ogoun
/
Zero
mirror of https://github.com/ogoun/Zero.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

HNSW

Browse Source 
Optimization for big data.
Quantization
pull/1/head
unknown 3 years ago
parent 1b6c56033f
commit 28ccefa255
21 changed files with 1459 additions and 1289 deletions
Show Stats Download Patch File Download Diff File

752
TestHNSW/HNSWDemo/Program.cs
Unescape View File

@ -99,77 +99,499 @@ namespace HNSWDemo
} }
} }
public class QVectorsDirectCompare
{
private const int HALF_LONG_BITS = 32;
private readonly IList<byte[]> _vectors;
private readonly Func<byte[], byte[], float> _distance;
public QVectorsDirectCompare(List<byte[]> vectors, Func<byte[], byte[], float> distance)
{
_vectors = vectors;
_distance = distance;
}
public IEnumerable<(int, float)> KNearest(byte[] v, int k)
{
var weights = new Dictionary<int, float>();
for (int i = 0; i < _vectors.Count; i++)
{
var d = _distance(v, _vectors[i]);
weights[i] = d;
}
return weights.OrderBy(p => p.Value).Take(k).Select(p => (p.Key, p.Value));
}
public List<HashSet<int>> DetectClusters()
{
var links = new SortedList<long, float>();
for (int i = 0; i < _vectors.Count; i++)
{
for (int j = i + 1; j < _vectors.Count; j++)
{
long k = (((long)(i)) << HALF_LONG_BITS) + j;
links.Add(k, _distance(_vectors[i], _vectors[j]));
}
}
// 1. Find R - bound between intra-cluster distances and out-of-cluster distances
var histogram = new Histogram(HistogramMode.SQRT, links.Values);
int threshold = histogram.OTSU();
var min = histogram.Bounds[threshold - 1];
var max = histogram.Bounds[threshold];
var R = (max + min) / 2;
// 2. Get links with distances less than R
var resultLinks = new SortedList<long, float>();
foreach (var pair in links)
{
if (pair.Value < R)
{
resultLinks.Add(pair.Key, pair.Value);
}
}
// 3. Extract clusters
List<HashSet<int>> clusters = new List<HashSet<int>>();
foreach (var pair in resultLinks)
{
var k = pair.Key;
var id1 = (int)(k >> HALF_LONG_BITS);
var id2 = (int)(k - (((long)id1) << HALF_LONG_BITS));
bool found = false;
foreach (var c in clusters)
{
if (c.Contains(id1))
{
c.Add(id2);
found = true;
break;
}
else if (c.Contains(id2))
{
c.Add(id1);
found = true;
break;
}
}
if (found == false)
{
var c = new HashSet<int>();
c.Add(id1);
c.Add(id2);
clusters.Add(c);
}
}
return clusters;
}
}
public class QLVectorsDirectCompare
{
private const int HALF_LONG_BITS = 32;
private readonly IList<long[]> _vectors;
private readonly Func<long[], long[], float> _distance;
public QLVectorsDirectCompare(List<long[]> vectors, Func<long[], long[], float> distance)
{
_vectors = vectors;
_distance = distance;
}
public IEnumerable<(int, float)> KNearest(long[] v, int k)
{
var weights = new Dictionary<int, float>();
for (int i = 0; i < _vectors.Count; i++)
{
var d = _distance(v, _vectors[i]);
weights[i] = d;
}
return weights.OrderBy(p => p.Value).Take(k).Select(p => (p.Key, p.Value));
}
public List<HashSet<int>> DetectClusters()
{
var links = new SortedList<long, float>();
for (int i = 0; i < _vectors.Count; i++)
{
for (int j = i + 1; j < _vectors.Count; j++)
{
long k = (((long)(i)) << HALF_LONG_BITS) + j;
links.Add(k, _distance(_vectors[i], _vectors[j]));
}
}
// 1. Find R - bound between intra-cluster distances and out-of-cluster distances
var histogram = new Histogram(HistogramMode.SQRT, links.Values);
int threshold = histogram.OTSU();
var min = histogram.Bounds[threshold - 1];
var max = histogram.Bounds[threshold];
var R = (max + min) / 2;
// 2. Get links with distances less than R
var resultLinks = new SortedList<long, float>();
foreach (var pair in links)
{
if (pair.Value < R)
{
resultLinks.Add(pair.Key, pair.Value);
}
}
// 3. Extract clusters
List<HashSet<int>> clusters = new List<HashSet<int>>();
foreach (var pair in resultLinks)
{
var k = pair.Key;
var id1 = (int)(k >> HALF_LONG_BITS);
var id2 = (int)(k - (((long)id1) << HALF_LONG_BITS));
bool found = false;
foreach (var c in clusters)
{
if (c.Contains(id1))
{
c.Add(id2);
found = true;
break;
}
else if (c.Contains(id2))
{
c.Add(id1);
found = true;
break;
}
}
if (found == false)
{
var c = new HashSet<int>();
c.Add(id1);
c.Add(id2);
clusters.Add(c);
}
}
return clusters;
}
}
public enum Gender public enum Gender
{ {
Unknown, Male, Feemale Unknown, Male, Feemale
}
public class Person
{
public Gender Gender { get; set; }
public int Age { get; set; }
public long Number { get; set; }
private static (float[], Person) Generate(int vector_size)
{
var rnd = new Random((int)Environment.TickCount);
var vector = new float[vector_size];
DefaultRandomGenerator.Instance.NextFloats(vector);
VectorUtils.NormalizeSIMD(vector);
var p = new Person();
p.Age = rnd.Next(15, 80);
var gr = rnd.Next(0, 3);
p.Gender = (gr == 0) ? Gender.Male : (gr == 1) ? Gender.Feemale : Gender.Unknown;
p.Number = CreateNumber(rnd);
return (vector, p);
}
public static List<(float[], Person)> GenerateRandom(int vectorSize, int vectorsCount)
{
var vectors = new List<(float[], Person)>();
for (int i = 0; i < vectorsCount; i++)
{
vectors.Add(Generate(vectorSize));
}
return vectors;
}
static HashSet<long> _exists = new HashSet<long>();
private static long CreateNumber(Random rnd)
{
long start_number;
do
{
start_number = 79600000000L;
start_number = start_number + rnd.Next(4, 8) * 10000000;
start_number += rnd.Next(0, 1000000);
}
while (_exists.Add(start_number) == false);
return start_number;
}
}
private static List<float[]> RandomVectors(int vectorSize, int vectorsCount)
{
var vectors = new List<float[]>();
for (int i = 0; i < vectorsCount; i++)
{
var vector = new float[vectorSize];
DefaultRandomGenerator.Instance.NextFloats(vector);
VectorUtils.NormalizeSIMD(vector);
vectors.Add(vector);
}
return vectors;
}
static void Main(string[] args)
{
QuantizatorTest();
Console.WriteLine("Completed");
Console.ReadKey();
}
static void QAccuracityTest()
{
int K = 200;
var count = 5000;
var testCount = 500;
var dimensionality = 128;
var totalHits = new List<int>();
var timewatchesNP = new List<float>();
var timewatchesHNSW = new List<float>();
var q = new Quantizator(-1f, 1f);
var samples = RandomVectors(dimensionality, count).Select(v => q.QuantizeToLong(v)).ToList();
var sw = new Stopwatch();
var test = new QLVectorsDirectCompare(samples, CosineDistance.NonOptimized);
var world = new SmallWorld<long[]>(NSWOptions<long[]>.Create(8, 12, 100, 100, CosineDistance.NonOptimized));
sw.Start();
var ids = world.AddItems(samples.ToArray());
sw.Stop();
Console.WriteLine($"Insert {ids.Length} items: {sw.ElapsedMilliseconds} ms");
Console.WriteLine("Start test");
var test_vectors = RandomVectors(dimensionality, testCount).Select(v => q.QuantizeToLong(v)).ToList();
foreach (var v in test_vectors)
{
sw.Restart();
var gt = test.KNearest(v, K).ToDictionary(p => p.Item1, p => p.Item2);
sw.Stop();
timewatchesNP.Add(sw.ElapsedMilliseconds);
sw.Restart();
var result = world.Search(v, K);
sw.Stop();
timewatchesHNSW.Add(sw.ElapsedMilliseconds);
var hits = 0;
foreach (var r in result)
{
if (gt.ContainsKey(r.Item1))
{
hits++;
}
}
totalHits.Add(hits);
}
Console.WriteLine($"MIN Accuracity: {totalHits.Min() * 100 / K}%");
Console.WriteLine($"AVG Accuracity: {totalHits.Average() * 100 / K}%");
Console.WriteLine($"MAX Accuracity: {totalHits.Max() * 100 / K}%");
Console.WriteLine($"MIN HNSW TIME: {timewatchesHNSW.Min()} ms");
Console.WriteLine($"AVG HNSW TIME: {timewatchesHNSW.Average()} ms");
Console.WriteLine($"MAX HNSW TIME: {timewatchesHNSW.Max()} ms");
Console.WriteLine($"MIN NP TIME: {timewatchesNP.Min()} ms");
Console.WriteLine($"AVG NP TIME: {timewatchesNP.Average()} ms");
Console.WriteLine($"MAX NP TIME: {timewatchesNP.Max()} ms");
}
static void QInsertTimeExplosionTest()
{
var count = 10000;
var iterationCount = 100;
var dimensionality = 128;
var sw = new Stopwatch();
var world = new SmallWorld<long[]>(NSWOptions<long[]>.Create(6, 12, 100, 100, CosineDistance.NonOptimized));
var q = new Quantizator(-1f, 1f);
for (int i = 0; i < iterationCount; i++)
{
var samples = RandomVectors(dimensionality, count);
sw.Restart();
var ids = world.AddItems(samples.Select(v => q.QuantizeToLong(v)).ToArray());
sw.Stop();
Console.WriteLine($"ITERATION: [{i.ToString("D4")}] COUNT: [{ids.Length}] ELAPSED [{sw.ElapsedMilliseconds} ms]");
}
}
static void AccuracityTest()
{
int K = 200;
var count = 3000;
var testCount = 500;
var dimensionality = 128;
var totalHits = new List<int>();
var timewatchesNP = new List<float>();
var timewatchesHNSW = new List<float>();
var samples = RandomVectors(dimensionality, count);
var sw = new Stopwatch();
var test = new VectorsDirectCompare(samples, CosineDistance.NonOptimized);
var world = new SmallWorld<float[]>(NSWOptions<float[]>.Create(8, 12, 100, 100, CosineDistance.NonOptimized));
sw.Start();
var ids = world.AddItems(samples.ToArray());
sw.Stop();
/*
byte[] dump;
using (var ms = new MemoryStream())
{
world.Serialize(ms);
dump = ms.ToArray();
}
Console.WriteLine($"Full dump size: {dump.Length} bytes");
ReadOnlySmallWorld<float[]> world;
using (var ms = new MemoryStream(dump))
{
world = SmallWorld.CreateReadOnlyWorldFrom<float[]>(NSWReadOnlyOption<float[]>.Create(100, CosineDistance.NonOptimized, true, true, selectionHeuristic: NeighbourSelectionHeuristic.SelectSimple), ms);
}
*/
Console.WriteLine($"Insert {ids.Length} items: {sw.ElapsedMilliseconds} ms");
Console.WriteLine("Start test");
var test_vectors = RandomVectors(dimensionality, testCount);
foreach (var v in test_vectors)
{
sw.Restart();
var gt = test.KNearest(v, K).ToDictionary(p => p.Item1, p => p.Item2);
sw.Stop();
timewatchesNP.Add(sw.ElapsedMilliseconds);
sw.Restart();
var result = world.Search(v, K);
sw.Stop();
timewatchesHNSW.Add(sw.ElapsedMilliseconds);
var hits = 0;
foreach (var r in result)
{
if (gt.ContainsKey(r.Item1))
{
hits++;
}
}
totalHits.Add(hits);
}
Console.WriteLine($"MIN Accuracity: {totalHits.Min() * 100 / K}%");
Console.WriteLine($"AVG Accuracity: {totalHits.Average() * 100 / K}%");
Console.WriteLine($"MAX Accuracity: {totalHits.Max() * 100 / K}%");
Console.WriteLine($"MIN HNSW TIME: {timewatchesHNSW.Min()} ms");
Console.WriteLine($"AVG HNSW TIME: {timewatchesHNSW.Average()} ms");
Console.WriteLine($"MAX HNSW TIME: {timewatchesHNSW.Max()} ms");
Console.WriteLine($"MIN NP TIME: {timewatchesNP.Min()} ms");
Console.WriteLine($"AVG NP TIME: {timewatchesNP.Average()} ms");
Console.WriteLine($"MAX NP TIME: {timewatchesNP.Max()} ms");
}
static void QuantizatorTest()
{
var samples = RandomVectors(128, 500000);
var min = samples.SelectMany(s => s).Min();
var max = samples.SelectMany(s => s).Max();
var q = new Quantizator(min, max);
var q_samples = samples.Select(s => q.QuantizeToLong(s)).ToArray();
// comparing
var list = new List<float>();
for (int i = 0; i < samples.Count - 1; i++)
{
var v1 = samples[i];
var v2 = samples[i + 1];
var dist = CosineDistance.NonOptimized(v1, v2);
var qv1 = q_samples[i];
var qv2 = q_samples[i + 1];
var qdist = CosineDistance.NonOptimized(qv1, qv2);
list.Add(Math.Abs(dist - qdist));
}
Console.WriteLine($"Min diff: {list.Min()}");
Console.WriteLine($"Avg diff: {list.Average()}");
Console.WriteLine($"Max diff: {list.Max()}");
} }
public class Person static void SaveRestoreTest()
{ {
public Gender Gender { get; set; } var count = 1000;
public int Age { get; set; } var dimensionality = 128;
public long Number { get; set; } var samples = RandomVectors(dimensionality, count);
var world = new SmallWorld<float[]>(NSWOptions<float[]>.Create(6, 15, 200, 200, CosineDistance.ForUnits));
var sw = new Stopwatch();
sw.Start();
var ids = world.AddItems(samples.ToArray());
sw.Stop();
Console.WriteLine($"Insert {ids.Length} items on {sw.ElapsedMilliseconds} ms");
Console.WriteLine("Start test");
private static (float[], Person) Generate(int vector_size) byte[] dump;
using (var ms = new MemoryStream())
{ {
var rnd = new Random((int)Environment.TickCount); world.Serialize(ms);
var vector = new float[vector_size]; dump = ms.ToArray();
DefaultRandomGenerator.Instance.NextFloats(vector);
VectorUtils.NormalizeSIMD(vector);
var p = new Person();
p.Age = rnd.Next(15, 80);
var gr = rnd.Next(0, 3);
p.Gender = (gr == 0) ? Gender.Male : (gr == 1) ? Gender.Feemale : Gender.Unknown;
p.Number = CreateNumber(rnd);
return (vector, p);
} }
Console.WriteLine($"Full dump size: {dump.Length} bytes");
public static List<(float[], Person)> GenerateRandom(int vectorSize, int vectorsCount) byte[] testDump;
var restoredWorld = new SmallWorld<float[]>(NSWOptions<float[]>.Create(6, 15, 200, 200, CosineDistance.ForUnits));
using (var ms = new MemoryStream(dump))
{ {
var vectors = new List<(float[], Person)>(); restoredWorld.Deserialize(ms);
for (int i = 0; i < vectorsCount; i++)
{
vectors.Add(Generate(vectorSize));
}
return vectors;
} }
static HashSet<long> _exists = new HashSet<long>(); using (var ms = new MemoryStream())
private static long CreateNumber(Random rnd)
