Use the InverseFunctions.inverse function in Functional transform (#227)

* Use the 'InverseFunctions.inverse' function in Functional transform

* Update tests

* Update docstring

* Update code

---------

Co-authored-by: Júlio Hoffimann <[email protected]>

Author: eliascarv

Project.toml

@@ -10,6 +10,7 @@ CoDa = "5900dafe-f573-5c72-b367-76665857777b"
 ColumnSelectors = "9cc86067-7e36-4c61-b350-1ac9833d277f"
 DataScienceTraits = "6cb2f572-2d2b-4ba6-bdb3-e710fa044d6c"
 Distributions = "31c24e10-a181-5473-b8eb-7969acd0382f"
+InverseFunctions = "3587e190-3f89-42d0-90ee-14403ec27112"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 NelderMead = "2f6b4ddb-b4ff-44c0-b59b-2ab99302f970"
 PrettyTables = "08abe8d2-0d0c-5749-adfa-8a2ac140af0d"
@@ -28,6 +29,7 @@ CoDa = "1.2"
 ColumnSelectors = "0.1"
 DataScienceTraits = "0.1"
 Distributions = "0.25"
+InverseFunctions = "0.1"
 LinearAlgebra = "1.9"
 NelderMead = "0.4"
 PrettyTables = "2"
@@ -36,6 +38,6 @@ Statistics = "1.9"
 StatsBase = "0.33, 0.34"
 Tables = "1.6"
 Transducers = "0.4"
-TransformsBase = "1.2"
+TransformsBase = "1.3"
 Unitful = "1.17"
 julia = "1.9"

src/TableTransforms.jl

@@ -21,13 +21,14 @@ using ColumnSelectors: AllSelector, Column, selector, selectsingle
 using DataScienceTraits: SciType, Continuous, Categorical, coerce
 using Unitful: AbstractQuantity, AffineQuantity, AffineUnits, Units
 using Distributions: ContinuousUnivariateDistribution, Normal
+using InverseFunctions: NoInverse, inverse as invfun
 using StatsBase: AbstractWeights, Weights, sample
 using Transducers: tcollect
 using NelderMead: optimise
 
 import Distributions: quantile, cdf
-import TransformsBase: assertions, isrevertible, preprocess
-import TransformsBase: apply, revert, reapply
+import TransformsBase: assertions, isrevertible, isinvertible
+import TransformsBase: apply, revert, reapply, preprocess, inverse
 
 include("assertions.jl")
 include("tabletraits.jl")

src/transforms/functional.jl

@@ -3,98 +3,71 @@
 # ------------------------------------------------------------------
 
 """
-    Functional(func)
+    Functional(fun)
 
-The transform that applies a `func` elementwise.
+The transform that applies a `fun` elementwise.
 
-    Functional(col₁ => func₁, col₂ => func₂, ..., colₙ => funcₙ)
+    Functional(col₁ => fun₁, col₂ => fun₂, ..., colₙ => funₙ)
 
-Apply the corresponding `funcᵢ` function to each `colᵢ` column.
+Apply the corresponding `funᵢ` function to each `colᵢ` column.
 
 # Examples
 
 ```julia
-Functional(cos)
-Functional(sin)
-Functional(1 => cos, 2 => sin)
-Functional(:a => cos, :b => sin)
-Functional("a" => cos, "b" => sin)
+Functional(exp)
+Functional(log)
+Functional(1 => exp, 2 => log)
+Functional(:a => exp, :b => log)
+Functional("a" => exp, "b" => log)
 ```
 """
 struct Functional{S<:ColumnSelector,F} <: StatelessFeatureTransform
   selector::S
-  func::F
+  fun::F
 end
 
-Functional(func) = Functional(AllSelector(), func)
+Functional(fun) = Functional(AllSelector(), fun)
 
 Functional(pairs::Pair{C}...) where {C<:Column} = Functional(selector(first.(pairs)), last.(pairs))
 
