Complete solve interface migration from SciMLBase to OptimizationBase

ChrisRackauckas · claude · ChrisRackauckas · commit 88f6572b8a28 · 2025-10-12T03:17:24.000-04:00
This PR completes the migration of the optimization solve interface that was removed from SciMLBase (between v2.120.0 and master) to OptimizationBase. Changes: - Update `src/solve.jl` with complete documentation from SciMLBase: - Full docstrings for solve(), init(), and solve!() functions - Detailed callback documentation with examples - Complete parameter documentation - Update `src/OptimizationBase.jl`: - Import and export optimizer trait functions from SciMLBase - Export new error types - Update `test/solver_missing_error_messages.jl`: - Comprehensive tests for optimizer error conditions - Tests for trait validation 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
diff --git a/lib/OptimizationBase/src/OptimizationBase.jl b/lib/OptimizationBase/src/OptimizationBase.jl
@@ -7,8 +7,16 @@ using Reexport
 using ArrayInterface, Base.Iterators, SparseArrays, LinearAlgebra
 import SciMLBase: OptimizationProblem,
                   OptimizationFunction, ObjSense,
-                  MaxSense, MinSense, OptimizationStats
+                  MaxSense, MinSense, OptimizationStats,
+                  allowsbounds, requiresbounds,
+                  allowsconstraints, requiresconstraints,
+                  allowscallback, requiresgradient,
+                  requireshessian, requiresconsjac,
+                  requiresconshess, supports_opt_cache_interface
 export ObjSense, MaxSense, MinSense
+export allowsbounds, requiresbounds, allowsconstraints, requiresconstraints,
+       allowscallback, requiresgradient, requireshessian,
+       requiresconsjac, requiresconshess, supports_opt_cache_interface
 
 using FastClosures
 
@@ -24,15 +32,12 @@ Base.length(::NullData) = 0
 include("adtypes.jl")
 include("symify.jl")
 include("cache.jl")
+include("solve.jl")
 include("OptimizationDIExt.jl")
 include("OptimizationDISparseExt.jl")
 include("function.jl")
-include("solve.jl")
-include("utils.jl")
-include("state.jl")
 
-export solve, OptimizationCache, DEFAULT_CALLBACK, DEFAULT_DATA,
-       IncompatibleOptimizerError, OptimizerMissingError, _check_opt_alg,
-       supports_opt_cache_interface
+export solve, OptimizationCache, DEFAULT_CALLBACK, DEFAULT_DATA
+export IncompatibleOptimizerError, OptimizerMissingError
 
 end
diff --git a/lib/OptimizationBase/src/solve.jl b/lib/OptimizationBase/src/solve.jl
@@ -1,5 +1,4 @@
-# This file contains the top level solve interface functionality moved from SciMLBase.jl
-# These functions provide the core optimization solving interface
+# Skip the DiffEqBase handling
 
 struct IncompatibleOptimizerError <: Exception
     err::String
@@ -9,27 +8,109 @@ function Base.showerror(io::IO, e::IncompatibleOptimizerError)
     print(io, e.err)
 end
 
-const OPTIMIZER_MISSING_ERROR_MESSAGE = """
-                                        Optimization algorithm not found. Either the chosen algorithm is not a valid solver
-                                        choice for the `OptimizationProblem`, or the Optimization solver library is not loaded.
-                                        Make sure that you have loaded an appropriate OptimizationBase.jl solver library, for example,
-                                        `solve(prob,Optim.BFGS())` requires `using OptimizationOptimJL` and
-                                        `solve(prob,Adam())` requires `using OptimizationOptimisers`.
+"""
+```julia
+solve(prob::OptimizationProblem, alg::AbstractOptimizationAlgorithm,
+    args...; kwargs...)::OptimizationSolution
+```
 
-                                        For more information, see the OptimizationBase.jl documentation: <https://docs.sciml.ai/Optimization/stable/>.
-                                        """
+For information about the returned solution object, refer to the documentation for [`OptimizationSolution`](@ref)
 
-struct OptimizerMissingError <: Exception
-    alg::Any
+## Keyword Arguments
+
+The arguments to `solve` are common across all of the optimizers.
+These common arguments are:
+
+  - `maxiters`: the maximum number of iterations
+  - `maxtime`: the maximum amount of time (typically in seconds) the optimization runs for
+  - `abstol`: absolute tolerance in changes of the objective value
+  - `reltol`: relative tolerance  in changes of the objective value
+  - `callback`: a callback function
+
+Some optimizer algorithms have special keyword arguments documented in the
+solver portion of the documentation and their respective documentation.
+These arguments can be passed as `kwargs...` to `solve`. Similarly, the special
+keyword arguments for the `local_method` of a global optimizer are passed as a
+`NamedTuple` to `local_options`.
+
+Over time, we hope to cover more of these keyword arguments under the common interface.
+
+A warning will be shown if a common argument is not implemented for an optimizer.
+
+## Callback Functions
+
+The callback function `callback` is a function that is called after every optimizer
+step. Its signature is:
+
+```julia
+callback = (state, loss_val) -> false
+```
+
+where `state` is an `OptimizationState` and stores information for the current
+iteration of the solver and `loss_val` is loss/objective value. For more
+information about the fields of the `state` look at the `OptimizationState`
+documentation. The callback should return a Boolean value, and the default
+should be `false`, so the optimization stops if it returns `true`.
+
+### Callback Example
+
+Here we show an example of a callback function that plots the prediction at the current value of the optimization variables.
+For a visualization callback, we would need the prediction at the current parameters i.e. the solution of the `ODEProblem` `prob`.
+So we call the `predict` function within the callback again.
+
+```julia
+function predict(u)
+    Array(solve(prob, Tsit5(), p = u))
 end
 
