Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add destructure, take II #54

Merged
merged 13 commits into from
Feb 14, 2022
2 changes: 1 addition & 1 deletion Project.toml
Original file line number Diff line number Diff line change
Expand Up @@ -12,7 +12,7 @@ Statistics = "10745b16-79ce-11e8-11f9-7d13ad32a3b2"

[compat]
ChainRulesCore = "1"
Functors = "0.2.7"
Functors = "0.2.8"
julia = "1.6"

[extras]
Expand Down
7 changes: 7 additions & 0 deletions docs/src/api.md
Original file line number Diff line number Diff line change
Expand Up @@ -42,6 +42,13 @@ optimiser to act on all suitable fields. To restrict this, define `trainable`:
Optimisers.trainable
```

Such restrictions are also obeyed by this function for flattening a model:

```@docs
Optimisers.destructure
Optimisers.Restructure
```

## Rule Definition

```@docs
Expand Down
5 changes: 4 additions & 1 deletion src/Optimisers.jl
Original file line number Diff line number Diff line change
Expand Up @@ -4,8 +4,11 @@ using Functors: functor, fmap, isleaf
using LinearAlgebra

include("interface.jl")
include("rules.jl")

include("destructure.jl")
export destructure, total, total2

include("rules.jl")
export Descent, ADAM, Momentum, Nesterov, RMSProp,
ADAGrad, AdaMax, ADADelta, AMSGrad, NADAM, ADAMW, RADAM, OADAM, AdaBelief,
WeightDecay, ClipGrad, ClipNorm, OptimiserChain
Expand Down
145 changes: 145 additions & 0 deletions src/destructure.jl
Original file line number Diff line number Diff line change
@@ -0,0 +1,145 @@

using ChainRulesCore: ChainRulesCore, NoTangent, ProjectTo, unthunk
const NoT = NoTangent()

base(dx::Tangent{<:Tangent}) = backing(dx).backing # might be needed for gradient(gradient(destructure))
base(dx::Tangent{Any, <:NamedTuple{(:backing,)}}) = base(backing(dx).backing) # Zygote version

"""
destructure(model) -> vector, reconstructor

Copies all [`trainable`](@ref), [`isnumeric`](@ref) parameters in the model
to a vector, and returns also a function which reverses this transformation.
Differentiable.

# Example
```jldoctest
julia> v, re = destructure((x=[1.0, 2.0], y=(sin, [3 + 4im])))
(ComplexF64[1.0 + 0.0im, 2.0 + 0.0im, 3.0 + 4.0im], Restructure(NamedTuple, ..., 3))

julia> re([3, 5-im, 7+11im])
(x = [3.0, 5.0], y = (sin, ComplexF64[7.0 + 11.0im]))
```
"""
function destructure(x)
flat, off, len = _flatten(x)
flat, Restructure(x, off, len)
end

"""
Restructure(Model, ..., length)

This is what [`destructure`](@ref) returns, and `re(p)` will re-build the model with
new parameters from vector `p`. If the model is callable, then `re(x, p) == re(p)(x)`.

# Example
```julia
julia> using Flux, Optimisers

julia> _, re = destructure(Dense([1 2; 3 4], [0, 0], sigmoid))
([1, 3, 2, 4, 0, 0], Restructure(Dense, ..., 6))

julia> m = re(-4:1)
Dense(2, 2, σ) # 6 parameters

julia> m([0.2, 0.3]) ≈ re([0.2, 0.3], -4:1)
true
```
"""
struct Restructure{T,S}
model::T
offsets::S
length::Int
end
(re::Restructure)(flat::AbstractVector) = _rebuild(re.model, re.offsets, flat, re.length)
(re::Restructure)(x, flat::AbstractVector) = re(flat)(x)
Base.show(io::IO, re::Restructure{T}) where T = print(io, "Restructure(", T.name.name, ", ..., ", re.length, ")")
Base.length(re::Restructure) = re.length

