StackOverflow Issue on Certain Functions

[jmurphy6895]

I'm adding some tests to some packages using DifferentiationInterface to make sure everything is differentiable and I'm occasionally running into a weird bug where I get a StackOverflow error

Here's a simple MWE where I run into it.

When trying to differentiate through this function --> the derivative is 0.0 everywhere

using DifferentiationInterface
import FiniteDiff, ForwardDiff

function get_Δat(JD::Number)
    # If JD is before ΔAT_Data[1,1], then return 10.0.
    @inbounds if JD < _ΔAT[1, 1]
        return 10.0
    else
        for i = 2:size(_ΔAT, 1)
            (JD < _ΔAT[i, 1]) && return _ΔAT[i - 1, 2]
        end
    end

    # In this case, JD is after ΔAT_Data[end,1].
    return @inbounds _ΔAT[end, 2]
end

I get a StackOverflow error when running this code (only on ForwardDiff not FiniteDiff):

jd_utc = date_to_jd(2004, 4, 6, 7, 51, 28.386009)

f_fd, df_fd = value_and_derivative(
    get_Δat,
    AutoFiniteDiff(),
    jd_utc
)

f_ad, df_ad = value_and_derivative(
    get_Δat,
    AutoForwardDiff(),
    jd_utc
)

But have no problem when running
ForwardDiff.derivative(get_Δat, jd_utc)

Do you have any idea of what might be happening? I copied the function I'm working with because it's short but to get the pure function you can add using SatelliteToolboxTransformations.

Pkg Status for completeness -- Julia v"1.10.0"
Project SatelliteToolboxTransformations v0.1.8
Status C:\Users\jmurp\.julia\dev\SatelliteToolboxTransformations\Project.toml
[a8cc5b0e] Crayons v4.1.1
[8bb1440f] DelimitedFiles v1.9.1
[a0c0ee7d] DifferentiationInterface v0.6.1
[6a86dc24] FiniteDiff v2.24.0
[f6369f11] ForwardDiff v0.10.36
[a98d9a8b] Interpolations v0.15.1
[189a3867] Reexport v1.2.2
[74f56ac7] ReferenceFrameRotations v3.0.2
[9e17983a] SatelliteToolboxBase v0.3.2
[6c6a2e73] Scratch v1.2.1
[90137ffa] StaticArrays v1.9.7
[ade2ca70] Dates
[f43a241f] Downloads v1.6.0
[37e2e46d] LinearAlgebra

[gdalle]

Hi, thanks for reporting this! Can you provide a self-sufficient runnable example, including the definition of what seem to be global variables in your code (ΔAT_Data for example)? I'm assuming the function date_to_jd comes from SatelliteToolboxTransformations? Are there any other imports I need?

[jmurphy6895]

Yeah, sorry about that here is a simpler self-contained version

using DifferentiationInterface
import ForwardDiff, FiniteDiff

using SatelliteToolboxTransformations

jd_utc = date_to_jd(2004, 4, 6, 7, 51, 28.386009)

f_fd, df_fd = value_and_derivative(
    get_Δat,
    AutoFiniteDiff(),
    jd_utc
)

f_ad, df_ad = value_and_derivative(
    get_Δat,
    AutoForwardDiff(),
    jd_utc
)

ForwardDiff.derivative(get_Δat, jd_utc)

[gdalle]

Okay I know what's happening, your function is piecewise constant and its output is not created from its input, so ForwardDiff doesn't get a Dual number in the end. I didn't plan for this case but there's an easy fix.

[gdalle]

Can you try installing DI from the branch in the PR #526 to see if it solves your issue?

[jmurphy6895]

No more StackOverflow, but there a new error message

ERROR: ArgumentError: Tuple contains 0 elements, must contain exactly 1 element
Stacktrace:
[1] only(x::Tuple{})
@ Base.Iterators .\iterators.jl:1537
[2] value_and_derivative(::typeof(get_Δat), ::DifferentiationInterfaceForwardDiffExt.ForwardDiffOneArgDerivativePrep{…}, ::AutoForwardDiff{…}, ::Float64)
@ DifferentiationInterfaceForwardDiffExt C:\Users\jmurp.julia\packages\DifferentiationInterface\dRZrJ\ext\DifferentiationInterfaceForwardDiffExt\onearg.jl:119
[3] value_and_derivative(::typeof(get_Δat), ::AutoForwardDiff{nothing, Nothing}, ::Float64)
@ DifferentiationInterface C:\Users\jmurp.julia\packages\DifferentiationInterface\dRZrJ\src\fallbacks\no_prep.jl:61
[4] top-level scope
@ Untitled-1:14
Some type information was truncated. Use show(err) to see complete types.

[gdalle]

How about now?

[jmurphy6895]

Yeah, it's working! Thanks so much for the quick help!

[gdalle]

I opened #527 to keep track of the tests I need to add, but I'll release the fix before that