use NNlib.within_gradient (#2152)

Author: mcabbott

Project.toml

@@ -30,7 +30,7 @@ ChainRulesCore = "1.12"
 Functors = "0.3, 0.4"
 MLUtils = "0.2, 0.3.1, 0.4"
 MacroTools = "0.5"
-NNlib = "0.8.9"
+NNlib = "0.8.14"
 NNlibCUDA = "0.2.4"
 OneHotArrays = "0.1, 0.2"
 Optimisers = "0.2.12"

src/cuda/cudnn.jl

@@ -8,7 +8,7 @@ function (BN::Flux.BatchNorm)(x::Union{CuArray{T,2},CuArray{T,4},CuArray{T,5}},
   @assert length(BN.β) == size(x, ndims(x)-1) "BatchNorm: input has wrong number of channels"
   return BN.λ.(batchnorm(BN.γ, BN.β, x, BN.μ, BN.σ², BN.momentum; 
                   cache=cache, alpha=1, beta=0, eps=BN.ϵ, 
-                  training=Flux._isactive(BN)))
+                  training=Flux._isactive(BN, x)))
 end
 
 function ChainRulesCore.rrule(::typeof(batchnorm), g, b, x, running_mean, running_var, momentum; kw...)

src/deprecations.jl

@@ -86,6 +86,17 @@ Base.@deprecate_binding Data Flux false "Sub-module Flux.Data has been removed.
 
 @deprecate rng_from_array() default_rng_value()
 
+function istraining()
+  Base.depwarn("Flux.istraining() is deprecated, use NNlib.within_gradient(x) instead", :istraining)
+  false
+end
+ChainRulesCore.rrule(::typeof(istraining)) = true, _ -> (NoTangent(),)
+
+function _isactive(m)
+  Base.depwarn("_isactive(m) is deprecated, use _isactive(m,x)", :_isactive, force=true)
+  _isactive(m, 1:0)
+end
+
 #=
   # Valid method in Optimise, old implicit style, is:
   train!(loss, ps::Params, data, opt::AbstractOptimiser; cb = () -> ())

src/layers/normalise.jl

@@ -1,8 +1,5 @@
-istraining() = false
 
-ChainRulesCore.rrule(::typeof(istraining)) = true, _ -> (NoTangent(),)
-
-_isactive(m) = isnothing(m.active) ? istraining() : m.active
+_isactive(m, x) = isnothing(m.active) ? NNlib.within_gradient(x) : m.active
 
 _dropout_shape(s, ::Colon) = size(s)
 _dropout_shape(s, dims) = tuple((i ∉ dims ? 1 : si for (i, si) ∈ enumerate(size(s)))...)
@@ -107,7 +104,7 @@ end
 trainable(a::Dropout) = (;)
 
 function (a::Dropout)(x)
-  _isactive(a) || return x
+  _isactive(a, x) || return x
   return dropout(a.rng, x, a.p; dims=a.dims, active=true)
 end
 
@@ -162,7 +159,7 @@ AlphaDropout(p; rng = default_rng_value()) = AlphaDropout(p, nothing, rng)
 trainable(a::AlphaDropout) = (;)
 
 function (a::AlphaDropout)(x::AbstractArray{T}) where T
-  _isactive(a) || return x
+  _isactive(a, x) || return x
   p = a.p
   iszero(p) && return x
   isone(p) && return sign.(x) .* T(0)
@@ -242,7 +239,7 @@ end
 function _norm_layer_forward(
   l, x::AbstractArray{T, N}; reduce_dims, affine_shape,
 ) where {T, N}
-  if !_isactive(l) && l.track_stats # testmode with tracked stats
+  if !_isactive(l, x) && l.track_stats # testmode with tracked stats
     stats_shape = ntuple(i -> i == N-1 ? size(x, N-1) : 1, N)
     μ = reshape(l.μ, stats_shape)
     σ² = reshape(l.σ², stats_shape)

test/layers/normalisation.jl

@@ -475,5 +475,20 @@ end
   # This was an error, https://github.com/FluxML/Flux.jl/issues/2122
   @test ForwardDiff.jacobian(bn, rand(Float32, 3, 4)) isa Matrix{Float32}
   @test !iszero(bn.μ)
+
+  # Easy case of 2122, gradient with x
+  x5 = rand(Float32, 5, 3)
+  bn1 = BatchNorm(5, relu)
+  bn2 = BatchNorm(5, relu)
+  g1 = Zygote.gradient(x -> sum(abs2, bn1(x)), x5)[1]
+  g2 = ForwardDiff.gradient(x -> sum(abs2, bn2(x)), x5)
+  @test g1 ≈ g2
+
+  # Harder case? 
+  v1, re1 = Flux.destructure(BatchNorm(5, relu));
+  g1 = Zygote.gradient(v -> sum(abs2, re1(v)(x5)), v1)[1]
+
+  v2, re2 = Flux.destructure(BatchNorm(5, relu));
+  g2 = ForwardDiff.gradient(v -> sum(abs2, re2(v)(x5)), v2)
 end