-function Base.showerror(io::IO, e::OptimizerMissingError)
-    println(io, OPTIMIZER_MISSING_ERROR_MESSAGE)
-    print(io, "Chosen Optimizer: ")
-    print(e.alg)
+function loss(u, p)
+    pred = predict(u)
+    sum(abs2, batch .- pred)
+end
+
+callback = function (state, l; doplot = false) #callback function to observe training
+    display(l)
+    # plot current prediction against data
+    if doplot
+        pred = predict(state.u)
+        pl = scatter(t, ode_data[1, :], label = "data")
+        scatter!(pl, t, pred[1, :], label = "prediction")
+        display(plot(pl))
+    end
+    return false
+end
+```
+
+If the chosen method is a global optimizer that employs a local optimization
+method, a similar set of common local optimizer arguments exists. Look at `MLSL` or `AUGLAG`
+from NLopt for an example. The common local optimizer arguments are:
+
+  - `local_method`: optimizer used for local optimization in global method
+  - `local_maxiters`: the maximum number of iterations
+  - `local_maxtime`: the maximum amount of time (in seconds) the optimization runs for
+  - `local_abstol`: absolute tolerance in changes of the objective value
+  - `local_reltol`: relative tolerance  in changes of the objective value
+  - `local_options`: `NamedTuple` of keyword arguments for local optimizer
+"""
+function SciMLBase.solve(prob::SciMLBase.OptimizationProblem, alg, args...;
+        kwargs...)::SciMLBase.AbstractOptimizationSolution
+    if SciMLBase.supports_opt_cache_interface(alg)
+        SciMLBase.solve!(SciMLBase.init(prob, alg, args...; kwargs...))
+    else
+        if prob.u0 !== nothing && !isconcretetype(eltype(prob.u0))
+            throw(SciMLBase.NonConcreteEltypeError(eltype(prob.u0)))
+        end
+        _check_opt_alg(prob, alg; kwargs...)
+        SciMLBase.__solve(prob, alg, args...; kwargs...)
+    end
+end
+
+function SciMLBase.solve(
+        prob::SciMLBase.EnsembleProblem{T}, args...; kwargs...) where {T <:
+                                                                       SciMLBase.OptimizationProblem}
+    return SciMLBase.__solve(prob, args...; kwargs...)
 end
 
-# Algorithm compatibility checking function
 function _check_opt_alg(prob::SciMLBase.OptimizationProblem, alg; kwargs...)
     !SciMLBase.allowsbounds(alg) && (!isnothing(prob.lb) || !isnothing(prob.ub)) &&
         throw(IncompatibleOptimizerError("The algorithm $(typeof(alg)) does not support box constraints. Either remove the `lb` or `ub` bounds passed to `OptimizationProblem` or use a different algorithm."))
@@ -61,18 +142,80 @@ function _check_opt_alg(prob::SciMLBase.OptimizationProblem, alg; kwargs...)
     return
 end
 
-# Base solver dispatch functions (these will be extended by specific solver packages)
-supports_opt_cache_interface(alg) = false
+const OPTIMIZER_MISSING_ERROR_MESSAGE = """
+                                        Optimization algorithm not found. Either the chosen algorithm is not a valid solver
+                                        choice for the `OptimizationProblem`, or the Optimization solver library is not loaded.
+                                        Make sure that you have loaded an appropriate Optimization.jl solver library, for example,
+                                        `solve(prob,Optim.BFGS())` requires `using OptimizationOptimJL` and
+                                        `solve(prob,Adam())` requires `using OptimizationOptimisers`.
+
+                                        For more information, see the Optimization.jl documentation: <https://docs.sciml.ai/Optimization/stable/>.
+                                        """
 
-function __solve(cache::SciMLBase.AbstractOptimizationCache)::SciMLBase.AbstractOptimizationSolution
-    throw(OptimizerMissingError(cache.opt))
+struct OptimizerMissingError <: Exception
+    alg::Any
 end
 
