Constraints for AutoFiniteDiff #318

DanielVandH · 2022-07-07T15:24:49Z

This defines a method for generating Jacobians and Hessians of the constraints when using AutoFiniteDiff (see also #130). Hopefully it's not too bad, I've left some comments in the code where I was unsure. I also have some comments below about issues I haven't been able to fix, that should probably be fixed before any action is taken on this pull request.

Line 50 of finitediff.jl: I had to take args... outside of the f.f call or else things were breaking. Any ideas why? You can run the AutoFiniteDiff section of the tests with args included to see the errors, e.g.

prob = OptimizationProblem(optprob, x0)
f = Optimization.instantiate_function(prob.f, prob.u0, prob.f.adtype, prob.p)
_f = (θ, args...) -> first(prob.f.f(θ, prob.p, args...))
_f(θ, args...)
ERROR: MethodError: no method matching rosenbrock(::Vector{Float64}, ::SciMLBase.NullParameters, ::SciMLBase.NullParameters)
Closest candidates are:
  rosenbrock(::Any, ::Any) at c:\Users\licer\.julia\dev\Optimization\test\ADtests.jl:6
  rosenbrock(::Any) at c:\Users\licer\.julia\dev\Optimization\test\ADtests.jl:6
Stacktrace:
 [1] (::OptimizationFunction{true, Main.Optimization.AutoFiniteDiff{Val{:forward}, Val{:forward}, Val{:hcentral}}, typeof(rosenbrock), Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing})(::Vector{Float64}, ::Vararg{Any})
   @ SciMLBase C:\Users\licer\.julia\packages\SciMLBase\IJbT7\src\scimlfunctions.jl:2887
 [2] (::var"#75#76")(θ::Vector{Float64}, args::SciMLBase.NullParameters)
   @ Main c:\Users\licer\.julia\dev\Optimization\test\ADtests.jl:167
 [3] top-level scope
   @ c:\Users\licer\.julia\dev\Optimization\test\ADtests.jl:168

Line 86 of finitediff.jl: I'm not sure I understand the way I should be using FiniteDiff.JacobianCache here. It was for some reason completely incompatible with the rectangular Jacobian tested in Line 166--180 of ADtests.jl, giving either an issue with m, n = size(J) not working for a vector J, or some parent error with _color = reshape(colorvec, axes(x)...).
Finally, I tried solving some problems, e.g. at the end of ADtests.jl I have

## Solving some problems 
#cons = (x, p) -> [x[1]^2 + x[2]^2]
#optf = OptimizationFunction(rosenbrock, Optimization.AutoFiniteDiff(); cons = cons)
#prob = OptimizationProblem(optf, x0, lcons = [0.2], rcons = [0.2])
#sol = solve(prob, Optim.BFGS()) ## not recognising gradients?

but optf was for some reason not recognising that the gradient is now defined. I figured I shouldn't tinker any further with this last problem since it's defined outside of Optimization.jl. Maybe there's a syntax error I just haven't recognised for whatever reason.

codecov · 2022-07-07T15:35:22Z

Codecov Report

Merging #318 (9ed5aae) into master (a9be183) will increase coverage by 4.68%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #318      +/-   ##
==========================================
+ Coverage   74.68%   79.37%   +4.68%     
==========================================
  Files           9        9              
  Lines         241      257      +16     
==========================================
+ Hits          180      204      +24     
+ Misses         61       53       -8

Impacted Files	Coverage Δ
src/function/finitediff.jl	`100.00% <100.00%> (ø)`
src/function/forwarddiff.jl	`100.00% <0.00%> (+20.51%)`	⬆️

📣 Codecov can now indicate which changes are the most critical in Pull Requests. Learn more

ChrisRackauckas · 2022-07-07T18:32:35Z

For a non-square Jacobian, you just need to provide the cache with the output arrays so it knows how they are sized.

FiniteDiff.JacobianCache(
              x1 ,
              fx ,
              fx1,
              fdtype     :: Type{T1} = Val{:central},
              returntype :: Type{T2} = eltype(fx),
              colorvec = 1:length(x),
              sparsity = nothing)

provide it with the fx and fx1 sized as out.

ChrisRackauckas · 2022-07-07T18:35:24Z

Line 50 of finitediff.jl: I had to take args... outside of the f.f call or else things were breaking. Any ideas why? You can run the AutoFiniteDiff section of the tests with args included to see the errors, e.g.

It needs access to the batch data. Dig through one step at a time, why does it get batch data as null parameters instead of the empty tuple?

ChrisRackauckas · 2022-07-07T18:36:20Z

BFGS doesn't support constraints. Try IPNewton.

DanielVandH · 2022-07-07T23:42:03Z

Line 50 of finitediff.jl: I had to take args... outside of the f.f call or else things were breaking. Any ideas why? You can run the AutoFiniteDiff section of the tests with args included to see the errors, e.g.

