/// <summary>
/// Select either the Top N or Bottom N items in sorted order from the given collection, serially (not in parallel).
///
/// This only performs a partial sort.
/// </summary>
/// <typeparam name="TElement">Type of element in the collection.</typeparam>
/// <param name="items">Collection of items to sort and select.</param>
/// <param name="topN">If true, find the Top N items in descending order, otherwise the Bottom N items in ascending order.</param>
/// <param name="k">Number of items to select.</param>
/// <param name="comparisonDelegate">If null, assume the items are IComparable and sort them according to their natural ordering.
/// If not null, use this in the comparisons to establish the ordering.</param>
/// <returns>The Top N or Bottom N items, as requested, sorted appropriately</returns>
public static IEnumerable <TElement> SelectSerial <TElement>(this IEnumerable <TElement> items, bool topN, int k,
IComparer <TElement> comparisonDelegate = null)
{
// Seems counterintuitive, but when looking for the Top N we use a Min Heap, and when
// looking for the Bottom N we use a Max Heap.
var heap = new BinaryHeap <TElement>(topN ? BinaryHeapType.MinHeap : BinaryHeapType.MaxHeap, k, comparisonDelegate);
foreach (var item in items)
{
heap.AddRemove(item);
}
var resultsCount = heap.Count;
for (var i = 0; i < resultsCount; i++)
{
yield return(heap.Remove());
}
}
/// <summary>
/// Search many Hilbert orderings of the points, each based on a different permutation of the dimensions, and
/// keep the ones yielding the best Metrics, likely those that estimate the lowest values
/// for the number of clusters.
/// </summary>
/// <param name="points">Points to index.</param>
/// <param name="indexCount">Number of the best indices to return.
/// For example, if this is 10, then the 10 indices with the lowest scores will be kept.</param>
/// <param name="startingPermutation">Starting permutation.</param>
/// <returns>The best indices found and the permutations that generated them.
/// THe first item in the returned list is the best of the best, and the last is the worst of the best.</returns>
public IList <PermutationFound> SearchMany(IReadOnlyList <UnsignedPoint> points, int indexCount, Permutation <uint> startingPermutation = null)
{
if (points.Count() < 10)
{
throw new ArgumentException("List has too few elements", nameof(points));
}
var queue = new BinaryHeap <PermutationFound>(BinaryHeapType.MaxHeap, indexCount);
int dimensions = points[0].Dimensions;
if (startingPermutation == null)
{
startingPermutation = new Permutation <uint>(dimensions);
}
List <UnsignedPoint> firstCurve;
if (ProblemSize(points.Count, dimensions, BitsPerDimension) < 1500000000L)
{
firstCurve = HilbertSort.Sort(points, BitsPerDimension, startingPermutation);
}
else
{
// Used for larger problems.
firstCurve = SmallBucketSort <UnsignedPoint> .Sort(points, point => point.Coordinates.HilbertIndex(BitsPerDimension));
}
// Measure our first index, then loop through random permutations
// looking for a better one, always accumulating the best in results.
var metricResults = Metric(firstCurve);
var bestResults = new PermutationFound(startingPermutation, firstCurve, metricResults.Item1, metricResults.Item2);
LowestCountSeen = Min(LowestCountSeen, bestResults.EstimatedClusterCount);
Logger.Info($"Cluster count Starts at: {bestResults}");
var startingCount = bestResults.EstimatedClusterCount;
if (MaxIterations <= 1)
{
return new List <PermutationFound> {
bestResults
}
}
;
queue.AddRemove(bestResults);
// Decide if we are to sample points or use them all
var sampledPoints = points;
var sampleSize = points.Count();
if (UseSample)
{
sampleSize = SampleSize(points, bestResults.EstimatedClusterCount);
sampledPoints = Sample(points, sampleSize);
Logger.Info($" Sample is {sampleSize} of {points.Count} points");
}
var rejectedSampleSizes = new HashSet <int>();
var iterationsWithoutImprovement = 0;
var parallelOpts = new ParallelOptions {
MaxDegreeOfParallelism = EstimateMaxDegreesOfParallelism(sampledPoints)
};
List <Permutation <uint> > allPermutations = null;
// If the number of dimensions is small, we might waste time trying the same randomly chosen permutations mutiple times.
// Instead, we will try all or many of them in order.
if (dimensions <= 7)
{
allPermutations = Permutation <uint> .AllPermutations(dimensions).ToList();
}
for (var iteration = 0; iteration < MaxIterations; iteration++)
{
var improvedCount = 0;
var startFromPermutation = bestResults.PermutationUsed;
Parallel.For(0, ParallelTrials, parallelOpts,
i =>
{
Permutation <uint> permutationToTry;
// This locking is needed because we use a static random number generator to create a new permutation.
