Update softmax with parameterization

src/basicfuns.jl

@@ -232,19 +232,18 @@ That is, `r` is overwritten with `exp.(x)`, normalized to sum to 1.
 
 See the [Wikipedia entry](https://en.wikipedia.org/wiki/Softmax_function)
 """
-function softmax!(r::AbstractArray{R}, x::AbstractArray{T}) where {R<:AbstractFloat,T<:Real}
+function softmax!(r::AbstractArray{<:Real}, x::AbstractArray{<:Real})
     n = length(x)
     length(r) == n || throw(DimensionMismatch("Inconsistent array lengths."))
     u = maximum(x)
-    s = 0.
+    s = zero(eltype(r))
     @inbounds for i = 1:n
         s += (r[i] = exp(x[i] - u))
     end
-    invs = convert(R, inv(s))
     @inbounds for i = 1:n
-        r[i] *= invs
+        r[i] /= s
     end
-    r
+    return r
 end
 
 """
@@ -261,4 +260,4 @@ $(SIGNATURES)
 Return the [`softmax transformation`](https://en.wikipedia.org/wiki/Softmax_function)
 applied to `x`.
 """
-softmax(x::AbstractArray{<:Real}) = softmax!(similar(x, Float64), x)
+softmax(x::AbstractArray{<:Real}) = softmax!(similar(x, float(eltype(x))), x)

test/basicfuns.jl

@@ -149,7 +149,22 @@ end
 @testset "softmax" begin
     x = [1.0, 2.0, 3.0]
     r = exp.(x) ./ sum(exp.(x))
-    @test softmax([1.0, 2.0, 3.0]) ≈ r
+    @test softmax(x) ≈ r
     softmax!(x)
     @test x ≈ r
+
+    x = [1, 2, 3]
+    r = exp.(x) ./ sum(exp.(x))
+    @test softmax(x) ≈ r
+    @test eltype(softmax(x)) == Float64
+
+    x = [1//2, 2//3, 3//4]
+    r = exp.(x) ./ sum(exp.(x))
+    @test softmax(x) ≈ r
+    @test eltype(softmax(x)) == Float64
+
+    x = Float32[1, 2, 3]
+    r = exp.(x) ./ sum(exp.(x))
+    @test softmax(x) ≈ r
+    @test eltype(softmax(x)) == Float32
 end