Add adjoint for pairs(::NamedTuple) and Pair

[YingboMa]

Fix SciML/DiffEqFlux.jl#105

[YingboMa]

julia> @nograd pairs

julia> Zygote.gradient(()->sum(neural_ode(dudt,x,tspan,Tsit5(),save_everystep=false,save_start=false,sensealg=BacksolveAdjoint())),p)
Grads(...)

[MikeInnes]

Thanks! FWIW, a minimal reproducer is

julia> foo(;kw...) = 1
foo (generic function with 1 method)

julia> gradient(() -> foo(a=1,b=2.0))
ERROR: Can't differentiate gc_preserve_end expression

(This would be good to have as a test case.)

This works if all kwargs are of the same type; it would be great to understand why that is, to make sure we've got the whole issue fixed.

[YingboMa]

CI failure looks unrelated. @MikeInnes could you review again?

[MikeInnes]

The patch looks good to me. I'd still like to understand what when wrong in detail before merging, but I can look into that myself if it's not already clear.

bors try

[bors]

try

Build succeeded

• ci/gitlab/gitlab.com

[CarloLucibello]

@MikeInnes what should we do here?

[cossio]

bump

[CarloLucibello]

related to #669 , but of the tests in this PR

    @test (x->10*pairs((a=x, b=2))[1])'(100) === 10
    @test (x->10*pairs((a=x, b=2))[2])'(100) === 0
    foo(;kw...) = 1
    @test gradient(() -> foo(a=1,b=2.0)) === ()
    @test (x->10*(x => 2)[1])'(100) === 10
    @test (x->10*(x => 2)[2])'(100) === 0

only the 3rd passes on current master, so we should merge this as well

[CarloLucibello]

@YingboMa could your rebase on master and remove the pairs adjoint from #669 ?

[CarloLucibello]

bors r+

[bors]

Build succeeded:

• ci/gitlab/gitlab.com

[Yansf677]

How to write the adjoint method for LinearAlgebra.diagm of Pair{Int64,Array{Float64,1}}