Array output seems to be initialized but not filled

[gdalle]

I caught this bug as part of my DifferentiationInterface.jl test suite. The first object returned by value_and_pullback!! looks like the result of Vector{Float64}(undef, n) instead of containing the actual function output. And it changes from run to run. Meanwhile the VJP is stable and correct:

julia> import Taped

julia> scalar_to_vector(x::Number) = [sin(x), sin(2x)]
scalar_to_vector (generic function with 1 method)

julia> r = Taped.build_rrule(scalar_to_vector, 1.0)
Taped.DerivedRule{Core.OpaqueClosure{Tuple{Taped.AugmentedRegister{Taped.CoDual{typeof(scalar_to_vector), Taped.NoTangent}, Taped.NoTangentRef}, Taped.AugmentedRegister{Taped.CoDual{Float64, Float64}, Base.RefArray{Float64, Vector{Float64}, Nothing}}}, Taped.AugmentedRegister{Taped.CoDual{Vector{Float64}, Vector{Float64}}, Base.RefArray{Vector{Float64}, Vector{Vector{Float64}}, Nothing}}}, Tuple{Taped.NoTangentStack, Taped.Stack{Float64}}, Core.OpaqueClosure{Tuple{Vector{Float64}, Taped.NoTangent, Float64}, Nothing}, Val{false}, Val{2}}((::Taped.AugmentedRegister{Taped.CoDual{typeof(scalar_to_vector), Taped.NoTangent}, Taped.NoTangentRef}, ::Taped.AugmentedRegister{Taped.CoDual{Float64, Float64}, Base.RefArray{Float64, Vector{Float64}, Nothing}})::Taped.AugmentedRegister{Taped.CoDual{Vector{Float64}, Vector{Float64}}, Base.RefArray{Vector{Float64}, Vector{Vector{Float64}}, Nothing}}->◌, (::Vector{Float64}, ::Taped.NoTangent, ::Float64)::Nothing->◌, (Taped.NoTangentStack(), Taped.Stack{Float64}(Float64[], 0)), Taped.Stack{Int32}(Int32[], 0), Val{false}(), Val{2}())

julia> Taped.value_and_pullback!!(r, [1.0, 2.0], scalar_to_vector, 3.0)
([0.0, 0.0], (Taped.NoTangent(), 2.8506886500010182))

julia> Taped.value_and_pullback!!(r, [1.0, 2.0], scalar_to_vector, 3.0)
([6.1801573873894e-310, 6.18027838103766e-310], (Taped.NoTangent(), 2.8506886500010182))

julia> Taped.value_and_pullback!!(r, [1.0, 2.0], scalar_to_vector, 3.0)
([5.0e-324, 0.0], (Taped.NoTangent(), 2.8506886500010182))

julia> 1.0 * cos(3.0) + 2.0 * 2cos(6.0)
2.8506886500010182

[willtebbutt]

This is not nice. Will investigate.

[gdalle]

FYI you can now use Taped with DifferentiationInterface.jl, the functionality has been merged in
gdalle/DifferentiationInterface.jl#92

It is not documented but that may allow you to catch more bugs if you want to try out our test suite

[willtebbutt]

Amazing, thanks. I'll definitely give your test suite a go -- I'm sure there are holes in mine.

[willtebbutt]

On this, could you point me towards the test suite that I should be looking at?

[gdalle]

1. Clone the repo https://github.com/gdalle/DifferentiationInterface.jl/
2. Open a REPL and activate the test environment with TestEnv.activate()
3. Run https://github.com/gdalle/DifferentiationInterface.jl/blob/main/test/taped.jl
4. The code you want to modify if things fail due to the bridging is at https://github.com/gdalle/DifferentiationInterface.jl/blob/main/ext/DifferentiationInterfaceTapedExt/DifferentiationInterfaceTapedExt.jl

[willtebbutt]

Okay. This should now be fixed on main.

[gdalle]

cool, thanks!
I'll beef up our test suite in the coming days, will let you know if I run into any more trouble

[willtebbutt]

Great. I'll close this. Please open more issues if you run into problems