# This flattens a model, and returns a web of offsets for later use:
function _flatten(x)
isnumeric(x) && return vcat(_vec(x)), 0, length(x) # trivial case
arrays = AbstractVector[]
len = Ref(0)
off = fmap(x; exclude = isnumeric, walk = (f, z) -> map(f, _trainable(z))) do y
push!(arrays, _vec(y))
o = len[]
len[] = o + length(y)
o
end
reduce(vcat, arrays), off, len[]
end

_vec(x::Number) = LinRange(x,x,1)
_vec(x::AbstractArray) = vec(x)

function ChainRulesCore.rrule(::typeof(_flatten), x)
flat, off, len = _flatten(x)
_flatten_back((dflat, _, _)) = (NoT, _rebuild(x, off, unthunk(dflat), len; walk = _Tangent_biwalk, prune = NoT))
(flat, off, len), _flatten_back
end

# This reconstructs either a model like x, or a gradient for it:
function _rebuild(x, off, flat::AbstractVector, len = length(flat); walk = _trainable_biwalk, kw...)
len == length(flat) || throw(DimensionMismatch("Rebuild expected a vector of length $len, got $(length(flat))"))
fmap(x, off; exclude = isnumeric, walk, kw...) do y, o
_getat(y, o, flat)
end
end

_getat(y::Number, o::Int, flat::AbstractVector) = ProjectTo(y)(flat[o + 1])
_getat(y::AbstractArray, o::Int, flat::AbstractVector) =
ProjectTo(y)(reshape(flat[o .+ (1:length(y))], axes(y))) # ProjectTo is just correcting eltypes

function _trainable_biwalk(f, x, aux)
ch, re = functor(typeof(x), x)
au, _ = functor(typeof(x), aux)
_trainmap(f, ch, _trainable(x), au) |> re
end

function _trainmap(f, ch, tr, aux)
map(ch, tr, aux) do c, t, a # isnothing(t) indicates non-trainable field, safe given isnumeric(c)
isnothing(t) ? c : f(t, a)
end
end

function _Tangent_biwalk(f, x, aux) # use with prune = NoT
ch, re = functor(typeof(x), x)
au, _ = functor(typeof(x), aux)
y = _trainmap(f, ch, _trainable(x), au)
y isa Tuple{} && return NoT
p = ProjectTo(x)
if p isa ProjectTo # e.g. Array, NamedTuple
p(y)
else # p === identity for unknown structs
Tangent{typeof(x), typeof(y)}(y)
end
end

function ChainRulesCore.rrule(::typeof(_rebuild), x, off, flat, len; kw...)
_rebuild_back(dx) = (NoT, NoT, NoT, _grad!(x, unthunk(dx), off, _zero(flat)), NoT)
_rebuild(x, off, flat, len; kw...), _rebuild_back
end

_zero(x) = map!(zero, similar(x, float(eltype(x))), x) # mutable zero array for _grad!
ChainRulesCore.@non_differentiable _zero(x)

# This is the gradient of model reconstruction, accumulating duplicates:
function _grad!(x, dx, off, flat::AbstractVector)
x′, _ = functor(typeof(x), x)
dx′, _ = functor(typeof(x), base(dx))
off′, _ = functor(typeof(x), off)
foreach((xᵢ, dxᵢ, oᵢ) -> _grad!(xᵢ, dxᵢ, oᵢ, flat), x′, dx′, off′)
flat
end
function _grad!(x, dx, off::Integer, flat::AbstractVector)
@views flat[off .+ (1:length(x))] .+= dx # must visit all tied nodes
flat
end
_grad!(x, dx::Zero, off, flat::AbstractVector) = dx
_grad!(x, dx::Zero, off::Integer, flat::AbstractVector) = dx # ambiguity

function ChainRulesCore.rrule(::typeof(_grad!), x, dx, off, flat)
_grad_back(dflat) = (NoT, NoT, _rebuild(x, off, unthunk(dflat); walk = _Tangent_biwalk, prune = NoT), NoT, NoT)
_grad!(x, dx, off, flat), _grad_back
end
3 changes: 2 additions & 1 deletion src/interface.jl
Original file line number Diff line number Diff line change
Expand Up @@ -70,7 +70,8 @@ trainable(x) = functor(x)[1]