{ {
long start_number; restoredWorld.Serialize(ms);
do testDump = ms.ToArray();
{
start_number = 79600000000L;
start_number = start_number + rnd.Next(4, 8) * 10000000;
start_number += rnd.Next(0, 1000000);
}
while (_exists.Add(start_number) == false);
return start_number;
} }
} if (testDump.Length != dump.Length)
private static List<float[]> RandomVectors(int vectorSize, int vectorsCount)
{
var vectors = new List<float[]>();
for (int i = 0; i < vectorsCount; i++)
{ {
var vector = new float[vectorSize]; Console.WriteLine($"Incorrect restored size. Got {testDump.Length}. Expected: {dump.Length}");
DefaultRandomGenerator.Instance.NextFloats(vector); return;
VectorUtils.NormalizeSIMD(vector);
vectors.Add(vector);
} }
return vectors;
} }
static void InsertTimeExplosionTest()
static void Main(string[] args)
{ {
InsertTimeExplosionTest(); var count = 10000;
Console.WriteLine("Completed"); var iterationCount = 100;
Console.ReadKey(); var dimensionality = 128;
var sw = new Stopwatch();
var world = new SmallWorld<float[]>(NSWOptions<float[]>.Create(6, 12, 100, 100, CosineDistance.NonOptimized));
for (int i = 0; i < iterationCount; i++)
{
var samples = RandomVectors(dimensionality, count);
sw.Restart();
var ids = world.AddItems(samples.ToArray());
sw.Stop();
Console.WriteLine($"ITERATION: [{i.ToString("D4")}] COUNT: [{ids.Length}] ELAPSED [{sw.ElapsedMilliseconds} ms]");
}
} }
/*
static void TestOnMnist() static void TestOnMnist()
{ {
int imageCount, rowCount, colCount; int imageCount, rowCount, colCount;
@ -227,77 +649,74 @@ namespace HNSWDemo
} }
static void AutoClusteringTest() static void AutoClusteringTest()
{ {
var vectors = RandomVectors(128, 3000); var vectors = RandomVectors(128, 3000);
var world = SmallWorld.CreateWorld<float[]>(NSWOptions<float[]>.Create(8, 16, 200, 200, Metrics.L2Euclidean, selectionHeuristic: NeighbourSelectionHeuristic.SelectSimple)); var world = SmallWorld.CreateWorld<float[]>(NSWOptions<float[]>.Create(8, 16, 200, 200, Metrics.L2Euclidean, selectionHeuristic: NeighbourSelectionHeuristic.SelectSimple));
world.AddItems(vectors); world.AddItems(vectors);
var clusters = AutomaticGraphClusterer.DetectClusters(world); var clusters = AutomaticGraphClusterer.DetectClusters(world);
Console.WriteLine($"Found {clusters.Count} clusters"); Console.WriteLine($"Found {clusters.Count} clusters");
for (int i = 0; i < clusters.Count; i++) for (int i = 0; i < clusters.Count; i++)
{ {
Console.WriteLine($"Cluster {i + 1} countains {clusters[i].Count} items"); Console.WriteLine($"Cluster {i + 1} countains {clusters[i].Count} items");
} }
} }
static void HistogramTest() static void HistogramTest()
{ {
var vectors = RandomVectors(128, 3000); var vectors = RandomVectors(128, 3000);
var world = SmallWorld.CreateWorld<float[]>(NSWOptions<float[]>.Create(8, 16, 200, 200, Metrics.L2Euclidean, selectionHeuristic: NeighbourSelectionHeuristic.SelectSimple)); var world = SmallWorld.CreateWorld<float[]>(NSWOptions<float[]>.Create(8, 16, 200, 200, Metrics.L2Euclidean, selectionHeuristic: NeighbourSelectionHeuristic.SelectSimple));
world.AddItems(vectors); world.AddItems(vectors);
var histogram = world.GetHistogram(); var histogram = world.GetHistogram();
int threshold = histogram.OTSU(); int threshold = histogram.OTSU();
var min = histogram.Bounds[threshold - 1]; var min = histogram.Bounds[threshold - 1];
var max = histogram.Bounds[threshold]; var max = histogram.Bounds[threshold];
var R = (max + min) / 2; var R = (max + min) / 2;
DrawHistogram(histogram, @"D:\hist.jpg"); DrawHistogram(histogram, @"D:\hist.jpg");
} }
static void DrawHistogram(Histogram histogram, string filename) static void DrawHistogram(Histogram histogram, string filename)
{ {
/* while (histogram.CountSignChanges() > 3) var wb = 1200 / histogram.Values.Length;
{ var k = 600.0f / (float)histogram.Values.Max();
histogram.Smooth();
}*/
var wb = 1200 / histogram.Values.Length;
var k = 600.0f / (float)histogram.Values.Max();
var maxes = histogram.GetMaximums().ToDictionary(m => m.Index, m => m); var maxes = histogram.GetMaximums().ToDictionary(m => m.Index, m => m);
int threshold = histogram.OTSU(); int threshold = histogram.OTSU();
using (var bmp = new Bitmap(1200, 600)) using (var bmp = new Bitmap(1200, 600))
{ {
using (var g = Graphics.FromImage(bmp)) { using (var g = Graphics.FromImage(bmp))
for (int i = 0; i < histogram.Values.Length; i++) {
for (int i = 0; i<histogram.Values.Length; i++)
{ {
var height = (int)(histogram.Values[i] * k); var height = (int)(histogram.Values[i] * k);
if (maxes.ContainsKey(i)) if (maxes.ContainsKey(i))
{ {
g.DrawRectangle(Pens.Red, i * wb, bmp.Height - height, wb, height); g.DrawRectangle(Pens.Red, i* wb, bmp.Height - height, wb, height);
g.DrawRectangle(Pens.Red, i * wb + 1, bmp.Height - height, wb - 1, height); g.DrawRectangle(Pens.Red, i* wb + 1, bmp.Height - height, wb - 1, height);
} }
else else
{ {
g.DrawRectangle(Pens.Blue, i * wb, bmp.Height - height, wb, height); g.DrawRectangle(Pens.Blue, i * wb, bmp.Height - height, wb, height);
} }
if (i == threshold) if (i == threshold)
{ {
g.DrawLine(Pens.Green, i * wb + wb / 2, 0, i * wb + wb / 2, bmp.Height); g.DrawLine(Pens.Green, i * wb + wb / 2, 0, i * wb + wb / 2, bmp.Height);
} }
} }
} }
bmp.Save(filename); bmp.Save(filename);
} }
} }
static void TransformToCompactWorldTest() static void TransformToCompactWorldTest()
{ {
var count = 10000; var count = 10000;
var dimensionality = 128; var dimensionality = 128;
var samples = RandomVectors(dimensionality, count); var samples = RandomVectors(dimensionality, count);
var world = new SmallWorld<float[]>(NSWOptions<float[]>.Create(6, 15, 200, 200, CosineDistance.ForUnits, true, true, selectionHeuristic: NeighbourSelectionHeuristic.SelectSimple)); var world = new SmallWorld<float[]>(NSWOptions<float[]>.Create(6, 15, 200, 200, CosineDistance.ForUnits));
var ids = world.AddItems(samples.ToArray()); var ids = world.AddItems(samples.ToArray());
Console.WriteLine("Start test"); Console.WriteLine("Start test");
@ -313,7 +732,7 @@ namespace HNSWDemo
ReadOnlySmallWorld<float[]> compactWorld; ReadOnlySmallWorld<float[]> compactWorld;
using (var ms = new MemoryStream(dump)) using (var ms = new MemoryStream(dump))
{ {
compactWorld = SmallWorld.CreateReadOnlyWorldFrom<float[]>(NSWReadOnlyOption<float[]>.Create(200, CosineDistance.ForUnits, true, true, selectionHeuristic: NeighbourSelectionHeuristic.SelectSimple), ms); compactWorld = SmallWorld.CreateReadOnlyWorldFrom<float[]>(NSWReadOnlyOption<float[]>.Create(200, CosineDistance.ForUnits), ms);
} }
// Compare worlds outputs // Compare worlds outputs
@ -379,7 +798,7 @@ namespace HNSWDemo
var samples = RandomVectors(dimensionality, count); var samples = RandomVectors(dimensionality, count);
var test = new VectorsDirectCompare(samples, CosineDistance.ForUnits); var test = new VectorsDirectCompare(samples, CosineDistance.ForUnits);
var world = new SmallWorld<float[]>(NSWOptions<float[]>.Create(6, 15, 200, 200, CosineDistance.ForUnits, true, true, selectionHeuristic: NeighbourSelectionHeuristic.SelectSimple)); var world = new SmallWorld<float[]>(NSWOptions<float[]>.Create(6, 15, 200, 200, CosineDistance.ForUnits));
var ids = world.AddItems(samples.ToArray()); var ids = world.AddItems(samples.ToArray());
Console.WriteLine("Start test"); Console.WriteLine("Start test");
@ -394,7 +813,7 @@ namespace HNSWDemo
ReadOnlySmallWorld<float[]> compactWorld; ReadOnlySmallWorld<float[]> compactWorld;
using (var ms = new MemoryStream(dump)) using (var ms = new MemoryStream(dump))
{ {
compactWorld = SmallWorld.CreateReadOnlyWorldFrom<float[]>(NSWReadOnlyOption<float[]>.Create(200, CosineDistance.ForUnits, true, true, selectionHeuristic: NeighbourSelectionHeuristic.SelectSimple), ms); compactWorld = SmallWorld.CreateReadOnlyWorldFrom<float[]>(NSWReadOnlyOption<float[]>.Create(200, CosineDistance.ForUnits), ms);
} }
// Compare worlds outputs // Compare worlds outputs
@ -493,63 +912,8 @@ namespace HNSWDemo
Console.WriteLine($"AVG HNSWCompact Accuracity: {totalHitsHNSWCompact.Average() * 100 / K}%"); Console.WriteLine($"AVG HNSWCompact Accuracity: {totalHitsHNSWCompact.Average() * 100 / K}%");
Console.WriteLine($"MAX HNSWCompact Accuracity: {totalHitsHNSWCompact.Max() * 100 / K}%"); Console.WriteLine($"MAX HNSWCompact Accuracity: {totalHitsHNSWCompact.Max() * 100 / K}%");
} }
static void SaveRestoreTest() static void FilterTest()
{
var count = 1000;
var dimensionality = 128;
var samples = RandomVectors(dimensionality, count);
var world = new SmallWorld<float[]>(NSWOptions<float[]>.Create(6, 15, 200, 200, CosineDistance.ForUnits, true, true, selectionHeuristic: NeighbourSelectionHeuristic.SelectSimple));
var sw = new Stopwatch();
sw.Start();
var ids = world.AddItems(samples.ToArray());
sw.Stop();
Console.WriteLine($"Insert {ids.Length} items on {sw.ElapsedMilliseconds} ms");
Console.WriteLine("Start test");
byte[] dump;
using (var ms = new MemoryStream())
{
world.Serialize(ms);
dump = ms.ToArray();
}
Console.WriteLine($"Full dump size: {dump.Length} bytes");
byte[] testDump;
var restoredWorld = new SmallWorld<float[]>(NSWOptions<float[]>.Create(6, 15, 200, 200, CosineDistance.ForUnits, true, true, selectionHeuristic: NeighbourSelectionHeuristic.SelectSimple));
using (var ms = new MemoryStream(dump))
{
restoredWorld.Deserialize(ms);
}
using (var ms = new MemoryStream())
{
restoredWorld.Serialize(ms);
testDump = ms.ToArray();
}
if (testDump.Length != dump.Length)
{
Console.WriteLine($"Incorrect restored size. Got {testDump.Length}. Expected: {dump.Length}");
return;
}
ReadOnlySmallWorld<float[]> compactWorld;
using (var ms = new MemoryStream(dump))
{
compactWorld = SmallWorld.CreateReadOnlyWorldFrom<float[]>(NSWReadOnlyOption<float[]>.Create(200, CosineDistance.ForUnits, true, true, selectionHeuristic: NeighbourSelectionHeuristic.SelectSimple), ms);
}
byte[] smallWorldDump;
using (var ms = new MemoryStream())
{
compactWorld.Serialize(ms);
smallWorldDump = ms.ToArray();
}
var p = smallWorldDump.Length * 100.0f / dump.Length;
Console.WriteLine($"Compact dump size: {smallWorldDump.Length} bytes. Decrease: {100 - p}%");
}
static void FilterTest()
{ {
var count = 1000; var count = 1000;
var testCount = 100; var testCount = 100;
@ -598,82 +962,6 @@ namespace HNSWDemo
Console.WriteLine($"SUCCESS: {hits}"); Console.WriteLine($"SUCCESS: {hits}");
Console.WriteLine($"ERROR: {miss}"); Console.WriteLine($"ERROR: {miss}");
} }
*/
static void AccuracityTest()
{
int K = 200;
var count = 2000;
var testCount = 1000;
var dimensionality = 128;
var totalHits = new List<int>();
var timewatchesNP = new List<float>();
var timewatchesHNSW = new List<float>();
var samples = RandomVectors(dimensionality, count);
var sw = new Stopwatch();
var test = new VectorsDirectCompare(samples, CosineDistance.NonOptimized);
var world = new SmallWorld<float[]>(NSWOptions<float[]>.Create(6, 12, 100, 100, CosineDistance.NonOptimized, true, true, selectionHeuristic: NeighbourSelectionHeuristic.SelectSimple));
sw.Start();
var ids = world.AddItems(samples.ToArray());
sw.Stop();
Console.WriteLine($"Insert {ids.Length} items: {sw.ElapsedMilliseconds} ms");
Console.WriteLine("Start test");
var test_vectors = RandomVectors(dimensionality, testCount);
foreach (var v in test_vectors)
{
sw.Restart();
var gt = test.KNearest(v, K).ToDictionary(p => p.Item1, p => p.Item2);
sw.Stop();
timewatchesNP.Add(sw.ElapsedMilliseconds);
sw.Restart();
var result = world.Search(v, K);
sw.Stop();
timewatchesHNSW.Add(sw.ElapsedMilliseconds);
var hits = 0;
foreach (var r in result)
{
if (gt.ContainsKey(r.Item1))
{
hits++;
}
}
totalHits.Add(hits);
}
Console.WriteLine($"MIN Accuracity: {totalHits.Min() * 100 / K}%");
Console.WriteLine($"AVG Accuracity: {totalHits.Average() * 100 / K}%");
Console.WriteLine($"MAX Accuracity: {totalHits.Max() * 100 / K}%");
Console.WriteLine($"MIN HNSW TIME: {timewatchesHNSW.Min()} ms");
Console.WriteLine($"AVG HNSW TIME: {timewatchesHNSW.Average()} ms");
Console.WriteLine($"MAX HNSW TIME: {timewatchesHNSW.Max()} ms");
Console.WriteLine($"MIN NP TIME: {timewatchesNP.Min()} ms");
Console.WriteLine($"AVG NP TIME: {timewatchesNP.Average()} ms");
Console.WriteLine($"MAX NP TIME: {timewatchesNP.Max()} ms");
}
static void InsertTimeExplosionTest()
{
var count = 20000;
var iterationCount = 100;
var dimensionality = 128;
var sw = new Stopwatch();
var world = new SmallWorld<float[]>(NSWOptions<float[]>.Create(6, 8, 150, 150, Metrics.L2Euclidean, true, true, selectionHeuristic: NeighbourSelectionHeuristic.SelectSimple));
for (int i = 0; i < iterationCount; i++)
{
var samples = RandomVectors(dimensionality, count);
sw.Restart();
var ids = world.AddItems(samples.ToArray());
sw.Stop();
Console.WriteLine($"ITERATION: [{i.ToString("D4")}] COUNT: [{ids.Length}] ELAPSED [{sw.ElapsedMilliseconds} ms]");
}
}
} }
} }

