safer gradients, less piracy (#2098)

Author: mcabbott

Project.toml

@@ -1,10 +1,9 @@
 name = "Flux"
 uuid = "587475ba-b771-5e3f-ad9e-33799f191a9c"
-version = "0.13.7"
+version = "0.13.8"
 
 [deps]
 Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
-ArrayInterface = "4fba245c-0d91-5ea0-9b3e-6abc04ee57a9"
 CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
 ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
 Functors = "d9f16b24-f501-4c13-a1f2-28368ffc5196"
@@ -27,7 +26,6 @@ Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 
 [compat]
 Adapt = "3.0"
-ArrayInterface = "3.1, 4, 5, 6"
 CUDA = "3"
 ChainRulesCore = "1.12"
 Functors = "0.3"

src/optimise/Optimise.jl

@@ -1,7 +1,6 @@
 module Optimise
 
 using LinearAlgebra
-import ArrayInterface
 
 export train!, update!,
 	Descent, Adam, Momentum, Nesterov, RMSProp,

src/optimise/train.jl

@@ -13,8 +13,8 @@ As a result, the parameters are mutated and the optimizer's internal state may c
 The gradient could be mutated as well.
 """
 function update!(opt::AbstractOptimiser, x, x̄)
-  x̄r = ArrayInterface.restructure(x, x̄) # address some cases where Zygote's
-                                          # output are not mutable, see #1510 
+  x̄r = copyto!(similar(x̄), x̄)  # Flux.Optimise assumes it can mutate the gradient. This is not
+                               # safe due to aliasing, nor guaranteed to be possible, e.g. Fill.
   x .-= apply!(opt, x, x̄r)
 end