//' Simple KMeans Train
//'
//' @export
//'
// [[Rcpp::depends(BH)]]
// [[Rcpp::export]]
List SharkKMeansTrain (NumericMatrix X, ssize_t k) {
// convert data
UnlabeledData<RealVector> data = NumericMatrixToUnlabeledData(X);
std::size_t elements = data.numberOfElements();
// compute centroids using k-means clustering
Centroids centroids;
kMeans (data, k, centroids);
// convert cluster centers/centroids
Data<RealVector> const& c = centroids.centroids();
NumericMatrix cM = DataRealVectorToNumericMatrix (c);
// cluster training data we are given and convert to Rcpp object
HardClusteringModel<RealVector> model(¢roids);
Data<unsigned> clusters = model(data);
NumericVector labels = LabelsToNumericVector (clusters);
// return solution found
Rcpp::List rl = R_NilValue;
rl = Rcpp::List::create(
Rcpp::Named("labels") = labels,
Rcpp::Named("centroids") = cM );
return (rl);
}
//' Simple KMeans Predict
//'
//' @export
//'
// [[Rcpp::depends(BH)]]
// [[Rcpp::export]]
List SharkKMeansPredict (NumericMatrix X, NumericMatrix centroids) {
// convert data
UnlabeledData<RealVector> data = NumericMatrixToUnlabeledData (X);
std::size_t elements = data.numberOfElements();
// convert centroids
Centroids c (NumericMatrixToDataRealVector (centroids));
// cluster data we are given and convert to Rcpp object
HardClusteringModel<RealVector> model (&c);
Data<unsigned> clusters = model (data);
NumericVector labels = LabelsToNumericVector (clusters);
// return solution found
Rcpp::List rl = R_NilValue;
rl = Rcpp::List::create(Rcpp::Named("labels") = labels);
return (rl);
}
void DiscreteLoss::defineBalancedCost(UnlabeledData<unsigned int> const& labels){
std::size_t classes = numberOfClasses(labels);
std::size_t ic = labels.numberOfElements();
std::vector<unsigned int> freq(classes);
BOOST_FOREACH(unsigned int label, labels.elements()){
freq[label]++;
}
m_cost.resize(classes, classes);
for (std::size_t i = 0; i!= classes; i++){
double c = (freq[i] == 0) ? 1.0 : ic / (double)(classes * freq[i]);
for ( std::size_t j = 0; j != classes; j++)
m_cost(i, j) = c;
m_cost(i, i) = 0.0;
}
}
void NormalizeComponentsWhitening::train(ModelType& model, UnlabeledData<RealVector> const& input){
std::size_t dc = dataDimension(input);
SHARK_CHECK(input.numberOfElements() >= dc + 1, "[NormalizeComponentsWhitening::train] input needs to contain more points than there are input dimensions");
SHARK_CHECK(m_targetVariance > 0.0, "[NormalizeComponentsWhitening::train] target variance must be positive");
// dense model with bias having input and output dimension equal to data dimension
model.setStructure(dc, dc, true);
RealVector mean;
RealMatrix covariance;
meanvar(input, mean, covariance);
RealMatrix whiteningMatrix = createWhiteningMatrix(covariance);
whiteningMatrix *= std::sqrt(m_targetVariance);
RealVector offset = -prod(trans(whiteningMatrix),mean);
model.setStructure(RealMatrix(trans(whiteningMatrix)), offset);
}
/// Set the input data, which is stored in the PCA object.
void PCA::setData(UnlabeledData<RealVector> const& inputs) {
SHARK_CHECK(inputs.numberOfElements() >= 2, "[PCA::train] input needs to contain at least two points");
m_l = inputs.numberOfElements(); ///< number of data points
PCAAlgorithm algorithm = m_algorithm;
m_n = dataDimension(inputs);
if(algorithm == AUTO) {
if(m_n > m_l) algorithm = SMALL_SAMPLE; // more attributes than data points
else algorithm = STANDARD;
}
// decompose covariance matrix
if(algorithm == STANDARD) { // standard case
RealMatrix S(m_n,m_n);//covariance matrix
meanvar(inputs,m_mean,S);
m_eigenvalues.resize(m_n);
m_eigenvectors.resize(m_n, m_n);
eigensymm(S, m_eigenvectors, m_eigenvalues);
} else {
//let X0 be the design matrix having all inputs as rows
//we want to avoid to form it directly but us it's batch represntation in the dataset
m_mean = shark::mean(inputs);
RealMatrix S(m_l,m_l,0.0);//S=X0 X0^T
//as every batch is only a fraction of all samples, we have to loop through all
//combinations of the batches of first and second argument and calculate a block of S
//so if X0^T = (B0^T,B1^T)
//than S = B0 B0^T B0 B1^T
// B1 B0^T B1 B1^T
std::size_t start1 = 0;
for(std::size_t b1 = 0; b1 != inputs.numberOfBatches(); ++b1){
std::size_t batchSize1 = inputs.batch(b1).size1();
RealMatrix X1 = inputs.batch(b1)-repeat(m_mean,batchSize1);
std::size_t start2 = 0;
//calculate off-diagonal blocks
//and the block X2 X1^T is the transpose of X1X2^T and thus can bee calculated for free.
for(std::size_t b2 = 0; b2 != b1; ++b2){
std::size_t batchSize2 = inputs.batch(b2).size1();
RealMatrix X2 = inputs.batch(b2)-repeat(m_mean,batchSize2);
RealSubMatrix X1X2T= subrange(S,start1,start1+batchSize1,start2,start2+batchSize2);
RealSubMatrix X2X1T= subrange(S,start2,start2+batchSize2,start1,start1+batchSize1);
noalias(X1X2T) = prod(X1,trans(X2));// X1 X2^T
noalias(X2X1T) = trans(X1X2T);// X2 X1^T
start2+=batchSize2;
}
//diagonal block
RealSubMatrix X1X1T= subrange(S,start1,start1+batchSize1,start1,start1+batchSize1);
symm_prod(X1,X1X1T);
start1+=batchSize1;
}
S /= m_l;
m_eigenvalues.resize(m_l);
m_eigenvectors.resize(m_n,m_l);
m_eigenvectors.clear();
RealMatrix U(m_l, m_l);
eigensymm(S, U, m_eigenvalues);
// compute true eigenvectors
//eigenv=X0^T U
std::size_t batchStart = 0;
for(std::size_t b = 0; b != inputs.numberOfBatches(); ++b){
std::size_t batchSize = inputs.batch(b).size1();
std::size_t batchEnd = batchStart+batchSize;
RealMatrix X = inputs.batch(b)-repeat(m_mean,batchSize);
noalias(m_eigenvectors) += prod(trans(X),rows(U,batchStart,batchEnd));
batchStart = batchEnd;
}
//normalize
for(std::size_t i=0; i != m_l; i++)
column(m_eigenvectors, i) /= norm_2(column(m_eigenvectors, i));
}
}