26
ZeroLevel.HNSW/HNSWMap.cs
Unescape View File

@ -1,5 +1,4 @@
using System.Collections.Concurrent; using System.Collections.Generic;
using System.Collections.Generic;
using ZeroLevel.Services.Serialization; using ZeroLevel.Services.Serialization;
namespace ZeroLevel.HNSW namespace ZeroLevel.HNSW
@ -10,8 +9,23 @@ namespace ZeroLevel.HNSW
public class HNSWMap<TFeature> public class HNSWMap<TFeature>
: IBinarySerializable : IBinarySerializable
{ {
private ConcurrentDictionary<TFeature, int> _map = new ConcurrentDictionary<TFeature, int>(); private Dictionary<TFeature, int> _map;
private ConcurrentDictionary<int, TFeature> _reverse_map = new ConcurrentDictionary<int, TFeature>(); private Dictionary<int, TFeature> _reverse_map;
public HNSWMap(int capacity = -1)
{
if (capacity > 0)
{
_map = new Dictionary<TFeature, int>(capacity);
_reverse_map = new Dictionary<int, TFeature>(capacity);
}
else
{
_map = new Dictionary<TFeature, int>();
_reverse_map = new Dictionary<int, TFeature>();
}
}
public void Append(TFeature feature, int vectorId) public void Append(TFeature feature, int vectorId)
{ {
@ -45,8 +59,8 @@ namespace ZeroLevel.HNSW
public void Deserialize(IBinaryReader reader) public void Deserialize(IBinaryReader reader)
{ {
this._map = reader.ReadDictionaryAsConcurrent<TFeature, int>(); this._map = reader.ReadDictionary<TFeature, int>();
this._reverse_map = reader.ReadDictionaryAsConcurrent<int, TFeature>(); this._reverse_map = reader.ReadDictionary<int, TFeature>();
} }
public void Serialize(IBinaryWriter writer) public void Serialize(IBinaryWriter writer)

5
ZeroLevel.HNSW/Model/Histogram.cs
Unescape View File

@ -21,12 +21,13 @@ namespace ZeroLevel.HNSW
public float[] Bounds { get; } public float[] Bounds { get; }
public int[] Values { get; } public int[] Values { get; }
public Histogram(HistogramMode mode, IList<float> data) public Histogram(HistogramMode mode, IEnumerable<float> data)
{ {
Mode = mode; Mode = mode;
Min = data.Min(); Min = data.Min();
Max = data.Max(); Max = data.Max();
int M = mode == HistogramMode.LOG ? (int)(1f + 3.2f * Math.Log(data.Count)) : (int)(Math.Sqrt(data.Count)); int count = data.Count();
int M = mode == HistogramMode.LOG ? (int)(1f + 3.2f * Math.Log(count)) : (int)(Math.Sqrt(count));
BoundsPeriod = (Max - Min) / M; BoundsPeriod = (Max - Min) / M;
Bounds = new float[M - 1]; Bounds = new float[M - 1];

21
ZeroLevel.HNSW/Model/NSWOptions.cs
Unescape View File

@ -21,12 +21,6 @@ namespace ZeroLevel.HNSW
/// </summary> /// </summary>
public readonly Func<TItem, TItem, float> Distance; public readonly Func<TItem, TItem, float> Distance;
public readonly bool ExpandBestSelection;
public readonly bool KeepPrunedConnections;
public readonly NeighbourSelectionHeuristic SelectionHeuristic;
public readonly int LayersCount; public readonly int LayersCount;
@ -34,29 +28,20 @@ namespace ZeroLevel.HNSW
int m, int m,
int ef, int ef,
int ef_construction, int ef_construction,
Func<TItem, TItem, float> distance, Func<TItem, TItem, float> distance)
bool expandBestSelection,
bool keepPrunedConnections,
NeighbourSelectionHeuristic selectionHeuristic)
{ {
LayersCount = layersCount; LayersCount = layersCount;
M = m; M = m;
EF = ef; EF = ef;
EFConstruction = ef_construction; EFConstruction = ef_construction;
Distance = distance; Distance = distance;
ExpandBestSelection = expandBestSelection;
KeepPrunedConnections = keepPrunedConnections;
SelectionHeuristic = selectionHeuristic;
} }
public static NSWOptions<TItem> Create(int layersCount, public static NSWOptions<TItem> Create(int layersCount,
int M, int M,
int EF, int EF,
int EF_construction, int EF_construction,
Func<TItem, TItem, float> distance, Func<TItem, TItem, float> distance) =>
bool expandBestSelection = false, new NSWOptions<TItem>(layersCount, M, EF, EF_construction, distance);
bool keepPrunedConnections = false,
NeighbourSelectionHeuristic selectionHeuristic = NeighbourSelectionHeuristic.SelectSimple) =>
new NSWOptions<TItem>(layersCount, M, EF, EF_construction, distance, expandBestSelection, keepPrunedConnections, selectionHeuristic);
} }
} }

44
ZeroLevel.HNSW/Model/NSWReadOnlyOption.cs
Unescape View File

@ -1,44 +0,0 @@
using System;
namespace ZeroLevel.HNSW
{
public sealed class NSWReadOnlyOption<TItem>
{
/// <summary>
/// Max search buffer
/// </summary>
public readonly int EF;
/// <summary>
/// Distance function beetween vectors
/// </summary>
public readonly Func<TItem, TItem, float> Distance;
public readonly bool ExpandBestSelection;
public readonly bool KeepPrunedConnections;
public readonly NeighbourSelectionHeuristic SelectionHeuristic;
private NSWReadOnlyOption(
int ef,
Func<TItem, TItem, float> distance,
bool expandBestSelection,
bool keepPrunedConnections,
NeighbourSelectionHeuristic selectionHeuristic)
{
EF = ef;
Distance = distance;
ExpandBestSelection = expandBestSelection;
KeepPrunedConnections = keepPrunedConnections;
SelectionHeuristic = selectionHeuristic;
}
public static NSWReadOnlyOption<TItem> Create(
int EF,
Func<TItem, TItem, float> distance,
bool expandBestSelection = false,
bool keepPrunedConnections = false,
NeighbourSelectionHeuristic selectionHeuristic = NeighbourSelectionHeuristic.SelectSimple) =>
new NSWReadOnlyOption<TItem>(EF, distance, expandBestSelection, keepPrunedConnections, selectionHeuristic);
}
}

18
ZeroLevel.HNSW/Model/NeighbourSelectionHeuristic.cs
Unescape View File

@ -1,18 +0,0 @@
namespace ZeroLevel.HNSW
{
/// <summary>
/// Type of heuristic to select best neighbours for a node.
/// </summary>
public enum NeighbourSelectionHeuristic
{
/// <summary>
/// Marker for the Algorithm 3 (SELECT-NEIGHBORS-SIMPLE) from the article. Implemented in <see cref="Algorithms.Algorithm3{TItem, TDistance}"/>
/// </summary>
SelectSimple,
/// <summary>
/// Marker for the Algorithm 4 (SELECT-NEIGHBORS-HEURISTIC) from the article. Implemented in <see cref="Algorithms.Algorithm4{TItem, TDistance}"/>
/// </summary>
SelectHeuristic
}
}

152
ZeroLevel.HNSW/ReadOnlySmallWorld.cs
Unescape View File

@ -1,152 +0,0 @@
using System;
using System.Collections.Generic;
using System.IO;
using System.Linq;
using ZeroLevel.Services.Serialization;
namespace ZeroLevel.HNSW
{
public class ReadOnlySmallWorld<TItem>
{
private readonly NSWReadOnlyOption<TItem> _options;
private ReadOnlyVectorSet<TItem> _vectors;
private ReadOnlyLayer<TItem>[] _layers;
private int EntryPoint = 0;
private int MaxLayer = 0;
private ReadOnlySmallWorld() { }
internal ReadOnlySmallWorld(NSWReadOnlyOption<TItem> options, Stream stream)
{
_options = options;
Deserialize(stream);
}
/// <summary>
/// Search in the graph K for vectors closest to a given vector
/// </summary>
/// <param name="vector">Given vector</param>
/// <param name="k">Count of elements for search</param>
/// <param name="activeNodes"></param>
/// <returns></returns>
public IEnumerable<(int, TItem, float)> Search(TItem vector, int k)
{
foreach (var pair in KNearest(vector, k))
{
yield return (pair.Item1, _vectors[pair.Item1], pair.Item2);
}
}
public IEnumerable<(int, TItem, float)> Search(TItem vector, int k, SearchContext context)
{
if (context == null)
{
foreach (var pair in KNearest(vector, k))
{
yield return (pair.Item1, _vectors[pair.Item1], pair.Item2);
}
}
else
{
foreach (var pair in KNearest(vector, k, context))
{
yield return (pair.Item1, _vectors[pair.Item1], pair.Item2);
}
}
}
/// <summary>
/// Algorithm 5
/// </summary>
private IEnumerable<(int, float)> KNearest(TItem q, int k)
{
if (_vectors.Count == 0)
{
return Enumerable.Empty<(int, float)>();
}
var distance = new Func<int, float>(candidate => _options.Distance(q, _vectors[candidate]));
// W ← ∅ // set for the current nearest elements
var W = new Dictionary<int, float>(k + 1);
// ep ← get enter point for hnsw
var ep = EntryPoint;
// L ← level of ep // top layer for hnsw
var L = MaxLayer;
// for lc ← L … 1
for (int layer = L; layer > 0; --layer)
{
// W ← SEARCH-LAYER(q, ep, ef = 1, lc)
_layers[layer].KNearestAtLayer(ep, distance, W, 1);
// ep ← get nearest element from W to q
ep = W.OrderBy(p => p.Value).First().Key;
W.Clear();
}
// W ← SEARCH-LAYER(q, ep, ef, lc =0)
_layers[0].KNearestAtLayer(ep, distance, W, k);
// return K nearest elements from W to q
return W.Select(p => (p.Key, p.Value));
}
private IEnumerable<(int, float)> KNearest(TItem q, int k, SearchContext context)
{
if (_vectors.Count == 0)
{
return Enumerable.Empty<(int, float)>();
}
var distance = new Func<int, float>(candidate => _options.Distance(q, _vectors[candidate]));
// W ← ∅ // set for the current nearest elements
var W = new Dictionary<int, float>(k + 1);
// ep ← get enter point for hnsw
var ep = EntryPoint;
// L ← level of ep // top layer for hnsw
var L = MaxLayer;
// for lc ← L … 1
for (int layer = L; layer > 0; --layer)
{
// W ← SEARCH-LAYER(q, ep, ef = 1, lc)
_layers[layer].KNearestAtLayer(ep, distance, W, 1);
// ep ← get nearest element from W to q
ep = W.OrderBy(p => p.Value).First().Key;
W.Clear();
}
// W ← SEARCH-LAYER(q, ep, ef, lc =0)
_layers[0].KNearestAtLayer(ep, distance, W, k, context);
// return K nearest elements from W to q
return W.Select(p => (p.Key, p.Value));
}
public void Serialize(Stream stream)
{
using (var writer = new MemoryStreamWriter(stream))
{
writer.WriteInt32(EntryPoint);
writer.WriteInt32(MaxLayer);
_vectors.Serialize(writer);
writer.WriteInt32(_layers.Length);
foreach (var l in _layers)
{
l.Serialize(writer);
}
}
}
public void Deserialize(Stream stream)
{
using (var reader = new MemoryStreamReader(stream))
{
this.EntryPoint = reader.ReadInt32();
this.MaxLayer = reader.ReadInt32();
_vectors = new ReadOnlyVectorSet<TItem>();
_vectors.Deserialize(reader);
var countLayers = reader.ReadInt32();
_layers = new ReadOnlyLayer<TItem>[countLayers];
for (int i = 0; i < countLayers; i++)
{
_layers[i] = new ReadOnlyLayer<TItem>(_vectors);
_layers[i].Deserialize(reader);
}
}
}
}
}

4
ZeroLevel.HNSW/Services/AutomaticGraphClusterer.cs
Unescape View File

@ -6,7 +6,7 @@ namespace ZeroLevel.HNSW.Services
{ {
private const int HALF_LONG_BITS = 32; private const int HALF_LONG_BITS = 32;
public static List<HashSet<int>> DetectClusters<T>(SmallWorld<T> world) /*public static List<HashSet<int>> DetectClusters<T>(SmallWorld<T> world)
{ {
var links = world.GetNSWLinks(); var links = world.GetNSWLinks();
// 1. Find R - bound between intra-cluster distances and out-of-cluster distances // 1. Find R - bound between intra-cluster distances and out-of-cluster distances
@ -60,6 +60,6 @@ namespace ZeroLevel.HNSW.Services
} }
} }
return clusters; return clusters;
} }*/
} }
} }

