Comments (3)
I expect this has an identical root cause as #5, though the specific error has come out differently. The same bandaid works – 0.5(x'*S*x)
.
julia> derivative(g, @SVector randn(4))
4-element SArray{Tuple{4},Float64,1,4}:
-8.43300310272326
28.03595720287608
-4.489528942268124
-6.965151312812216
As an aside, it's very pleasing that custom array types already work nicely like this; it's not something I had tested up to now. StaticArrays are an interesting case, because for things like this they should be very fast.
from zygote.jl.
Feel free to close this issue in favor of the old one. This works for me.
Thanks for the great work here - I'm excitedly following the development of this and Capstan.
from zygote.jl.
Right now, this code is type stable, but much slower than ForwardDiff (or the analytic gradient):
julia> using StaticArrays, BenchmarkTools, ForwardDiff
julia> import Zygote: gradient #always re-precompiles?
[ Info: Precompiling Zygote [e88e6eb3-aa80-5325-afca-941959d7151f]
[ Info: Precompiling IRTools [7869d1d1-7146-5819-86e3-90919afe41df]
julia> const S = (@SMatrix randn(6,4)) |> x -> x' * x;
julia> g(x) = 0.5 * (x' * S * x);
julia> x = @SVector randn(4);
julia> g(x)
12.515760061124144
julia> x' * S
1×4 LinearAlgebra.Adjoint{Float64,SArray{Tuple{4},Float64,1,4}}:
-1.82664 -13.4078 -5.39434 1.61724
julia> gradient(g, x)
([-1.82664, -13.4078, -5.39434, 1.61724],)
julia> ForwardDiff.gradient(g, x)'
1×4 LinearAlgebra.Adjoint{Float64,SArray{Tuple{4},Float64,1,4}}:
-1.82664 -13.4078 -5.39434 1.61724
julia> @benchmark gradient(g, $x)
BenchmarkTools.Trial:
memory estimate: 1.64 KiB
allocs estimate: 22
--------------
minimum time: 1.327 μs (0.00% GC)
median time: 1.374 μs (0.00% GC)
mean time: 2.254 μs (35.01% GC)
maximum time: 5.420 ms (99.91% GC)
--------------
samples: 10000
evals/sample: 10
julia> @benchmark ForwardDiff.gradient(g, $x)
BenchmarkTools.Trial:
memory estimate: 0 bytes
allocs estimate: 0
--------------
minimum time: 47.753 ns (0.00% GC)
median time: 48.698 ns (0.00% GC)
mean time: 48.696 ns (0.00% GC)
maximum time: 70.348 ns (0.00% GC)
--------------
samples: 10000
evals/sample: 988
julia> @benchmark $x' * S
BenchmarkTools.Trial:
memory estimate: 0 bytes
allocs estimate: 0
--------------
minimum time: 4.029 ns (0.00% GC)
median time: 4.034 ns (0.00% GC)
mean time: 4.209 ns (0.00% GC)
maximum time: 18.588 ns (0.00% GC)
--------------
samples: 10000
evals/sample: 1000
julia> @code_warntype gradient(g, x)
Body::Tuple{SArray{Tuple{4},Float64,1,4}}
32 1 ── %1 = (getfield)(args, 1)::SArray{Tuple{4},Float64,1,4} │
│ %2 = Zygote.nothing::Nothing │╻╷ forward
│ %3 = %new(Zygote.Context, %2)::Zygote.Context ││┃││ _forward
│ %4 = %new(LinearAlgebra.Adjoint{Float64,SArray{Tuple{4},Float64,1,4}}, %1)::LinearAlgebra.Adjoint{Float64,SArray{Tuple{4},Float64,1,4}}
│ %5 = Main.S::SArray{Tuple{4,4},Float64,2,16} ││││
