ForwardDiffSensitivity() and TrackerAdjoint() give different answers

using DiffEqSensitivity, OrdinaryDiffEq, Zygote, StochasticDiffEq

# similar to documentation, but out-of-place 
function fiip(u,p,t)
  du1 = dx = p[1]*u[1] - p[2]*u[1]*u[2]
  du2 = dy = -p[3]*u[2] + p[4]*u[1]*u[2]
  return [du1,du2]
end
function fiip_s(u,p,t)
  du1 = dx = 0.001
  du2 = dy = 0.001
  return [du1,du2]
end

p = [1.5,1.0,3.0,1.0]; u0 = [1.0;1.0]
probSDE = SDEProblem(fiip,fiip_s,u0,(0.0,1.0),p) # changed terminal time to 1.0 
solSDE = solve(probSDE,SOSRI())

# correct answer
du01SDE,dp1SDE = Zygote.gradient((u0,p)->sum(solve(probSDE,SOSRI(),u0=u0,p=p,saveat=0.1,sensealg=TrackerAdjoint())),u0,p)
# gradient with respect to u0 = nothing, which is wrong
du01SDE,dp1SDE = Zygote.gradient((u0,p)->sum(solve(probSDE,SOSRI(),u0=u0,p=p,saveat=0.1,sensealg=ForwardDiffSensitivity())),u0,p) 
# error but docs say it is supported?
du01SDE,dp1SDE = Zygote.gradient((u0,p)->sum(solve(probSDE,SOSRI(),u0=u0,p=p,saveat=0.1,sensealg=ReverseDiffAdjoint())),u0,p) 

FWIW, they all give the same answer. The forward-mode one just doesn’t return the u0 part of the derivative (track the issue here: https://github.com/SciML/DiffEqSensitivity.jl/issues/156), but nothing is correct for Zygote and if you try to use that gradient value you’ll properly get an error because right now our forward-mode overload doesn’t compute it.

ReverseDiffAdjoint just needs DistributionsAD and a ReverseDiff tag. I put those in motion and by tomorrow morning it should be good again. Sorry about that.

1 Like

Thank you! Very helpful.

I had interpreted nothing as a “generalized zero” as explained in this issue: https://github.com/FluxML/Zygote.jl/issues/329 so I thought it was “wrong” but now I see nothing can also mean “not yet implemented”.

Thanks again!

Yeah, that issue is about how nothing and Zero() could (and should) be different. But Zygote just has nothing so you do have to be a bit careful.