CWeightedDegreeStringKernel::CWeightedDegreeStringKernel(SGVector<float64_t> w)
: CStringKernel<char>(10)
{
init();
type=E_EXTERNAL;
degree=w.vlen;
weights=SG_MALLOC(float64_t, degree*(1+max_mismatch));
weights_degree=degree;
weights_length=(1+max_mismatch);
for (int32_t i=0; i<degree*(1+max_mismatch); i++)
weights[i]=w.vector[i];
w.free_vector();
}
bool CGaussianNaiveBayes::train(CFeatures* data)
{
// init features with data if necessary and assure type is correct
if (data)
{
if (!data->has_property(FP_DOT))
SG_ERROR("Specified features are not of type CDotFeatures\n");
set_features((CDotFeatures*) data);
}
// get int labels to train_labels and check length equality
ASSERT(m_labels);
SGVector<int32_t> train_labels = m_labels->get_int_labels();
ASSERT(m_features->get_num_vectors()==train_labels.vlen);
// init min_label, max_label and loop variables
int32_t min_label = train_labels.vector[0];
int32_t max_label = train_labels.vector[0];
int i,j;
// find minimal and maximal label
for (i=1; i<train_labels.vlen; i++)
{
min_label = CMath::min(min_label, train_labels.vector[i]);
max_label = CMath::max(max_label, train_labels.vector[i]);
}
// subtract minimal label from all labels
for (i=0; i<train_labels.vlen; i++)
train_labels.vector[i]-= min_label;
// get number of classes, minimal label and dimensionality
m_num_classes = max_label-min_label+1;
m_min_label = min_label;
m_dim = m_features->get_dim_feature_space();
// allocate memory for distributions' parameters and a priori probability
m_means.matrix = SG_MALLOC(float64_t, m_num_classes*m_dim);
m_means.num_rows = m_dim;
m_means.num_cols = m_num_classes;
m_variances.matrix = SG_MALLOC(float64_t, m_num_classes*m_dim);
m_variances.num_rows = m_dim;
m_variances.num_cols = m_num_classes;
m_label_prob.vector = SG_MALLOC(float64_t, m_num_classes);
m_label_prob.vlen = m_num_classes;
// allocate memory for label rates
m_rates.vector = SG_MALLOC(float64_t, m_num_classes);
m_rates.vlen = m_num_classes;
// assure that memory is allocated
ASSERT(m_means.matrix);
ASSERT(m_variances.matrix);
ASSERT(m_rates.vector);
ASSERT(m_label_prob.vector);
// make arrays filled by zeros before using
m_means.zero();
m_variances.zero();
m_label_prob.zero();
m_rates.zero();
// number of iterations in all cycles
int32_t max_progress = 2 * train_labels.vlen + 2 * m_num_classes;
// current progress
int32_t progress = 0;
SG_PROGRESS(progress, 0, max_progress);
// get sum of features among labels
for (i=0; i<train_labels.vlen; i++)
{
SGVector<float64_t> fea = m_features->get_computed_dot_feature_vector(i);
for (j=0; j<m_dim; j++)
m_means(j, train_labels.vector[i]) += fea.vector[j];
fea.free_vector();
m_label_prob.vector[train_labels.vector[i]]+=1.0;
progress++;
SG_PROGRESS(progress, 0, max_progress);
}
// get means of features of labels
for (i=0; i<m_num_classes; i++)
{
for (j=0; j<m_dim; j++)
m_means(j, i) /= m_label_prob.vector[i];
progress++;
SG_PROGRESS(progress, 0, max_progress);
}
// compute squared residuals with means available
for (i=0; i<train_labels.vlen; i++)
{
SGVector<float64_t> fea = m_features->get_computed_dot_feature_vector(i);
for (j=0; j<m_dim; j++)
{
m_variances(j, train_labels.vector[i]) +=
CMath::sq(fea[j]-m_means(j, train_labels.vector[i]));
}
fea.free_vector();
progress++;
SG_PROGRESS(progress, 0, max_progress);
}
// get variance of features of labels
for (i=0; i<m_num_classes; i++)
{
for (j=0; j<m_dim; j++)
m_variances(j, i) /= m_label_prob.vector[i] > 1 ? m_label_prob.vector[i]-1 : 1;
// get a priori probabilities of labels
m_label_prob.vector[i]/= m_num_classes;
progress++;
SG_PROGRESS(progress, 0, max_progress);
}
SG_DONE();
train_labels.free_vector();
return true;
}