152
ZeroLevel.HNSW/Services/CompactBiDirectionalLinksSet.cs
Unescape View File

@ -30,75 +30,6 @@ namespace ZeroLevel.HNSW
internal int Count => _set.Count; internal int Count => _set.Count;
/// <summary>
/// Разрывает связи id1 - id2 и id2 - id1, и строит новые id1 - id, id - id1
/// </summary>
internal void Relink(int id1, int id2, int id, float distance)
{
long k1old = (((long)(id1)) << HALF_LONG_BITS) + id2;
long k2old = (((long)(id2)) << HALF_LONG_BITS) + id1;
long k1new = (((long)(id1)) << HALF_LONG_BITS) + id;
long k2new = (((long)(id)) << HALF_LONG_BITS) + id1;
_rwLock.EnterWriteLock();
try
{
_set.Remove(k1old);
_set.Remove(k2old);
if (!_set.ContainsKey(k1new))
_set.Add(k1new, distance);
if (!_set.ContainsKey(k2new))
_set.Add(k2new, distance);
}
finally
{
_rwLock.ExitWriteLock();
}
}
/// <summary>
/// Разрывает связи id1 - id2 и id2 - id1, и строит новые id1 - id, id - id1, id2 - id, id - id2
/// </summary>
internal void Relink(int id1, int id2, int id, float distanceToId1, float distanceToId2)
{
long k_id1_id2 = (((long)(id1)) << HALF_LONG_BITS) + id2;
long k_id2_id1 = (((long)(id2)) << HALF_LONG_BITS) + id1;
long k_id_id1 = (((long)(id)) << HALF_LONG_BITS) + id1;
long k_id1_id = (((long)(id1)) << HALF_LONG_BITS) + id;
long k_id_id2 = (((long)(id)) << HALF_LONG_BITS) + id2;
long k_id2_id = (((long)(id2)) << HALF_LONG_BITS) + id;
_rwLock.EnterWriteLock();
try
{
_set.Remove(k_id1_id2);
_set.Remove(k_id2_id1);
if (!_set.ContainsKey(k_id_id1))
{
_set.Add(k_id_id1, distanceToId1);
}
if (!_set.ContainsKey(k_id1_id))
{
_set.Add(k_id1_id, distanceToId1);
}
if (!_set.ContainsKey(k_id_id2))
{
_set.Add(k_id_id2, distanceToId2);
}
if (!_set.ContainsKey(k_id2_id))
{
_set.Add(k_id2_id, distanceToId2);
}
}
finally
{
_rwLock.ExitWriteLock();
}
}
internal IEnumerable<(int, int, float)> FindLinksForId(int id) internal IEnumerable<(int, int, float)> FindLinksForId(int id)
{ {
_rwLock.EnterReadLock(); _rwLock.EnterReadLock();
@ -147,43 +78,6 @@ namespace ZeroLevel.HNSW
} }
} }
internal void RemoveIndex(int id)
{
long[] forward;
long[] backward;
_rwLock.EnterReadLock();
try
{
forward = Search(_set, id).Select(pair => pair.Item1).ToArray();
backward = forward.Select(k =>
{
var id1 = k >> HALF_LONG_BITS;
var id2 = k - (id1 << HALF_LONG_BITS);
return (id2 << HALF_LONG_BITS) + id1;
}).ToArray();
}
finally
{
_rwLock.ExitReadLock();
}
_rwLock.EnterWriteLock();
try
{
foreach (var k in forward)
{
_set.Remove(k);
}
foreach (var k in backward)
{
_set.Remove(k);
}
}
finally
{
_rwLock.ExitWriteLock();
}
}
internal void RemoveIndex(int id1, int id2) internal void RemoveIndex(int id1, int id2)
{ {
long k1 = (((long)(id1)) << HALF_LONG_BITS) + id2; long k1 = (((long)(id1)) << HALF_LONG_BITS) + id2;
@ -230,40 +124,46 @@ namespace ZeroLevel.HNSW
return false; return false;
} }
/*
function binary_search(A, n, T) is
L := 0
R := n − 1
while L ≤ R do
m := floor((L + R) / 2)
if A[m] < T then
L := m + 1
else if A[m] > T then
R := m − 1
else:
return m
return unsuccessful
*/
private static IEnumerable<(long, float)> Search(SortedList<long, float> set, int index) private static IEnumerable<(long, float)> Search(SortedList<long, float> set, int index)
{ {
long k = ((long)index) << HALF_LONG_BITS; long k = ((long)index) << HALF_LONG_BITS; // T
int left = 0; int left = 0;
int right = set.Count - 1; int right = set.Count - 1;
int mid; int mid;
long test; long test;
while (left < right) while (left <= right)
{ {
mid = (right + left) / 2; mid = (int)Math.Floor((right + left) / 2d);
test = (set.Keys[mid] >> HALF_LONG_BITS) << HALF_LONG_BITS; test = (set.Keys[mid] >> HALF_LONG_BITS) << HALF_LONG_BITS; // A[m]
if (left == mid || right == mid) if (test < k)
{ {
if (test == k) left = mid + 1;
{
return SearchByPosition(set, k, mid);
}
break;
} }
if (test < k) else if (test > k)
{ {
left = mid; right = mid - 1;
} }
else else
{ {
if (test == k) return SearchByPosition(set, k, mid);
{
return SearchByPosition(set, k, mid);
}
else
{
right = mid;
}
} }
} }
return Enumerable.Empty<(long, float)>(); return Enumerable.Empty<(long, float)>();

