very basic step! implementation

Manifest.toml

@@ -281,7 +281,7 @@ version = "0.8.0"
 
 [[Zygote]]
 deps = ["DiffRules", "ForwardDiff", "IRTools", "InteractiveUtils", "LinearAlgebra", "MacroTools", "NNlib", "NaNMath", "Random", "Requires", "SpecialFunctions"]
-git-tree-sha1 = "7e99e2a6c5287fe658273fdd1723726ff8a211d9"
+git-tree-sha1 = "029cbc1d784d4a2e3f2d26d9b1631d89c2a0afb2"
 repo-rev = "master"
 repo-url = "https://github.com/FluxML/Zygote.jl.git"
 uuid = "e88e6eb3-aa80-5325-afca-941959d7151f"

src/optimise/Optimise.jl

@@ -1,11 +1,12 @@
 module Optimise
 
-export train!,
+export train!, step!,
 	SGD, Descent, ADAM, Momentum, Nesterov, RMSProp,
 	ADAGrad, AdaMax, ADADelta, AMSGrad, NADAM, ADAMW,
 	InvDecay, ExpDecay, WeightDecay, stop, Optimiser
 
 include("optimisers.jl")
+include("update.jl")
 include("train.jl")
 
 end

src/optimise/optimisers.jl

@@ -4,8 +4,6 @@ using MacroTools: @forward
 
 const ϵ = 1e-8
 
-# TODO: should use weak refs
-
 """
     Descent(η)
 
@@ -18,8 +16,8 @@ end
 
 Descent() = Descent(0.1)
 
-function apply!(o::Descent, x, Δ)
+function apply(o::Descent, x, x̄, state = nothing)
-  Δ .*= o.eta
+  x̄ .* o.eta, state
 end
 
 """

src/optimise/train.jl

@@ -1,16 +1,26 @@
 using Juno
-import Zygote: Params, gradient
+import Zygote: Context, Params, _forward, gradient
 
-function update!(opt, x, x̄)
+# Training step
-  update!(x, -apply!(opt, x, x̄))
+
+function losscheck(x)
+  x isa Real || error("Function output is not scalar")
+  isinf(x) && error("Loss is infinite")
+  isnan(x) && error("Loss is NaN")
 end
 
-function update!(opt, xs::Params, gs)
+function step!(f, opt, x...)
-  for x in xs
+  cx = Context()
-    update!(opt, x, gs[x])
+  y, ∂f = _forward(cx, f, x...)
-  end
+  losscheck(y)
+  f̄ = ∂f(1)[1] # TODO update f
+  ḡ = Globals(cx)
+  update!(opt, nothing, ḡ)
+  return y
 end
 
+# Training loop
+
 # Callback niceties
 call(f, xs...) = f(xs...)
 runall(f) = f

src/optimise/update.jl

@@ -0,0 +1,56 @@
+using Zygote: Context, globals
+
+const Param{T<:Number} = Union{AbstractArray{T},T}
+
+struct Globals{T}
+  gs::T
+end
+
+Globals(cx::Context) = Globals(globals(cx))
+
+_apply(opt, x, x̄, state) = apply(opt, x, x̄, state)
+_apply(opt, x, x̄, ::Nothing) = apply(opt, x, x̄)
+
+# Immutable updates
+
+function update(opt, x::Param, x̄::Param, state = nothing)
+  Δ, state = _apply(opt, x, x̄, state)
+  return x .- Δ, state
+end
+
+# Mutable updates
+
+# Figure out if we can do in-place
+inplace(x, y) = false
+inplace(x, y::Nothing) = true
+inplace(x::AbstractArray, x̄::AbstractArray) = true
+inplace(x, x̄::NamedTuple) = all(inplace(getfield(x, f), getfield(x̄, f)) for f in fieldnames(typeof(x̄)))
+
+function update!(opt, x::AbstractArray{<:Number}, x̄::AbstractArray, state = nothing)
+  Δ, state = _apply(opt, x, x̄, state)
+  x .-= Δ
+  return state
+end
+
+function update!(opt, x, x̄::NamedTuple)
+  for f in fieldnames(typeof(x̄))
+    f̄ = getfield(x̄, f)
+    f̄ === nothing || update!(opt, getfield(x, f), f̄)
+  end
+end
+
+setglobal!(mod::Module, name::Symbol, x) =
+  ccall(:jl_set_global, Cvoid, (Any, Any, Any), mod, name, x)
+
+function update!(opt, ::Nothing, gs::Globals)
+  for (id, x̄) in gs.gs
+    x = getfield(id.mod, id.name)
+    if inplace(x, x̄)
+      update!(opt, x, x̄)
+    else
+      isconst(id.mod, id.name) && error("Can't update constant $id")
+      x′, state = update(opt, x, x̄)
+      setglobal!(id.mod, id.name, x′)
+    end
+  end
+end