Failure to handle nothing output from Zygote.jl

[MilesCranmer]

julia> using Zygote, DifferentiationInterface

julia> safe_log(x) = x > zero(x) ? log(x) : convert(typeof(x), NaN)
safe_log (generic function with 1 method)

julia> derivative(safe_log, AutoZygote(), 0.0)
ERROR: MethodError: no method matching iterate(::Nothing)
The function `iterate` exists, but no method is defined for this combination of argument types.

Closest candidates are:
  iterate(::Combinatorics.Combinations)
   @ Combinatorics ~/.julia/packages/Combinatorics/Udg6X/src/combinations.jl:13
  iterate(::Combinatorics.Combinations, ::Any)
   @ Combinatorics ~/.julia/packages/Combinatorics/Udg6X/src/combinations.jl:13
  iterate(::Tables.DictRowTable)
   @ Tables ~/.julia/packages/Tables/8p03y/src/dicts.jl:122
  ...

Stacktrace:
 [1] dot(x::Float64, y::Nothing)
   @ LinearAlgebra ~/.julia/juliaup/julia-1.11.2+0.aarch64.apple.darwin14/share/julia/stdlib/v1.11/LinearAlgebra/src/generic.jl:855
 [2] 
   @ DifferentiationInterface ~/.julia/packages/DifferentiationInterface/gjT8p/src/first_order/pushforward.jl:161
 [3] #6
   @ ~/.julia/packages/DifferentiationInterface/gjT8p/src/first_order/pushforward.jl:192 [inlined]
 [4] ntuple
   @ ./ntuple.jl:48 [inlined]
 [5] value_and_pushforward
   @ ~/.julia/packages/DifferentiationInterface/gjT8p/src/first_order/pushforward.jl:191 [inlined]
 [6] pushforward
   @ ~/.julia/packages/DifferentiationInterface/gjT8p/src/first_order/pushforward.jl:220 [inlined]
 [7] derivative
   @ ~/.julia/packages/DifferentiationInterface/gjT8p/src/first_order/derivative.jl:124 [inlined]
 [8] derivative(::typeof(safe_log), ::AutoZygote, ::Float64)
   @ DifferentiationInterface ~/.julia/packages/DifferentiationInterface/gjT8p/src/fallbacks/no_prep.jl:49
 [9] top-level scope
   @ REPL[13]:1
Some type information was truncated. Use `show(err)` to see complete types.

Zygote.jl returns a nothing given an invalid input, so I think this should be correctly handled. Returning a NaN seems like an option?

For AutoForwardDiff, this gives us:

julia> derivative(safe_log, AutoForwardDiff(), 0.0)
0.0

[gdalle]

Hi @MilesCranmer, sorry for the delay. I am aware of the issue, it is more or less a duplicate of #604.
The trouble with replacing nothing with something like NaN is that it would substitute a non-existing rule for an existing one. If the function happened to fix NaNs later on, Zygote's failure to differentiate part of the function would be completely hidden. Therefore, I don't think we should get rid of these nothings, because they denote a missing or faulty rule which needs to be corrected and not ignored.
However, I agree that we should add a better error message. Essentially what happens here is that I try to do y, pb = Zygote.pullback(f, x), which fails for y, pb = nothing. Here I could throw a custom ZygoteNothingError for example?

[MilesCranmer]

Yeah that sounds good

[MilesCranmer]

Actually maybe I could suggest another workaround – could we have an opt-in setting in the AutoZygote backend that does replace nothing with NaN or 0.0? For my use-case (SymbolicRegression.jl) this is basically the only way I can use it via the DifferentiationInterface API. My current workaround is to handle the nothings explicitly: https://github.com/SymbolicML/DynamicExpressions.jl/blob/15849184f38636631a7a5aee7a04c02bcb7dde33/ext/DynamicExpressionsZygoteExt.jl#L6-L30

[MilesCranmer]

Actually a 0.0 output would be preferred here – maybe the setting could be like AutoZygote(replace_nothing_with=(x, dx) -> dx) (the default; just leaves nothing as-is) and could also be set to

AugoZygote(replace_nothing_with=(x, dx) -> zero(x))

or

AutoZygote(replace_nothing_with=(x, dx) -> convert(typeof(x), NaN)

so then I have full awareness of this being a potential footgun

[gdalle]

I get where you're coming from but I'm very skeptical about this. A nothing with Zygote denotes a missing rule, and it's not DI's place to go around and tinker with the output of missing rules. Otherwise we'd have to offer the same for every backend: "your function doesn't support dual numbers, do you want me to output NaN instead?". The proper fix for a missing rule will always be the AD backend itself and not DI.

[MilesCranmer]

Wait so does ForwardDiff output 0.0 for missing rules?

[gdalle]

No, that's my point. Missing rules error for ForwardDiff, for Enzyme too, I don't see why we should introduce a special shortcut to zero them out for Zygote only

[MilesCranmer]

Oh I see. I thought ForwardDiff was outputting 0.0 for missing rules. I'm surprised that it covers so many functions!