ForwardDiff.gradient(x->log(1/max(1f-20,x[1])), zeros(Float32,1))) returns [NaN32]

[micklat]

Hi and thanks for ForwardDiff!

I've ran into what seems like a bug:

import ForwardDiff

println("This should be 0: ",
    ForwardDiff.gradient(x->log(1/max(1f-10,x[1])), zeros(Float32,1)))

println("This should be 0 too: ",
    ForwardDiff.gradient(x->log(1/max(1f-20,x[1])), zeros(Float32,1)))

output is:

This should be 0: Float32[-0.0]
This should be 0 too: Float32[NaN32]

I guess this looks like some kind of overflow problem, but the number isn't that big, so I wouldn't expect overflow here.

ForwardDiff version is 0.7
julia version is 0.6.2 (2017-12-13 18:08 UTC) x86_64-linux-gnu.

HTH.

[fredrikekre]

You can set

ForwardDiff.jl/src/prelude.jl

Line 1 in 49292ee

const NANSAFE_MODE_ENABLED = false

to true, see #179 and linked issues

[micklat]

Thanks, that worked around it! I wonder how nans got into this in the first place.

[micklat]

Oh I see, the drivative of 1/x overflows when x=1f-20.

[jrevels]

Closing since this seems resolved, but feel to reopen if there's still an issue here!