│ %6 = invoke Zygote._forward(%3::Zygote.Context, Main.:*::typeof(*), %4::LinearAlgebra.Adjoint{Float64,SArray{Tuple{4},Float64,1,4}}, $(QuoteNode([4.22863 0.218509 -1.17347 1.24761; 0.218509 8.47895 2.104 0.21028; -1.17347 2.104 3.31319 -1.31538; 1.24761 0.21028 -1.31538 3.23605]))::SArray{Tuple{4,4},Float64,2,16}, %1::SArray{Tuple{4},Float64,1,4})::Tuple{Float64,Zygote.J{Tuple{typeof(*),LinearAlgebra.Adjoint{Float64,SArray{Tuple{4},Float64,1,4}},SArray{Tuple{4,4},Float64,2,16},SArray{Tuple{4},Float64,1,4}},Tuple{typeof(*),LinearAlgebra.Adjoint{Float64,SArray{Tuple{4},Float64,1,4}},SArray{Tuple{4,4},Float64,2,16},SArray{Tuple{4},Float64,1,4},Tuple{},getfield(Zygote, Symbol("##165#back2#123")){getfield(Zygote, Symbol("##121#122")){Tuple{Tuple{Nothing,Nothing},Tuple{}},Zygote.J{Tuple{typeof(Base.afoldl),typeof(*),Float64},Tuple{typeof(Base.afoldl),typeof(*),Float64}}}},getfield(Zygote, Symbol("##147#back2#110")){typeof(identity)},getfield(Zygote, Symbol("##1014#back2#564")){getfield(Zygote, Symbol("##562#563")){LinearAlgebra.Adjoint{Float64,SArray{Tuple{4},Float64,1,4}},SArray{Tuple{4},Float64,1,4}}},getfield(Zygote, Symbol("##1014#back2#564")){getfield(Zygote, Symbol("##562#563")){LinearAlgebra.Adjoint{Float64,SArray{Tuple{4},Float64,1,4}},SArray{Tuple{4,4},Float64,2,16}}}}}}
│ %7 = (Base.getfield)(%6, 1, true)::Float64 ││││╻ getindex
│ %8 = (Base.getfield)(%6, 2, true)::Zygote.J{Tuple{typeof(*),LinearAlgebra.Adjoint{Float64,SArray{Tuple{4},Float64,1,4}},SArray{Tuple{4,4},Float64,2,16},SArray{Tuple{4},Float64,1,4}},Tuple{typeof(*),LinearAlgebra.Adjoint{Float64,SArray{Tuple{4},Float64,1,4}},SArray{Tuple{4,4},Float64,2,16},SArray{Tuple{4},Float64,1,4},Tuple{},getfield(Zygote, Symbol("##165#back2#123")){getfield(Zygote, Symbol("##121#122")){Tuple{Tuple{Nothing,Nothing},Tuple{}},Zygote.J{Tuple{typeof(Base.afoldl),typeof(*),Float64},Tuple{typeof(Base.afoldl),typeof(*),Float64}}}},getfield(Zygote, Symbol("##147#back2#110")){typeof(identity)},getfield(Zygote, Symbol("##1014#back2#564")){getfield(Zygote, Symbol("##562#563")){LinearAlgebra.Adjoint{Float64,SArray{Tuple{4},Float64,1,4}},SArray{Tuple{4},Float64,1,4}}},getfield(Zygote, Symbol("##1014#back2#564")){getfield(Zygote, Symbol("##562#563")){LinearAlgebra.Adjoint{Float64,SArray{Tuple{4},Float64,1,4}},SArray{Tuple{4,4},Float64,2,16}}}}}
└─── (Base.mul_float)(0.5, %7) │││││╻╷ macro expansion
34 2 ┄─ (Base.mul_float)(1.0, %7) │╻╷╷╷╷╷ #66
│ %11 = (Base.mul_float)(1.0, 0.5)::Float64 ││╻ g
│ %12 = (Base.getfield)(%8, :t)::Tuple{typeof(*),LinearAlgebra.Adjoint{Float64,SArray{Tuple{4},Float64,1,4}},SArray{Tuple{4,4},Float64,2,16},SArray{Tuple{4},Float64,1,4},Tuple{},getfield(Zygote, Symbol("##165#back2#123")){getfield(Zygote, Symbol("##121#122")){Tuple{Tuple{Nothing,Nothing},Tuple{}},Zygote.J{Tuple{typeof(Base.afoldl),typeof(*),Float64},Tuple{typeof(Base.afoldl),typeof(*),Float64}}}},getfield(Zygote, Symbol("##147#back2#110")){typeof(identity)},getfield(Zygote, Symbol("##1014#back2#564")){getfield(Zygote, Symbol("##562#563")){LinearAlgebra.Adjoint{Float64,SArray{Tuple{4},Float64,1,4}},SArray{Tuple{4},Float64,1,4}}},getfield(Zygote, Symbol("##1014#back2#564")){getfield(Zygote, Symbol("##562#563")){LinearAlgebra.Adjoint{Float64,SArray{Tuple{4},Float64,1,4}},SArray{Tuple{4,4},Float64,2,16}}}}
