How to create training data for libsvm (as an svm_node struct)

问题

I am trying to train an svm for a simple xor problem programmatically using libsvm to understand how the library works. The problem (i think) seems to be that i construct svm_node incorrectly; maybe i have trouble understanding the whole pointers to pointers thing. Could anybody help with this? I first construct a matrix for the xor problem then try to assign values from the matrix to svm_node (i am using 2 steps here because my real data will be in matrix format).

When testing the model i get incorrect values (always -1).

In a previous question i got help with the parameters C and gamma; these should be OK now since i got correct classifications for the xor problem using other code. Thanks again Pedrom!

I have searched in several places for answers, e.g. the Readme and in the SvmToy example; no luck however.

Here is the code that outputs incorrect classifications...

Thanks in advance!

//Parameters---------------------------------------------------------------------
svm_parameter param;
param.svm_type = C_SVC;
param.kernel_type = RBF;
param.degree = 3;
param.gamma = 0.5;
param.coef0 = 0;
param.nu = 0.5;
param.cache_size = 100;
param.C = 1;
param.eps = 1e-3;
param.p = 0.1;
param.shrinking = 1;
param.probability = 0;
param.nr_weight = 0;
param.weight_label = NULL;
param.weight = NULL;


//Problem definition-------------------------------------------------------------
svm_problem prob;

//Length, 4 examples
prob.l = 4;

//x values matrix of xor values
QVector< QVector<double> >matrix;
QVector<double>row(2);

row[0] = 1;row[1] = 1;
matrix.push_back(row);
row[0] = 1;row[1] = 0;
matrix.push_back(row);
row[0] = 0;row[1] = 1;
matrix.push_back(row);
row[0] = 0;row[1] = 0;
matrix.push_back(row);

//This part i have trouble understanding
svm_node* x_space = new svm_node[3];
svm_node** x = new svm_node *[prob.l];

//Trying to assign from matrix to svm_node training examples
for (int row = 0;row < matrix.size(); row++){
    for (int col = 0;col < 2;col++){
        x_space[col].index = col;
        x_space[col].value = matrix[row][col];
    }
    x_space[2].index = -1;      //Each row of properties should be terminated with a -1 according to the readme
    x[row] = x_space;
}

prob.x = x;

//yvalues
prob.y = new double[prob.l];
prob.y[0] = -1;
prob.y[1] = 1;
prob.y[2] = 1;
prob.y[3] = -1;

//Train model---------------------------------------------------------------------
svm_model *model = svm_train(&prob,&param);


//Test model----------------------------------------------------------------------
svm_node* testnode = new svm_node[3];
testnode[0].index = 0;
testnode[0].value = 1;
testnode[1].index = 1;
testnode[1].value = 0;
testnode[2].index = -1;

//Should return 1 but returns -1
double retval = svm_predict(model,testnode);
qDebug()<<retval;

回答1:

It seems you've been trying to get this example to work for weeks. I followed the style in svm-train.c that comes with libsvm. I used your values for C and gamma. It is working. I tried all points in the XOR example and it is giving correct results.

The summary of the problem you're having is that you're not allocating space for the 4 data points you train with, so you just over-write the data. This is a typical mistake with pointers in C. It may help you brushed up on pointers in C/C++.

Here's the code:

#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <ctype.h>
#include <errno.h>
#include "svm.h"
#define Malloc(type,n) (type *)malloc((n)*sizeof(type))

struct svm_parameter param;     // set by parse_command_line
struct svm_problem prob;        // set by read_problem
struct svm_model *model;
struct svm_node *x_space;

int main(int argc, char **argv)
{
    char input_file_name[1024];
    char model_file_name[1024];
    const char *error_msg;

    param.svm_type = C_SVC;
    param.kernel_type = RBF;
    param.degree = 3;
    param.gamma = 0.5;
    param.coef0 = 0;
    param.nu = 0.5;
    param.cache_size = 100;
    param.C = 1;
    param.eps = 1e-3;
    param.p = 0.1;
    param.shrinking = 1;
    param.probability = 0;
    param.nr_weight = 0;
    param.weight_label = NULL;
    param.weight = NULL;


    //Problem definition-------------------------------------------------------------
    prob.l = 4;

    //x values matrix of xor values
    double matrix[prob.l][2];
    matrix[0][0] = 1;
    matrix[0][1] = 1;

    matrix[1][0] = 1;
    matrix[1][1] = 0;

    matrix[2][0] = 0;
    matrix[2][1] = 1;

    matrix[3][0] = 0;
    matrix[3][1] = 0;


    //This part i have trouble understanding
    svm_node** x = Malloc(svm_node*,prob.l);

    //Trying to assign from matrix to svm_node training examples
    for (int row = 0;row <prob.l; row++){
        svm_node* x_space = Malloc(svm_node,3);
        for (int col = 0;col < 2;col++){
            x_space[col].index = col;
            x_space[col].value = matrix[row][col];
        }
        x_space[2].index = -1;      //Each row of properties should be terminated with a -1 according to the readme
        x[row] = x_space;
    }

    prob.x = x;

    //yvalues
    prob.y = Malloc(double,prob.l);
    prob.y[0] = -1;
    prob.y[1] = 1;
    prob.y[2] = 1;
    prob.y[3] = -1;

    //Train model---------------------------------------------------------------------
    svm_model *model = svm_train(&prob,&param);


    //Test model----------------------------------------------------------------------
    svm_node* testnode = Malloc(svm_node,3);
    testnode[0].index = 0;
    testnode[0].value = 1;
    testnode[1].index = 1;
    testnode[1].value = 0;
    testnode[2].index = -1;

    //This works correctly:
    double retval = svm_predict(model,testnode);
    printf("retval: %f\n",retval);


    svm_destroy_param(&param);
    free(prob.y);
    free(prob.x);
    free(x_space);

    return 0;
}

来源：https://stackoverflow.com/questions/21293332/how-to-create-training-data-for-libsvm-as-an-svm-node-struct