It needs access to the batch data. Dig through one step at a time, why does it get batch data as null parameters instead of the empty tuple?

This, for a reason that I'm not sure of, seems to just a problem with how I'm working with this repo I think. When I use ] test from the REPL and keep args..., it seems to work, so no problems here it seems.

For a non-square Jacobian, you just need to provide the cache with the output arrays so it knows how they are sized.
FiniteDiff.JacobianCache(
              x1 ,
              fx ,
              fx1,
              fdtype     :: Type{T1} = Val{:central},
              returntype :: Type{T2} = eltype(fx),
              colorvec = 1:length(x),
              sparsity = nothing)
provide it with the fx and fx1 sized as out.

Fixed now - thank you, I didn't realise fx and fx1 should be the same.

BFGS doesn't support constraints. Try IPNewton.

Doesn't seem to work, still. (The extra lb and ub were just to test if it's related to that at all, doesn't seem to be.)

julia> cons = (x, p) -> [x[1]^2 + x[2]^2]
       optf = OptimizationFunction(rosenbrock, Optimization.AutoFiniteDiff(); cons = cons)
       prob = OptimizationProblem(optf, x0, lcons = [0.2], rcons = [0.2], lb = [0.0, 0.0], ub = [1.0, 1.0])
       sol = solve(prob, IPNewton())
ERROR: This optimizer requires derivative definitions for nonlinear constraints. If the problem does not have nonlinear constraints, choose a different optimizer. Otherwise define the derivative for cons using OptimizationFunction either directly or automatically generate them with one of the autodiff backends
Stacktrace:
 [1] error(s::String)
   @ Base .\error.jl:35
 [2] ___solve(prob::OptimizationProblem{true, OptimizationFunction{true, Main.Optimization.AutoFiniteDiff{Val{:forward}, Val{:forward}, Val{:hcentral}}, typeof(rosenbrock), Nothing, Nothing, Nothing, var"#56#57", Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing}, Vector{Float64}, SciMLBase.NullParameters, Vector{Float64}, Vector{Float64}, Nothing, Nothing, Base.Pairs{Symbol, Vector{Float64}, Tuple{Symbol}, NamedTuple{(:rcons,), Tuple{Vector{Float64}}}}}, opt::IPNewton{typeof(Optim.backtrack_constrained_grad), Symbol}, data::Base.Iterators.Cycle{Tuple{Optimization.NullData}}; callback::Function, maxiters::Nothing, maxtime::Nothing, abstol::Nothing, reltol::Nothing, progress::Bool, kwargs::Base.Pairs{Symbol, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
   @ Main.OptimizationOptimJL c:\Users\licer\.julia\dev\Optimization\lib\OptimizationOptimJL\src\OptimizationOptimJL.jl:261
 [3] ___solve(prob::OptimizationProblem{true, OptimizationFunction{true, Main.Optimization.AutoFiniteDiff{Val{:forward}, Val{:forward}, Val{:hcentral}}, typeof(rosenbrock), Nothing, Nothing, Nothing, var"#56#57", Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing}, Vector{Float64}, SciMLBase.NullParameters, Vector{Float64}, Vector{Float64}, Nothing, Nothing, Base.Pairs{Symbol, Vector{Float64}, Tuple{Symbol}, NamedTuple{(:rcons,), Tuple{Vector{Float64}}}}}, opt::IPNewton{typeof(Optim.backtrack_constrained_grad), Symbol}, data::Base.Iterators.Cycle{Tuple{Optimization.NullData}})
   @ Main.OptimizationOptimJL c:\Users\licer\.julia\dev\Optimization\lib\OptimizationOptimJL\src\OptimizationOptimJL.jl:224
 [4] __solve(prob::OptimizationProblem{true, OptimizationFunction{true, Main.Optimization.AutoFiniteDiff{Val{:forward}, Val{:forward}, Val{:hcentral}}, typeof(rosenbrock), Nothing, Nothing, Nothing, var"#56#57", Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing}, Vector{Float64}, SciMLBase.NullParameters, Vector{Float64}, Vector{Float64}, Nothing, Nothing, Base.Pairs{Symbol, Vector{Float64}, Tuple{Symbol}, NamedTuple{(:rcons,), Tuple{Vector{Float64}}}}}, opt::IPNewton{typeof(Optim.backtrack_constrained_grad), Symbol}, data::Base.Iterators.Cycle{Tuple{Optimization.NullData}}; kwargs::Base.Pairs{Symbol, 
Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
   @ Main.OptimizationOptimJL c:\Users\licer\.julia\dev\Optimization\lib\OptimizationOptimJL\src\OptimizationOptimJL.jl:56
 [5] __solve(prob::OptimizationProblem{true, OptimizationFunction{true, Main.Optimization.AutoFiniteDiff{Val{:forward}, Val{:forward}, Val{:hcentral}}, typeof(rosenbrock), Nothing, Nothing, Nothing, var"#56#57", Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing}, Vector{Float64}, SciMLBase.NullParameters, Vector{Float64}, Vector{Float64}, Nothing, Nothing, Base.Pairs{Symbol, Vector{Float64}, Tuple{Symbol}, NamedTuple{(:rcons,), Tuple{Vector{Float64}}}}}, opt::IPNewton{typeof(Optim.backtrack_constrained_grad), Symbol}, data::Base.Iterators.Cycle{Tuple{Optimization.NullData}}) (repeats 2 times)
   @ Main.OptimizationOptimJL c:\Users\licer\.julia\dev\Optimization\lib\OptimizationOptimJL\src\OptimizationOptimJL.jl:40
 [6] solve(::OptimizationProblem{true, OptimizationFunction{true, Main.Optimization.AutoFiniteDiff{Val{:forward}, Val{:forward}, Val{:hcentral}}, typeof(rosenbrock), Nothing, Nothing, Nothing, var"#56#57", Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing}, Vector{Float64}, SciMLBase.NullParameters, Vector{Float64}, Vector{Float64}, Nothing, Nothing, Base.Pairs{Symbol, Vector{Float64}, Tuple{Symbol}, NamedTuple{(:rcons,), Tuple{Vector{Float64}}}}}, ::IPNewton{typeof(Optim.backtrack_constrained_grad), Symbol}; kwargs::Base.Pairs{Symbol, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
   @ SciMLBase C:\Users\licer\.julia\packages\SciMLBase\IJbT7\src\solve.jl:71
 [7] solve(::OptimizationProblem{true, OptimizationFunction{true, Main.Optimization.AutoFiniteDiff{Val{:forward}, Val{:forward}, Val{:hcentral}}, typeof(rosenbrock), Nothing, Nothing, Nothing, var"#56#57", Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing}, Vector{Float64}, SciMLBase.NullParameters, Vector{Float64}, Vector{Float64}, Nothing, Nothing, Base.Pairs{Symbol, Vector{Float64}, Tuple{Symbol}, NamedTuple{(:rcons,), Tuple{Vector{Float64}}}}}, ::IPNewton{typeof(Optim.backtrack_constrained_grad), Symbol})
   @ SciMLBase C:\Users\licer\.julia\packages\SciMLBase\IJbT7\src\solve.jl:70
 [8] top-level scope
   @ c:\Users\licer\.julia\dev\Optimization\test\ADtests.jl:204