// It is more expensive to make the random number generator threadsafe than to make this loop threadsafe.
if (dimensions > 7)
{
lock (startFromPermutation)
{
permutationToTry = PermutationStrategy(startFromPermutation, dimensions, iteration);
}
}
else
{
lock (allPermutations)
{
if (!allPermutations.Any())
{
return;
}
permutationToTry = allPermutations.Last();
allPermutations.RemoveAt(allPermutations.Count - 1);
}
}
IReadOnlyList <UnsignedPoint> sampledPointsToUse;
lock (points)
{
sampledPointsToUse = sampledPoints;
}
var curveToTry = HilbertSort.Sort(sampledPointsToUse, BitsPerDimension, permutationToTry);
metricResults = Metric(curveToTry);
var resultsToTry = new PermutationFound(permutationToTry, curveToTry, metricResults.Item1, metricResults.Item2);
lock (queue)
{
if (resultsToTry.EstimatedClusterCount < startingCount / 4 &&
UseSample && sampleSize != points.Count())
{
// If the cluster count has improved too much and we are sampled,
// reject it and increase the sample size.
// Why? If the clusters are irregular, sampling can break
// them into so many small pieces that most points end up in outliers.
// This leads to a false low count.
if (!rejectedSampleSizes.Contains(curveToTry.Count))
{
sampleSize = Math.Min(points.Count(), 3 * curveToTry.Count / 2);
Logger.Info($"Increasing sample size to {sampleSize} because estimated K = {resultsToTry.EstimatedClusterCount} (not trusted)");
var newSampledPoints = Sample(points, sampleSize);
lock (points)
{
sampledPoints = newSampledPoints;
}
rejectedSampleSizes.Add(curveToTry.Count);
}
}
else
{
queue.AddRemove(resultsToTry);
var improved = resultsToTry.IsBetterThan(bestResults);
if (improved)
{
bestResults = resultsToTry;
Interlocked.Add(ref improvedCount, 1);
LowestCountSeen = Math.Min(LowestCountSeen, bestResults.EstimatedClusterCount);
Logger.Info($"Cluster count Improved to: {bestResults}");
}
}
}
});
if (improvedCount > 0)
{
iterationsWithoutImprovement = 0;
}
else
{
iterationsWithoutImprovement++;
}
if (iterationsWithoutImprovement >= MaxIterationsWithoutImprovement)
{
break;
}
if (bestResults.EstimatedClusterCount <= 2)
{
break; // No point in continuing!
}
}
var indicesFound = queue.RemoveAll().Reverse().ToList();
if (sampledPoints.Count < points.Count)
{
// Results are based on Sampled set of points. Now we need to recreate these indices using the
// full set of points.
//TODO: "Unsample" the indices.
var unsampledIndices = indicesFound.Select(i => Unsample(points, i)).ToList();
Logger.Info($"Final, unsampled Cluster count: {unsampledIndices[0]}");
return(unsampledIndices);
}
else
{
return(indicesFound);
}
}
/// <summary>
/// Select either the Top N or Bottom N items in sorted order from the given collection, in parallel.
///
/// This only performs a partial sort.
/// </summary>
/// <typeparam name="TElement">Type of element in the collection.</typeparam>
/// <param name="items">Collection of items to sort and select.</param>
/// <param name="topN">If true, find the Top N items in descending order, otherwise the Bottom N items in ascending order.</param>
/// <param name="k">Number of items to select.</param>
/// <param name="comparisonDelegate">If null, assume the items are IComparable and sort them according to their natural ordering.
/// If not null, use this in the comparisons to establish the ordering.</param>
/// <param name="options">If null, use the default values, otherwise use these options to control the parallelism.</param>
/// <returns>The Top N or Bottom N items, as requested, sorted appropriately</returns>
static IEnumerable <TElement> SelectParallel <TElement>(IEnumerable <TElement> items, bool topN, int k,
IComparer <TElement> comparisonDelegate = null, SelectParallelOptions options = null)
{
options = options ?? new SelectParallelOptions();
// If we are only dedicating a single task to the operation, do it serially to save on Task overhead.
if (options.TaskCount == 1)
{
return(SelectSerial(items, topN, k, comparisonDelegate));
}
var tasks = new Task[options.TaskCount];
var extremeItems = new List <TElement>();
var enumerator = items.GetEnumerator();
for (var i = 0; i < options.TaskCount; i++)
{
var iTask = i;
var batch = new TElement[options.BatchSize];
tasks[iTask] = Task.Factory.StartNew(() =>
{
var heap = new BinaryHeap <TElement>(topN ? BinaryHeapType.MinHeap : BinaryHeapType.MaxHeap, k + 1, comparisonDelegate);
var moreItems = true;
var batchSize = options.BatchSize;
while (moreItems)
{
var iReadCount = 0;
lock (enumerator)
{
for (var iBatch = 0; iBatch < batchSize && moreItems; iBatch++)
{
if (enumerator.MoveNext())
{
batch[iReadCount++] = enumerator.Current;
}
else
{
moreItems = false;
}
}
}
for (var iBatch = 0; iBatch < iReadCount; iBatch++)
{
var item = batch[iBatch];
if (k + 1 > heap.Count)
{
heap.Add(item);
}
else if (heap.IsLessExtreme(item))
{
heap.Remove();
heap.Add(item);
}
}
}
lock (extremeItems)
{
extremeItems.AddRange(heap.RemoveAll());
}
});
}
Task.WaitAll(tasks);
// At this point we have as many as k*TaskCount items left. Take the k most extreme.
return(SelectSerial(extremeItems, topN, k, comparisonDelegate));
}