/// <summary>
/// Adds two vectors of the same size together.
/// </summary>
/// <param name="lhs">The left hand vector</param>
/// <param name="rhs">The right hand vector</param>
/// <returns>The sum of the two vectors</returns>
public static LabelVector operator +(LabelVector lhs, LabelVector rhs)
{
LabelVector result = lhs.Clone() as LabelVector;
lhs.DoAdd(rhs, result);
return(result);
}
/// <summary>
/// Multiplies all of the values of the vector by a scalar value.
/// </summary>
/// <param name="x">The vector</param>
/// <param name="scalar">The scalar value</param>
/// <returns>The product of a vector and a scalar</returns>
public static LabelVector operator *(LabelVector x, double scalar)
{
LabelVector result = x.Clone() as LabelVector;
x.DoMultiply(scalar, result);
return(result);
}
/// <summary>
/// Subtracts two vectors of the same size together.
/// </summary>
/// <param name="lhs">The left hand vector</param>
/// <param name="rhs">The right hand vector</param>
/// <returns>The difference of the two vectors</returns>
public static LabelVector operator -(LabelVector lhs, LabelVector rhs)
{
LabelVector result = lhs.Clone() as LabelVector;
lhs.DoSubtract(rhs, result);
return(result);
}
/// <summary>
/// Returns an exact copy of the data point in its current state.
/// </summary>
/// <returns>A copy of the point</returns>
public new object Clone()
{
double[] values = new double[Count];
Array.Copy(Values, values, Count);
LabelVector result = new LabelVector(values);
result.Label = Label;
result.FeatureValue = FeatureValue;
result.Weight = Weight;
return(result);
}
private void calculatePhi(object arg)
{
List <LabelVector> centers = _centers.Select(o => (LabelVector)o.Clone()).ToList();
Job job = arg as Job;
for (int i = job.StartIndex; i < job.EndIndex; i++)
{
LabelVector x = _data[i];
LabelVector center = findNearest(x, centers);
LabelVector dx = x - center;
x.FeatureValue = (float)dx.DotProduct(dx);
}
}
private LabelVector findNearest(LabelVector x, List <LabelVector> centers)
{
LabelVector nearest = centers[0];
double minDist = x.SquaredDistance(nearest);
for (int i = 1; i < centers.Count; i++)
{
double test = x.SquaredDistance(centers[i], minDist);
if (test < minDist)
{
minDist = test;
nearest = centers[i];
}
}
return(nearest);
}
private bool assignCenters(object arg)
{
List <LabelVector> centers = _centers.Select(o => (LabelVector)o.Clone()).ToList();
bool change = false;
Job job = arg as Job;
for (int i = job.StartIndex; i < job.EndIndex; i++)
{
LabelVector x = _data[i];
LabelVector center = findNearest(x, centers);
if (x.Label != center.Label)
{
x.Label = center.Label;
change = true;
}
}
return(change);
}
private List <LabelVector> findCenterCandidates(object arg)
{
Job job = arg as Job;
List <LabelVector> result = new List <LabelVector>();
double norm = OversampleRate / _phi;
for (int i = job.StartIndex; i < job.EndIndex; i++)
{
LabelVector x = _data[i];
double sample = x.FeatureValue * norm;
if (ThreadsafeRandom.Test(sample))
{
LabelVector center = x.Clone() as LabelVector;
result.Add(center);
}
}
return(result);
}
private void seedCenters()
{
UpdateManager.WriteLine("Performing K-Means++ initialization...");
LabelVector center = _data.SelectRandom().Clone() as LabelVector;
center.Label = 0;
_centers.Add(center);
Stopwatch sw = new Stopwatch();
for (int i = 1; i < _k; i++)
{
sw.Reset();
UpdateManager.Write("Recalculating phi...");
sw.Start();
runJobs(o => calculatePhi(o));
sw.Stop();
UpdateManager.WriteLine("Done [{0}ms]", sw.ElapsedMilliseconds);
float[] dist = _data.Select(o => o.FeatureValue).ToArray().Normalize();
LabelVector x = _data[dist.Sample()];
_centers.Add(x.Clone() as LabelVector);
}
}
/// <summary>
/// Fit the centers to the provided data.
