Revert "simplify default_rng etc"

This reverts commit 0e396a6.

Author: mcabbott

src/Flux.jl

@@ -2,7 +2,6 @@ module Flux
 
 using Base: tail
 using LinearAlgebra, Statistics, Random  # standard lib
-using Random: default_rng
 using MacroTools, Reexport, ProgressLogging, SpecialFunctions
 using MacroTools: @forward
 

src/deprecations.jl

@@ -84,6 +84,8 @@ Base.@deprecate_binding ADADelta AdaDelta
 # Remove sub-module Data, while making sure Flux.Data.DataLoader keeps working
 Base.@deprecate_binding Data Flux false "Sub-module Flux.Data has been removed. The only thing it contained may be accessed as Flux.DataLoader"
 
+@deprecate rng_from_array() default_rng_value()
+
 function istraining()
   Base.depwarn("Flux.istraining() is deprecated, use NNlib.within_gradient(x) instead", :istraining)
   false
@@ -183,8 +185,17 @@ function update!(opt::Optimise.AbstractOptimiser, ::Params, grads::Union{Tuple,
     """)
 end
 
-@deprecate rng_from_array() default_rng_value()
-@deprecate default_rng_value() Random.default_rng()
+
+function dropout(rng, x, p; dims=:, active::Bool=true)
+  if active
+    NNlib.dropout(rng, x, p; dims)
+  else
+    Base.depwarn("Flux.dropout(...; active=false) is deprecated. Please branch outside the function, or call dropout(x, 0) if you must.", :dropout)
+    return x
+  end
+end
+dropout(x, p; kwargs...) = dropout(NNlib._rng_from_array(x), x, p; kwargs...)
+
 
 # v0.14 deprecations
 

src/layers/normalise.jl

@@ -2,7 +2,7 @@
 _isactive(m, x) = isnothing(m.active) ? NNlib.within_gradient(x) : m.active
 
 """
-    Dropout(p; dims=:, rng = default_rng())
+    Dropout(p; dims=:, rng = default_rng_value())
 
 Layer implementing [dropout](https://arxiv.org/abs/1207.0580) with the given probability.
 This is used as a regularisation, i.e. to reduce overfitting.
@@ -61,9 +61,9 @@ mutable struct Dropout{F<:Real,D,R<:AbstractRNG}
   active::Union{Bool, Nothing}
   rng::R
 end
-Dropout(p::Real, dims, active) = Dropout(p, dims, active, default_rng())
+Dropout(p::Real, dims, active) = Dropout(p, dims, active, default_rng_value())
 
-function Dropout(p::Real; dims=:, rng = default_rng())
+function Dropout(p::Real; dims=:, rng = default_rng_value())
   0 ≤ p ≤ 1 || throw(ArgumentError("Dropout expexts 0 ≤ p ≤ 1, got p = $p"))
   if p isa Integer  # Dropout(0)
     return p==0 ? identity : zero
@@ -92,7 +92,7 @@ function Base.show(io::IO, d::Dropout)
 end
 
 """
-    AlphaDropout(p; rng = default_rng())
+    AlphaDropout(p; rng = default_rng_value())
 
 A dropout layer. Used in
 [Self-Normalizing Neural Networks](https://arxiv.org/abs/1706.02515).
@@ -126,8 +126,8 @@ mutable struct AlphaDropout{F,R<:AbstractRNG}
     new{typeof(p), typeof(rng)}(p, active, rng)
   end
 end
-AlphaDropout(p, active) = AlphaDropout(p, active, default_rng())
-AlphaDropout(p; rng = default_rng()) = AlphaDropout(p, nothing, rng)
+AlphaDropout(p, active) = AlphaDropout(p, active, default_rng_value())
+AlphaDropout(p; rng = default_rng_value()) = AlphaDropout(p, nothing, rng)
 
 @functor AlphaDropout
 trainable(a::AlphaDropout) = (;)

src/utils.jl

@@ -36,12 +36,32 @@ epseltype(x) = eps(float(eltype(x)))
 """
     rng_from_array([x])
 
-Create an instance of the RNG most appropriate for array `x`.
-If `x isa CuArray` then this is `CUDA.default_rng()`,
-otherwise `Random.default_rng()`.
+Create an instance of the RNG most appropriate for `x`.
+The current defaults are:
+- `x isa CuArray`: `CUDA.default_rng()`, else:
+- `x isa AbstractArray`, or no `x` provided:
+  - Julia version is < 1.7: `Random.GLOBAL_RNG`
+  - Julia version is >= 1.7: `Random.default_rng()`
+"""
+rng_from_array(::AbstractArray) = default_rng_value()
+rng_from_array(::CuArray) = CUDA.default_rng()
+
+@non_differentiable rng_from_array(::Any)
+
+if VERSION >= v"1.7"
+  default_rng_value() = Random.default_rng()
+else
+  default_rng_value() = Random.GLOBAL_RNG
+end
+
 """
-rng_from_array(x::AbstractArray) = NNlib._rng_from_array(x)
+    default_rng_value()
 
+Create an instance of the default RNG depending on Julia's version.
+- Julia version is < 1.7: `Random.GLOBAL_RNG`
+- Julia version is >= 1.7: `Random.default_rng()`
+"""
+default_rng_value
 
 """
     glorot_uniform([rng = default_rng_value()], size...; gain = 1) -> Array

test/layers/normalisation.jl

@@ -56,10 +56,10 @@ evalwgrad(f, x...) = pullback(f, x...)[1]
     y = m(x)
     @test count(a->a == 0, y) > 50
 
-    y = Flux.dropout(values(rng_kwargs)..., x, 0.9) # , active=true)
+    y = Flux.dropout(values(rng_kwargs)..., x, 0.9, active=true)
     @test count(a->a == 0, y) > 50
 
-    y = Flux.dropout(values(rng_kwargs)..., x, 0.9 * 0) #, active=false)
+    y = Flux.dropout(values(rng_kwargs)..., x, 0.9, active=false)
     @test count(a->a == 0, y) == 0
 
     # CPU RNGs map onto CPU ok