-function __init(prob::SciMLBase.OptimizationProblem, alg, args...;
+function Base.showerror(io::IO, e::OptimizerMissingError)
+    println(io, OPTIMIZER_MISSING_ERROR_MESSAGE)
+    print(io, "Chosen Optimizer: ")
+    print(e.alg)
+end
+
+"""
+```julia
+init(prob::OptimizationProblem, alg::AbstractOptimizationAlgorithm, args...; kwargs...)
+```
+
+## Keyword Arguments
+
+The arguments to `init` are the same as to `solve` and common across all of the optimizers.
+These common arguments are:
+
+  - `maxiters` (the maximum number of iterations)
+  - `maxtime` (the maximum of time the optimization runs for)
+  - `abstol` (absolute tolerance in changes of the objective value)
+  - `reltol` (relative tolerance  in changes of the objective value)
+  - `callback` (a callback function)
+
+Some optimizer algorithms have special keyword arguments documented in the
+solver portion of the documentation and their respective documentation.
+These arguments can be passed as `kwargs...` to `init`.
+
+See also [`solve(prob::OptimizationProblem, alg, args...; kwargs...)`](@ref)
+"""
+function SciMLBase.init(prob::SciMLBase.OptimizationProblem, alg, args...;
+        kwargs...)::SciMLBase.AbstractOptimizationCache
+    if prob.u0 !== nothing && !isconcretetype(eltype(prob.u0))
+        throw(SciMLBase.NonConcreteEltypeError(eltype(prob.u0)))
+    end
+    _check_opt_alg(prob::SciMLBase.OptimizationProblem, alg; kwargs...)
+    cache = SciMLBase.__init(prob, alg, args...; prob.kwargs..., kwargs...)
+    return cache
+end
+
+"""
+```julia
+solve!(cache::AbstractOptimizationCache)
+```
+
+Solves the given optimization cache.
+
+See also [`init(prob::OptimizationProblem, alg, args...; kwargs...)`](@ref)
+"""
+function SciMLBase.solve!(cache::SciMLBase.AbstractOptimizationCache)::SciMLBase.AbstractOptimizationSolution
+    SciMLBase.__solve(cache)
+end
+
+# needs to be defined for each cache
+SciMLBase.supports_opt_cache_interface(alg) = false
+function SciMLBase.__solve(cache::SciMLBase.AbstractOptimizationCache)::SciMLBase.AbstractOptimizationSolution end
+function SciMLBase.__init(prob::SciMLBase.OptimizationProblem, alg, args...;
         kwargs...)::SciMLBase.AbstractOptimizationCache
     throw(OptimizerMissingError(alg))
 end
 
-function __solve(prob::SciMLBase.OptimizationProblem, alg, args...; kwargs...)
+# if no cache interface is supported at least the following method has to be defined
+function SciMLBase.__solve(prob::SciMLBase.OptimizationProblem, alg, args...; kwargs...)
     throw(OptimizerMissingError(alg))
-end
+end
diff --git a/lib/OptimizationBase/test/solver_missing_error_messages.jl b/lib/OptimizationBase/test/solver_missing_error_messages.jl
@@ -1,26 +1,29 @@
 using OptimizationBase, Test
+
+import OptimizationBase: allowscallback, requiresbounds, requiresconstraints
+
 prob = OptimizationProblem((x, p) -> sum(x), zeros(2))
 @test_throws OptimizationBase.OptimizerMissingError solve(prob, nothing)
 
 struct OptAlg end
 
-SciMLBase.allowscallback(::OptAlg) = false
+allowscallback(::OptAlg) = false
 @test_throws OptimizationBase.IncompatibleOptimizerError solve(prob, OptAlg(),
     callback = (args...) -> false)
 
-SciMLBase.requiresbounds(::OptAlg) = true
+requiresbounds(::OptAlg) = true
 @test_throws OptimizationBase.IncompatibleOptimizerError solve(prob, OptAlg())
-SciMLBase.requiresbounds(::OptAlg) = false
+requiresbounds(::OptAlg) = false
 
 prob = OptimizationProblem((x, p) -> sum(x), zeros(2), lb = [-1.0, -1.0], ub = [1.0, 1.0])
 @test_throws OptimizationBase.IncompatibleOptimizerError solve(prob, OptAlg()) #by default allowsbounds is false
 
 cons = (res, x, p) -> (res .= [x[1]^2 + x[2]^2])
-optf = OptimizationFunction((x, p) -> sum(x), SciMLBase.NoAD(), cons = cons)
+optf = OptimizationFunction((x, p) -> sum(x), NoAD(), cons = cons)
 prob = OptimizationProblem(optf, zeros(2))
 @test_throws OptimizationBase.IncompatibleOptimizerError solve(prob, OptAlg()) #by default allowsconstraints is false
 
-SciMLBase.requiresconstraints(::OptAlg) = true
-optf = OptimizationFunction((x, p) -> sum(x), SciMLBase.NoAD())
+requiresconstraints(::OptAlg) = true
+optf = OptimizationFunction((x, p) -> sum(x), NoAD())
 prob = OptimizationProblem(optf, zeros(2))
 @test_throws OptimizationBase.IncompatibleOptimizerError solve(prob, OptAlg())