 Functional() = throw(ArgumentError("cannot create Functional transform without arguments"))
 
-# known invertible functions
-inverse(::typeof(log)) = exp
-inverse(::typeof(exp)) = log
-inverse(::typeof(cos)) = acos
-inverse(::typeof(acos)) = cos
-inverse(::typeof(sin)) = asin
-inverse(::typeof(asin)) = sin
-inverse(::typeof(cosd)) = acosd
-inverse(::typeof(acosd)) = cosd
-inverse(::typeof(sind)) = asind
-inverse(::typeof(asind)) = sind
-inverse(::typeof(identity)) = identity
+isrevertible(transform::Functional) = isinvertible(transform)
 
-# fallback to nothing
-inverse(::Any) = nothing
+_hasinverse(f) = !(invfun(f) isa NoInverse)
 
-isrevertible(transform::Functional{AllSelector}) = !isnothing(inverse(transform.func))
+isinvertible(transform::Functional{AllSelector}) = _hasinverse(transform.fun)
+isinvertible(transform::Functional) = all(_hasinverse, transform.fun)
 
-isrevertible(transform::Functional) = all(!isnothing, inverse.(transform.func))
+inverse(transform::Functional{AllSelector}) = Functional(transform.selector, invfun(transform.fun))
+inverse(transform::Functional) = Functional(transform.selector, invfun.(transform.fun))
 
-_funcdict(func, names) = Dict(nm => func for nm in names)
-_funcdict(func::Tuple, names) = Dict(names .=> func)
+_fundict(transform::Functional{AllSelector}, names) = Dict(nm => transform.fun for nm in names)
+_fundict(transform::Functional, names) = Dict(zip(names, transform.fun))
 
 function applyfeat(transform::Functional, feat, prep)
   cols = Tables.columns(feat)
   names = Tables.columnnames(cols)
   snames = transform.selector(names)
-  funcs = _funcdict(transform.func, snames)
+  fundict = _fundict(transform, snames)
 
-  columns = map(names) do nm
-    x = Tables.getcolumn(cols, nm)
-    if nm ∈ snames
-      func = funcs[nm]
-      y = func.(x)
+  columns = map(names) do name
+    x = Tables.getcolumn(cols, name)
+    if name ∈ snames
+      fun = fundict[name]
+      map(fun, x)
     else
-      y = x
+      x
     end
-    y
   end
 
   𝒯 = (; zip(names, columns)...)
   newfeat = 𝒯 |> Tables.materializer(feat)
-  return newfeat, (snames, funcs)
+
+  newfeat, nothing
 end
 
 function revertfeat(transform::Functional, newfeat, fcache)
-  cols = Tables.columns(newfeat)
-  names = Tables.columnnames(cols)
-
-  snames, funcs = fcache
-
-  columns = map(names) do nm
-    y = Tables.getcolumn(cols, nm)
-    if nm ∈ snames
-      func = funcs[nm]
-      invfunc = inverse(func)
-      x = invfunc.(y)
-    else
-      x = y
-    end
-    x
-  end
-
-  𝒯 = (; zip(names, columns)...)
-  𝒯 |> Tables.materializer(newfeat)
+  ofeat, _ = applyfeat(inverse(transform), newfeat, nothing)
+  ofeat
 end

test/metadata.jl

@@ -24,7 +24,7 @@
     mtₒ = revert(T, mn, mc)
     @test mtₒ == mt
 
-    T = Functional(sin)
+    T = Functional(exp)
     mn, mc = apply(T, mt)
     tn, tc = apply(T, t)
     @test mn.meta == m
@@ -33,7 +33,7 @@
     @test mtₒ.meta == mt.meta
     @test Tables.matrix(mtₒ.table) ≈ Tables.matrix(mt.table)
 