│ %13 = (Base.getfield)(%12, 8, true)::getfield(Zygote, Symbol("##1014#back2#564")){getfield(Zygote, Symbol("##562#563")){LinearAlgebra.Adjoint{Float64,SArray{Tuple{4},Float64,1,4}},SArray{Tuple{4},Float64,1,4}}}
│ %14 = (Base.getfield)(%12, 9, true)::getfield(Zygote, Symbol("##1014#back2#564")){getfield(Zygote, Symbol("##562#563")){LinearAlgebra.Adjoint{Float64,SArray{Tuple{4},Float64,1,4}},SArray{Tuple{4,4},Float64,2,16}}}
│ %15 = π (%11, Float64) │││││╻ #121
│ %16 = (Core.getfield)(%13, Symbol("#1013#back"))::getfield(Zygote, Symbol("##562#563")){LinearAlgebra.Adjoint{Float64,SArray{Tuple{4},Float64,1,4}},SArray{Tuple{4},Float64,1,4}}
│ %17 = (Core.getfield)(%16, :b)::SArray{Tuple{4},Float64,1,4} │││││╻ #562
│ (Base.ifelse)(false, 0, 4) ││││││╻╷╷╷╷╷╷╷╷╷╷╷ *
│ %19 = (Base.getfield)(%17, :data)::NTuple{4,Float64} │││││││╻╷╷╷ broadcast
│ %20 = (Base.getfield)(%19, 1, false)::Float64 ││││││││╻ broadcast
│ %21 = (Base.mul_float)(%15, %20)::Float64 │││││││││╻ materialize
│ %22 = (Base.getfield)(%17, :data)::NTuple{4,Float64} ││││││││││╻ copy
│ %23 = (Base.getfield)(%22, 2, false)::Float64 │││││││││││╻ _broadcast
│ %24 = (Base.mul_float)(%15, %23)::Float64 ││││││││││││╻ macro expansion
│ %25 = (Base.getfield)(%17, :data)::NTuple{4,Float64} │││││││││││││╻ getindex
│ %26 = (Base.getfield)(%25, 3, false)::Float64 ││││││││││││││╻ getindex
│ %27 = (Base.mul_float)(%15, %26)::Float64 ││││││││││││││╻ *
│ %28 = (Base.getfield)(%17, :data)::NTuple{4,Float64} ││││││││││││││╻ getproperty
│ %29 = (Base.getfield)(%28, 4, false)::Float64 ││││││││││││││╻ getindex
│ %30 = (Base.mul_float)(%15, %29)::Float64 ││││││││││││││╻ *
│ %31 = (StaticArrays.tuple)(%21, %24, %27, %30)::NTuple{4,Float64} │││││││││││││
│ %32 = %new(SArray{Tuple{4},Float64,1,4}, %31)::SArray{Tuple{4},Float64,1,4}│││││││││││╻ Type
└─── goto #4 │││││││││││││
3 ── $(Expr(:unreachable)) │││││││││││││
4 ┄─ goto #5 │││││││││││
5 ── goto #6 ││││││││││
6 ── goto #7 │││││││││
7 ── %38 = %new(LinearAlgebra.Transpose{Float64,SArray{Tuple{4},Float64,1,4}}, %32)::LinearAlgebra.Transpose{Float64,SArray{Tuple{4},Float64,1,4}}
└─── goto #8 ││││││││
8 ── goto #9 │││││││
9 ── %41 = (Core.getfield)(%16, :a)::LinearAlgebra.Adjoint{Float64,SArray{Tuple{4},Float64,1,4}}
│ %42 = (Base.getfield)(%41, :parent)::SArray{Tuple{4},Float64,1,4} │││││││╻ getproperty
│ (Base.ifelse)(false, 0, 4) │││││││╻╷╷╷╷╷╷╷╷╷ broadcast
│ %44 = (Base.getfield)(%42, :data)::NTuple{4,Float64} ││││││││╻╷╷╷ materialize
│ %45 = (Base.getfield)(%44, 1, false)::Float64 │││││││││╻ copy
