag = save_model(model_path.c_str(), model_); return flag; } void CxLibLinear::free_model() { if (model_ != NULL) { free_and_destroy_model(&model_); destroy_param(¶m); if (prob.y != NULL) { free(prob.y); prob.y = NULL; } if (prob.x != NULL) { free(prob.x); prob.x = NULL; } if (x_space != NULL) { free(x_space); x_space = NULL; } } } main.cpp #include "stdafx.h" #include "CxLibLinear.h" #include #include using namespace std; void gen_train_sample(vector>& x, vector& y, long sample_num, long dim, double scale); void gen_test_sample(vector& x, long sample_num, long dim, double scale); int _tmain(int argc, _TCHAR* argv[]) { //初始化liblinear对象 CxLibLinear linear; linear.init_linear_param(linear.param); /*1、准备训练数据*/ vector> x; //样本集 vector y; //样本类别标签集 gen_train_sample(x, y, 2000, 288, 1); /*1、交叉验证*/ int fold = 10; linear.do_cross_validation(x, y, linear.param, fold); /*2、训练*/ linear.train_model(x, y, linear.param); /*3、保存模型*/ string model_path = "linear_model.txt"; linear.save_linear_model(model_path); /*4、导入模型*/ linear.load_linear_model(model_path); /*5、预测*/ //生成随机测试数据 vector x_test; gen_test_sample(x_test, 2000, 288, -1); double prob_est; //预测 int value = linear.do_predict(x_test, prob_est); //打印预测类别和概率 printf("label: %d ", value); return 0; } void gen_train_sample(vector>& x, vector& y, long sample_num, long dim, double scale) { srand((unsigned)time(NULL));//随机数 //生成随机的正类样本 for (int i = 0; i < sample_num; i++) { vector rx; for (int j = 0; j < dim; j++) { rx.push_back(scale*(rand() % dim)); } x.push_back(rx); y.push_back(1); //printf("y = %d ",(int)y[i]); } //生成随机的负类样本 for (int m = 0; m < sample_num; m++) { vector rx; for (int n = 0; n < dim; n++) { rx.push_back(-scale*(rand() % dim)); } x.push_back(rx); y.push_back(2); //printf("y = %d ",(int)y[sample_num + m]); } } void gen_test_sample(vector& x, long sample_num, long dim, double scale) { srand((unsigned)time(NULL));//随机数 //生成随机的正类样本 for (int j = 0; j < dim; j++) { x.push_back(-scale*(rand() % dim)); } }