void KMeansTrainer::initializeMeans(math::ConstMatrixReference<double, math::MatrixLayout::columnMajor> X)
{
size_t N = X.NumColumns();
size_t K = _numClusters;
size_t choice = rand() % N;
_means.GetColumn(0).CopyFrom(X.GetColumn(choice));
math::ColumnVector<double> minimumDistance(X.NumColumns());
for (int k = 1; k < _numClusters; ++k)
{
// distance to previously selected mean
auto D = pairwiseDistance(X, _means.GetSubMatrix(0, k - 1, _means.NumRows(), 1));
auto distanceToPreviousMean = D.GetColumn(0);
if (k == 1)
{
minimumDistance.CopyFrom(distanceToPreviousMean);
}
else
{
// distance to closest center
for (int i = 0; i < minimumDistance.Size(); ++i)
minimumDistance[i] = std::min(minimumDistance[i], distanceToPreviousMean[i]);
}
choice = weightedSample(minimumDistance);
_means.GetColumn(k).CopyFrom(X.GetColumn(choice));
}
}
double KMeansTrainer::assignClosestCenter(math::ConstMatrixReference<double, math::MatrixLayout::columnMajor> X, math::VectorReference<size_t, math::VectorOrientation::column> clusterAssignment)
{
auto D = pairwiseDistance(X, _means);
double totalDist = 0;
for (int i = 0; i < D.NumRows(); ++i)
{
auto dist = D.GetRow(i);
auto minElement = std::min_element(dist.GetDataPointer(), dist.GetDataPointer() + dist.Size());
clusterAssignment[i] = minElement - dist.GetDataPointer();
totalDist += *minElement;
}
return totalDist;
}
Structured( const RMatrixXf & lbl, const VectorXf & weight, int n_struc_samples ): Split( project1D(pairwiseDistance(lbl,n_struc_samples),&rep_label_), weight ) {
}