365
ZeroLevel.HNSW/Services/Layer.cs
Unescape View File

@ -2,6 +2,7 @@
using System.Collections.Generic; using System.Collections.Generic;
using System.Linq; using System.Linq;
using ZeroLevel.HNSW.Services; using ZeroLevel.HNSW.Services;
using ZeroLevel.Services.Pools;
using ZeroLevel.Services.Serialization; using ZeroLevel.Services.Serialization;
namespace ZeroLevel.HNSW namespace ZeroLevel.HNSW
@ -14,146 +15,80 @@ namespace ZeroLevel.HNSW
{ {
private readonly NSWOptions<TItem> _options; private readonly NSWOptions<TItem> _options;
private readonly VectorSet<TItem> _vectors; private readonly VectorSet<TItem> _vectors;
private readonly CompactBiDirectionalLinksSet _links; private readonly LinksSet _links;
internal SortedList<long, float> Links => _links.Links; //internal SortedList<long, float> Links => _links.Links;
/// <summary> /// <summary>
/// There are links е the layer /// There are links е the layer
/// </summary> /// </summary>
internal bool HasLinks => (_links.Count > 0); internal bool HasLinks => (_links.Count > 0);
private int GetM(bool nswLayer)
{
return nswLayer ? 2 * _options.M : _options.M;
}
/// <summary> /// <summary>
/// HNSW layer /// HNSW layer
/// </summary> /// </summary>
/// <param name="options">HNSW graph options</param> /// <param name="options">HNSW graph options</param>
/// <param name="vectors">General vector set</param> /// <param name="vectors">General vector set</param>
internal Layer(NSWOptions<TItem> options, VectorSet<TItem> vectors) internal Layer(NSWOptions<TItem> options, VectorSet<TItem> vectors, bool nswLayer)
{ {
_options = options; _options = options;
_vectors = vectors; _vectors = vectors;
_links = new CompactBiDirectionalLinksSet(); _links = new LinksSet(GetM(nswLayer), (id1, id2) => options.Distance(_vectors[id1], _vectors[id2]));
} }
/// <summary> internal int FindEntryPointAtLayer(Func<int, float> targetCosts)
/// Adding new bidirectional link
/// </summary>
/// <param name="q">New node</param>
/// <param name="p">The node with which the connection will be made</param>
/// <param name="qpDistance"></param>
/// <param name="isMapLayer"></param>
/*internal void AddBidirectionallConnections(int q, int p, float qpDistance, bool isMapLayer)
{ {
// поиск в ширину ближайших узлов к найденному if (_links.Count == 0) return EntryPoint;
var nearest = _links.FindLinksForId(p).ToArray(); var set = new HashSet<int>(_links.Items().Select(p => p.Item1));
// если у найденного узла максимальное количество связей int minId = -1;
// if │eConn│ > Mmax // shrink connections of e float minDist = float.MaxValue;
if (nearest.Length >= (isMapLayer ? _options.M * 2 : _options.M)) foreach (var id in set)
{
// ищем связь с самой большой дистанцией
float distance = nearest[0].Item3;
int index = 0;
for (int ni = 1; ni < nearest.Length; ni++)
{
// Если осталась ссылка узла на себя, удаляем ее в первую очередь
if (nearest[ni].Item1 == nearest[ni].Item2)
{
index = ni;
break;
}
if (nearest[ni].Item3 > distance)
{
index = ni;
distance = nearest[ni].Item3;
}
}
// делаем перелинковку вставляя новый узел между найденными
var id1 = nearest[index].Item1;
var id2 = nearest[index].Item2;
_links.Relink(id1, id2, q, qpDistance, _options.Distance(_vectors[id2], _vectors[q]));
}
else
{ {
if (nearest.Length == 1 && nearest[0].Item1 == nearest[0].Item2) var d = targetCosts(id);
{ if (d < minDist && Math.Abs(d) > float.Epsilon)
// убираем связи на самих себя
var id1 = nearest[0].Item1;
var id2 = nearest[0].Item2;
_links.Relink(id1, id2, q, qpDistance, _options.Distance(_vectors[id2], _vectors[q]));
}
else
{ {
// добавляем связь нового узла к найденному minDist = d;
_links.Add(q, p, qpDistance); minId = id;
} }
} }
}*/ return minId;
internal void AddBidirectionallConnections(int q, int p, float qpDistance)
{
_links.Add(q, p, qpDistance);
} }
internal void TrimLinks(int q, bool isMapLayer) internal void AddBidirectionallConnections(int q, int p)
{ {
var M = (isMapLayer ? _options.M * 2 : _options.M); if (q == p)
// поиск в ширину ближайших узлов к найденному
var nearest = _links.FindLinksForId(q).ToArray();
if (nearest.Length <= M && nearest.Length > 1)
{ {
foreach (var l in nearest) if (EntryPoint >= 0)
{ {
if (l.Item1 == l.Item2) _links.Add(q, EntryPoint);
{
_links.RemoveIndex(l.Item1, l.Item2);
}
} }
} else
else if (nearest.Length > M)
{
var removeCount = nearest.Length - M;
foreach (var l in nearest.OrderByDescending(n => n.Item3).Take(removeCount))
{ {
_links.RemoveIndex(l.Item1, l.Item2); EntryPoint = q;
} }
} }
else
{
_links.Add(q, p);
}
} }
private int EntryPoint = -1;
/// <summary> internal void Trim(int id) => _links.Trim(id);
/// Adding a node with a connection to itself
/// </summary>
/// <param name="q"></param>
internal void Append(int q)
{
_links.Add(q, q, 0);
}
#region Implementation of https://arxiv.org/ftp/arxiv/papers/1603/1603.09320.pdf #region Implementation of https://arxiv.org/ftp/arxiv/papers/1603/1603.09320.pdf
internal int FindEntryPointAtLayer(Func<int, float> targetCosts)
{
var set = new HashSet<int>(_links.Items().Select(p => p.Item1));
int minId = -1;
float minDist = float.MaxValue;
foreach (var id in set)
{
var d = targetCosts(id);
if (d < minDist && Math.Abs(d) > float.Epsilon)
{
minDist = d;
minId = id;
}
}
return minId;
}
/// <summary> /// <summary>
/// Algorithm 2 /// Algorithm 2
/// </summary> /// </summary>
/// <param name="q">query element</param> /// <param name="q">query element</param>
/// <param name="ep">enter points ep</param> /// <param name="ep">enter points ep</param>
/// <returns>Output: ef closest neighbors to q</returns> /// <returns>Output: ef closest neighbors to q</returns>
internal IEnumerable<(int, float)> KNearestAtLayer(int entryPointId, Func<int, float> targetCosts, IEnumerable<(int, float)> w, int ef) internal IEnumerable<(int, float)> KNearestAtLayer(int entryPointId, Func<int, float> targetCosts, int ef)
{ {
/* /*
* v ← ep // set of visited elements * v ← ep // set of visited elements
@ -175,55 +110,60 @@ namespace ZeroLevel.HNSW
* remove furthest element from W to q * remove furthest element from W to q
* return W * return W
*/ */
var v = new VisitedBitSet(_vectors.Count, _options.M);
// v ← ep // set of visited elements
v.Add(entryPointId);
var W = new MaxHeap(ef + 1);
foreach (var i in w) W.Push(i);
int farthestId;
float farthestDistance;
var d = targetCosts(entryPointId); var d = targetCosts(entryPointId);
// C ← ep // set of candidates
var v = new VisitedBitSet(_vectors.Count, _options.M);
// * v ← ep // set of visited elements
v.Add(entryPointId);
// * C ← ep // set of candidates
var C = new MinHeap(ef); var C = new MinHeap(ef);
C.Push((entryPointId, d)); C.Push((entryPointId, d));
// W ← ep // dynamic list of found nearest neighbors // * W ← ep // dynamic list of found nearest neighbors
var W = new MaxHeap(ef + 1);
W.Push((entryPointId, d)); W.Push((entryPointId, d));
int farthestId; // * while │C│ > 0
float farthestDistance;
// run bfs
while (C.Count > 0) while (C.Count > 0)
{ {
// get next candidate to check and expand // * c ← extract nearest element from C to q
var toExpand = C.Pop(); var c = C.Pop();
if (W.TryPeek(out _, out farthestDistance) && toExpand.Item2 > farthestDistance) // * f ← get furthest element from W to q
// * if distance(c, q) > distance(f, q)
if (W.TryPeek(out _, out farthestDistance) && c.Item2 > farthestDistance)
{ {
// the closest candidate is farther than farthest result // * break // all elements in W are evaluated
break; break;
} }
// expand candidate // * for each e ∈ neighbourhood(c) at layer lc // update C and W
var neighboursIds = GetNeighbors(toExpand.Item1).ToArray(); foreach (var e in GetNeighbors(c.Item1))
for (int i = 0; i < neighboursIds.Length; ++i)
{ {
int neighbourId = neighboursIds[i]; // * if e ∉ v
if (!v.Contains(neighbourId)) if (!v.Contains(e))
{ {
// enqueue perspective neighbours to expansion list // * v ← v ⋃ e
v.Add(e);
// * f ← get furthest element from W to q
W.TryPeek(out farthestId, out farthestDistance); W.TryPeek(out farthestId, out farthestDistance);
var neighbourDistance = targetCosts(neighbourId); var eDistance = targetCosts(e);
if (W.Count < ef || (farthestId >= 0 && neighbourDistance < farthestDistance)) // * if distance(e, q) < distance(f, q) or │W│ < ef
if (W.Count < ef || (farthestId >= 0 && eDistance < farthestDistance))
{ {
C.Push((neighbourId, neighbourDistance)); // * C ← C ⋃ e
C.Push((e, eDistance));
W.Push((neighbourId, neighbourDistance)); // * W ← W ⋃ e
W.Push((e, eDistance));
// * if │W│ > ef
if (W.Count > ef) if (W.Count > ef)
{ {
// * remove furthest element from W to q
W.Pop(); W.Pop();
} }
} }
v.Add(neighbourId);
} }
} }
} }
@ -238,40 +178,22 @@ namespace ZeroLevel.HNSW
/// <param name="q">query element</param> /// <param name="q">query element</param>
/// <param name="ep">enter points ep</param> /// <param name="ep">enter points ep</param>
/// <returns>Output: ef closest neighbors to q</returns> /// <returns>Output: ef closest neighbors to q</returns>
internal IEnumerable<(int, float)> KNearestAtLayer(int entryPointId, Func<int, float> targetCosts, IEnumerable<(int, float)> w, int ef, SearchContext context) /*
internal IEnumerable<(int, float)> KNearestAtLayer(int entryPointId, Func<int, float> targetCosts, int ef, SearchContext context)
{ {
/* int farthestId;
* v ← ep // set of visited elements float farthestDistance;
* C ← ep // set of candidates var d = targetCosts(entryPointId);
* W ← ep // dynamic list of found nearest neighbors
* while │C│ > 0
* c ← extract nearest element from C to q
* f ← get furthest element from W to q
* if distance(c, q) > distance(f, q)
* break // all elements in W are evaluated
* for each e ∈ neighbourhood(c) at layer lc // update C and W
* if e ∉ v
* v ← v ⋃ e
* f ← get furthest element from W to q
* if distance(e, q) < distance(f, q) or │W│ < ef
* C ← C ⋃ e
* W ← W ⋃ e
* if │W│ > ef
* remove furthest element from W to q
* return W
*/
var v = new VisitedBitSet(_vectors.Count, _options.M); var v = new VisitedBitSet(_vectors.Count, _options.M);
// v ← ep // set of visited elements // v ← ep // set of visited elements
v.Add(entryPointId); v.Add(entryPointId);
var W = new MaxHeap(ef + 1);
foreach (var i in w) W.Push(i);
// C ← ep // set of candidates // C ← ep // set of candidates
var C = new MinHeap(ef); var C = new MinHeap(ef);
var d = targetCosts(entryPointId);
C.Push((entryPointId, d)); C.Push((entryPointId, d));
// W ← ep // dynamic list of found nearest neighbors // W ← ep // dynamic list of found nearest neighbors
var W = new MaxHeap(ef + 1);
// W ← ep // dynamic list of found nearest neighbors
if (context.IsActiveNode(entryPointId)) if (context.IsActiveNode(entryPointId))
{ {
W.Push((entryPointId, d)); W.Push((entryPointId, d));
@ -281,14 +203,10 @@ namespace ZeroLevel.HNSW
{ {
// get next candidate to check and expand // get next candidate to check and expand
var toExpand = C.Pop(); var toExpand = C.Pop();
if (W.Count > 0) if (W.TryPeek(out _, out farthestDistance) && toExpand.Item2 > farthestDistance)
{ {
if (W.TryPeek(out _, out var dist)) // the closest candidate is farther than farthest result
if (toExpand.Item2 > dist) break;
{
// the closest candidate is farther than farthest result
break;
}
} }
// expand candidate // expand candidate
@ -298,11 +216,12 @@ namespace ZeroLevel.HNSW
int neighbourId = neighboursIds[i]; int neighbourId = neighboursIds[i];
if (!v.Contains(neighbourId)) if (!v.Contains(neighbourId))
{ {
W.TryPeek(out farthestId, out farthestDistance);
// enqueue perspective neighbours to expansion list // enqueue perspective neighbours to expansion list
var neighbourDistance = targetCosts(neighbourId); var neighbourDistance = targetCosts(neighbourId);
if (context.IsActiveNode(neighbourId)) if (context.IsActiveNode(neighbourId))
{ {
if (W.Count < ef || (W.Count > 0 && (W.TryPeek(out _, out var dist) && neighbourDistance < dist))) if (W.Count < ef || (farthestId >= 0 && neighbourDistance < farthestDistance))
{ {
W.Push((neighbourId, neighbourDistance)); W.Push((neighbourId, neighbourDistance));
if (W.Count > ef) if (W.Count > ef)
@ -311,7 +230,7 @@ namespace ZeroLevel.HNSW
} }
} }
} }
if (W.Count < ef) if (W.TryPeek(out _, out farthestDistance) && neighbourDistance < farthestDistance)
{ {
C.Push((neighbourId, neighbourDistance)); C.Push((neighbourId, neighbourDistance));
} }
@ -323,6 +242,7 @@ namespace ZeroLevel.HNSW
v.Clear(); v.Clear();
return W; return W;
} }
*/
/// <summary> /// <summary>
/// Algorithm 2, modified for LookAlike /// Algorithm 2, modified for LookAlike
@ -330,28 +250,9 @@ namespace ZeroLevel.HNSW
/// <param name="q">query element</param> /// <param name="q">query element</param>
/// <param name="ep">enter points ep</param> /// <param name="ep">enter points ep</param>
/// <returns>Output: ef closest neighbors to q</returns> /// <returns>Output: ef closest neighbors to q</returns>
/*
internal IEnumerable<(int, float)> KNearestAtLayer(IEnumerable<(int, float)> w, int ef, SearchContext context) internal IEnumerable<(int, float)> KNearestAtLayer(IEnumerable<(int, float)> w, int ef, SearchContext context)
{ {
/*
* v ← ep // set of visited elements
* C ← ep // set of candidates
* W ← ep // dynamic list of found nearest neighbors
* while │C│ > 0
* c ← extract nearest element from C to q
* f ← get furthest element from W to q
* if distance(c, q) > distance(f, q)
* break // all elements in W are evaluated
* for each e ∈ neighbourhood(c) at layer lc // update C and W
* if e ∉ v
* v ← v ⋃ e
* f ← get furthest element from W to q
* if distance(e, q) < distance(f, q) or │W│ < ef
* C ← C ⋃ e
* W ← W ⋃ e
* if │W│ > ef
* remove furthest element from W to q
* return W
*/
// v ← ep // set of visited elements // v ← ep // set of visited elements
var v = new VisitedBitSet(_vectors.Count, _options.M); var v = new VisitedBitSet(_vectors.Count, _options.M);
// C ← ep // set of candidates // C ← ep // set of candidates
@ -445,106 +346,10 @@ namespace ZeroLevel.HNSW
v.Clear(); v.Clear();
return W; return W;
} }
*/
/// <summary>
/// Algorithm 3
/// </summary>
internal MaxHeap SELECT_NEIGHBORS_SIMPLE(IEnumerable<(int, float)> w, int M)
{
var W = new MaxHeap(w.Count());
foreach (var i in w) W.Push(i);
var bestN = M;
if (W.Count > bestN)
{
while (W.Count > bestN)
{
W.Pop();
}
}
return W;
}
/// <summary>
/// Algorithm 4
/// </summary>
/// <param name="q">base element</param>
/// <param name="C">candidate elements</param>
/// <param name="extendCandidates">flag indicating whether or not to extend candidate list</param>
/// <param name="keepPrunedConnections">flag indicating whether or not to add discarded elements</param>
/// <returns>Output: M elements selected by the heuristic</returns>
internal MaxHeap SELECT_NEIGHBORS_HEURISTIC(Func<int, float> distance, IEnumerable<(int, float)> w, int M)
{
// R ← ∅
var R = new MaxHeap(_options.EFConstruction);
// W ← C // working queue for the candidates
var W = new MaxHeap(_options.EFConstruction + 1);
foreach (var i in w) W.Push(i);
// if extendCandidates // extend candidates by their neighbors
if (_options.ExpandBestSelection)
{
var extendBuffer = new HashSet<int>();
// for each e ∈ C
foreach (var e in W)
{
var neighbors = GetNeighbors(e.Item1);
// for each e_adj ∈ neighbourhood(e) at layer lc
foreach (var e_adj in neighbors)
{
// if eadj ∉ W
if (extendBuffer.Contains(e_adj) == false)
{
extendBuffer.Add(e_adj);
}
}
}
// W ← W ⋃ eadj
foreach (var id in extendBuffer)
{
W.Push((id, distance(id)));
}
}
// Wd ← ∅ // queue for the discarded candidates
var Wd = new MinHeap(_options.EFConstruction);
// while │W│ > 0 and │R│< M
while (W.Count > 0 && R.Count < M)
{
// e ← extract nearest element from W to q
var (e, ed) = W.Pop();
var (fe, fd) = R.Pop();
// if e is closer to q compared to any element from R
if (R.Count == 0 ||
ed < fd)
{
// R ← R ⋃ e
R.Push((e, ed));
}
else
{
// Wd ← Wd ⋃ e
Wd.Push((e, ed));
}
}
// if keepPrunedConnections // add some of the discarded // connections from Wd
if (_options.KeepPrunedConnections)
{
// while │Wd│> 0 and │R│< M
while (Wd.Count > 0 && R.Count < M)
{
// R ← R ⋃ extract nearest element from Wd to q
var nearest = Wd.Pop();
R.Push((nearest.Item1, nearest.Item2));
}
}
// return R
return R;
}
#endregion #endregion
private IEnumerable<int> GetNeighbors(int id) => _links.FindLinksForId(id).Select(d => d.Item2); private IEnumerable<int> GetNeighbors(int id) => _links.FindNeighbors(id);
public void Serialize(IBinaryWriter writer) public void Serialize(IBinaryWriter writer)
{ {
@ -556,6 +361,6 @@ namespace ZeroLevel.HNSW
_links.Deserialize(reader); _links.Deserialize(reader);
} }
internal Histogram GetHistogram(HistogramMode mode) => _links.CalculateHistogram(mode); // internal Histogram GetHistogram(HistogramMode mode) => _links.CalculateHistogram(mode);
} }
} }

279
ZeroLevel.HNSW/Services/LinksSet.cs
Unescape View File

@ -0,0 +1,279 @@
using System;
using System.Collections.Concurrent;
using System.Collections.Generic;
using System.Linq;
using ZeroLevel.Services.Serialization;
namespace ZeroLevel.HNSW
{
/*
internal struct Link
: IEquatable<Link>
{
public int Id;
public float Distance;
public override int GetHashCode()
{
return Id.GetHashCode();
}
public override bool Equals(object obj)
{
if (obj is Link)
return this.Equals((Link)obj);
return false;
}
public bool Equals(Link other)
{
return this.Id == other.Id;
}
}
public class LinksSetWithCachee
{
private ConcurrentDictionary<int, HashSet<Link>> _set = new ConcurrentDictionary<int, HashSet<Link>>();
internal int Count => _set.Count;
private readonly int _M;
private readonly Func<int, int, float> _distance;
public LinksSetWithCachee(int M, Func<int, int, float> distance)
{
_distance = distance;
_M = M;
}
internal IEnumerable<int> FindNeighbors(int id)
{
if (_set.ContainsKey(id))
{
return _set[id].Select(l=>l.Id);
}
return Enumerable.Empty<int>();
}
internal void RemoveIndex(int id1, int id2)
{
var link1 = new Link { Id = id1 };
var link2 = new Link { Id = id2 };
_set[id1].Remove(link2);
_set[id2].Remove(link1);
}
internal bool Add(int id1, int id2, float distance)
{
if (!_set.ContainsKey(id1))
{
_set[id1] = new HashSet<Link>();
}
if (!_set.ContainsKey(id2))
{
_set[id2] = new HashSet<Link>();
}
var r1 = _set[id1].Add(new Link { Id = id2, Distance = distance });
var r2 = _set[id2].Add(new Link { Id = id1, Distance = distance });
//TrimSet(_set[id1]);
TrimSet(id2, _set[id2]);
return r1 || r2;
}
internal void Trim(int id) => TrimSet(id, _set[id]);
private void TrimSet(int id, HashSet<Link> set)
{
if (set.Count > _M)
{
var removeCount = set.Count - _M;
var removeLinks = set.OrderByDescending(n => n.Distance).Take(removeCount).ToArray();
foreach (var l in removeLinks)
{
set.Remove(l);
}
}
}
public void Dispose()
{
_set.Clear();
_set = null;
}
private const int HALF_LONG_BITS = 32;
public void Serialize(IBinaryWriter writer)
{
writer.WriteBoolean(false); // true - set with weights
writer.WriteInt32(_set.Sum(pair => pair.Value.Count));
foreach (var record in _set)
{
var id = record.Key;
foreach (var r in record.Value)
{
var key = (((long)(id)) << HALF_LONG_BITS) + r;
writer.WriteLong(key);
}
}
}
public void Deserialize(IBinaryReader reader)
{
if (reader.ReadBoolean() == false)
{
throw new InvalidOperationException("Incompatible data format. The set does not contain weights.");
}
_set.Clear();
_set = null;
var count = reader.ReadInt32();
_set = new ConcurrentDictionary<int, HashSet<int>>();
for (int i = 0; i < count; i++)
{
var key = reader.ReadLong();
var id1 = (int)(key >> HALF_LONG_BITS);
var id2 = (int)(key - (((long)id1) << HALF_LONG_BITS));
if (!_set.ContainsKey(id1))
{
_set[id1] = new HashSet<int>();
}
_set[id1].Add(id2);
}
}
}
*/
public class LinksSet
{
private ConcurrentDictionary<int, HashSet<int>> _set = new ConcurrentDictionary<int, HashSet<int>>();
internal IDictionary<int, HashSet<int>> Links => _set;
internal int Count => _set.Count;
private readonly int _M;
private readonly Func<int, int, float> _distance;
public LinksSet(int M, Func<int, int, float> distance)
{
_distance = distance;
_M = M;
}
internal IEnumerable<(int, int)> FindLinksForId(int id)
{
if (_set.ContainsKey(id))
{
return _set[id].Select(v => (id, v));
}
return Enumerable.Empty<(int, int)>();
}
internal IEnumerable<int> FindNeighbors(int id)
{
if (_set.ContainsKey(id))
{
return _set[id];
}
return Enumerable.Empty<int>();
}
internal IEnumerable<(int, int)> Items()
{
return _set
.SelectMany(pair => _set[pair.Key]
.Select(v => (pair.Key, v)));
}
internal void RemoveIndex(int id1, int id2)
{
_set[id1].Remove(id2);
_set[id2].Remove(id1);
}
internal bool Add(int id1, int id2)
{
if (!_set.ContainsKey(id1))
{
_set[id1] = new HashSet<int>(_M + 1);
}
if (!_set.ContainsKey(id2))
{
_set[id2] = new HashSet<int>(_M + 1);
}
var r1 = _set[id1].Add(id2);
var r2 = _set[id2].Add(id1);
TrimSet(id1, _set[id1]);
TrimSet(id2, _set[id2]);
return r1 || r2;
}
internal void Trim(int id) => TrimSet(id, _set[id]);
private void TrimSet(int id, HashSet<int> set)
{
if (set.Count > _M)
{
var removeCount = set.Count - _M;
var removeLinks = set.OrderByDescending(n => _distance(id, n)).Take(removeCount).ToArray();
foreach (var l in removeLinks)
{
set.Remove(l);
}
}
}
public void Dispose()
{
_set.Clear();
_set = null;
}
private const int HALF_LONG_BITS = 32;
public void Serialize(IBinaryWriter writer)
{
writer.WriteBoolean(false); // true - set with weights
writer.WriteInt32(_set.Sum(pair => pair.Value.Count));
foreach (var record in _set)
{
var id = record.Key;
foreach (var r in record.Value)
{
var key = (((long)(id)) << HALF_LONG_BITS) + r;
writer.WriteLong(key);
}
}
}
public void Deserialize(IBinaryReader reader)
{
if (reader.ReadBoolean() == false)
{
throw new InvalidOperationException("Incompatible data format. The set does not contain weights.");
}
_set.Clear();
_set = null;
var count = reader.ReadInt32();
_set = new ConcurrentDictionary<int, HashSet<int>>();
for (int i = 0; i < count; i++)
{
var key = reader.ReadLong();
var id1 = (int)(key >> HALF_LONG_BITS);
var id2 = (int)(key - (((long)id1) << HALF_LONG_BITS));
if (!_set.ContainsKey(id1))
{
_set[id1] = new HashSet<int>();
}
_set[id1].Add(id2);
}
}
}
}