_trainable(x) = _trainable(functor(x)[1], trainable(x))
_trainable(ch::NamedTuple, tr::NamedTuple) = merge(map(_ -> nothing, ch), tr)
_trainable(ch::Tuple, tr::Tuple) = tr
_trainable(ch::Tuple{Vararg{Any,N}}, tr::Tuple{Vararg{Any,N}}) where N = tr
_trainable(ch::AbstractArray, tr::AbstractArray) = tr
function _trainable(ch::NamedTuple, tr::Tuple) # for old Flux-style no-names tuple
@warn "trainable(x) should now return a NamedTuple with the field names, not a Tuple"
map(c -> c in tr ? c : nothing, ch)
Expand Down
166 changes: 166 additions & 0 deletions test/destructure.jl
Original file line number Diff line number Diff line change
@@ -0,0 +1,166 @@

m1 = collect(1:3.0)
m2 = (collect(1:3.0), collect(4:6.0))
m3 = (x = m1, y = sin, z = collect(4:6.0))
m4 = (x = m1, y = m1, z = collect(4:6.0)) # tied
m5 = (a = (m3, true), b = (m1, false), c = (m4, true))
m6 = (a = m1, b = [4.0 + im], c = m1)
m7 = TwoThirds((sin, collect(1:3.0)), (cos, collect(4:6.0)), (tan, collect(7:9.0)))
m8 = [Foo(m1, m1), (a = true, b = Foo([4.0], false), c = ()), [[5.0]]]

@testset "flatten & rebuild" begin
@test destructure(m1)[1] isa Vector{Float64}
@test destructure(m1)[1] == 1:3
@test destructure(m2)[1] == 1:6
@test destructure(m3)[1] == 1:6
@test destructure(m4)[1] == 1:6
@test destructure(m5)[1] == vcat(1:6, 4:6)
@test destructure(m6)[1] == vcat(1:3, 4 + im)

@test destructure(m1)[2](7:9) == [7,8,9]
@test destructure(m2)[2](4:9) == ([4,5,6], [7,8,9])
@test destructure(m3)[2](4:9) == (x = [4,5,6], y = sin, z = [7,8,9])
m4′ = destructure(m4)[2](4:9)
@test m4′ == (x = [4,5,6], y = [4,5,6], z = [7,8,9])
@test m4′.x === m4′.y
m5′ = destructure(m5)[2](reverse(1:9))
@test m5′.a[1].x === m5′.b[1]
@test m5′.b[2] === false
m6′ = destructure(m6)[2]((4:7) .+ (1:4) .* im)
@test m6′.a isa Vector{Float64}
@test m6′.a == 4:6
@test m6′.a === m6′.c
@test m6′.b == [7 + 4im]

# struct, trainable
@test destructure(m7)[1] == 1:3
m7′ = destructure(m7)[2]([10,20,30])
@test m7′.a == (sin, [10,20,30])
@test m7′.b == (cos, [4,5,6])
@test m7′.c == (tan, [7,8,9])

@test destructure(m8)[1] == 1:5
m8′ = destructure(m8)[2](1:5)
@test m8′[1].x === m8′[1].y
@test m8′[2].b.y === false
@test m8′[3][1] == [5.0]

# errors
@test_throws Exception destructure(m7)[2]([10,20])
@test_throws Exception destructure(m7)[2]([10,20,30,40])
end

@testset "gradient of flatten" begin
@test gradient(m -> destructure(m)[1][1], m1)[1] == [1,0,0]
@test gradient(m -> destructure(m)[1][2], m2)[1] == ([0,1,0], [0,0,0])
@test gradient(m -> destructure(m)[1][3], (m1, m1))[1] == ([0,0,1], nothing)
@test gradient(m -> destructure(m)[1][1], m3)[1] == (x = [1,0,0], y = nothing, z = [0,0,0])
@test gradient(m -> destructure(m)[1][2], m4)[1] == (x = [0,1,0], y = nothing, z = [0,0,0])

g5 = gradient(m -> destructure(m)[1][3], m5)[1]
@test g5.a[1].x == [0,0,1]
@test g5.a[2] === nothing

g6 = gradient(m -> imag(destructure(m)[1][4]), m6)[1]
@test g6.a == [0,0,0]
@test g6.a isa Vector{Float64}
@test g6.b == [0+im]

g8 = gradient(m -> sum(abs2, destructure(m)[1]), m8)[1]
@test g8[1].x == [2,4,6]
@test g8[2].b.x == [8]
@test g8[3] == [[10.0]]

