Merge #1089

1089: Fix broadcasting `-` with booleans  r=mcabbott a=mcabbott

Closes #1086

The issue is that `unbroadcast` collapses the gradient of a non-differentiable argument to `nothing`, Zygote's marker for such things, which the rule then tried to negate. (This is why CRC defines its own zero types for this purpose, instead of using nothing). I presume the original reason for `-unbroadcast(y, Δ)` not `unbroadcast(y, -Δ)` is to save an allocation when `y` is a scalar etc.

I did not see any other cases where further operations are done after `unbroadcast`. 

Co-authored-by: Michael Abbott <32575566+mcabbott@users.noreply.github.com>

Project.toml

@@ -1,6 +1,6 @@
 name = "Zygote"
 uuid = "e88e6eb3-aa80-5325-afca-941959d7151f"
-version = "0.6.25"
+version = "0.6.26"
 
 [deps]
 AbstractFFTs = "621f4979-c628-5d54-868e-fcf4e3e8185c"

src/lib/broadcast.jl

@@ -72,7 +72,9 @@ unbroadcast(x::AbstractArray, x̄::Nothing) = nothing
   broadcast(+, xs...), ȳ -> (nothing, map(x -> unbroadcast(x, ȳ), xs)...)
 
 @adjoint broadcasted(::typeof(-), x::Numeric, y::Numeric) = x .- y,
-  Δ -> (nothing, unbroadcast(x, Δ), -unbroadcast(y, Δ))
+  Δ -> (nothing, unbroadcast(x, Δ), _minus(unbroadcast(y, Δ)))
+_minus(Δ) = -Δ
+_minus(::Nothing) = nothing
 
 @adjoint broadcasted(::typeof(*), x::Numeric, y::Numeric) = x.*y,
    Δ -> (nothing, unbroadcast(x, Δ .* conj.(y)), unbroadcast(y, Δ .* conj.(x)))

test/features.jl

@@ -570,6 +570,11 @@ end
   @test gradient(x -> sum(_f.(x)), [1,2,3]) == ([0.5, 0.5, 0.5],)
   @test gradient(x -> sum(map(_f, x)), [1,2,3]) == ([0.5, 0.5, 0.5],)
 
+  # with Bool
+  @test gradient(x -> sum(1 .- (x .> 0)), randn(5)) == (nothing,)
+  @test gradient(x -> sum((y->1-y).(x .> 0)), randn(5)) == (nothing,)
+  @test gradient(x -> sum(x .- (x .> 0)), randn(5)) == ([1,1,1,1,1],)
+
   @test gradient(x -> sum(x ./ [1,2,4]), [1,2,pi]) == ([1.0, 0.5, 0.25],)
   @test gradient(x -> sum(map(/, x, [1,2,4])), [1,2,pi]) == ([1.0, 0.5, 0.25],)