private static IEnumerable <Assignment> AngularCluster(IList <Matrix <double> > features, ulong numberClusters)
{
if (features == null)
{
throw new ArgumentNullException(nameof(features));
}
var size = features.Count;
if (size == 0)
{
throw new ArgumentException("The dataset can't be empty", nameof(features));
}
var featureSize = features[0].Size;
for (var index = 0; index < size; ++index)
{
if (features[index].Size != featureSize)
{
throw new ArgumentException("All feature vectors must have the same length.", nameof(features));
}
}
// find the centroid of feats
Matrix <double> tmp;
var m = Matrix <double> .CreateTemplateParameterizeMatrix(0, 1);
for (var index = 0; index < size; ++index)
{
tmp = m + features[index];
m.Dispose();
m = tmp;
}
tmp = m;
m = tmp / size;
tmp.Dispose();
// Now center feats and then project onto the unit sphere. The reason for projecting
// onto the unit sphere is so pick_initial_centers() works in a sensible way.
for (var index = 0; index < size; ++index)
{
tmp = features[index] - m;
features[index].Dispose();
features[index] = tmp;
var length = Dlib.Length(features[index]);
if (Math.Abs(length) > double.Epsilon)
{
tmp = features[index] / length;
features[index].Dispose();
features[index] = tmp;
}
}
// now do angular clustering of the points
var linearKernel = new LinearKernel <Matrix <double> >(0, 1);
var tempCenters = Dlib.PickInitialCenters((int)numberClusters, features, linearKernel, 0.05).ToArray();
var centers = Dlib.FindClustersUsingAngularKMeans(features, tempCenters).ToArray();
foreach (var center in tempCenters)
{
center.Dispose();
}
linearKernel.Dispose();
// and then report the resulting assignments
var assignments = new List <Assignment>(size);
for (var index = 0; index < size; ++index)
{
var temp = new Assignment();
temp.C = Dlib.NearestCenter(centers, features[index]);
using (var temp2 = features[index] - centers[temp.C])
temp.Distance = Dlib.Length(temp2);
temp.Index = (ulong)index;
assignments.Add(temp);
}
return(assignments);
}