-    T = (Functional(sin) → MinMax()) ⊔ Center()
+    T = (Functional(exp) → MinMax()) ⊔ Center()
     mn, mc = apply(T, mt)
     tn, tc = apply(T, t)
     @test mn.meta == m
@@ -68,7 +68,7 @@
     mtₒ = revert(T, mn, mc)
     @test mtₒ == mt
 
-    T = Functional(cos)
+    T = Functional(exp)
     mn, mc = apply(T, mt)
     tn, tc = apply(T, t)
     @test mn.meta == VarMeta(m.data .+ 2)
@@ -79,7 +79,7 @@
 
     # first revertible branch has two transforms,
     # so metadata is increased by 2 + 2 = 4
-    T = (Functional(sin) → MinMax()) ⊔ Center()
+    T = (Functional(exp) → MinMax()) ⊔ Center()
     mn, mc = apply(T, mt)
     tn, tc = apply(T, t)
     @test mn.meta == VarMeta(m.data .+ 4)
@@ -90,7 +90,7 @@
 
     # first revertible branch has one transform,
     # so metadata is increased by 2
-    T = Center() ⊔ (Functional(sin) → MinMax())
+    T = Center() ⊔ (Functional(exp) → MinMax())
     mn, mc = apply(T, mt)
     tn, tc = apply(T, t)
     @test mn.meta == VarMeta(m.data .+ 2)

test/shows.jl

@@ -317,18 +317,18 @@
   end
 
   @testset "Functional" begin
-    T = Functional(sin)
+    T = Functional(log)
 
     # compact mode
     iostr = sprint(show, T)
-    @test iostr == "Functional(all, sin)"
+    @test iostr == "Functional(all, log)"
 
     # full mode
     iostr = sprint(show, MIME("text/plain"), T)
     @test iostr == """
     Functional transform
     ├─ selector = all
-    └─ func = sin"""
+    └─ fun = log"""
   end
 
   @testset "EigenAnalysis" begin
@@ -419,31 +419,31 @@
   @testset "ParallelTableTransform" begin
     t1 = Scale(low=0.3, high=0.6)
     t2 = EigenAnalysis(:VDV)
-    t3 = Functional(cos)
+    t3 = Functional(exp)
     pipeline = t1 ⊔ t2 ⊔ t3
 
     # compact mode
     iostr = sprint(show, pipeline)
-    @test iostr == "Scale(all, 0.3, 0.6) ⊔ EigenAnalysis(:VDV, nothing, 1.0) ⊔ Functional(all, cos)"
+    @test iostr == "Scale(all, 0.3, 0.6) ⊔ EigenAnalysis(:VDV, nothing, 1.0) ⊔ Functional(all, exp)"
 
     # full mode
     iostr = sprint(show, MIME("text/plain"), pipeline)
     @test iostr == """
     ParallelTableTransform
     ├─ Scale(all, 0.3, 0.6)
     ├─ EigenAnalysis(:VDV, nothing, 1.0)
-    └─ Functional(all, cos)"""
+    └─ Functional(all, exp)"""
 
     # parallel and sequential
     f1 = ZScore()
     f2 = Scale()
-    f3 = Functional(cos)
+    f3 = Functional(exp)
     f4 = Interquartile()
     pipeline = (f1 → f2) ⊔ (f3 → f4)
 
     # compact mode
     iostr = sprint(show, pipeline)
-    @test iostr == "ZScore(all) → Scale(all, 0.25, 0.75) ⊔ Functional(all, cos) → Scale(all, 0.25, 0.75)"
+    @test iostr == "ZScore(all) → Scale(all, 0.25, 0.75) ⊔ Functional(all, exp) → Scale(all, 0.25, 0.75)"
 
     # full mode
     iostr = sprint(show, MIME("text/plain"), pipeline)
@@ -453,7 +453,7 @@
     │  ├─ ZScore(all)
     │  └─ Scale(all, 0.25, 0.75)
     └─ SequentialTransform
-       ├─ Functional(all, cos)
+       ├─ Functional(all, exp)
        └─ Scale(all, 0.25, 0.75)"""
   end
 end