I will have to come back to this a bit later. All the tests do pass now though, except for this error.

DanielVandH · 2022-07-08T00:32:10Z

Also: Should we be using the prototype/colorvec options from f in these caches? I initially chose not to because they were set back to nothing in ForwardDiff, so I kept that here, but thinking about it some more that seems wrong. edit: Working on this now to see.

DanielVandH · 2022-07-08T14:25:50Z

Tested some problems now with IPNewton. I think this is all working now, no more changes incoming :).

src/function/finitediff.jl

Vaibhavdixit02 · 2022-07-08T14:54:41Z

test/ADtests.jl

+    @test sol1.u ≈ sol2.u
+
+    optf1 = OptimizationFunction(rosenbrock, Optimization.AutoFiniteDiff(); cons = consf)
+    prob1 = OptimizationProblem(optf, [0.3, 0.5], lb = [0.2, 0.4], ub = [0.6, 0.8], lcons = [0.2], ucons = [0.55])


The constraint's bounds here need to be different for the two elements, this is the failure on CI currently

What do you mean?

I'll fix that. Forgot that con2_c has two constraints in it when making this loop.

Co-authored-by: Vaibhav Kumar Dixit <[email protected]>

ChrisRackauckas · 2022-07-08T15:09:20Z

JuliaRegistries/General#63877 inc.

ChrisRackauckas · 2022-07-08T15:11:00Z

src/function/finitediff.jl

+
+    if cons !== nothing && f.cons_h === nothing
+        cons_h = function (res, θ)
+            for i in 1:num_cons#note: colorvecs not yet supported by FiniteDiff for Hessians 


We should start integrating/testing JuliaDiff/SparseDiffTools.jl#190 and then do similar to the FiniteDiff side (cc @ElOceanografo)

ChrisRackauckas · 2022-07-08T15:12:26Z

I think the test issue is the only thing left. If with those bounds it passes then I think this is good.

DanielVandH · 2022-07-08T15:44:27Z

I don't really understand why it's still failing? It works for me locally...

