private int[] compute(double[][] points)
{
// first, select initial points
double[][] seeds = createSeeds(points, 2 * Bandwidth);
var maxcandidates = new ConcurrentStack <double[]>();
// construct map of the data
tree = KDTree.FromData <int>(points, distance);
// now, for each initial point
if (UseParallelProcessing)
{
Parallel.For(0, seeds.Length, (index) =>
iterate(seeds, maxcandidates, index));
}
else
{
for (int index = 0; index < seeds.Length; index++)
{
iterate(seeds, maxcandidates, index);
}
}
// suppress non-maximum points
double[][] maximum = cut ? maxcandidates.ToArray() : supress(seeds);
// create a decision map using seeds
int[] seedLabels = classifySeeds(seeds, maximum);
tree = KDTree.FromData(seeds, seedLabels, distance);
// create the cluster structure
clusters = new MeanShiftClusterCollection(tree, maximum);
// label each point
return(clusters.Nearest(points));
}
/// <summary>
/// Initializes a new instance of the <see cref="MeanShiftCluster"/> class.
/// </summary>
///
/// <param name="owner">The owner.</param>
/// <param name="index">The cluster index.</param>
///
public MeanShiftCluster(MeanShiftClusterCollection owner, int index)
{
this.owner = owner;
this.index = index;
}
/// <summary>
/// Learns a model that can map the given inputs to the desired outputs.
/// </summary>
/// <param name="x">The model inputs.</param>
/// <param name="weights">The weight of importance for each input sample.</param>
/// <returns>A model that has learned how to produce suitable outputs
/// given the input data <paramref name="x" />.</returns>
public MeanShiftClusterCollection Learn(double[][] x, int[] weights = null)
{
if (weights == null)
{
weights = Vector.Ones <int>(x.Length);
}
if (x.Length != weights.Length)
{
throw new DimensionMismatchException("weights",
"The weights and points vector must have the same dimension.");
}
// First of all, construct map of the original points. We will
// be saving the weight of every point in the node of the tree.
KDTree <int> tree = KDTree.FromData(x, weights, Distance);
// Let's sample some points in the problem surface
double[][] seeds = createSeeds(x, 2 * Bandwidth);
// Now, we will duplicate those points and make them "move"
// into this surface in the direction of the surface modes.
double[][] current = seeds.MemberwiseClone();
// We will store any modes that we find here
var maxima = new ConcurrentStack <double[]>();
// Optimization for uniform kernel
Action <ICollection <NodeDistance <KDTreeNode <int> > >, double[]> func;
if (kernel is UniformKernel)
{
func = uniform;
}
else
{
func = general;
}
// For each seed
if (ParallelOptions.MaxDegreeOfParallelism != 1)
{
Parallel.For(0, current.Length, ParallelOptions, i =>
move(tree, current, i, maxima, func));
for (int i = 0; i < current.Length; i++)
{
supress(current, i, maxima);
}
}
else
{
for (int i = 0; i < current.Length; i++)
{
move(tree, current, i, maxima, func);
}
}
var modes = maxima.ToArray();
// At this point, the current points have moved into
// the location of the modes of the surface. Now we
// have to backtrack and check, for each mode, from
// where those points departed from.
int[] labels = classify(modes: modes, points: current);
// Now we create a decision map using the original seed positions
tree = KDTree.FromData(seeds, labels, Distance, inPlace: true);
clusters = new MeanShiftClusterCollection(this, modes.Length, tree, modes);
if (ComputeLabels || ComputeProportions)
{
int sum = 0;
int[] counts = new int[modes.Length];
labels = new int[x.Length];
for (int i = 0; i < labels.Length; i++)
{
int j = tree.Nearest(x[i]).Value;
labels[i] = j;
counts[j] += weights[i];
sum += weights[i];
}
for (int i = 0; i < counts.Length; i++)
{
clusters.Proportions[i] = counts[i] / (double)sum;
}
}
return(clusters);
}
/// <summary>
/// Divides the input data into clusters.
/// </summary>
///
/// <param name="points">The data where to compute the algorithm.</param>
/// <param name="threshold">The relative convergence threshold
/// for the algorithm. Default is 1e-3.</param>
/// <param name="maxIterations">The maximum number of iterations. Default is 100.</param>
///
public int[] Compute(double[][] points, double threshold, int maxIterations = 100)
{
// first, select initial points
double[][] seeds = createSeeds(points, 2 * Bandwidth);
var maxcandidates = new ConcurrentStack <double[]>();
// construct map of the data
tree = KDTree.FromData <int>(points, distance);
// now, for each initial point
global::Accord.Threading.Tasks.Parallel.For(0, seeds.Length,
#if DEBUG
new ParallelOptions()
{
MaxDegreeOfParallelism = 1
},
#endif
(index) =>
{
double[] point = seeds[index];
double[] mean = new double[point.Length];
double[] delta = new double[point.Length];
// we will keep moving it in the
// direction of the density modes
int iterations = 0;
// until convergence or max iterations reached
while (iterations < maxIterations)
{
iterations++;
// compute the shifted mean
computeMeanShift(point, mean);
// extract the mean shift vector
for (int j = 0; j < mean.Length; j++)
{
delta[j] = point[j] - mean[j];
}
// update the point towards a mode
for (int j = 0; j < mean.Length; j++)
{
point[j] = mean[j];
}
// Check if we are already near any maximum point
if (cut && nearest(point, maxcandidates) != null)
{
break;
}
// check for convergence: magnitude of the mean shift
// vector converges to zero (Comaniciu 2002, page 606)
if (Norm.Euclidean(delta) < threshold * Bandwidth)
{
break;
}
}
if (cut)
{
double[] match = nearest(point, maxcandidates);
if (match != null)
{
seeds[index] = match;
}
else
{
maxcandidates.Push(point);
}
}
});
// suppress non-maximum points
double[][] maximum = cut ? maxcandidates.ToArray() : supress(seeds);
// create a decision map using seeds
int[] seedLabels = classifySeeds(seeds, maximum);
tree = KDTree.FromData(seeds, seedLabels, distance);
// create the cluster structure
clusters = new MeanShiftClusterCollection(tree, maximum);
// label each point
return(clusters.Nearest(points));
}
