Fix more test results

Author: tecosaur

src/DecisionTree.jl

@@ -27,7 +27,7 @@ export InfoNode, InfoLeaf, wrap
 ########## Types ##########
 
 struct Leaf{T, N}
-    features :: NTuple{N, T}
+    classes :: NTuple{N, T}
     majority :: Int
     values   :: NTuple{N, Int}
     total    :: Int
@@ -54,15 +54,20 @@ struct Ensemble{S, T, N}
     featim  :: Vector{Float64}
 end
 
-Leaf(features::NTuple{T, N}) where {T, N} =
+Leaf(features::NTuple{N, T}) where {T, N} =
     Leaf(features, 0, Tuple(zeros(T, N)), 0)
+Leaf(features::NTuple{N, T}, frequencies::NTuple{N, Int}) where {T, N} =
+    Leaf(features, argmax(frequencies), frequencies, sum(frequencies))
+Leaf(features::Union{<:AbstractVector, <:Tuple},
+     frequencies::Union{<:AbstractVector{Int}, <:Tuple}) =
+    Leaf(Tuple(features), Tuple(frequencies))
 
 is_leaf(l::Leaf) = true
 is_leaf(n::Node) = false
 
 _zero(::Type{String}) = ""
 _zero(x::Any) = zero(x)
-convert(::Type{Node{S, T, N}}, lf::Leaf{T, N}) where {S, T, N} = Node(0, _zero(S), lf, Leaf(lf.features))
+convert(::Type{Node{S, T, N}}, lf::Leaf{T, N}) where {S, T, N} = Node(0, _zero(S), lf, Leaf(lf.classes))
 convert(::Type{Root{S, T, N}}, node::LeafOrNode{S, T, N}) where {S, T, N} = Root{S, T, N}(node, 0, Float64[])
 convert(::Type{LeafOrNode{S, T, N}}, tree::Root{S, T, N}) where {S, T, N} = tree.node
 promote_rule(::Type{Node{S, T, N}}, ::Type{Leaf{T, N}}) where {S, T, N} = Node{S, T, N}
@@ -101,7 +106,7 @@ depth(tree::Node) = 1 + max(depth(tree.left), depth(tree.right))
 depth(tree::Root) = depth(tree.node)
 
 function print_tree(io::IO, leaf::Leaf, depth=-1, indent=0; sigdigits=4, feature_names=nothing)
-    println(io, leaf.features[leaf.majority], " : ",
+    println(io, leaf.classes[leaf.majority], " : ",
             leaf.values[leaf.majority], '/', leaf.total)
 end
 function print_tree(leaf::Leaf, depth=-1, indent=0; sigdigits=4, feature_names=nothing)
@@ -165,7 +170,7 @@ end
 
 function show(io::IO, leaf::Leaf)
     println(io, "Decision Leaf")
-    println(io, "Majority: ", leaf.features[leaf.majority])
+    println(io, "Majority: ", leaf.classes[leaf.majority])
     print(io,   "Samples:  ", leaf.total)
 end
 

src/classification/main.jl

@@ -117,6 +117,7 @@ function build_stump(
         labels              :: AbstractVector{T},
         features            :: AbstractMatrix{S},
         weights              = nothing;
+        n_classes           :: Int = length(unique(labels)),
         rng                  = Random.GLOBAL_RNG,
         impurity_importance :: Bool = true) where {S, T}
 
@@ -133,7 +134,7 @@ function build_stump(
         min_purity_increase = 0.0;
         rng                 = rng)
 
-    return _build_tree(t, labels, size(features, 2), size(features, 1), impurity_importance)
+    return _build_tree(t, labels, n_classes, size(features, 2), size(features, 1), impurity_importance)
 end
 
 function build_tree(
@@ -144,6 +145,7 @@ function build_tree(
         min_samples_leaf     = 1,
         min_samples_split    = 2,
         min_purity_increase  = 0.0;
+        n_classes           :: Int = length(unique(labels)),
         loss                 = util.entropy :: Function,
         rng                  = Random.GLOBAL_RNG,
         impurity_importance :: Bool = true) where {S, T}
@@ -168,18 +170,18 @@ function build_tree(
         min_purity_increase = Float64(min_purity_increase),
         rng                 = rng)
 
-    return _build_tree(t, labels, size(features, 2), size(features, 1), impurity_importance)
+    return _build_tree(t, labels, n_classes, size(features, 2), size(features, 1), impurity_importance)
 end
 
 function _build_tree(
     tree::treeclassifier.Tree{S, T},
     labels::AbstractVector{T},
+    n_classes::Int,
     n_features,
     n_samples,
     impurity_importance::Bool
 ) where {S, T}
     node = _convert(tree.root, tree.list, labels[tree.labels])
-    n_classes = unique(labels) |> length
     if !impurity_importance
         return Root{S, T, n_classes}(node, n_features, Float64[])
     else
@@ -237,15 +239,15 @@ function prune_tree(
             if !isempty(fi)
                 update_pruned_impurity!(tree, fi, ntt, loss)
             end
-            return Leaf{T, N}(tree.left.features, majority, combined, total)
+            return Leaf{T, N}(tree.left.classes, majority, combined, total)
         else
             return tree
         end
     end
     function _prune_run(tree::Root{S, T, N}, purity_thresh::Real) where {S, T, N}
         fi = deepcopy(tree.featim) ## recalculate feature importances
         node = _prune_run(tree.node, purity_thresh, fi)
-        return Root{S, T, N}(node, fi)
+        return Root{S, T, N}(node, tree.n_feat, fi)
     end
     function _prune_run(
         tree::LeafOrNode{S, T, N},
@@ -273,7 +275,7 @@ function prune_tree(
 end
 
