Make round(::Type{Bool}, ::AbstractFloat) work

[oxinabox]

Bool is the only integer subtype that never works with round, with a Float.

From #gripes on slack:

Lyndon White

I feel like round(Bool, 0.7) should return true. Not give an error

Alex Arslan

Why?
An error seems much more reasonable IMO

Lyndon White

Because Bool <: Number
we basically use bools as UInt1

Alex Arslan

Bool as a Number type is like the 2 of prime numbers

Lyndon White

accuracy = mean(predictions==ground_truth)
works great
Bool(round(UInt8, 0.7)) works without error
as round(Bool, 1) works (edited)
Bool is like the only integer subtype we are apparently never allowed to round a float to.

Simon Byrne

I'd be on board with that
We generally take round(T, x) to be equivalent to convert(T, round(x)).

Jacob Quinn

yeah, seems reasonable to me.

Stefan Karpinski

As long as Bool is an integer type it should

[ararslan]

round(::Type{Bool}, x::AbstractFloat) = Bool(round(UInt8, x))

seems like a fairly reasonable definition.

[oxinabox]

Need to pass:

@testset "Round Bool" begin
    @test_throws InexactError round(Bool, -4.1)
    @test_throws InexactError round(Bool, 7.3)
    @test true == round(Bool, 1.0)
    @test false == round(Bool, 0.0)
    @test true == round(Bool, 0.6)
    @test false == round(Bool, 0.4)
end

Which

round(::Type{Bool}, x::AbstractFloat) = Bool(round(UInt8, x))`

Does do so.
Seems best.

Base.round(::Type{Bool}, x::AbstractFloat) = x>0.5 fails the first two tests.