287
ZeroLevel.HNSW/Services/Pool.cs
Unescape View File

@ -0,0 +1,287 @@
using System;
using System.Collections.Generic;
using System.Diagnostics;
using System.Threading;
namespace ZeroLevel.Services.Pools
{
public enum LoadingMode { Eager, Lazy, LazyExpanding };
public enum AccessMode { FIFO, LIFO, Circular };
public sealed class Pool<T> : IDisposable
{
private bool isDisposed;
private Func<Pool<T>, T> factory;
private LoadingMode loadingMode;
private IItemStore itemStore;
private int size;
private int count;
private Semaphore sync;
public Pool(int size, Func<Pool<T>, T> factory)
: this(size, factory, LoadingMode.Lazy, AccessMode.FIFO)
{
}
public Pool(int size, Func<Pool<T>, T> factory,
LoadingMode loadingMode, AccessMode accessMode)
{
if (size <= 0)
throw new ArgumentOutOfRangeException("size", size,
"Argument 'size' must be greater than zero.");
if (factory == null)
throw new ArgumentNullException("factory");
this.size = size;
this.factory = factory;
sync = new Semaphore(size, size);
this.loadingMode = loadingMode;
this.itemStore = CreateItemStore(accessMode, size);
if (loadingMode == LoadingMode.Eager)
{
PreloadItems();
}
}
public T Acquire()
{
sync.WaitOne();
switch (loadingMode)
{
case LoadingMode.Eager:
return AcquireEager();
case LoadingMode.Lazy:
return AcquireLazy();
default:
Debug.Assert(loadingMode == LoadingMode.LazyExpanding,
"Unknown LoadingMode encountered in Acquire method.");
return AcquireLazyExpanding();
}
}
public void Release(T item)
{
lock (itemStore)
{
itemStore.Store(item);
}
sync.Release();
}
public void Dispose()
{
if (isDisposed)
{
return;
}
isDisposed = true;
if (typeof(IDisposable).IsAssignableFrom(typeof(T)))
{
lock (itemStore)
{
while (itemStore.Count > 0)
{
IDisposable disposable = (IDisposable)itemStore.Fetch();
disposable.Dispose();
}
}
}
sync.Close();
}
#region Acquisition
private T AcquireEager()
{
lock (itemStore)
{
return itemStore.Fetch();
}
}
private T AcquireLazy()
{
lock (itemStore)
{
if (itemStore.Count > 0)
{
return itemStore.Fetch();
}
}
Interlocked.Increment(ref count);
return factory(this);
}
private T AcquireLazyExpanding()
{
bool shouldExpand = false;
if (count < size)
{
int newCount = Interlocked.Increment(ref count);
if (newCount <= size)
{
shouldExpand = true;
}
else
{
// Another thread took the last spot - use the store instead
Interlocked.Decrement(ref count);
}
}
if (shouldExpand)
{
return factory(this);
}
else
{
lock (itemStore)
{
return itemStore.Fetch();
}
}
}
private void PreloadItems()
{
for (int i = 0; i < size; i++)
{
T item = factory(this);
itemStore.Store(item);
}
count = size;
}
#endregion
#region Collection Wrappers
interface IItemStore
{
T Fetch();
void Store(T item);
int Count { get; }
}
private IItemStore CreateItemStore(AccessMode mode, int capacity)
{
switch (mode)
{
case AccessMode.FIFO:
return new QueueStore(capacity);
case AccessMode.LIFO:
return new StackStore(capacity);
default:
Debug.Assert(mode == AccessMode.Circular,
"Invalid AccessMode in CreateItemStore");
return new CircularStore(capacity);
}
}
class QueueStore : Queue<T>, IItemStore
{
public QueueStore(int capacity) : base(capacity)
{
}
public T Fetch()
{
return Dequeue();
}
public void Store(T item)
{
Enqueue(item);
}
}
class StackStore : Stack<T>, IItemStore
{
public StackStore(int capacity) : base(capacity)
{
}
public T Fetch()
{
return Pop();
}
public void Store(T item)
{
Push(item);
}
}
class CircularStore : IItemStore
{
private List<Slot> slots;
private int freeSlotCount;
private int position = -1;
public CircularStore(int capacity)
{
slots = new List<Slot>(capacity);
}
public T Fetch()
{
if (Count == 0)
throw new InvalidOperationException("The buffer is empty.");
int startPosition = position;
do
{
Advance();
Slot slot = slots[position];
if (!slot.IsInUse)
{
slot.IsInUse = true;
--freeSlotCount;
return slot.Item;
}
} while (startPosition != position);
throw new InvalidOperationException("No free slots.");
}
public void Store(T item)
{
Slot slot = slots.Find(s => object.Equals(s.Item, item));
if (slot == null)
{
slot = new Slot(item);
slots.Add(slot);
}
slot.IsInUse = false;
++freeSlotCount;
}
public int Count
{
get { return freeSlotCount; }
}
private void Advance()
{
position = (position + 1) % slots.Count;
}
class Slot
{
public Slot(T item)
{
this.Item = item;
}
public T Item { get; private set; }
public bool IsInUse { get; set; }
}
}
#endregion
public bool IsDisposed
{
get { return isDisposed; }
}
}
}

86
ZeroLevel.HNSW/Services/Quantizator.cs
Unescape View File

@ -0,0 +1,86 @@
using System;
namespace ZeroLevel.HNSW.Services
{
public class Quantizator
{
private readonly float _min;
private readonly float _max;
private readonly float _diff;
public Quantizator(float min, float max)
{
_min = min;
_max = max;
_diff = _max - _min;
}
public byte[] Quantize(float[] v)
{
var result = new byte[v.Length];
for (int i = 0; i < v.Length; i++)
{
result[i] = _quantizeInRange(v[i]);
}
return result;
}
public int[] QuantizeToInt(float[] v)
{
if (v.Length % 4 != 0)
{
throw new ArgumentOutOfRangeException("v.Length % 4 must be zero!");
}
var result = new int[v.Length / 4];
byte[] buf = new byte[4];
for (int i = 0; i < v.Length; i += 4)
{
buf[0] = _quantizeInRange(v[i]);
buf[1] = _quantizeInRange(v[i + 1]);
buf[2] = _quantizeInRange(v[i + 2]);
buf[3] = _quantizeInRange(v[i + 3]);
result[(i >> 2)] = BitConverter.ToInt32(buf);
}
return result;
}
public long[] QuantizeToLong(float[] v)
{
if (v.Length % 8 != 0)
{
throw new ArgumentOutOfRangeException("v.Length % 8 must be zero!");
}
var result = new long[v.Length / 8];
byte[] buf = new byte[8];
for (int i = 0; i < v.Length; i += 8)
{
buf[0] = _quantizeInRange(v[i + 0]);
buf[1] = _quantizeInRange(v[i + 1]);
buf[2] = _quantizeInRange(v[i + 2]);
buf[3] = _quantizeInRange(v[i + 3]);
buf[4] = _quantizeInRange(v[i + 4]);
buf[5] = _quantizeInRange(v[i + 5]);
buf[6] = _quantizeInRange(v[i + 6]);
buf[7] = _quantizeInRange(v[i + 7]);
result[(i >> 3)] = BitConverter.ToInt64(buf);
}
return result;
}
//Map x in [0,1] to {0, 1, ..., 255}
private byte _quantize(float x)
{
x = (int)Math.Floor(256 * x);
if (x < 0) return 0;
else if (x > 255) return 255;
else return (byte)x;
}
//Map x in [min,max] to {0, 1, ..., 255}
private byte _quantizeInRange(float x)
{
return _quantize((x - _min) / (_diff));
}
}
}

107
ZeroLevel.HNSW/Services/ReadOnly/ReadOnlyCompactBiDirectionalLinksSet.cs
Unescape View File

@ -1,107 +0,0 @@
using System;
using System.Collections.Generic;
using System.Linq;
using ZeroLevel.Services.Serialization;
namespace ZeroLevel.HNSW
{
internal sealed class ReadOnlyCompactBiDirectionalLinksSet
: IBinarySerializable, IDisposable
{
private const int HALF_LONG_BITS = 32;
private Dictionary<int, int[]> _set = new Dictionary<int, int[]>();
internal int Count => _set.Count;
internal IEnumerable<int> FindLinksForId(int id)
{
if (_set.ContainsKey(id))
{
return _set[id];
}
return Enumerable.Empty<int>();
}
public void Dispose()
{
_set.Clear();
_set = null;
}
public void Serialize(IBinaryWriter writer)
{
writer.WriteBoolean(false); // false - set without weights
writer.WriteInt32(_set.Count);
foreach (var record in _set)
{
writer.WriteInt32(record.Key);
writer.WriteCollection(record.Value);
}
}
public void Deserialize(IBinaryReader reader)
{
_set.Clear();
_set = null;
if (reader.ReadBoolean() == false)
{
var count = reader.ReadInt32();
_set = new Dictionary<int, int[]>(count);
for (int i = 0; i < count; i++)
{
var key = reader.ReadInt32();
var value = reader.ReadInt32Array();
_set.Add(key, value);
}
}
else
{
var count = reader.ReadInt32();
_set = new Dictionary<int, int[]>(count);
// hack, We know that an sortedset has been saved
long key;
int id1, id2;
var prevId = -1;
var set = new HashSet<int>();
for (int i = 0; i < count; i++)
{
key = reader.ReadLong();
id1 = (int)(key >> HALF_LONG_BITS);
id2 = (int)(key - (((long)id1) << HALF_LONG_BITS));
reader.ReadFloat(); // SKIP
if (prevId == -1)
{
prevId = id1;
if (id1 != id2)
{
set.Add(id2);
}
}
else if (prevId != id1)
{
_set.Add(prevId, set.ToArray());
set.Clear();
prevId = id1;
}
else
{
if (id1 != id2)
{
set.Add(id2);
}
}
}
if (set.Count > 0)
{
_set.Add(prevId, set.ToArray());
set.Clear();
}
}
}
}
}