@testset "second derivative" begin
@test gradient([1,2,3.0]) do v
sum(abs2, gradient(m -> sum(abs2, destructure(m)[1]), (v, [4,5,6.0]))[1][1])
end[1] ≈ [8,16,24]
# With Diffractor, non-leaf _grad!(x, dx, off, flat::AbstractVector) gets double-wrapped dx:
# off = (0, 3), dx = Tangent{Tangent{Tuple{Vector{Float64}, Vector{Float64}}, ...
# until you add explicit double-unwrap: base(dx::Tangent{<:Tangent}) = backing(dx).backing
# With Zygote, instead:
# dx = Tangent{Any}(backing = Tangent{Any}([4.0, 8.0, 12.0], ZeroTangent()),)

@test gradient([1,2,3.0]) do v
sum(gradient(m -> sum(destructure(m)[1])^3, (v, [4,5,6.0]))[1][1])
end[1] == [378, 378, 378]

@test_broken gradient([1,2,3.0]) do v
sum(abs2, gradient(m -> sum(abs2, destructure(m)[1]), (x = v, y = sin, z = [4,5,6.0]))[1][1])
end[1] ≈ [8,16,24]
# Zygote error in (::typeof(∂(canonicalize)))(Δ::NamedTuple{(:backing,), Tuple{NamedTuple{(:x, :y, :z)
# Diffractor error in perform_optic_transform
end
end

@testset "gradient of rebuild" begin
re1 = destructure(m1)[2]
@test gradient(x -> re1(x)[1], rand(3))[1] == [1,0,0]
re2 = destructure(m2)[2]
@test gradient(x -> re2(x)[1][2], rand(6))[1] == [0,1,0,0,0,0]
re3 = destructure(m3)[2]
@test gradient(x -> re3(x).x[3], rand(6))[1] == [0,0,1,0,0,0]
@test gradient(x -> re3(x).z[1], rand(6))[1] == [0,0,0,1,0,0]

re4 = destructure(m4)[2]
@test gradient(x -> re4(x).x[1], rand(6))[1] == [1,0,0,0,0,0]
@test gradient(x -> re4(x).y[2], rand(6))[1] == [0,1,0,0,0,0]
@test gradient(rand(6)) do x
m = re4(x)
m.x[1] + 2*m.y[2] + 3*m.z[3]
end[1] == [1,2,0, 0,0,3]

re7 = destructure(m7)[2]
@test gradient(x -> re7(x).a[2][3], rand(3))[1] == [0,0,1]
@test gradient(x -> re7(x).b[2][2], rand(3))[1] == [0,0,0]
@test gradient(x -> re7(x).c[2][1], rand(3))[1] == [0,0,0]

v8, re8 = destructure(m8)
@test gradient(x -> sum(abs2, re8(x)[1].y), v8)[1] == [2,4,6,0,0]
@test gradient(x -> only(sum(re8(x)[3]))^2, v8)[1] == [0,0,0,0,10]

@testset "second derivative" begin
@test_broken gradient(collect(1:6.0)) do y
sum(abs2, gradient(x -> sum(abs2, re2(x)[1]), y)[1])
end[1] ≈ [8,16,24,0,0,0]
# ERROR: Need an adjoint for constructor ChainRulesCore.Tangent{Any, Tuple{Vector{Float64}, ChainRulesCore.ZeroTangent}}. Gradient is of type Tuple{Vector{Float64}, Vector{Float64}}
# with Zygote, which can be fixed by:
# Zygote.@adjoint Tangent{T,B}(x::Tuple) where {T,B<:Tuple} = Tangent{T,B}(x), dx -> (dx,)
Comment on lines +122 to +128
Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think the remaining question on this PR is whether and how much to care about 2nd derivatives. Some work, some don't. I convinced myself there is no bug in the basic logic. But in the details of when to wrap what in a Tangent, or unwrap it for Zygote, there might be bugs, here or upstream.

