Skip to content

rm adjoint & transpose adjoints #1259

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.

Sign up for GitHub

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

Merged
mcabbott merged 2 commits into from
Jul 13, 2022
Merged

rm adjoint & transpose adjoints #1259

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
bdefedc
rm adjoint + transpose adjoint
mcabbott Jul 3, 2022
26ae24b
restore parent adjoints
mcabbott Jul 3, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
28 changes: 0 additions & 28 deletions src/lib/array.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -337,34 +337,6 @@ end
# LinearAlgebra
# =============

@adjoint function transpose(x)
back(Δ) = (transpose(Δ),)
back(Δ::NamedTuple{(:parent,)}) = (Δ.parent,)
return transpose(x), back
end

@adjoint function LinearAlgebra.Transpose(x)
back(Δ) = (LinearAlgebra.Transpose(Δ),)
back(Δ::NamedTuple{(:parent,)}) = (Δ.parent,)
return LinearAlgebra.Transpose(x), back
end


@adjoint function Base.adjoint(x)
back(Δ) = (Δ',)
back(Δ::NamedTuple{(:parent,)}) = (Δ.parent,)
return x', back
end

@adjoint function LinearAlgebra.Adjoint(x)
back(Δ) = (LinearAlgebra.Adjoint(Δ),)
back(Δ::NamedTuple{(:parent,)}) = (Δ.parent,)
return LinearAlgebra.Adjoint(x), back
end

@adjoint parent(x::LinearAlgebra.Adjoint) = parent(x), ȳ -> (LinearAlgebra.Adjoint(ȳ),)
@adjoint parent(x::LinearAlgebra.Transpose) = parent(x), ȳ -> (LinearAlgebra.Transpose(ȳ),)
Comment on lines 365 to 366
Copy link
Member Author

@mcabbott mcabbott Jul 3, 2022

Choose a reason for hiding this comment

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

I forgot that JuliaDiff/ChainRulesCore.jl#446 isn't merged, so removing these lines probably isn't a good idea.

Before:

julia> gradient(x -> first(parent(x)), [1,2,3]')[1]
1×3 adjoint(::Vector{Float64}) with eltype Float64:
 1.0  0.0  0.0

This PR, with tagged CRC:

julia> gradient(x -> first(parent(x)), [1,2,3]')[1]
(parent = [1.0, 0.0, 0.0],)

Copy link
Member

@ToucheSir ToucheSir Jul 3, 2022

Choose a reason for hiding this comment

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

One alternative idea is to desguar the @adjoint and overload _pullback directly. That'll have the benefit of removing some unnecessary layers, while making it easier to do stuff like FluxML/ZygoteRules.jl#22.

Copy link
Member Author

@mcabbott mcabbott Jul 3, 2022

Choose a reason for hiding this comment

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

Glad to see there were tests for this, at least: fails on isapprox(::NamedTuple{(:parent,), ...}, ::Matrix{Float64}).


function _kron(mat1::AbstractMatrix,mat2::AbstractMatrix)
m1, n1 = size(mat1)
mat1_rsh = reshape(mat1,(1,m1,1,n1))
Expand Down