209
ZeroLevel.HNSW/Services/ReadOnly/ReadOnlyLayer.cs
Unescape View File

@ -1,209 +0,0 @@
using System;
using System.Collections.Generic;
using System.Linq;
using ZeroLevel.Services.Serialization;
namespace ZeroLevel.HNSW
{
/// <summary>
/// NSW graph
/// </summary>
internal sealed class ReadOnlyLayer<TItem>
: IBinarySerializable
{
private readonly ReadOnlyVectorSet<TItem> _vectors;
private readonly ReadOnlyCompactBiDirectionalLinksSet _links;
/// <summary>
/// HNSW layer
/// </summary>
/// <param name="vectors">General vector set</param>
internal ReadOnlyLayer(ReadOnlyVectorSet<TItem> vectors)
{
_vectors = vectors;
_links = new ReadOnlyCompactBiDirectionalLinksSet();
}
#region Implementation of https://arxiv.org/ftp/arxiv/papers/1603/1603.09320.pdf
/// <summary>
/// Algorithm 2
/// </summary>
/// <param name="q">query element</param>
/// <param name="ep">enter points ep</param>
/// <returns>Output: ef closest neighbors to q</returns>
internal void KNearestAtLayer(int entryPointId, Func<int, float> targetCosts, IDictionary<int, float> W, int ef)
{
/*
* v ← ep // set of visited elements
* C ← ep // set of candidates
* W ← ep // dynamic list of found nearest neighbors
* while │C│ > 0
* c ← extract nearest element from C to q
* f ← get furthest element from W to q
* if distance(c, q) > distance(f, q)
* break // all elements in W are evaluated
* for each e ∈ neighbourhood(c) at layer lc // update C and W
* if e ∉ v
* v ← v ⋃ e
* f ← get furthest element from W to q
* if distance(e, q) < distance(f, q) or │W│ < ef
* C ← C ⋃ e
* W ← W ⋃ e
* if │W│ > ef
* remove furthest element from W to q
* return W
*/
var v = new VisitedBitSet(_vectors.Count, 1);
// v ← ep // set of visited elements
v.Add(entryPointId);
// C ← ep // set of candidates
var C = new Dictionary<int, float>();
C.Add(entryPointId, targetCosts(entryPointId));
// W ← ep // dynamic list of found nearest neighbors
W.Add(entryPointId, C[entryPointId]);
var popCandidate = new Func<(int, float)>(() => { var pair = C.OrderBy(e => e.Value).First(); C.Remove(pair.Key); return (pair.Key, pair.Value); });
var fartherFromResult = new Func<(int, float)>(() => { var pair = W.OrderByDescending(e => e.Value).First(); return (pair.Key, pair.Value); });
var fartherPopFromResult = new Action(() => { var pair = W.OrderByDescending(e => e.Value).First(); W.Remove(pair.Key); });
// run bfs
while (C.Count > 0)
{
// get next candidate to check and expand
var toExpand = popCandidate();
var farthestResult = fartherFromResult();
if (toExpand.Item2 > farthestResult.Item2)
{
// the closest candidate is farther than farthest result
break;
}
// expand candidate
var neighboursIds = GetNeighbors(toExpand.Item1).ToArray();
for (int i = 0; i < neighboursIds.Length; ++i)
{
int neighbourId = neighboursIds[i];
if (!v.Contains(neighbourId))
{
// enqueue perspective neighbours to expansion list
farthestResult = fartherFromResult();
var neighbourDistance = targetCosts(neighbourId);
if (W.Count < ef || neighbourDistance < farthestResult.Item2)
{
C.Add(neighbourId, neighbourDistance);
W.Add(neighbourId, neighbourDistance);
if (W.Count > ef)
{
fartherPopFromResult();
}
}
v.Add(neighbourId);
}
}
}
C.Clear();
v.Clear();
}
/// <summary>
/// Algorithm 2
/// </summary>
/// <param name="q">query element</param>
/// <param name="ep">enter points ep</param>
/// <returns>Output: ef closest neighbors to q</returns>
internal void KNearestAtLayer(int entryPointId, Func<int, float> targetCosts, IDictionary<int, float> W, int ef, SearchContext context)
{
/*
* v ← ep // set of visited elements
* C ← ep // set of candidates
* W ← ep // dynamic list of found nearest neighbors
* while │C│ > 0
* c ← extract nearest element from C to q
* f ← get furthest element from W to q
* if distance(c, q) > distance(f, q)
* break // all elements in W are evaluated
* for each e ∈ neighbourhood(c) at layer lc // update C and W
* if e ∉ v
* v ← v ⋃ e
* f ← get furthest element from W to q
* if distance(e, q) < distance(f, q) or │W│ < ef
* C ← C ⋃ e
* W ← W ⋃ e
* if │W│ > ef
* remove furthest element from W to q
* return W
*/
var v = new VisitedBitSet(_vectors.Count, 1);
// v ← ep // set of visited elements
v.Add(entryPointId);
// C ← ep // set of candidates
var C = new Dictionary<int, float>();
C.Add(entryPointId, targetCosts(entryPointId));
// W ← ep // dynamic list of found nearest neighbors
if (context.IsActiveNode(entryPointId))
{
W.Add(entryPointId, C[entryPointId]);
}
var popCandidate = new Func<(int, float)>(() => { var pair = C.OrderBy(e => e.Value).First(); C.Remove(pair.Key); return (pair.Key, pair.Value); });
var farthestDistance = new Func<float>(() => { var pair = W.OrderByDescending(e => e.Value).First(); return pair.Value; });
var fartherPopFromResult = new Action(() => { var pair = W.OrderByDescending(e => e.Value).First(); W.Remove(pair.Key); });
// run bfs
while (C.Count > 0)
{
// get next candidate to check and expand
var toExpand = popCandidate();
if (W.Count > 0)
{
if (toExpand.Item2 > farthestDistance())
{
// the closest candidate is farther than farthest result
break;
}
}
// expand candidate
var neighboursIds = GetNeighbors(toExpand.Item1).ToArray();
for (int i = 0; i < neighboursIds.Length; ++i)
{
int neighbourId = neighboursIds[i];
if (!v.Contains(neighbourId))
{
// enqueue perspective neighbours to expansion list
var neighbourDistance = targetCosts(neighbourId);
if (context.IsActiveNode(neighbourId))
{
if (W.Count < ef || (W.Count > 0 && neighbourDistance < farthestDistance()))
{
W.Add(neighbourId, neighbourDistance);
if (W.Count > ef)
{
fartherPopFromResult();
}
}
}
if (W.Count < ef)
{
C.Add(neighbourId, neighbourDistance);
}
v.Add(neighbourId);
}
}
}
C.Clear();
v.Clear();
}
#endregion
private IEnumerable<int> GetNeighbors(int id) => _links.FindLinksForId(id);
public void Serialize(IBinaryWriter writer)
{
_links.Serialize(writer);
}
public void Deserialize(IBinaryReader reader)
{
_links.Deserialize(reader);
}
}
}

33
ZeroLevel.HNSW/Services/ReadOnly/ReadOnlyVectorSet.cs
Unescape View File

@ -1,33 +0,0 @@
using System.Collections.Generic;
using ZeroLevel.Services.Serialization;
namespace ZeroLevel.HNSW
{
internal sealed class ReadOnlyVectorSet<T>
: IBinarySerializable
{
private List<T> _set = new List<T>();
internal T this[int index] => _set[index];
internal int Count => _set.Count;
public void Deserialize(IBinaryReader reader)
{
int count = reader.ReadInt32();
_set = new List<T>(count + 1);
for (int i = 0; i < count; i++)
{
_set.Add(reader.ReadCompatible<T>());
}
}
public void Serialize(IBinaryWriter writer)
{
writer.WriteInt32(_set.Count);
foreach (var r in _set)
{
writer.WriteCompatible<T>(r);
}
}
}
}

54
ZeroLevel.HNSW/SmallWorld.cs
Unescape View File

@ -17,7 +17,6 @@ namespace ZeroLevel.HNSW
private int MaxLayer = 0; private int MaxLayer = 0;
private readonly ProbabilityLayerNumberGenerator _layerLevelGenerator; private readonly ProbabilityLayerNumberGenerator _layerLevelGenerator;
private ReaderWriterLockSlim _lockGraph = new ReaderWriterLockSlim(); private ReaderWriterLockSlim _lockGraph = new ReaderWriterLockSlim();
internal SortedList<long, float> GetNSWLinks() => _layers[0].Links;
public SmallWorld(NSWOptions<TItem> options) public SmallWorld(NSWOptions<TItem> options)
{ {
@ -27,7 +26,7 @@ namespace ZeroLevel.HNSW
_layerLevelGenerator = new ProbabilityLayerNumberGenerator(_options.LayersCount, _options.M); _layerLevelGenerator = new ProbabilityLayerNumberGenerator(_options.LayersCount, _options.M);
for (int i = 0; i < _options.LayersCount; i++) for (int i = 0; i < _options.LayersCount; i++)
{ {
_layers[i] = new Layer<TItem>(_options, _vectors); _layers[i] = new Layer<TItem>(_options, _vectors, i == 0);
} }
} }
@ -51,7 +50,7 @@ namespace ZeroLevel.HNSW
yield return (pair.Item1, _vectors[pair.Item1], pair.Item2); yield return (pair.Item1, _vectors[pair.Item1], pair.Item2);
} }
} }
/*
public IEnumerable<(int, TItem, float)> Search(TItem vector, int k, SearchContext context) public IEnumerable<(int, TItem, float)> Search(TItem vector, int k, SearchContext context)
{ {
if (context == null) if (context == null)
@ -84,6 +83,7 @@ namespace ZeroLevel.HNSW
} }
} }
} }
*/
/// <summary> /// <summary>
/// Adding vectors batch /// Adding vectors batch
@ -119,21 +119,23 @@ namespace ZeroLevel.HNSW
var W = new MinHeap(_options.EFConstruction + 1); var W = new MinHeap(_options.EFConstruction + 1);
// ep ← get enter point for hnsw // ep ← get enter point for hnsw
var ep = _layers[MaxLayer].FindEntryPointAtLayer(distance); var ep = _layers[MaxLayer].FindEntryPointAtLayer(distance);
if (ep == -1) ep = EntryPoint; if (ep == -1)
ep = EntryPoint;
var epDist = distance(ep); var epDist = distance(ep);
// L ← level of ep // top layer for hnsw // L ← level of ep // top layer for hnsw
var L = MaxLayer; var L = MaxLayer;
// l ← ⌊-ln(unif(0..1))∙mL⌋ // new element’s level // l ← ⌊-ln(unif(0..1))∙mL⌋ // new element’s level
int l = _layerLevelGenerator.GetRandomLayer(); int l = _layerLevelGenerator.GetRandomLayer();
// for lc ← L … l+1
// Проход с верхнего уровня до уровня где появляется элемент, для нахождения точки входа // Проход с верхнего уровня до уровня где появляется элемент, для нахождения точки входа
int id; int id;
float value; float value;
// for lc ← L … l+1
for (int lc = L; lc > l; --lc) for (int lc = L; lc > l; --lc)
{ {
// W ← SEARCH-LAYER(q, ep, ef = 1, lc) // W ← SEARCH-LAYER(q, ep, ef = 1, lc)
foreach (var i in _layers[lc].KNearestAtLayer(ep, distance, W, 1)) foreach (var i in _layers[lc].KNearestAtLayer(ep, distance, 1))
{ {
W.Push(i); W.Push(i);
} }
@ -143,7 +145,6 @@ namespace ZeroLevel.HNSW
ep = id; ep = id;
epDist = value; epDist = value;
} }
_layers[lc].TrimLinks(q, lc == 0);
W.Clear(); W.Clear();
} }
//for lc ← min(L, l) … 0 //for lc ← min(L, l) … 0
@ -152,12 +153,12 @@ namespace ZeroLevel.HNSW
{ {
if (_layers[lc].HasLinks == false) if (_layers[lc].HasLinks == false)
{ {
_layers[lc].Append(q); _layers[lc].AddBidirectionallConnections(q, q);
} }
else else
{ {
// W ← SEARCH - LAYER(q, ep, efConstruction, lc) // W ← SEARCH - LAYER(q, ep, efConstruction, lc)
foreach (var i in _layers[lc].KNearestAtLayer(ep, distance, W, _options.EFConstruction)) foreach (var i in _layers[lc].KNearestAtLayer(ep, distance, _options.EFConstruction))
{ {
W.Push(i); W.Push(i);
} }
@ -170,13 +171,13 @@ namespace ZeroLevel.HNSW
} }
// neighbors ← SELECT-NEIGHBORS(q, W, M, lc) // alg. 3 or alg. 4 // neighbors ← SELECT-NEIGHBORS(q, W, M, lc) // alg. 3 or alg. 4
var neighbors = SelectBestForConnecting(lc, distance, W); var neighbors = SelectBestForConnecting(lc, W);
// add bidirectionall connectionts from neighbors to q at layer lc // add bidirectionall connectionts from neighbors to q at layer lc
// for each e ∈ neighbors // shrink connections if needed // for each e ∈ neighbors // shrink connections if needed
foreach (var e in neighbors) foreach (var e in neighbors)
{ {
// eConn ← neighbourhood(e) at layer lc // eConn ← neighbourhood(e) at layer lc
_layers[lc].AddBidirectionallConnections(q, e.Item1, e.Item2); _layers[lc].AddBidirectionallConnections(q, e.Item1);
// if distance from newNode to newNeighbour is better than to bestPeer => update bestPeer // if distance from newNode to newNeighbour is better than to bestPeer => update bestPeer
if (e.Item2 < epDist) if (e.Item2 < epDist)
{ {
@ -184,8 +185,6 @@ namespace ZeroLevel.HNSW
epDist = e.Item2; epDist = e.Item2;
} }
} }
_layers[lc].TrimLinks(q, lc == 0);
W.Clear(); W.Clear();
} }
} }
@ -218,11 +217,11 @@ namespace ZeroLevel.HNSW
return layer == 0 ? 2 * _options.M : _options.M; return layer == 0 ? 2 * _options.M : _options.M;
} }
private IEnumerable<(int, float)> SelectBestForConnecting(int layer, Func<int, float> distance, IEnumerable<(int, float)> candidates) private IEnumerable<(int, float)> SelectBestForConnecting(int layer, MinHeap candidates)
{ {
if (_options.SelectionHeuristic == NeighbourSelectionHeuristic.SelectSimple) int count = GetM(layer);
return _layers[layer].SELECT_NEIGHBORS_SIMPLE(candidates, GetM(layer)); while (count >= 0 && candidates.Count > 0)
return _layers[layer].SELECT_NEIGHBORS_HEURISTIC(distance, candidates, GetM(layer)); yield return candidates.Pop();
} }
/// <summary> /// <summary>
@ -252,7 +251,7 @@ namespace ZeroLevel.HNSW
for (int layer = L; layer > 0; --layer) for (int layer = L; layer > 0; --layer)
{ {
// W ← SEARCH-LAYER(q, ep, ef = 1, lc) // W ← SEARCH-LAYER(q, ep, ef = 1, lc)
foreach (var i in _layers[layer].KNearestAtLayer(ep, distance, W, 1)) foreach (var i in _layers[layer].KNearestAtLayer(ep, distance, 1))
{ {
W.Push(i); W.Push(i);
} }
@ -264,7 +263,7 @@ namespace ZeroLevel.HNSW
W.Clear(); W.Clear();
} }
// W ← SEARCH-LAYER(q, ep, ef, lc =0) // W ← SEARCH-LAYER(q, ep, ef, lc =0)
foreach (var i in _layers[0].KNearestAtLayer(ep, distance, W, k)) foreach (var i in _layers[0].KNearestAtLayer(ep, distance, k))
{ {
W.Push(i); W.Push(i);
} }
@ -276,7 +275,7 @@ namespace ZeroLevel.HNSW
_lockGraph.ExitReadLock(); _lockGraph.ExitReadLock();
} }
} }
/*
private IEnumerable<(int, float)> KNearest(TItem q, int k, SearchContext context) private IEnumerable<(int, float)> KNearest(TItem q, int k, SearchContext context)
{ {
_lockGraph.EnterReadLock(); _lockGraph.EnterReadLock();
@ -300,7 +299,7 @@ namespace ZeroLevel.HNSW
for (int layer = L; layer > 0; --layer) for (int layer = L; layer > 0; --layer)
{ {
// W ← SEARCH-LAYER(q, ep, ef = 1, lc) // W ← SEARCH-LAYER(q, ep, ef = 1, lc)
foreach (var i in _layers[layer].KNearestAtLayer(ep, distance, W, 1)) foreach (var i in _layers[layer].KNearestAtLayer(ep, distance, 1))
{ {
W.Push(i); W.Push(i);
} }
@ -312,7 +311,7 @@ namespace ZeroLevel.HNSW
W.Clear(); W.Clear();
} }
// W ← SEARCH-LAYER(q, ep, ef, lc =0) // W ← SEARCH-LAYER(q, ep, ef, lc =0)
foreach (var i in _layers[0].KNearestAtLayer(ep, distance, W, k, context)) foreach (var i in _layers[0].KNearestAtLayer(ep, distance, k, context))
{ {
W.Push(i); W.Push(i);
} }
@ -324,7 +323,9 @@ namespace ZeroLevel.HNSW
_lockGraph.ExitReadLock(); _lockGraph.ExitReadLock();
} }
} }
*/
/*
private IEnumerable<(int, float)> KNearest(int k, SearchContext context) private IEnumerable<(int, float)> KNearest(int k, SearchContext context)
{ {
_lockGraph.EnterReadLock(); _lockGraph.EnterReadLock();
@ -349,6 +350,7 @@ namespace ZeroLevel.HNSW
_lockGraph.ExitReadLock(); _lockGraph.ExitReadLock();
} }
} }
*/
#endregion #endregion
public void Serialize(Stream stream) public void Serialize(Stream stream)
@ -378,13 +380,13 @@ namespace ZeroLevel.HNSW
_layers = new Layer<TItem>[countLayers]; _layers = new Layer<TItem>[countLayers];
for (int i = 0; i < countLayers; i++) for (int i = 0; i < countLayers; i++)
{ {
_layers[i] = new Layer<TItem>(_options, _vectors); _layers[i] = new Layer<TItem>(_options, _vectors, i == 0);
_layers[i].Deserialize(reader); _layers[i].Deserialize(reader);
} }
} }
} }
public Histogram GetHistogram(HistogramMode mode = HistogramMode.SQRT) /*public Histogram GetHistogram(HistogramMode mode = HistogramMode.SQRT)
=> _layers[0].GetHistogram(mode); => _layers[0].GetHistogram(mode);*/
} }
} }

