uses objects instead of names to identify function errors

sadit · Jan 15, 2022 · c5cbbde · c5cbbde · sadit · Jan 15, 2022
1 parent 34f0f0e
commit c5cbbde
Show file tree

Hide file tree

Showing 3 changed files with 22 additions and 19 deletions.
diff --git a/Project.toml b/Project.toml
@@ -1,7 +1,7 @@
 name = "SimilaritySearch"
 uuid = "053f045d-5466-53fd-b400-a066f88fe02a"
 authors = ["Eric S. Tellez <donsadit@gmail.com>"]
-version = "0.8.6"
+version = "0.8.7"
 
 [deps]
 Dates = "ade2ca70-3891-5945-98fb-dc099432e06a"

diff --git a/src/graph/opt.jl b/src/graph/opt.jl
@@ -3,8 +3,12 @@
 using SearchModels, Random
 using StatsBase
 import SearchModels: combine, mutate
-export OptimizeParameters, optimize!, BeamSearchSpace
+export OptimizeParameters, optimize!, BeamSearchSpace, MinRecall, ParetoRecall, ParetoRadius
 
+abstract type ErrorFunction end
+struct MinRecall <: ErrorFunction end
+struct ParetoRecall <: ErrorFunction end
+struct ParetoRadius <: ErrorFunction end
 
 @with_kw struct BeamSearchSpace <: AbstractSolutionSpace
     bsize = 8:8:64
@@ -31,12 +35,12 @@ function mutate(space::BeamSearchSpace, c::BeamSearch, iter)
 end
 
 @with_kw mutable struct OptimizeParameters <: Callback
-    kind = :pareto_recall_searchtime # :pareto_distance_searchtime, :pareto_recall_searchtime, :minimum_recall_searchtime
+    kind::ErrorFunction = ParetoRecall()
     initialpopulation = 16
     params = SearchParams(maxpopulation=16, bsize=4, mutbsize=16, crossbsize=8, tol=-1.0, maxiters=16)
     ksearch::Int32 = 10
     numqueries::Int32 = 64
-    minrecall = 0.9  # used with :minimum_recall_searchtime
+    minrecall = 0.9  # used with MinRecall()
     space::BeamSearchSpace = BeamSearchSpace()
 end
 
@@ -130,9 +134,8 @@ function optimize!(
         queries = SubDatabase(index.db, sample)
     end
 
-    recall_options = (:pareto_recall_searchtime, :minimum_recall_searchtime)
     knnlist = [KnnResult(opt.ksearch) for i in eachindex(queries)]
-    gold = if opt.kind in recall_options
+    gold = if opt.kind isa ParetoRecall || opt.kind isa MinRecall
         db = @view index.db[1:length(index)]
         seq = ExhaustiveSearch(index.dist, db)
         searchbatch(seq, queries, knnlist; parallel=true)
@@ -156,9 +159,9 @@ function optimize!(
 
     function geterr(p)
         cost = p.visited[2] / M[]
-        if opt.kind === :pareto_recall_searchtime    
+        if opt.kind isa ParetoRecall 
             cost^2 + (1.0 - p.recall)^2
-        elseif opt.kind === :minimum_recall_searchtime
+        elseif opt.kind isa MinRecall
             p.recall < opt.minrecall ? 3.0 - 2 * p.recall : cost
         else
             _kfun(cost) + _kfun(p.radius[2] / R[])

diff --git a/test/testsearchgraph.jl b/test/testsearchgraph.jl
@@ -38,32 +38,32 @@ using Test
         @info "===="
     end
 
-    @info "--- Optimizing parameters :pareto_distance_searchtime ---"
+    @info "--- Optimizing parameters ParetoRadius ---"
     graph = SearchGraph(; dist, search_algo=BeamSearch(bsize=2), verbose=false)
     graph.neighborhood.reduce = SatNeighborhood()
     append!(graph, db)
-    @info "---- starting :pareto_distance_searchtime optimization ---"
+    @info "---- starting ParetoRadius optimization ---"
     optimize!(graph, OptimizeParameters())
     I, D, searchtime = timedsearchbatch(graph, queries, ksearch)
     recall = macrorecall(goldI, I)
-    @info ":pareto_distance_search_time:> queries per second: ", 1/searchtime, ", recall:", recall
+    @info "ParetoRadius:> queries per second: ", 1/searchtime, ", recall:", recall
     @info graph.search_algo
     @test recall >= 0.6
 
 
-    @info "---- starting :pareto_recall_searchtime optimization ---"
-    optimize!(graph, OptimizeParameters(kind=:pareto_recall_searchtime))
+    @info "---- starting ParetoRecall optimization ---"
+    optimize!(graph, OptimizeParameters(kind=ParetoRecall()))
     I, D, searchtime = timedsearchbatch(graph, queries, ksearch)
     recall = macrorecall(goldI, I)
-    @info ":pareto_recall_search_time:> queries per second: ", 1/searchtime, ", recall:", recall
+    @info "ParetoRecall:> queries per second: ", 1/searchtime, ", recall:", recall
     @info graph.search_algo
     @test recall >= 0.6
 
     @info "========================= Callback optimization ======================"
-    @info "--- Optimizing parameters :pareto_distance_searchtime ---"
+    @info "--- Optimizing parameters ParetoRadius ---"
     graph = SearchGraph(; db, dist, search_algo=BeamSearch(bsize=2), verbose=false)
     graph.neighborhood.reduce = SatNeighborhood()
-    push!(graph.callbacks, OptimizeParameters(kind=:pareto_distance_searchtime))
+    push!(graph.callbacks, OptimizeParameters(kind=ParetoRadius()))
     index!(graph)
     I, D, searchtime = timedsearchbatch(graph, queries, ksearch)
     recall = macrorecall(goldI, I)
@@ -72,17 +72,17 @@ using Test
     @test recall >= 0.6
 
     @info "#############=========== Callback optimization 2 ==========###########"
-    @info "--- Optimizing parameters :pareto_distance_searchtime L2 ---"
+    @info "--- Optimizing parameters ParetoRadius L2 ---"
     dim = 4
     db = MatrixDatabase(ceil.(Int32, rand(Float32, dim, n) .* 100))
     queries = VectorDatabase(ceil.(Int32, rand(Float32, dim, m) .* 100))
     seq = ExhaustiveSearch(dist, db)
     goldI, goldD = searchbatch(seq, queries, ksearch)
     graph = SearchGraph(; db, dist, search_algo=BeamSearch(bsize=2), verbose=false)
     graph.neighborhood.reduce = SatNeighborhood()
-    push!(graph.callbacks, OptimizeParameters(kind=:pareto_recall_searchtime))
+    push!(graph.callbacks, OptimizeParameters(kind=ParetoRecall()))
     index!(graph)
-    #optimize!(graph, OptimizeParameters(kind=:minimum_recall_searchtime, minrecall=0.7))
+    #optimize!(graph, OptimizeParameters(kind=MinRecall(), minrecall=0.7))
     I, D, searchtime = timedsearchbatch(graph, queries, ksearch)
     recall = macrorecall(goldI, I)
     @info "testing without additional optimizations> queries per second:", 1/searchtime, ", recall: ", recall