│ %46 = (Base.mul_float)(%45, %15)::Float64 ││││││││││╻ _broadcast
│ %47 = (Base.getfield)(%42, :data)::NTuple{4,Float64} │││││││││││╻ macro expansion
│ %48 = (Base.getfield)(%47, 2, false)::Float64 ││││││││││││╻ getindex
│ %49 = (Base.mul_float)(%48, %15)::Float64 ││││││││││││╻ *
│ %50 = (Base.getfield)(%42, :data)::NTuple{4,Float64} │││││││││││││╻ getproperty
│ %51 = (Base.getfield)(%50, 3, false)::Float64 │││││││││││││╻ getindex
│ %52 = (Base.mul_float)(%51, %15)::Float64 ││││││││││││╻ *
│ %53 = (Base.getfield)(%42, :data)::NTuple{4,Float64} │││││││││││││╻ getproperty
│ %54 = (Base.getfield)(%53, 4, false)::Float64 │││││││││││││╻ getindex
│ %55 = (Base.mul_float)(%54, %15)::Float64 ││││││││││││╻ *
└─── goto #11 ││││││││││││
10 ─ $(Expr(:unreachable)) ││││││││││││
11 ┄ goto #12 ││││││││││
12 ─ goto #13 │││││││││
13 ─ goto #14 ││││││││
14 ─ goto #15 │││││││
15 ─ goto #16 ││││││
16 ─ goto #17 │││││
17 ─ %64 = (Core.getfield)(%14, Symbol("#1013#back"))::getfield(Zygote, Symbol("##562#563")){LinearAlgebra.Adjoint{Float64,SArray{Tuple{4},Float64,1,4}},SArray{Tuple{4,4},Float64,2,16}}
│ %65 = invoke %64(%38::LinearAlgebra.Transpose{Float64,SArray{Tuple{4},Float64,1,4}})::Tuple{LinearAlgebra.Transpose{Float64,SArray{Tuple{4},Float64,1,4}},SArray{Tuple{4,4},Float64,2,16}}
│ %66 = (getfield)(%65, 1)::LinearAlgebra.Transpose{Float64,SArray{Tuple{4},Float64,1,4}} _gradtuple
│ %67 = (getfield)(%65, 2)::SArray{Tuple{4,4},Float64,2,16} ││││││
└─── goto #18 ││││
18 ─ invoke Zygote.accum_param(%3::Zygote.Context, %5::SArray{Tuple{4,4},Float64,2,16}, %67::SArray{Tuple{4,4},Float64,2,16})
│ %70 = (Base.getfield)(%66, :parent)::SArray{Tuple{4},Float64,1,4} ││││╻╷╷ #568
│ (Base.ifelse)(false, 0, 4) ││││╻╷╷╷╷╷╷╷╷ materialize
│ (Base.ifelse)(false, 0, 4) │││││╻╷╷╷╷╷ instantiate
│ %73 = (4 === 4)::Bool ││││││╻╷╷╷╷ combine_axes
│ %74 = (Base.and_int)(true, %73)::Bool │││││││╻ broadcast_shape
└─── goto #20 if not %74 ││││││││┃││ _bcs
19 ─ goto #21 │││││││││┃│ _bcs1
20 ─ goto #21 ││││││││││┃ _bcsm
21 ┄ %78 = φ (#19 => %74, #20 => false)::Bool ││││││││││
└─── goto #23 if not %78 ││││││││││
22 ─ goto #29 ││││││││││
23 ─ %81 = (4 === 4)::Bool │││││││││││╻╷ ==
│ %82 = (Base.and_int)(true, %81)::Bool ││││││││││││╻ &
└─── goto #25 if not %82 │││││││││││
24 ─ goto #26 │││││││││││
25 ─ goto #26 │││││││││││
26 ┄ %86 = φ (#24 => %82, #25 => false)::Bool ││││││││││
└─── goto #28 if not %86 ││││││││││
27 ─ goto #29 ││││││││││
28 ─ %89 = %new(Base.DimensionMismatch, "arrays could not be broadcast to a common size")::DimensionMismatchpe
│ (Base.Broadcast.throw)(%89) ││││││││││
└─── $(Expr(:unreachable)) ││││││││││
29 ┄ goto #30 │││││││││
30 ─ goto #31 ││││││││
31 ─ goto #32 │││││││
32 ─ goto #33 │││││╻ instantiate
33 ─ %96 = (Base.getfield)(%70, :data)::NTuple{4,Float64} ││││││╻╷╷╷ _broadcast
│ %97 = (Base.getfield)(%96, 1, false)::Float64 │││││││╻ macro expansion
