ANN: Nonconvex.jl a toolbox for AD-based constrained non-convex optimization

I don’t think projection will work for MMA without generalising the algorithm in some way. May be possible but it’s not trivial.

Can we use this package to optimize over parameters which are organized in structs (instead of having all parameters arranged linearly in a vector or an array)?

Given that this uses Zygote, I am thinking it might not be so difficult to make it work with more general parameter structures, just like Flux.jl does (via the handy params construct).

What are your thoughts on this? From the README it seems at the moment Nonconvex.jl only handles having all parameters in a single array, am I correct?

For now it’s only vectors. I do plan to have conversion functions between vectors and other data structures. Supporting non-vector data structures natively may be possible for pure Julia solvers but not for the other ones. The conversion function approach is more general and can be used to support sparse matrices, complex numbers, etc.

Can you please provide an example of how one would one use ForwardDiff instead of Zygote for a specific objective function with this package?

I added an example to the README GitHub - mohamed82008/Nonconvex.jl: Toolbox for non-convex constrained optimization.. @oxinabox may not approve though.


I approve of this, it looks like the right way to do it to me.

If you do it to a function you don’t own it is type-piracy.
But I mean its probably not the wrong thing to do.
Since if you want it to be evaluated in forward mode for optimization purposes you porbably in general want it to be evaluated in forward mode even if used as part of a larger system that a reverse-mode AD is doing.
And now your AD is mixed-mode.


You might be interested in ParameterHandling.jl which has doing this as it’s main purpose via flatten

It’s @willtebbutt’s work.

Similar code also exists in FiniteDifferences.jl as to_vec.
though we would like to remove it and replace it with ChainRule’s differnetial types which naturally can perform addition with arbitary data-types without having to push it down to a vectors and back-up.
But you probably can’t do that isn’t IPOPT probably demands a Vector, right?

1 Like

You might be interested in ParameterHandling.jl which has doing this as it’s main purpose via flatten

Awesome! I will look into it, thanks.


Thanks. This is fine as a workaround, but since many packages implement some kind of AD shim that takes a real-valued function and provides a gradient (eg GalacticOptim, LogDensityProblems, many more), it would be great to arrive at a single, lightweight package that does this for all AD frameworks.

That was the point of GitHub - JuliaDiff/AbstractDifferentiation.jl: An abstract interface for automatic differentiation.. I need to finish that PR though.


Update - Release v0.5.1

Mixed integer optimization is now available in Nonconvex.jl using Juniper and Ipopt. To my knowledge, this is the first mixed integer nonlinear programming package in Julia that uses Zygote for automatic differentiation and doesn’t require an explicit JuMP model definition. Happy optimization!


How would I modify that for the Jacobian (of the constraint)?

function ChainRulesCore.rrule(::typeof(constraint), x::AbstractVector)
    val = constraint(x)
    jac = ForwardDiff.jacobian(constraint, x)
    val, Δ -> (NO_FIELDS, Δ * jac)

does not work (DimensionMismatch).

Δ * jacjac' * Δ

1 Like