 
-apply_tree(leaf::Leaf, feature::AbstractVector) = leaf.features[leaf.majority]
+apply_tree(leaf::Leaf, feature::AbstractVector) = leaf.classes[leaf.majority]
 apply_tree(
     tree::Root{S, T},
     features::AbstractVector{S}
@@ -369,10 +371,11 @@ function build_forest(
 
     t_samples = length(labels)
     n_samples = floor(Int, partial_sampling * t_samples)
+    n_classes = length(unique(labels))
 
     forest = impurity_importance ?
-        Vector{Root{S, T}}(undef, n_trees) :
-        Vector{LeafOrNode{S, T}}(undef, n_trees)
+        Vector{Root{S, T, n_classes}}(undef, n_trees) :
+        Vector{LeafOrNode{S, T, n_classes}}(undef, n_trees)
 
     entropy_terms = util.compute_entropy_terms(n_samples)
     loss = (ns, n) -> util.entropy(ns, n, entropy_terms)
@@ -392,7 +395,8 @@ function build_forest(
                 max_depth,
                 min_samples_leaf,
                 min_samples_split,
-                min_purity_increase,
+                min_purity_increase;
+                n_classes,
                 loss = loss,
                 rng = _rng,
                 impurity_importance = impurity_importance)
@@ -408,7 +412,8 @@ function build_forest(
                 max_depth,
                 min_samples_leaf,
                 min_samples_split,
-                min_purity_increase,
+                min_purity_increase;
+                n_classes,
                 loss = loss,
                 impurity_importance = impurity_importance)
         end
@@ -418,13 +423,13 @@ function build_forest(
 end
 
 function _build_forest(
-        forest              :: Vector{<: Union{Root{S, T}, LeafOrNode{S, T}}},
+        forest              :: Vector{<: Union{Root{S, T, N}, LeafOrNode{S, T, N}}},
         n_features          ,
         n_trees             ,
-        impurity_importance :: Bool) where {S, T}
+        impurity_importance :: Bool) where {S, T, N}
 
     if !impurity_importance
-        return Ensemble{S, T}(forest, n_features, Float64[])
+        return Ensemble{S, T, N}(forest, n_features, Float64[])
     else
         fi = zeros(Float64, n_features)
         for tree in forest
@@ -434,12 +439,12 @@ function _build_forest(
             end
         end
 
-        forest_new = Vector{LeafOrNode{S, T}}(undef, n_trees)
+        forest_new = Vector{LeafOrNode{S, T, N}}(undef, n_trees)
         Threads.@threads for i in 1:n_trees
             forest_new[i] = forest[i].node
         end
 
-        return Ensemble{S, T}(forest_new, n_features, fi ./ n_trees)
+        return Ensemble{S, T, N}(forest_new, n_features, fi ./ n_trees)
     end
 end
 
@@ -516,11 +521,13 @@ function build_adaboost_stumps(
     stumps = Node{S, T}[]
     coeffs = Float64[]
     n_features = size(features, 2)
+    n_classes = length(unique(labels))
     for i in 1:n_iterations
         new_stump = build_stump(
             labels,
             features,
             weights;
+            n_classes,
             rng=mk_rng(rng),
             impurity_importance=false
         )

test/miscellaneous/abstract_trees_test.jl

@@ -17,9 +17,9 @@ clabel_pattern(clabel)  = "─ " * clabel * " ("      # class labels are embedde
 check_occurence(str_tree, pool, pattern) = count(map(elem -> occursin(pattern(elem), str_tree), pool)) == length(pool)
 
 @info("Test base functionality")    
-l1 = Leaf(1, [1,1,2])
-l2 = Leaf(2, [1,2,2])
-l3 = Leaf(3, [3,3,1])
+l1 = Leaf((1,2,3), 1, (2, 1, 0), 3)
+l2 = Leaf((1,2,3), 2, (1, 2, 0), 3)
+l3 = Leaf((1,2,3), 3, (1, 0, 2), 3)
 n2 = Node(2, 0.5, l2, l3)
 n1 = Node(1, 0.7, l1, n2)
 feature_names = ["firstFt", "secondFt"]
@@ -81,4 +81,4 @@ end
 traverse_tree(leaf::InfoLeaf) = nothing
 
 traverse_tree(wrapped_tree)
-end
+end

test/miscellaneous/convert.jl

@@ -2,7 +2,7 @@
 
 @testset "convert.jl" begin
 
-lf = Leaf(1, [1])
+lf = Leaf([1], [1])
 nv = Node{Int, Int}[]
 rv = Root{Int, Int}[]
 push!(nv, lf)
@@ -22,7 +22,7 @@ push!(rv, nv[1])
 @test apply_tree(rv[1], [0]) == 1.0
 @test apply_tree(rv[2], [0]) == 1.0
 
-lf = Leaf("A", ["B", "A"])
+lf = Leaf(["A", "B"], [2, 1])
 nv = Node{Int, String}[]
 rv = Root{Int, String}[]
 push!(nv, lf)