public static AllResults GroupIntoClusters(AllResults loaded, int imagePerCandidateMaxTreshold = 2000, int imagesPerCandidateMinTreshold = 4)
{
imgColors.Clear();
totalColors = 0;
var directional = loaded.Rows.SelectMany(x => x.Hits.Select(y => new { x.Query, y.Hit })).Distinct().ToList();
neighbours = directional.Union(directional.Select(a => new { Query = a.Hit, Hit = a.Query }))
.ToLookup(x => x.Query, x => x.Hit);
stack = new Stack <Patch>(neighbours.Select(x => x.Key));
while (stack.Count > 0)
{
Visit(stack.Pop());
}
var stats = imgColors.GroupBy(x => x.Value)
.Where(x => x.Select(z => z.Key.ImageId).Distinct().Count() < imagePerCandidateMaxTreshold &&
x.Select(y => y.Key.ImageId).Distinct().Count() >= imagesPerCandidateMinTreshold);
Console.WriteLine("Having " + stats.Count() + " nice clusters");
var rareImgs = new HashSet <Patch>(stats.SelectMany(x => x.Select(y => y.Key)));
var artificalResults = new AllResults {
ImageEncoding = loaded.ImageEncoding, PatchEncoding = loaded.PatchEncoding
};
var colorGroups = loaded.Rows
.Where(r => rareImgs.Contains(r.Query))
.GroupBy(x => imgColors[x.Query]);
foreach (var g in colorGroups)
{
artificalResults.Rows.Add(new ResultsRow
{
Query = g.First().Query,
Hits = g.SelectMany(x => x.Hits.Concat(new[] { new SearchHit {
Hit = x.Query, Distance = g.Max(colorGroup => colorGroup.Hits.Min(h => h.Distance))
} }))
.GroupBy(x => x.Hit)
.Select(x => new SearchHit {
Hit = x.Key, Distance = x.Min(y => y.Distance)
})
.Distinct()
.ToArray()
});
}
artificalResults.RefreshReferenceMap();
return(artificalResults);
}
private static string Filter(AllResults loadedFile, string file)
{
for (int i = 0; i < 31; i++)
{
Console.WriteLine($"iteration {i} starting");
loadedFile.RefreshReferenceMap();
loadedFile.RefBasedShrink();
}
var newName = file.Replace(".bin", "-refShrink" + 30 + ".bin");
loadedFile.Save(newName);
return(newName);
}
public static void AgglomerativeClustering(AllResults loaded)
{
var imgName = "241666.jpg"; //"159161.jpg";
var imgId = loaded.ImageEncoding[imgName];
var relevantRows = loaded.Rows
//.Where(r => r.Query.ImageId == imgId)
.ToList();
// cluster all of them together? Include query into dissimilarity function then
// Or product by product, filter down to big elements, and offer to transitively load more and more?
var metric = new ResultsRowSetBasedDistance();
var linkage = new AverageLinkage <ResultsRow>(metric);
var algorithm = new AgglomerativeClusteringAlgorithm <ResultsRow>(linkage);
var clusters = algorithm.GetClustering(new HashSet <ResultsRow>(relevantRows));
clusters.SaveToCsv(@"G:\siret\zoot\protobuf\clustertest.csv");
//RenderData();
var dummyResults = new AllResults {
ImageEncoding = loaded.ImageEncoding, PatchEncoding = loaded.PatchEncoding
};
var clusterQueue = new Queue <Cluster <ResultsRow> >(new[] { clusters.SingleCluster });
while (clusterQueue.Count > 0)
{
var item = clusterQueue.Dequeue();
if (item.Dissimilarity <= 0.70 && item.Count < 50)
{
dummyResults.Rows.Add(new ResultsRow
{
Query = item.First().Query,
Hits = item.SelectMany(x => x.Hits)
.GroupBy(x => x.Hit)
.Select(x => new SearchHit {
Hit = x.Key, Distance = x.Min(y => y.Distance)
})
.Concat(item.Select(i => new SearchHit {
Hit = i.Query, Distance = -1
}))
.ToArray()
});
}
else
{
clusterQueue.Enqueue(item.Parent1);
clusterQueue.Enqueue(item.Parent2);
}
}
loaded.RefreshReferenceMap();
foreach (var k in AllResults.ReferenceMap.Keys)
{
AllResults.ReferenceMap[k][k] = 1;
}
using (var sw = new StreamWriter(@"G:\siret\zoot\protobuf\clusteringTestMega.html", append: false))
{
dummyResults.Render(sw);
}
}
