Skip to content

Product of Inf terms leading to NaNs #745

Closed
#777
Closed
Product of Inf terms leading to NaNs#745
#777
Labels
nan-safe-mode

Description

@Uroc327
Uroc327
opened on Apr 9, 2025

ForwardDiff (v1.0.1 as well as v0.10.38) fails to compute the gradient when the inputs are too large for the following function.

julia> using ForwardDiff

julia> foo(a) = a[1] * exp(-a[2])
foo (generic function with 1 method)

julia> ForwardDiff.gradient(foo, [1., -1e3])
2-element Vector{Float64}:
 NaN
 NaN

The correct gradient should be [Inf, -Inf]. This works for values a2 small enough such that exp(a2) is finite.

Activity

  1. Uroc327
    changed the title [-]Product Rule and Infinities leading to NaNs[/-] [+]Product of Inf terms leading to NaNs[/+] on Apr 9, 2025

  2. devmotion commented on Oct 1, 2025

    @devmotion
    devmotion
    on Oct 1, 2025
    Member

    The more general issue is #774. As mentioned in #774, until the problem is fixed you can work around this problem by enabling NaN-safe mode AND restricting the chunksize to 1:

    julia> ForwardDiff.gradient(foo, [1., -1e3], ForwardDiff.GradientConfig(foo, [1., -1e3], ForwardDiff.Chunk{1}()))
2-element Vector{Float64}:
  Inf
 -Inf
  3. devmotion
    added
    nan-safe-mode
    on Oct 1, 2025
  4. devmotion
    linked a pull request that will close this issueFix NaN-safe mode #777on Oct 1, 2025
  5. devmotion
    closed this as completedin #777on Oct 7, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    nan-safe-mode

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions