add new kde sampler

Project.toml

@@ -4,6 +4,7 @@ authors = ["Sikorski <[email protected]> and contributors"]
 version = "1.0.0"
 
 [deps]
+AverageShiftedHistograms = "77b51b56-6f8f-5c3a-9cb4-d71f9594ea6e"
 Bonito = "824d6782-a2ef-11e9-3a09-e5662e0c26f8"
 CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
 ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
@@ -52,6 +53,7 @@ Molly = "aa0f7f06-fcc0-5ec4-a7f3-a573f33f9c4c"
 MollyExt = "Molly"
 
 [compat]
+AverageShiftedHistograms = "0.8.9"
 Bonito = "3.1.2"
 CUDA = "5.2.0"
 ChainRulesCore = "1.23.0"

src/iso.jl

@@ -6,7 +6,7 @@
     data::D
     transform
     losses = Float64[]
-    loggers = [autoplot(1)]
+    loggers = Any[autoplot(1)]
     minibatch = 100
 end
 
@@ -20,7 +20,7 @@ function Iso(data;
     gpu=CUDA.has_cuda(),
     autoplot=0,
     validation=nothing,
-    loggers::Vector{Any}=[],
+    loggers=[],
     kwargs...)
 
     opt = Flux.setup(opt, model)

src/simulation.jl

@@ -170,17 +170,20 @@ function chistratcoords(d::SimulationData, model, n; keepedges=false)
 end
 
 
-function resample_kde(data::SimulationData, model, n; bandwidth=0.02, unique=true)
+function resample_kde(data::SimulationData, model, n; bandwidth=0.02, unique=false)
     n == 0 && return data
 
-    sampled = Set(eachcol(data.coords[1]))
-    selinds = unique ? [i for (i, c) in enumerate(eachcol(values(data.coords[2]) |> flattenlast)) if !(c in sampled)] : (:)
-
+    selinds = if unique
+        sampled = Set(eachcol(data.coords[1]))
+        [i for (i, c) in enumerate(eachcol(values(data.coords[2]) |> flattenlast)) if !(c in sampled)]
+    else
+        (:)
+    end
 
     chix = data.features[1] |> model |> vec |> cpu
     chiy = data.features[2] |> flattenlast |> x -> getindex(x, :, selinds) |> model |> vec |> cpu
 
-    iy = resample_kde(chix, chiy, n; bandwidth)
+    iy = resample_kde_ash(chix, chiy, n)
 
     ys = values(data.coords[2]) |> flattenlast |> x -> getindex(x, :, selinds)
     newdata = addcoords(data, ys[:, iy])

src/subsample.jl

@@ -101,6 +101,8 @@ end
 to_pdf(f::Function) = f
 to_pdf(d::Distributions.Distribution) = x -> Distributions.pdf(d, x)
 
+import AverageShiftedHistograms
+
 function kde_needles(xs, n=10; bandwidth, target=Distributions.Uniform())
     xs = copy(xs)
     needles = similar(xs, 0)
@@ -115,3 +117,25 @@ function kde_needles(xs, n=10; bandwidth, target=Distributions.Uniform())
     end
     return needles
 end
+
+function resample_kde_ash(xs, ys, n=10; m=50, target=Distributions.Uniform())
+    iys = zeros(Int, n)
+    rng = 0:0.001:1
+    kde = AverageShiftedHistograms.ash(xs; rng, m)
+    #display(kde)
+    target = to_pdf(target)(rng)
+    for i in 1:n
+        @show chi = rng[argmax(target - kde.density)] # position of maximal difference to target pdf
+        min = Inf
+        local iy
+        for j in 1:length(ys)
+            if abs(ys[j] - chi) < min && !(j in iys)
+                min = abs(ys[j] - chi)
+                iy = j
+            end
+        end
+        AverageShiftedHistograms.ash!(kde, ys[iy])
+        iys[i] = iy
+    end
+    return iys
+end