1
ZeroLevel.HNSW/SmallWorldFactory.cs
Unescape View File

@ -6,6 +6,5 @@ namespace ZeroLevel.HNSW
{ {
public static SmallWorld<TItem> CreateWorld<TItem>(NSWOptions<TItem> options) => new SmallWorld<TItem>(options); public static SmallWorld<TItem> CreateWorld<TItem>(NSWOptions<TItem> options) => new SmallWorld<TItem>(options);
public static SmallWorld<TItem> CreateWorldFrom<TItem>(NSWOptions<TItem> options, Stream stream) => new SmallWorld<TItem>(options, stream); public static SmallWorld<TItem> CreateWorldFrom<TItem>(NSWOptions<TItem> options, Stream stream) => new SmallWorld<TItem>(options, stream);
public static ReadOnlySmallWorld<TItem> CreateReadOnlyWorldFrom<TItem>(NSWReadOnlyOption<TItem> options, Stream stream) => new ReadOnlySmallWorld<TItem>(options, stream);
} }
} }

115
ZeroLevel.HNSW/Utils/CosineDistance.cs
Unescape View File

@ -47,6 +47,121 @@ namespace ZeroLevel.HNSW
return 1 - similarity; return 1 - similarity;
} }
public static float NonOptimized(byte[] u, byte[] v)
{
if (u.Length != v.Length)
{
throw new ArgumentException("Vectors have non-matching dimensions");
}
float dot = 0.0f;
float nru = 0.0f;
float nrv = 0.0f;
for (int i = 0; i < u.Length; ++i)
{
dot += (float)(u[i] * v[i]);
nru += (float)(u[i] * u[i]);
nrv += (float)(v[i] * v[i]);
}
var similarity = dot / (float)(Math.Sqrt(nru) * Math.Sqrt(nrv));
return 1 - similarity;
}
public static float NonOptimized(int[] u, int[] v)
{
if (u.Length != v.Length)
{
throw new ArgumentException("Vectors have non-matching dimensions");
}
float dot = 0.0f;
float nru = 0.0f;
float nrv = 0.0f;
byte[] bu;
byte[] bv;
for (int i = 0; i < u.Length; ++i)
{
bu = BitConverter.GetBytes(u[i]);
bv = BitConverter.GetBytes(v[i]);
dot += (float)(bu[0] * bv[0]);
nru += (float)(bu[0] * bu[0]);
nrv += (float)(bv[0] * bv[0]);
dot += (float)(bu[1] * bv[1]);
nru += (float)(bu[1] * bu[1]);
nrv += (float)(bv[1] * bv[1]);
dot += (float)(bu[2] * bv[2]);
nru += (float)(bu[2] * bu[2]);
nrv += (float)(bv[2] * bv[2]);
dot += (float)(bu[3] * bv[3]);
nru += (float)(bu[3] * bu[3]);
nrv += (float)(bv[3] * bv[3]);
}
var similarity = dot / (float)(Math.Sqrt(nru) * Math.Sqrt(nrv));
return 1 - similarity;
}
public static float NonOptimized(long[] u, long[] v)
{
if (u.Length != v.Length)
{
throw new ArgumentException("Vectors have non-matching dimensions");
}
float dot = 0.0f;
float nru = 0.0f;
float nrv = 0.0f;
byte[] bu;
byte[] bv;
for (int i = 0; i < u.Length; ++i)
{
bu = BitConverter.GetBytes(u[i]);
bv = BitConverter.GetBytes(v[i]);
dot += (float)(bu[0] * bv[0]);
nru += (float)(bu[0] * bu[0]);
nrv += (float)(bv[0] * bv[0]);
dot += (float)(bu[1] * bv[1]);
nru += (float)(bu[1] * bu[1]);
nrv += (float)(bv[1] * bv[1]);
dot += (float)(bu[2] * bv[2]);
nru += (float)(bu[2] * bu[2]);
nrv += (float)(bv[2] * bv[2]);
dot += (float)(bu[3] * bv[3]);
nru += (float)(bu[3] * bu[3]);
nrv += (float)(bv[3] * bv[3]);
dot += (float)(bu[4] * bv[4]);
nru += (float)(bu[4] * bu[4]);
nrv += (float)(bv[4] * bv[4]);
dot += (float)(bu[5] * bv[5]);
nru += (float)(bu[5] * bu[5]);
nrv += (float)(bv[5] * bv[5]);
dot += (float)(bu[6] * bv[6]);
nru += (float)(bu[6] * bu[6]);
nrv += (float)(bv[6] * bv[6]);
dot += (float)(bu[7] * bv[7]);
nru += (float)(bu[7] * bu[7]);
nrv += (float)(bv[7] * bv[7]);
}
var similarity = dot / (float)(Math.Sqrt(nru) * Math.Sqrt(nrv));
return 1 - similarity;
}
/// <summary> /// <summary>
/// Calculates cosine distance with assumption that u and v are unit vectors. /// Calculates cosine distance with assumption that u and v are unit vectors.
/// </summary> /// </summary>

28
ZeroLevel.sln
Unescape View File

@ -63,10 +63,6 @@ Project("{9A19103F-16F7-4668-BE54-9A1E7A4F7556}") = "ZeroLevel.HNSW", "ZeroLevel
EndProject EndProject
Project("{9A19103F-16F7-4668-BE54-9A1E7A4F7556}") = "HNSWDemo", "TestHNSW\HNSWDemo\HNSWDemo.csproj", "{E0E9EC21-B958-4018-AE30-67DB88EFCB90}" Project("{9A19103F-16F7-4668-BE54-9A1E7A4F7556}") = "HNSWDemo", "TestHNSW\HNSWDemo\HNSWDemo.csproj", "{E0E9EC21-B958-4018-AE30-67DB88EFCB90}"
EndProject EndProject
Project("{FAE04EC0-301F-11D3-BF4B-00C04F79EFBC}") = "temp", "temp\temp.csproj", "{DFE59EBC-B6BC-450C-9D81-394CCAE30498}"
EndProject
Project("{FAE04EC0-301F-11D3-BF4B-00C04F79EFBC}") = "temp2", "temp2\temp2.csproj", "{DEBF2F14-E7F8-40A4-A4A8-87C9033D52A4}"
EndProject
Global Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution GlobalSection(SolutionConfigurationPlatforms) = preSolution
Debug|Any CPU = Debug|Any CPU Debug|Any CPU = Debug|Any CPU
@ -317,30 +313,6 @@ Global
{E0E9EC21-B958-4018-AE30-67DB88EFCB90}.Release|x64.Build.0 = Release|Any CPU {E0E9EC21-B958-4018-AE30-67DB88EFCB90}.Release|x64.Build.0 = Release|Any CPU
{E0E9EC21-B958-4018-AE30-67DB88EFCB90}.Release|x86.ActiveCfg = Release|Any CPU {E0E9EC21-B958-4018-AE30-67DB88EFCB90}.Release|x86.ActiveCfg = Release|Any CPU
{E0E9EC21-B958-4018-AE30-67DB88EFCB90}.Release|x86.Build.0 = Release|Any CPU {E0E9EC21-B958-4018-AE30-67DB88EFCB90}.Release|x86.Build.0 = Release|Any CPU
{DFE59EBC-B6BC-450C-9D81-394CCAE30498}.Debug|Any CPU.ActiveCfg = Debug|Any CPU
{DFE59EBC-B6BC-450C-9D81-394CCAE30498}.Debug|Any CPU.Build.0 = Debug|Any CPU
{DFE59EBC-B6BC-450C-9D81-394CCAE30498}.Debug|x64.ActiveCfg = Debug|Any CPU
{DFE59EBC-B6BC-450C-9D81-394CCAE30498}.Debug|x64.Build.0 = Debug|Any CPU
{DFE59EBC-B6BC-450C-9D81-394CCAE30498}.Debug|x86.ActiveCfg = Debug|Any CPU
{DFE59EBC-B6BC-450C-9D81-394CCAE30498}.Debug|x86.Build.0 = Debug|Any CPU
{DFE59EBC-B6BC-450C-9D81-394CCAE30498}.Release|Any CPU.ActiveCfg = Release|Any CPU
{DFE59EBC-B6BC-450C-9D81-394CCAE30498}.Release|Any CPU.Build.0 = Release|Any CPU
{DFE59EBC-B6BC-450C-9D81-394CCAE30498}.Release|x64.ActiveCfg = Release|Any CPU
{DFE59EBC-B6BC-450C-9D81-394CCAE30498}.Release|x64.Build.0 = Release|Any CPU
{DFE59EBC-B6BC-450C-9D81-394CCAE30498}.Release|x86.ActiveCfg = Release|Any CPU
{DFE59EBC-B6BC-450C-9D81-394CCAE30498}.Release|x86.Build.0 = Release|Any CPU
{DEBF2F14-E7F8-40A4-A4A8-87C9033D52A4}.Debug|Any CPU.ActiveCfg = Debug|Any CPU
{DEBF2F14-E7F8-40A4-A4A8-87C9033D52A4}.Debug|Any CPU.Build.0 = Debug|Any CPU
{DEBF2F14-E7F8-40A4-A4A8-87C9033D52A4}.Debug|x64.ActiveCfg = Debug|Any CPU
{DEBF2F14-E7F8-40A4-A4A8-87C9033D52A4}.Debug|x64.Build.0 = Debug|Any CPU
{DEBF2F14-E7F8-40A4-A4A8-87C9033D52A4}.Debug|x86.ActiveCfg = Debug|Any CPU
{DEBF2F14-E7F8-40A4-A4A8-87C9033D52A4}.Debug|x86.Build.0 = Debug|Any CPU
{DEBF2F14-E7F8-40A4-A4A8-87C9033D52A4}.Release|Any CPU.ActiveCfg = Release|Any CPU
{DEBF2F14-E7F8-40A4-A4A8-87C9033D52A4}.Release|Any CPU.Build.0 = Release|Any CPU
{DEBF2F14-E7F8-40A4-A4A8-87C9033D52A4}.Release|x64.ActiveCfg = Release|Any CPU
{DEBF2F14-E7F8-40A4-A4A8-87C9033D52A4}.Release|x64.Build.0 = Release|Any CPU
{DEBF2F14-E7F8-40A4-A4A8-87C9033D52A4}.Release|x86.ActiveCfg = Release|Any CPU
{DEBF2F14-E7F8-40A4-A4A8-87C9033D52A4}.Release|x86.Build.0 = Release|Any CPU
EndGlobalSection EndGlobalSection
GlobalSection(SolutionProperties) = preSolution GlobalSection(SolutionProperties) = preSolution
HideSolutionNode = FALSE HideSolutionNode = FALSE

10
ZeroLevel/Services/Network/Utils/RequestBuffer.cs
Unescape View File

@ -8,11 +8,11 @@ namespace ZeroLevel.Network
internal sealed class RequestBuffer internal sealed class RequestBuffer
{ {
private ConcurrentDictionary<long, RequestInfo> _requests = new ConcurrentDictionary<long, RequestInfo>(); private ConcurrentDictionary<long, RequestInfo> _requests = new ConcurrentDictionary<long, RequestInfo>();
private static ObjectPool<RequestInfo> _ri_pool = new ObjectPool<RequestInfo>(() => new RequestInfo()); private static Pool<RequestInfo> _ri_pool = new Pool<RequestInfo>(128, (p) => new RequestInfo());
public void RegisterForFrame(int identity, Action<byte[]> callback, Action<string> fail = null) public void RegisterForFrame(int identity, Action<byte[]> callback, Action<string> fail = null)
{ {
var ri = _ri_pool.Allocate(); var ri = _ri_pool.Acquire();
ri.Reset(callback, fail); ri.Reset(callback, fail);
_requests[identity] = ri; _requests[identity] = ri;
} }
@ -23,7 +23,7 @@ namespace ZeroLevel.Network
if (_requests.TryRemove(frameId, out ri)) if (_requests.TryRemove(frameId, out ri))
{ {
ri.Fail(message); ri.Fail(message);
_ri_pool.Free(ri); _ri_pool.Release(ri);
} }
} }
@ -33,7 +33,7 @@ namespace ZeroLevel.Network
if (_requests.TryRemove(frameId, out ri)) if (_requests.TryRemove(frameId, out ri))
{ {
ri.Success(data); ri.Success(data);
_ri_pool.Free(ri); _ri_pool.Release(ri);
} }
} }
@ -53,7 +53,7 @@ namespace ZeroLevel.Network
{ {
if (_requests.TryRemove(frameIds[i], out ri)) if (_requests.TryRemove(frameIds[i], out ri))
{ {
_ri_pool.Free(ri); _ri_pool.Release(ri);
} }
} }
} }

Write Preview
Loading…
Cancel
Save

Powered by TurnKey Linux.