JuliaDiff / JuliaDiff/ForwardDiff.jl

Supporting copysign and flipsign

Open
#344 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Julia
Stars
1k
Forks
160
PR merge metrics
No merged PRs in 30d

Description

I have some code that contains usages of copysign and flipsign; rather than replacing these calls with sign-based equivalents I was hoping to add support for copysign and flipsign to ForwardDiff. Here's what I've tried:

Status quo on 0.7.0
  | | |_| | | | (_| |  |  Version 0.7.0 (2018-08-08 06:46 UTC)
 _/ |\__'_|_|_|\__'_|  |  
|__/                   |  x86_64-linux-gnu

julia> using ForwardDiff

julia> ForwardDiff.derivative(x -> flipsign(1., x), 1.)
ERROR: MethodError: no method matching Float64(::ForwardDiff.Dual{ForwardDiff.Tag{getfield(Main, Symbol("##3#4")),Float64},Float64,1})
Closest candidates are:
  Float64(::Real, ::RoundingMode) where T<:AbstractFloat at rounding.jl:173
  Float64(::T<:Number) where T<:Number at boot.jl:725
  Float64(::Int8) at float.jl:60
  ...
Stacktrace:
 [1] Float64(::ForwardDiff.Dual{ForwardDiff.Tag{getfield(Main, Symbol("##3#4")),Float64},Float64,1}) at ./deprecated.jl:468
 [2] flipsign(::Float64, ::ForwardDiff.Dual{ForwardDiff.Tag{getfield(Main, Symbol("##3#4")),Float64},Float64,1}) at ./floatfuncs.jl:13
Hacking around the missing method
julia> Core.Float64(x::ForwardDiff.Dual) = convert(Float64, x)

julia> ForwardDiff.derivative(x -> flipsign(1., x), 1.)
ERROR: StackOverflowError:
Stacktrace:
 [1] Float64(::ForwardDiff.Dual{ForwardDiff.Tag{getfield(Main, Symbol("##5#6")),Float64},Float64,1}) at ./REPL[4]:1
 [2] convert(::Type{Float64}, ::ForwardDiff.Dual{ForwardDiff.Tag{getfield(Main, Symbol("##5#6")),Float64},Float64,1}) at ./number.jl:7
 ... (the last 2 lines are repeated 39998 more times)
 [79999] Float64(::ForwardDiff.Dual{ForwardDiff.Tag{getfield(Main, Symbol("##5#6")),Float64},Float64,1}) at ./REPL[4]:1
Adding rules to DiffRules.jl
# within DiffRules.jl/src/rules.jl
@define_diffrule Base.copysign(x,y)   = :( signbit($y) == signbit($x) ? one($x) : -one($x)         ), :( zero($y)                                                                )
@define_diffrule Base.flipsign(x,y)   = :( signbit($y) ? -one($x) : one($x)                        ), :( zero($y)                                                                )

yields

julia> ForwardDiff.derivative(x -> flipsign(1., x), 1.)
ERROR: MethodError: flipsign(::Float64, ::ForwardDiff.Dual{ForwardDiff.Tag{getfield(Main, Symbol("##3#4")),Float64},Float64,1}) is ambiguous. Candidates:
  flipsign(x::AbstractFloat, y::ForwardDiff.Dual{Ty,V,N} where N where V<:Real) where Ty in ForwardDiff at /home/schmrlng/.julia/packages/ForwardDiff/OXtu9/src/dual.jl:114
  flipsign(x::Real, y::ForwardDiff.Dual{Ty,V,N} where N where V<:Real) where Ty in ForwardDiff at /home/schmrlng/.julia/packages/ForwardDiff/OXtu9/src/dual.jl:114
  flipsign(x::Float64, y::Real) in Base at floatfuncs.jl:13
Possible fix, define
  flipsign(::Float64, ::ForwardDiff.Dual{Ty,V,N} where N where V<:Real)
Stacktrace:
 [1] (::getfield(Main, Symbol("##3#4")))(::ForwardDiff.Dual{ForwardDiff.Tag{getfield(Main, Symbol("##3#4")),Float64},Float64,1}) at ./REPL[2]:1
 [2] derivative(::getfield(Main, Symbol("##3#4")), ::Float64) at /home/schmrlng/.julia/packages/ForwardDiff/OXtu9/src/derivative.jl:14

It looks like ForwardDiff already has some method definitions designed to avoid ambiguities as above (see https://github.com/JuliaDiff/ForwardDiff.jl/blob/master/src/dual.jl?utf8=%E2%9C%93#L114); should we consider also adding the concrete FloatXX types to AMBIGUOUS_TYPES? Or is there some "better"-practices way for me to approach differentiating through the second argument of copysign and flipsign?

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start with the reproducer for differentiating flipsign, then read src/dual.jl around the existing AMBIGUOUS_TYPES methods and derivative.jl. Review the proposed DiffRules.jl/src/rules.jl rules and determine how copysign and flipsign should handle differentiation through their second argument without ambiguity. Done means both functions are supported with unambiguous methods and the reported derivative case works.

Written by the indexing model from the issue text.

Assessment

Tech stack
julia
Domain
tooling
Issue type
Feature
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Mostly clear
Newbie friendliness
42/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.