fft; hints; typos

Project.toml

@@ -1,7 +1,7 @@
 name = "SimilaritySearch"
 uuid = "053f045d-5466-53fd-b400-a066f88fe02a"
 authors = ["Eric S. Tellez <donsadit@gmail.com>"]
-version = "0.11.7"
+version = "0.11.8"
 
 [deps]
 Dates = "ade2ca70-3891-5945-98fb-dc099432e06a"

src/SimilaritySearch.jl

@@ -248,7 +248,6 @@ function __init__()
     DEFAULT_SEARCH_GRAPH_CONTEXT[] = SearchGraphContext()
 end
 
-
 using PrecompileTools
 
 
@@ -279,5 +278,4 @@ using PrecompileTools
         end
     end
 end
-
 end  # end SimilaritySearch module

src/fft.jl

@@ -26,10 +26,10 @@ function fft(dist::SemiMetric, X::AbstractDatabase, k::Integer; verbose=true)
     dmax::Float32 = typemax(Float32)
     N == 0 && return (; centers, nn, dists=nndists, dmax)
 
-    @inbounds for i in 1:N
+    @inbounds for i in 1:k
         push!(dmaxlist, dmax)
         push!(centers, imax)
-        verbose && println(stderr, "computing fartest point $(length(centers)), dmax: $dmax, imax: $imax, n: $(length(X))")
+        verbose && println(stderr, "computing farthest point $(length(centers)), dmax: $dmax, imax: $imax, n: $(length(X))")
 
         pivot = X[imax]
         @batch minbatch=getminbatch(0, N) for i in 1:N
@@ -41,7 +41,6 @@ function fft(dist::SemiMetric, X::AbstractDatabase, k::Integer; verbose=true)
         end
 
         dmax, imax = findmax(nndists)
-        length(dmaxlist) < k || break
     end
 
     (; centers, nn, dists=nndists, dmax)

src/searchgraph/SearchGraph.jl

@@ -5,7 +5,7 @@ export index!, push_item!
 export Neighborhood, IdentityNeighborhood, DistalSatNeighborhood, SatNeighborhood
 export find_neighborhood
 export BeamSearch, BeamSearchSpace, Callback
-export KDisjointHints, DisjointHints, RandomHints, EpsilonHints
+export KDisjointHints, DisjointHints, RandomHints, EpsilonHints, KCentersHints
 export RandomPruning, KeepNearestPruning, SatPruning, prune!
 
 """

src/searchgraph/context.jl

@@ -100,7 +100,8 @@ end
 function SearchGraphContext(;
         logger=InformativeLog(),
         neighborhood=Neighborhood(SatNeighborhood(0f0)),
-        hints_callback=DisjointHints(),
+        #hints_callback=DisjointHints(),
+        hints_callback=KCentersHints(kfun=x->log(1.2, x)),
         #hints_callback=EpsilonHints(quantile=1/64),
         hyperparameters_callback=OptimizeParameters(),
         parallel_block=4Threads.nthreads(),

src/searchgraph/hints.jl

@@ -152,3 +152,28 @@ function execute_callback(index::SearchGraph, ctx::SearchGraphContext, opt::Epsi
     index.hints .= v
 end
 
+"""
+    mutable struct KCentersHints
+
+Indicates that hints are a small set of objects having a minimal distance between them 
+"""
+mutable struct KCentersHints <: Callback
+    samplesize::Function
+    kfun::Function
+end
+
+KCentersHints(; samplesize=sqrt, kfun=n->log(1.15)) = KCentersHints(samplesize, kfun)
+
+function execute_callback(index::SearchGraph, ctx::SearchGraphContext, opt::KCentersHints)
+    n = length(index)
+    m = min(n, ceil(Int, opt.samplesize(n)))
+    s = rand(1:n, m) |> unique! |> sort!
+    k = min(ceil(Int, opt.kfun(n)), m-1)
+
+    D = SubDatabase(database(index), s)
+    A = fft(distance(index), D, k)
+    @show n, m, k, length(A.centers)
+    resize!(index.hints, k)
+    index.hints .= D.map[A.centers]
+end
+

test/testfft.jl

@@ -4,9 +4,12 @@ using Test, SimilaritySearch, LinearAlgebra
 
 @testset "farthest first traversal" begin 
     dist = L2Distance()
-    X = rand(Float32, 4, 300)
-    res = fft(dist, MatrixDatabase(X), 30)
+    X = rand(Float32, 4, 30)
+    k = 10
+    res = fft(dist, MatrixDatabase(X), k)
+    @test k == length(res.centers)
     @test Set(res.centers) == Set(res.nn)
     @test all(res.dmax .>= res.dists)
 end
 
+