│ %98 = (Base.add_float)(%46, %97)::Float64 ││││││││╻ accum
│ %99 = (Base.getfield)(%70, :data)::NTuple{4,Float64} │││││││││╻ getproperty
│ %100 = (Base.getfield)(%99, 2, false)::Float64 │││││││││╻ getindex
│ %101 = (Base.add_float)(%49, %100)::Float64 │││││││││╻ +
│ %102 = (Base.getfield)(%70, :data)::NTuple{4,Float64} │││││││││╻ getproperty
│ %103 = (Base.getfield)(%102, 3, false)::Float64 │││││││││╻ getindex
│ %104 = (Base.add_float)(%52, %103)::Float64 │││││││││╻ +
│ %105 = (Base.getfield)(%70, :data)::NTuple{4,Float64} │││││││││╻ getproperty
│ %106 = (Base.getfield)(%105, 4, false)::Float64 │││││││││╻ getindex
│ %107 = (Base.add_float)(%55, %106)::Float64 │││││││││╻ +
│ %108 = (StaticArrays.tuple)(%98, %101, %104, %107)::NTuple{4,Float64} ││││││││
│ %109 = %new(SArray{Tuple{4},Float64,1,4}, %108)::SArray{Tuple{4},Float64,1,4}│││││╻ Type
└─── goto #35 ││││││││
34 ─ $(Expr(:unreachable)) ││││││││
35 ┄ goto #36 ││││││
36 ─ goto #37 │││││
37 ─ goto #38 ││││
38 ─ goto #39 │││
39 ─ %116 = (Core.tuple)(%109)::Tuple{SArray{Tuple{4},Float64,1,4}} │││╻ tail
└─── goto #40 ││
40 ─ return %116 │
41 ─ goto #2
I also do not see any change from Zygote.refresh()
:
julia> using Zygote
julia> Zygote.refresh()
julia> gradient(g, x)
([-1.82664, -13.4078, -5.39434, 1.61724],)
julia> @benchmark gradient(g, $x)
BenchmarkTools.Trial:
memory estimate: 1.64 KiB
allocs estimate: 22
--------------
minimum time: 1.346 μs (0.00% GC)
median time: 1.393 μs (0.00% GC)
mean time: 2.312 μs (35.93% GC)
maximum time: 5.704 ms (99.92% GC)
--------------
samples: 10000
evals/sample: 10
from zygote.jl.
Related Issues (20)
- `repeat(X; outer, inner)` triggers scalar indexing error with CUDA HOT 1
- Missing support for muladd in case of brodcasting with a complex argument HOT 1
- `nothing` in output of a `pullback` HOT 2
- Assignment to multiple arrays is not differentiable on GPU since Zygote.jl 0.6.67 HOT 5
- Spurious "Output is complex, so the gradient is not defined" error HOT 2
- NaN in gradient of abs() on complex 0 HOT 1
- Pullback on mean() gives illegal memory access code 700 HOT 32
- test
- Type unstable gradients (@code_warntype) HOT 1
- Type unstable gradients HOT 1
- Zygote gradients different from ForwardDiff/ReverseDiff on Julia 1.10-rc2 HOT 3
- try/catch is not supported when attempting to use `remake` with Zygote HOT 1
- gradient of SVD not working for complex input HOT 1
- `Zygote` doesn't properly work with `Metal.jl` and half precision. HOT 4
- `gradient` broken for `(*)(::Diagonal{Real}, ::Matrix{Complex}, ::Diagonal{Real})` when updating Julia 1.8 -> 1.9 HOT 6
- Method ambiguities reported by Aqua
- slow/high allocation gradient with mapreduce and iterators HOT 11
- error in summation of product iterator HOT 2
- `sort(x; rev=true)` is not supported HOT 1
- Incorrect gradients for `plan_rfft(x) * x` HOT 2
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from zygote.jl.