use explicit converts

src/layers/stateless.jl

@@ -2,16 +2,16 @@ using NNlib: logsoftmax, logσ
 
 # Cost functions
 
-mse(ŷ, y) = sum((ŷ .- y).^2)/length(y)
+mse(ŷ, y; efftype = eltype(ŷ)) = sum((ŷ .- y).^2)/convert(efftype, length(y))
 
-function crossentropy(ŷ::AbstractVecOrMat, y::AbstractVecOrMat; weight = 1)
+function crossentropy(ŷ::AbstractVecOrMat, y::AbstractVecOrMat; weight = 1, efftype = eltype(ŷ))
-  -sum(y .* log.(ŷ) .* weight) / size(y, 2)
+  -sum(y .* log.(ŷ) .* weight) / convert(efftype, size(y, 2))
 end
 
 @deprecate logloss(x, y) crossentropy(x, y)
 
-function logitcrossentropy(logŷ::AbstractVecOrMat, y::AbstractVecOrMat; weight = 1)
+function logitcrossentropy(logŷ::AbstractVecOrMat, y::AbstractVecOrMat; weight = 1, efftype = eltype(ŷ))
-  return -sum(y .* logsoftmax(logŷ) .* weight) / size(y, 2)
+  return -sum(y .* logsoftmax(logŷ) .* weight) / convert(efftype, size(y, 2))
 end
 
 """

src/tracker/lib/real.jl

@@ -72,7 +72,7 @@ for (M, f, arity) in DiffRules.diffrules()
   f = :($M.$f)
   @eval begin
     @grad $f(a::TrackedReal, b::TrackedReal) = $f(data(a), data(b)), Δ -> (Δ * $da, Δ * $db)
-    @grad $f(a::TrackedReal, b::Real) = $f(data(a), b), Δ -> (Δ * convert(TrackedReal{eltype(Δ)}, $da), _zero(b))
+    @grad $f(a::TrackedReal, b::Real) = $f(data(a), b), Δ -> (Δ * $da, _zero(b))
     @grad $f(a::Real, b::TrackedReal) = $f(a, data(b)), Δ -> (_zero(a), Δ * $db)
     $f(a::TrackedReal, b::TrackedReal)  = track($f, a, b)
     $f(a::TrackedReal, b::Real) = track($f, a, b)