Merge pull request #226 from xinadi/dev

Float64 replaced by AbstractFloat for regression

Project.toml

@@ -2,7 +2,7 @@ name = "DecisionTree"
 uuid = "7806a523-6efd-50cb-b5f6-3fa6f1930dbb"
 license = "MIT"
 desc = "Julia implementation of Decision Tree (CART) and Random Forest algorithms"
-version = "0.12.3"
+version = "0.12.4"
 
 [deps]
 AbstractTrees = "1520ce14-60c1-5f80-bbc7-55ef81b5835c"

src/classification/main.jl

@@ -93,7 +93,7 @@ function update_pruned_impurity!(
     feature_importance::Vector{Float64},
     ntt::Int,
     loss::Function=mean_squared_error,
-) where {S,T<:Float64}
+) where {S,T<:AbstractFloat}
     μl = mean(tree.left.values)
     nl = length(tree.left.values)
     μr = mean(tree.right.values)
@@ -220,7 +220,7 @@ See also [`build_tree`](@ref).
 function prune_tree(
     tree::Union{Root{S,T},LeafOrNode{S,T}},
     purity_thresh=1.0,
-    loss::Function=T <: Float64 ? mean_squared_error : util.entropy,
+    loss::Function=T <: AbstractFloat ? mean_squared_error : util.entropy,
 ) where {S,T}
     if purity_thresh >= 1.0
         return tree
@@ -293,11 +293,7 @@ function apply_tree(tree::LeafOrNode{S,T}, features::AbstractMatrix{S}) where {S
     for i in 1:N
         predictions[i] = apply_tree(tree, features[i, :])
     end
-    if T <: Float64
-        return Float64.(predictions)
-    else
-        return predictions
-    end
+    return predictions
 end
 
 """
@@ -343,7 +339,7 @@ end
 Train a random forest model, built on standard CART decision trees, using the specified
 `labels` (target) and `features` (patterns). Here:
 
-- `labels` is any `AbstractVector`. If the element type is `Float64`, regression is
+- `labels` is any `AbstractVector`. If the element type is `AbstractFloat`, regression is
   applied, and otherwise classification is applied.
 
 - `features` is any `AbstractMatrix{T}` where `T` supports ordering with `<` (unordered
@@ -619,7 +615,7 @@ function apply_forest(forest::Ensemble{S,T}, features::AbstractVector{S}) where
         votes[i] = apply_tree(forest.trees[i], features)
     end
 
-    if T <: Float64
+    if T <: AbstractFloat
         return mean(votes)
     else
         return majority_vote(votes)

src/measures.jl

@@ -269,7 +269,7 @@ function _nfoldCV(
     args...;
     verbose,
     rng,
-) where {T<:Float64}
+) where {T<:AbstractFloat}
     _rng = mk_rng(rng)::Random.AbstractRNG
     nfolds = args[1]
     if nfolds < 2
@@ -361,7 +361,7 @@ function nfoldCV_tree(
     min_purity_increase::Float64=0.0;
     verbose::Bool=true,
     rng=Random.GLOBAL_RNG,
-) where {S,T<:Float64}
+) where {S,T<:AbstractFloat}
     _nfoldCV(
         :tree,
         labels,
@@ -389,7 +389,7 @@ function nfoldCV_forest(
     min_purity_increase::Float64=0.0;
     verbose::Bool=true,
     rng=Random.GLOBAL_RNG,
-) where {S,T<:Float64}
+) where {S,T<:AbstractFloat}
     _nfoldCV(
         :forest,
         labels,

src/regression/main.jl

@@ -1,6 +1,6 @@
 include("tree.jl")
 
-function _convert(node::treeregressor.NodeMeta{S}, labels::Array{T}) where {S,T<:Float64}
+function _convert(node::treeregressor.NodeMeta{S}, labels::Array{T}) where {S,T<:AbstractFloat}
     if node.is_leaf
         return Leaf{T}(node.label, labels[node.region])
     else
@@ -27,7 +27,7 @@ function build_stump(
     features::AbstractMatrix{S};
     rng=Random.GLOBAL_RNG,
     impurity_importance::Bool=true,
-) where {S,T<:Float64}
+) where {S,T<:AbstractFloat}
     return build_tree(labels, features, 0, 1; rng, impurity_importance)
 end
 
@@ -41,7 +41,7 @@ function build_tree(
     min_purity_increase=0.0;
     rng=Random.GLOBAL_RNG,
     impurity_importance::Bool=true,
-) where {S,T<:Float64}
+) where {S,T<:AbstractFloat}
     if max_depth == -1
         max_depth = typemax(Int)
     end
@@ -85,7 +85,7 @@ function build_forest(
     min_purity_increase=0.0;
     rng::Union{Integer,AbstractRNG}=Random.GLOBAL_RNG,
     impurity_importance::Bool=true,
-) where {S,T<:Float64}
+) where {S,T<:AbstractFloat}
     if n_trees < 1
         throw("the number of trees must be >= 1")
     end

src/regression/tree.jl

@@ -47,7 +47,7 @@ end
 # (max_depth, min_samples_split, min_purity_increase)
 function _split!(
     X::AbstractMatrix{S}, # the feature array
-    Y::AbstractVector{Float64}, # the label array
+    Y::AbstractVector{T}, # the label array
     W::AbstractVector{U},
     node::NodeMeta{S}, # the node to split
     max_features::Int, # number of features to consider
@@ -59,10 +59,10 @@ function _split!(
     # we split using samples in indX[node.region]
     # the two arrays below are given for optimization purposes
     Xf::AbstractVector{S},
-    Yf::AbstractVector{Float64},
+    Yf::AbstractVector{T},
     Wf::AbstractVector{U},
     rng::Random.AbstractRNG,
-) where {S,U}
+) where {S,T<:AbstractFloat,U}
     region = node.region
     n_samples = length(region)
     r_start = region.start - 1
@@ -245,18 +245,18 @@ end
 
 function _fit(
     X::AbstractMatrix{S},
-    Y::AbstractVector{Float64},
+    Y::AbstractVector{T},
     W::AbstractVector{U},
     max_features::Int,
     max_depth::Int,
     min_samples_leaf::Int,
     min_samples_split::Int,
     min_purity_increase::Float64,
     rng=Random.GLOBAL_RNG::Random.AbstractRNG,
-) where {S,U}
+) where {S,T<:AbstractFloat,U}
     n_samples, n_features = size(X)
 
-    Yf = Array{Float64}(undef, n_samples)
+    Yf = Array{T}(undef, n_samples)
     Xf = Array{S}(undef, n_samples)
     Wf = Array{U}(undef, n_samples)
 
@@ -293,7 +293,7 @@ end
 
 function fit(;
     X::AbstractMatrix{S},
-    Y::AbstractVector{Float64},
+    Y::AbstractVector{<:AbstractFloat},
     W::Union{Nothing,AbstractVector{U}},
     max_features::Int,
     max_depth::Int,

test/regression/low_precision.jl

@@ -113,6 +113,9 @@
     model = build_forest(labels, features)
     preds = apply_forest(model, features)
     @test typeof(preds) == Vector{Float16}
+    # Verify that the `preds` were calculated based on `labels` of the same type.
+    # If the code at some point converts the numbers to, say, `Float64`, then this test will fail.
+    @test !all(x->(x in labels), preds)
 
     preds_MT = apply_forest(model, features; use_multithreading=true)
     @test typeof(preds_MT) == Vector{Float16}