ChrisRackauckas · 2022-07-08T16:50:31Z

Check that the dependencies are the same. It looks like it got FiniteDiff patch this time, but was there another patch you had somewhere?

DanielVandH · 2022-07-08T22:55:55Z

Check that the dependencies are the same. It looks like it got FiniteDiff patch this time, but was there another patch you had somewhere?

Seemed to be related - it keeps giving me the old version without any of my changes and I didn't realise. Still learning how to use other packages like this 😅 Also forgot to use the correct optf's when I changed to the for loop and it seems to work now... hopefully the CI does too.

Well, it fails but at least it's not erroring. I'll have to come back to this a bit later and see why.

DanielVandH · 2022-07-09T04:44:17Z

Should be all good now, finally.

ChrisRackauckas · 2022-07-09T04:45:40Z

I'm not waiting on the DiffEqFlux downstream since it doesn't use AutoFiniteDiff anyways.

🎉

Vaibhavdixit02 · 2022-07-09T06:31:54Z

test/ADtests.jl

    optf2 = OptimizationFunction(rosenbrock, Optimization.AutoForwardDiff(); cons = consf)
    prob2 = OptimizationProblem(optf2, [0.3, 0.5], lb = [0.2, 0.4], ub = [0.6, 0.8], lcons = lcons, ucons = ucons)
-    sol2 = solve(prob2,IPNewton())
-    @test sol1.minimum ≈ sol2.minimum 
+    sol2 = solve(prob2,Optim.SAMIN(), maxiters = 10000)


SAMIN doesn't use the constraints though?

But it doesn't error if constraints are given? Optim just lets them through?

I thought it did? Based on what is given here http://optimization.sciml.ai/stable/optimization_packages/optim/#With-Constraint-Equations "The following method in Optim is performing global optimization on problems with constraint equation"

I did miss that SAMIN is derivative-free though (PSO too), so we should probably mark that in the docs.

DanielVandH added 4 commits July 7, 2022 21:40

test

ad98f57

Construct necessary constraint functions

16ebabf

Support for constraints

77a555c

autofinitediff

156fa53

Fix JacobianCache and use copy()

869aba5

DanielVandH added 2 commits July 8, 2022 20:59

Support prototypes and colorvecs

fe3f0c9

Test that we can actually solve problems now with constrained finitediff

fc4188e

Vaibhavdixit02 reviewed Jul 8, 2022

View reviewed changes

src/function/finitediff.jl Outdated Show resolved Hide resolved

Vaibhavdixit02 reviewed Jul 8, 2022

View reviewed changes

src/function/finitediff.jl Show resolved Hide resolved

Vaibhavdixit02 reviewed Jul 8, 2022

View reviewed changes

Update src/function/finitediff.jl

b9301bb

Co-authored-by: Vaibhav Kumar Dixit <[email protected]>

ChrisRackauckas reviewed Jul 8, 2022

View reviewed changes

Fix lcons/ucons for con2_c

8cefccf

ChrisRackauckas approved these changes Jul 8, 2022

View reviewed changes

optf -> optf1, optf2

d8896c9

DanielVandH added 2 commits July 9, 2022 09:50

IPNewton -> SAMIN?

ba40958

Use test_broken instead of commenting out the tests

9ed5aae

ChrisRackauckas merged commit fd343ab into SciML:master Jul 9, 2022

Vaibhavdixit02 reviewed Jul 9, 2022

View reviewed changes

Uh oh!

Constraints for AutoFiniteDiff #318

Constraints for AutoFiniteDiff #318

Uh oh!

Conversation

DanielVandH commented Jul 7, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

codecov bot commented Jul 7, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

ChrisRackauckas commented Jul 7, 2022

Uh oh!

ChrisRackauckas commented Jul 7, 2022

Uh oh!

ChrisRackauckas commented Jul 7, 2022

Uh oh!

DanielVandH commented Jul 7, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

DanielVandH commented Jul 8, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

DanielVandH commented Jul 8, 2022

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ChrisRackauckas commented Jul 8, 2022

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ChrisRackauckas commented Jul 8, 2022

Uh oh!

DanielVandH commented Jul 8, 2022

Uh oh!

ChrisRackauckas commented Jul 8, 2022

Uh oh!

DanielVandH commented Jul 8, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

DanielVandH commented Jul 9, 2022

Uh oh!

ChrisRackauckas commented Jul 9, 2022

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

DanielVandH Jul 9, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

DanielVandH commented Jul 7, 2022 •

edited

Loading

codecov bot commented Jul 7, 2022 •

edited

Loading

DanielVandH commented Jul 7, 2022 •

edited

Loading

DanielVandH commented Jul 8, 2022 •

edited

Loading

DanielVandH commented Jul 8, 2022 •

edited

Loading

DanielVandH Jul 9, 2022 •

edited

Loading