void LinearSVM::distribution(InputSet& inputSet,
unsigned int sample,
scalar_t* distr) const {
// Push the selected features if required
if(!indices_.empty()) {
inputSet.pushFeatures(indices_);
}
// Get the number of features and labels
const unsigned int nbFeatures = inputSet.nbFeatures();
const unsigned int nbLabels = inputSet.nbLabels();
// Make sure that we have a model
assert(model_);
// Make sure that there is the same number of labels
assert(static_cast<unsigned int>(get_nr_class(model_)) <= nbLabels);
// Create a node
feature_node node;
node.dim = nbFeatures;
node.values = const_cast<scalar_t*>(inputSet.features(sample));
// The predicted labels
std::vector<double> predictions(nbLabels);
predict_values(model_, &node, &predictions[0]);
// Update the ditribution according to the predictions
for(unsigned int l = 0; l < nbLabels; ++l) {
distr[map_[l]] = predictions[l];
}
// Pop the selected features if required
if(!indices_.empty()) {
inputSet.popFeatures();
}
}
void LinearSVM::train(InputSet& inputSet) {
// Sample features from every heuristic in order for the matrix of features
// to fit in memory
if(inputSet.nbFeatures() > NB_FEATURES_MAX / inputSet.nbSamples()) {
inputSet.sampleFeatures(NB_FEATURES_MAX / inputSet.nbSamples(), indices_);
inputSet.pushFeatures(indices_);
}
else {
indices_.clear();
}
// Get the number of features, samples and labels
const unsigned int nbFeatures = inputSet.nbFeatures();
const unsigned int nbSamples = inputSet.nbSamples();
const unsigned int nbLabels = inputSet.nbLabels();
// Delete the previous model
if(model_) {
free_and_destroy_model(&model_);
}
// Create a new problem
problem prob;
// Recopy the number of samples
prob.l = nbSamples;
prob.n = nbFeatures;
// Recopy the labels
std::vector<int> labels(inputSet.labels(), inputSet.labels() + nbSamples);
prob.y = &labels[0];
// Recopy the features (as we want to normalize them)
std::vector<scalar_t> matrix;
inputSet.swapFeatures(matrix);
// Create samples as expected by liblinear
std::vector<feature_node> samples(nbSamples);
prob.x = &samples[0];
// Compute the mean norm
scalar_t meanNorm = 0;
for(unsigned int s = 0; s < nbSamples; ++s) {
samples[s].dim = nbFeatures;
samples[s].values = &matrix[s * nbFeatures];
// Add the mean norm of that sample
meanNorm += nrm2(samples[s].dim, samples[s].values, 1);
}
// Divide the sum of the norms by the number of samples
meanNorm /= nbSamples;
std::cout << "[LinearSVM::train] mean(norm): " << meanNorm << '.'
<< std::endl;
// Rescale the features so that their mean norm is 1
std::transform(matrix.begin(), matrix.end(), matrix.begin(),
std::bind2nd(std::divides<scalar_t>(), meanNorm));
// Sets the bias to the default value (liblinear doesn't seem to handle the
// bias parameter value correctly)
prob.bias = -1;
// Make sure that the parameters are correct
bool crossValidate = parameters_.C < 0;
// Sets C to a default value in order to pass the parameter check
if(crossValidate) {
parameters_.C = 1;
}
// There is a problem with the parameters
assert(!check_parameter(&prob, ¶meters_));
// If C is below zero, use 5-folds cross-validation to determine it
if(crossValidate) {
std::vector<int> target(nbSamples); // The predicted labels
unsigned int nbErrorsMin = nbSamples + 1; // Initialize past the maximum
for(parameters_.C = 1000; parameters_.C >= 0.01; parameters_.C /= 10) {
cross_validation(&prob, ¶meters_, 5, &target[0]);
// Count the number of errors
unsigned int nbErrors = 0;
for(unsigned int s = 0; s < nbSamples; ++s) {
if(target[s] != labels[s]) {
++nbErrors;
}
}
std::cout << "[LinearSVM::train] 5 folds cross-validation error "
"for C = " << parameters_.C << ": "
<< nbErrors * 100.0f / nbSamples << "%." << std::endl;
// The new C is better than the previous one
if(nbErrors < nbErrorsMin) {
nbErrorsMin = nbErrors;
}
// The optimal C was found
else {
break;
}
}
// C got divided one time too much
parameters_.C *= 10;
// Print C to the log
std::cout << "[LinearSVM::train] optimal C as determined by 5 folds "
"cross-validation: " << parameters_.C << '.' << std::endl;
}
// Train the svm
model_ = ::train(&prob, ¶meters_);
assert(model_);
// Reset C so that it will be cross-validated again
if(crossValidate) {
parameters_.C = -1;
}
// Save libsvm labels
map_.clear();
map_.resize(nbLabels);
get_labels(model_, &map_[0]);
// Pop the selected features if required
if(!indices_.empty()) {
inputSet.popFeatures();
}
}