If we want to be pedantic we could make all 2nd derivatives an error, rather than risk any being wrong. Or a warning.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

At least a warning sounds good to me.

Copy link
Member Author

@mcabbott mcabbott Feb 13, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done!

All warnings are maxlog=3, so as not to be too annoying if something does actually work.

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Good to go?


@test_broken gradient(collect(1:6.0)) do y
sum(abs2, gradient(x -> sum(abs2, re3(x).z), y)[1])
end[1] ≈ [0,0,0,32,40,48]
# Not fixed by this:
# Zygote.@adjoint Tangent{T,B}(x::NamedTuple) where {T,B<:NamedTuple} = Tangent{T,B}(x), dx -> (dx,)
end
end

@testset "Flux issue 1826" begin
v, re = destructure((x=[1,2.0], y=[3,4,5.0]))
@test gradient(zero(v)) do w
m = re(w)
5 * sum(m.x) + 7 * sum(m[2]) # uses both x and y
end == ([5.0, 5.0, 7.0, 7.0, 7.0],)
# This, using only x, was broken on Flux:
@test gradient(w -> sum(re(w).x), zero(v)) == ([1.0, 1.0, 0.0, 0.0, 0.0],)

sh = [7,7.0];
v, re = destructure((x=sh, y=[3.0,4.0], z=sh)) # shared array in the model
@test v == [7, 7, 3, 4]
@test re([1,10,100,1000]) == (x = [1, 10], y = [100, 1000], z = [1, 10])

@test gradient(zero(v)) do w
m = re(w)
3 * sum(m.x) + 13 * sum(m.z) # no dependence on y, but two distinct gradient arrays
end == ([16, 16, 0, 0],) # Flux gave ([3.0, 3.0, 13.0, 13.0],)

@test gradient(zero(v)) do w
m = re(w)
4(sum(m.x) + sum(m.z)) # now two gradients are ===, so it eliminates one
end == ([8,8,0,0],)

@test gradient(zero(v)) do w
m = re(w)
4(sum(m.x) + sum(m.y)) + 13*sum(m.z) # again two gradients are ===, so it eliminates one
end == ([17,17,4,4],) # Flux gave ([4.0, 4.0, 13.0, 13.0],)
end
8 changes: 4 additions & 4 deletions test/rules.jl
Original file line number Diff line number Diff line change
Expand Up @@ -44,7 +44,7 @@ end
end
end

@testset verbose=true "simple sum" begin
@testset "simple sum" begin
empty!(LOG)
@testset "$(name(o))" for o in RULES
m = shuffle!(reshape(1:64, 8, 8) .+ 0.0)
Expand Down Expand Up @@ -79,7 +79,7 @@ end
end
end

@testset verbose=true "StaticArrays" begin
@testset "StaticArrays" begin
empty!(LOG)
@testset "$(name(o))" for o in RULES
W1 = @SMatrix randn(10, 10)
Expand Down Expand Up @@ -157,7 +157,7 @@ end
end
end

@testset verbose=true "mutation check" begin
@testset "mutation check" begin
# If @lazy captures a matrix which is later mutated, the results won't agree here:
@testset "$(name(o))" for o in RULES
model = Float64.(rand(Int8, 8))
Expand All @@ -174,7 +174,7 @@ end
end
end

@testset "with complex numebers: Flux#1776" begin
@testset "with complex numbers: Flux#1776" begin
empty!(LOG)
@testset "$(name(opt))" for opt in [
# The Flux PR had 1e-2 for all. But ADADelta(ρ) needs ρ≈0.9 not small. And it helps to make ε not too small too:
Expand Down
Loading