/// </summary>
/// <param name="data">The data to fit</param>
/// <param name="weights">Weights for the data. If not provided, all points are equally weighted.</param>
public void Fit(List <LabelVector> data, float[] weights = null)
{
UpdateManager.WriteLine("Fitting centers using k-means clustering...");
_data = data;
_minDist = new float[data.Count];
_jobs = new List <Job>();
if (data.Count > JobSize)
{
int jobCount = _data.Count / JobSize;
if (_data.Count % JobSize != 0)
{
jobCount++;
}
for (int i = 0; i < _data.Count; i += JobSize)
{
_jobs.Add(new Job {
StartIndex = i, EndIndex = Math.Min(i + JobSize, _data.Count)
});
}
}
else
{
_jobs.Add(new Job {
StartIndex = 0, EndIndex = data.Count
});
}
UpdateManager.WriteLine("Data partitioned into {0} job{1}.", _jobs.Count, _jobs.Count == 1 ? "" : "s");
_centers = new List <LabelVector>();
UpdateManager.WriteLine("Initializing centers...");
UpdateManager.AddIndent();
if (data.Count > PARALLEL_THRESHOLD)
{
seedCentersParallel();
}
else
{
seedCenters();
}
UpdateManager.RemoveIndent();
for (short label = 0; label < _centers.Count; label++)
{
_centers[label].Label = label;
}
UpdateManager.WriteLine("Running Lloyd's algorithm...");
Stopwatch sw = new Stopwatch();
UpdateManager.Write("Building initial clusters...");
sw.Start();
while (runJobs(o => assignCenters(o)))
{
sw.Stop();
UpdateManager.WriteLine("Done [{0}ms]", sw.ElapsedMilliseconds);
for (int i = 0; i < _centers.Count; i++)
{
_centers[i] = new LabelVector(_data[0].Count);
}
sw.Reset();
UpdateManager.Write("Recomputing cluster centers...");
sw.Start();
for (int i = 0; i < data.Count; i++)
{
LabelVector x = data[i];
if (weights != null)
{
float featureValue = _centers[x.Label].FeatureValue;
_centers[x.Label] += weights[i] * x;
_centers[x.Label].FeatureValue += weights[i];
}
else
{
float featureValue = _centers[x.Label].FeatureValue;
_centers[x.Label] += x;
_centers[x.Label].FeatureValue += 1;
}
}
for (short i = 0; i < _centers.Count; i++)
{
_centers[i] /= _centers[i].FeatureValue;
_centers[i].Label = i;
}
sw.Stop();
UpdateManager.WriteLine("Done [{0}ms]", sw.ElapsedMilliseconds);
sw.Reset();
UpdateManager.Write("Rebuilding clusters...");
sw.Start();
}
sw.Stop();
UpdateManager.WriteLine("Done and done [{0}ms]", sw.ElapsedMilliseconds);
UpdateManager.WriteLine("Fitting complete.");
}
private void seedCentersParallel()
{
UpdateManager.WriteLine("Performing scalable K-Means++ initialization...");
LabelVector center = _data.SelectRandom().Clone() as LabelVector;
center.Label = 0;
_centers.Add(center);
runJobs(o => calculatePhi(o));
_phi = _data.Sum(o => o.FeatureValue);
int N = (int)Math.Log(_phi);
if (MaxIterations != 0)
{
N = MaxIterations;
}
Stopwatch sw = new Stopwatch();
for (int i = 0; i < N; i++)
{
sw.Reset();
UpdateManager.Write("Gathering center candidates ({0} of {1})...", i + 1, N);
sw.Start();
List <LabelVector> newCenters = runJobs(new Func <object, List <LabelVector> >(findCenterCandidates));
_centers.AddRange(newCenters);
for (short s = 0; s < _centers.Count; s++)
{
_centers[s].Label = s;
}
sw.Stop();
UpdateManager.WriteLine("Done [{0}ms]", sw.ElapsedMilliseconds);
sw.Reset();
UpdateManager.Write("Recalculating phi...");
sw.Start();
runJobs(o => calculatePhi(o));
_phi = _data.Sum(o => o.FeatureValue);
sw.Stop();
UpdateManager.WriteLine("Done [{0}ms]", sw.ElapsedMilliseconds);
}
sw.Reset();
UpdateManager.Write("Assigning centers...");
sw.Start();
for (short s = 0; s < _centers.Count; s++)
{
_centers[s].Label = s;
}
runJobs(o => assignCenters(o));
sw.Stop();
UpdateManager.WriteLine("Done [{0}ms]", sw.ElapsedMilliseconds);
float[] weights = new float[_centers.Count];
foreach (LabelVector x in _data)
{
weights[x.Label] += 1;
}
KMeans clusterCenters = new KMeans(_k);
UpdateManager.AddIndent();
clusterCenters.Fit(_centers, weights);
UpdateManager.RemoveIndent();
_centers = clusterCenters.Centers;
}
