Added epsilon term to binarycrossentropy

src/layers/stateless.jl

@@ -15,9 +15,9 @@ function logitcrossentropy(logŷ::AbstractVecOrMat, y::AbstractVecOrMat; weight
 end
 
 """
-    binarycrossentropy(ŷ, y)
+    binarycrossentropy(ŷ, y; ϵ)
 
-Return `-y*log(ŷ) - (1-y)*log(1-ŷ)`.
+Return `-y*log(ŷ + ϵ) - (1-y)*log(1-ŷ + ϵ)`. The ϵ term provides numerical stability.
 
     julia> binarycrossentropy.(σ.([-1.1491, 0.8619, 0.3127]), [1, 1, 0.])
     3-element Array{Float64,1}:
@@ -25,7 +25,7 @@ Return `-y*log(ŷ) - (1-y)*log(1-ŷ)`.
     0.352317
     0.86167
 """
-binarycrossentropy(ŷ, y) = -y*log(ŷ) - (1 - y)*log(1 - ŷ)
+binarycrossentropy(ŷ, y; ϵ=1e-7) = -y*log(ŷ + ϵ) - (1 - y)*log(1 - ŷ + ϵ)
 
 """
     logitbinarycrossentropy(logŷ, y)

test/layers/stateless.jl

@@ -2,6 +2,8 @@ using Base.Test
 using Flux: onehotbatch, mse, crossentropy, logitcrossentropy,
             σ, binarycrossentropy, logitbinarycrossentropy
 
+const ϵ = 1e-7
+
 @testset "losses" begin
   # First, regression-style y's
   y = [1, 1, 0, 0]
@@ -40,10 +42,11 @@ using Flux: onehotbatch, mse, crossentropy, logitcrossentropy,
 
   logŷ, y = randn(3), rand(3)
   @testset "binarycrossentropy" begin
-    @test binarycrossentropy.(σ.(logŷ), y) ≈ -y.*log.(σ.(logŷ)) - (1 - y).*log.(1 - σ.(logŷ))
+    @test binarycrossentropy.(σ.(logŷ), y; ϵ=0) ≈ -y.*log.(σ.(logŷ)) - (1 - y).*log.(1 - σ.(logŷ))
+    @test binarycrossentropy.(σ.(logŷ), y) ≈ -y.*log.(σ.(logŷ) + 1e-7) - (1 - y).*log.(1 - σ.(logŷ) + 1e-7)
   end
 
   @testset "logitbinarycrossentropy" begin
-    @test logitbinarycrossentropy.(logŷ, y) ≈ binarycrossentropy.(σ.(logŷ), y)
+    @test logitbinarycrossentropy.(logŷ, y) ≈ binarycrossentropy.(σ.(logŷ), y; ϵ=0)
   end
 end