libsvm: giving all the test points a label of 0

Vote count:

0

I am using Java to run a minimal example of a two class SVM model. All the training points are in a text file formatted in the recommended way by lbsvm. label index:value....

0 1:0.0 2:1.5336402654647827 3:0.0 4:0.0 5:0.0 6:1.5336402654647827 7:1.5336402654647827 8:0.0 
1 1:0.0 2:0.0 3:0.0 4:1.8472979068756104 5:1.8472979068756104 6:0.0 7:0.0 8:1.8472979068756104 
...

When I use the libsvm via command line (svm-train train.txt), I get the following output results for my model

optimization finished, #iter = 11
nu = 0.021188
obj = -1.409060, rho = 0.034411
nSV = 4, nBSV = 0
Total nSV = 4

The test set has decent results and classifies the points correctly. I used the default parameters as defined in the README.

But when I run it in my Java program, I get the following results for the model training,

optimization finished, #iter = 56
nu = 0.8421052631578947
obj = -112.0, rho = -1.0
nSV = 112, nBSV = 112
Total nSV = 112

I read the same data file in using the following method:

public void readSampleData()
{
    BufferedReader bufReader = null;

    FileReader reader = new FileReader("libsvmTrain.txt");
    bufReader = new BufferedReader(reader);
    String line;
    svmProb = new svm_problem();
    svmProb.l = 133;
    svmProb.x = new svm_node[133][];
    svmProb.y = new double[133];
    try {
        int trainPt = 0;
        while ((line = bufReader.readLine()) != null)
        {
            String[] coords = line.split(" ");
            //First one is label
            svmProb.y[trainPt] = Doubles.tryParse(coords[0]);
            svmProb.x[trainPt] = new svm_node[coords.length-1];

            for (int i = 1; i < coords.length; i++)
            {
                svm_node node = new svm_node();
                String[] point = coords[i].split(":");
                node.index = Integer.parseInt(point[0]);
                node.value = Doubles.tryParse(point[1]);
                svmProb.x[trainPt][i-1] = node;
            }

            trainPt++;
        }
        reader.close();

And below are the default parameters:

svmParam = new svm_parameter();
    // default values
    svmParam.svm_type = svm_parameter.C_SVC;
    svmParam.kernel_type = svm_parameter.RBF;
    svmParam.degree = 3;
    svmParam.gamma = 0; // 1/num_features
    svmParam.coef0 = 0;
    svmParam.nu = 0.5;
    svmParam.cache_size = 100;
    svmParam.C = 1;
    svmParam.eps = 1e-3;
    svmParam.p = 0.1;
    svmParam.shrinking = 1;
    svmParam.probability = 0;
    svmParam.nr_weight = 0;
    svmParam.weight_label = null;
    svmParam.weight = null;

I use svm.svm_train(svmProb, svmParam). Not only does the training result in an output very different but it classifies everything in the test set with a 0 label.

Is there some bug in libsvm regarding the command line version versus the java library?

asked 28 secs ago

nEm

1,576