LIBSVM is in C++ and it returns probability estimates if you specify the -b option.

SVMs do not directly return probability estimates, but there are some turn arounds like using a logistic fitting. There are also SVMs scores which can be estimated using the distance of the example to the hyperplane. As Anjith commented both options are implemented in ScikitLearn in Python (check link below).

http://scikit-learn.org/stable/modules/svm.html#scores-probabilities

@Juan Carlos Gomez  Thanks but I need a C++ implementation since I am implementing my work there. Any ideas?

LIBSVM is in C++ and it returns probability estimates if you specify the -b option.

@Jafet Morales. Do you have any sample code for classification? I basically wrote everything using dlib and for sure LIBSVM will be new to me. An example code would really help. Thanks.

This is the webpage of LibSVM which provide you with the compilied version of the library and the source code in C++:

https://www.csie.ntu.edu.tw/~cjlin/libsvm/

The library is very straight to use. You first train a model using a dataset

svm-train training_set

This creates a training_set.model

and then you use the model for predictions in the test dataset (svmguide3.t):

svm-predict test_set training_set.model test_set.predict

This creates the output file test_set.predict with the results

The data in training_set and test_set must be in LibSVM format:

: : : …

...

where each line is an example, with:

is the class of the example that has to be a number

: is a feature value. is the index of the feature, and has to be an integer starting from 1 and is a real number indicating the value for that feature.

Example:

1 1:-1.382400 2:0.062875 3:0.622518 4:1.612060

1 1:0.651922 2:2.296340 3:2.296340 4:1.649750

2 1:0.654087 2:1.413430 3:0.225929 4:-0.067186

Check the source code of svm-predict.c and svm-train.c to see how to integrete them in your code (in case you don't want to use the standalone library).

You can use several options when training the model: -b 1 allows you to compute probabilities (just note that the probabilities estimated some times are inconsistent with the classification output).

@Juan Carlos Gomez  I cant still crack the data formatting for training and testing I set!

Hi. Sorry I do not understand what you mean by "crack the data formatting". I recommend first to read all the documentation in the LibSVM webpage.

@Juan Carlos Gomez I meant data inputs to the SVM problem. Below is a sample of my code:

//====SVM in LIBSVM

const int rows = mImage.rows;

const int cols = mImage.cols;

size_t nNodes = rows * cols;

const int nFeatures = 3;//mImage.channels();

const long nFolds = 3;

struct svm_parameter param;

param.svm_type = 0; //C-SVC support vector classification

param.kernel_type = 2; //RBF

param.C = 10; //Cost Penalty parameter

param.gamma = 0.2;

param.multiclass_type = 0; //one against one

param.probability = 0;

param.eps = 0.00001; //termination criterion (default 0.001)

struct svm_problem prob;

size_t nCount = 0;

#pragma omp parallel for //private(mTrainSet)

for(int iii = 0; iii < rows; iii++){

for(int jjj =0; jjj < cols; jjj++){

if(mTrainSet.at(iii, jjj) < nStates) nCount+=1;

}

}

prob.l = nCount; //Number of labels

mTrainSet.convertTo(mTrainSet, CV_64FC1); //Convert training set tp double

prob.y = new double[prob.l]; //label storage equal number of labels

svm_node *x_space = new svm_node[nFeatures + 1]; //data

prob.x = new svm_node *[prob.l];

mImage.convertTo(mImage, CV_MAKE_TYPE(CV_64F,nFeatures));

size_t nodeIndex = 0;

#pragma omp parallel for //private(mTrainSet)

for(int iii = 0; iii < rows; iii++){

double *pdouble = mImage.ptr(iii);

const double *pGT = mTrainSet.ptr(iii);

for(int jjj =0; jjj < cols; jjj++){

double nGT = pGT[jjj]; //Label

if(nGT < nStates){

nodeIndex++;

prob.y[nodeIndex] = nGT; //Node label

for(int kkk = 1; kkk < nFeatures+1; kkk++){

double dPixel = pdouble[jjj * nFeatures + kkk];

x_space[kkk].index = kkk;

x_space[kkk].value = dPixel;

}

x_space[nFeatures+1].index = -1; //-1 indicates the end of one vector

prob.x[nodeIndex] = x_space;

}

}

}

svm_model *model = svm_train(&prob, ¶m);

In my work, I use multi-class SVM by constructing one binary SVM classifier for each class. To get the probability estimates you just add -b 1 to the SVM parameters, here is my code in Matlab

load TrainSet;

load TrainClass;

load TestSet;

load TestClass;

numLabels=;

numTest=;

%# train one-against-all models

model = cell(numLabels,1);

matlabpool('local',4);

parfor k=1:numLabels

model{k} = svmtrain(double(TrainClass==k), TrainSet, '-t 0 -c 1 -g 0.2 -b 1 -q');

end

%# get probability estimates of test instances using each model

prob = zeros(numTest,numLabels);

parfor k=1:numLabels

[~,~,p] = svmpredict(double(TestClass==k), TestSet, model{k}, '-b 1');

prob(:,k) = p(:,model{k}.Label==1); %# probability of class==k

end

matlabpool('close');

%# predict the class with the highest probability

[~,pred] = max(prob,[],2);

acc = (sum(pred == TestClass) ./ numel(TestClass))*100 %# accuracy

% C = confusionmat(TestClass, pred) %# confusion matrix

I agree with Jafet Morales answer. I have already used this option.

@Krupa Parikh Do you have example code for training and estimation of probabilities? Kindly please share with me .