Continuous is broken in the case of non-analytic gradients

[MikeInnes]

If t is not provided:

julia> adjoint_sensitivities(sol,Vern9(),dg)
ERROR: type ODEAdjointSensitivityFunction has no field g_gradient_config

It would be useful for me to understand how the continuous case is different from the discrete one; at present it doesn't appear to be tested and has bitrotted.

[ChrisRackauckas]

Except it's not broken and it is tested? https://github.com/JuliaDiffEq/DiffEqSensitivity.jl/blob/master/test/adjoint.jl#L73-L117

What code did you run?

[MikeInnes]

julia> function lotka_volterra(du,u,p,t)
         x, y = u
         α, β, δ, γ = p
         du[1] = dx = (α - β*y)x*t
         du[2] = dy = (δ*x - γ)y
       end

julia> p = [2.2, 1.0, 2.0, 0.4]
julia> prob = ODEProblem(lotka_volterra,[1.0,1.0],(0.0,10.0), p)
julia> sol = solve(prob,Vern9())

julia> function dg(out,u,p,t,i)
         out .= 1 .- u
       end

julia> adjoint_sensitivities(sol,Vern9(),dg)
ERROR: type ODEAdjointSensitivityFunction has no field g_gradient_config

My example might be bad, in which case it's just a case of throwing a different error. But this line refers to a field that doesn't exist, so something has gone wrong in that code.

[ChrisRackauckas]

You didn't supply a g, and dg is in the form for a discrete sensitivity. This is just a case of RTFM

[ChrisRackauckas]

If t is not provided: julia> adjoint_sensitivities(sol,Vern9(),dg)

For adjoint_sensitivities, the syntax for continuous is adjoint_sensitivities(sol,Vern9(),g,nothing,dg) where g(u,p,t) is the functional and dg(out,u,p,t) is its gradient w.r.t. u.

[MikeInnes]

I didn't open this issue because I want a continuous sensitivity, I opened it because this example (even if not correct) triggers obviously broken code in the package, rather than throwing a sensible error. Why is that branch even there if it only gets triggered by incorrect user input? It should clearly either be removed entirely or fixed and tested.

[ChrisRackauckas]

Okay, it looks like https://github.com/JuliaDiffEq/DiffEqSensitivity.jl/commit/3e101221bee9889caf64c5685d9303f1108412a2 broke continuous adjoint sensitivities in the case where the user doesn't provide a gradient function. the gradient config needs to be re-added. The deletes went too far and deleted more than just the ReverseDiff usage.

[MikeInnes]

That commit deleted g_grad_config, which was unused, not g_gradient_config, which never existed in the first place. So it was broken before, though of course the correct fix would have been to rename one to the other.

[MikeInnes]

Should be easy to use Flux here at the point where we need the gradient. If you can provide a test case I'm also happy to look at it myself.