Files
QIDISlicer/src/libslic3r/Optimize/NLoptOptimizer.hpp
2023-09-16 16:26:29 +08:00

413 lines
14 KiB
C++

#ifndef NLOPTOPTIMIZER_HPP
#define NLOPTOPTIMIZER_HPP
#ifdef _MSC_VER
#pragma warning(push)
#pragma warning(disable: 4244)
#pragma warning(disable: 4267)
#endif
#include <nlopt.h>
#ifdef _MSC_VER
#pragma warning(pop)
#endif
#include <utility>
#include <libslic3r/libslic3r.h>
#include "Optimizer.hpp"
namespace Slic3r { namespace opt {
namespace detail {
// Helper types for NLopt algorithm selection in template contexts
template<nlopt_algorithm alg> struct NLoptAlg {};
// NLopt can combine multiple algorithms if one is global an other is a local
// method. This is how template specializations can be informed about this fact.
template<nlopt_algorithm gl_alg, nlopt_algorithm lc_alg = NLOPT_LN_NELDERMEAD>
struct NLoptAlgComb {};
template<class M> struct IsNLoptAlg {
static const constexpr bool value = false;
};
template<nlopt_algorithm a> struct IsNLoptAlg<NLoptAlg<a>> {
static const constexpr bool value = true;
};
template<nlopt_algorithm a1, nlopt_algorithm a2>
struct IsNLoptAlg<NLoptAlgComb<a1, a2>> {
static const constexpr bool value = true;
};
// NLopt can wrap any of its algorithms to use the augmented lagrangian method
// for deriving an object function from all equality and inequality constraints
// This way one can use algorithms that do not support these constraints natively
template<class Alg> struct NLoptAUGLAG {};
template<nlopt_algorithm a1, nlopt_algorithm a2>
struct IsNLoptAlg<NLoptAUGLAG<NLoptAlgComb<a1, a2>>> {
static const constexpr bool value = true;
};
template<nlopt_algorithm a> struct IsNLoptAlg<NLoptAUGLAG<NLoptAlg<a>>> {
static const constexpr bool value = true;
};
template<class M, class T = void>
using NLoptOnly = std::enable_if_t<IsNLoptAlg<M>::value, T>;
template<class M> struct GetNLoptAlg_ {
static constexpr nlopt_algorithm Local = NLOPT_NUM_ALGORITHMS;
static constexpr nlopt_algorithm Global = NLOPT_NUM_ALGORITHMS;
static constexpr bool IsAUGLAG = false;
};
template<nlopt_algorithm a> struct GetNLoptAlg_<NLoptAlg<a>> {
static constexpr nlopt_algorithm Local = NLOPT_NUM_ALGORITHMS;
static constexpr nlopt_algorithm Global = a;
static constexpr bool IsAUGLAG = false;
};
template<nlopt_algorithm g, nlopt_algorithm l>
struct GetNLoptAlg_<NLoptAlgComb<g, l>> {
static constexpr nlopt_algorithm Local = l;
static constexpr nlopt_algorithm Global = g;
static constexpr bool IsAUGLAG = false;
};
template<class M> constexpr nlopt_algorithm GetNLoptAlg_Global = GetNLoptAlg_<remove_cvref_t<M>>::Global;
template<class M> constexpr nlopt_algorithm GetNLoptAlg_Local = GetNLoptAlg_<remove_cvref_t<M>>::Local;
template<class M> constexpr bool IsAUGLAG = GetNLoptAlg_<remove_cvref_t<M>>::IsAUGLAG;
template<class M> struct GetNLoptAlg_<NLoptAUGLAG<M>> {
static constexpr nlopt_algorithm Local = GetNLoptAlg_Local<M>;
static constexpr nlopt_algorithm Global = GetNLoptAlg_Global<M>;
static constexpr bool IsAUGLAG = true;
};
enum class OptDir { MIN, MAX }; // Where to optimize
struct NLoptRAII { // Helper RAII class for nlopt_opt
nlopt_opt ptr = nullptr;
template<class...A> explicit NLoptRAII(A&&...a)
{
ptr = nlopt_create(std::forward<A>(a)...);
}
NLoptRAII(const NLoptRAII&) = delete;
NLoptRAII(NLoptRAII&&) = delete;
NLoptRAII& operator=(const NLoptRAII&) = delete;
NLoptRAII& operator=(NLoptRAII&&) = delete;
~NLoptRAII() { nlopt_destroy(ptr); }
};
// Wrap each element of the tuple tup into a wrapper class W and return
// a new tuple with each element being of type W<T_i> where T_i is the type of
// i-th element of tup.
template<template<class> class W, class...Args>
auto wrap_tup(const std::tuple<Args...> &tup)
{
return std::tuple<W<Args>...>(tup);
}
template<class M, class = NLoptOnly<M>>
class NLoptOpt {
StopCriteria m_stopcr;
StopCriteria m_loc_stopcr;
OptDir m_dir = OptDir::MIN;
static constexpr double ConstraintEps = 1e-6;
template<class Fn> struct OptData {
Fn fn;
NLoptOpt *self = nullptr;
nlopt_opt opt_raw = nullptr;
OptData(const Fn &f): fn{f} {}
OptData(const Fn &f, NLoptOpt *s, nlopt_opt nlopt_raw)
: fn{f}, self{s}, opt_raw{nlopt_raw} {}
};
template<class Fn, size_t N>
static double optfunc(unsigned n, const double *params,
double *gradient, void *data)
{
assert(n == N);
auto tdata = static_cast<OptData<Fn>*>(data);
if (tdata->self->m_stopcr.stop_condition())
nlopt_force_stop(tdata->opt_raw);
auto funval = to_arr<N>(params);
double scoreval = 0.;
using RetT = decltype(tdata->fn(funval));
if constexpr (std::is_convertible_v<RetT, ScoreGradient<N>>) {
ScoreGradient<N> score = tdata->fn(funval);
for (size_t i = 0; i < n; ++i) gradient[i] = (*score.gradient)[i];
scoreval = score.score;
} else {
scoreval = tdata->fn(funval);
}
return scoreval;
}
template<class Fn, size_t N>
static double constrain_func(unsigned n, const double *params,
double *gradient, void *data)
{
assert(n == N);
auto tdata = static_cast<OptData<Fn>*>(data);
auto funval = to_arr<N>(params);
return tdata->fn(funval);
}
template<size_t N>
static void set_up(NLoptRAII &nl,
const Bounds<N> &bounds,
const StopCriteria &stopcr)
{
std::array<double, N> lb, ub;
for (size_t i = 0; i < N; ++i) {
lb[i] = bounds[i].min();
ub[i] = bounds[i].max();
}
nlopt_set_lower_bounds(nl.ptr, lb.data());
nlopt_set_upper_bounds(nl.ptr, ub.data());
double abs_diff = stopcr.abs_score_diff();
double rel_diff = stopcr.rel_score_diff();
double stopval = stopcr.stop_score();
if(!std::isnan(abs_diff)) nlopt_set_ftol_abs(nl.ptr, abs_diff);
if(!std::isnan(rel_diff)) nlopt_set_ftol_rel(nl.ptr, rel_diff);
if(!std::isnan(stopval)) nlopt_set_stopval(nl.ptr, stopval);
if(stopcr.max_iterations() > 0)
nlopt_set_maxeval(nl.ptr, stopcr.max_iterations());
}
template<class Fn, size_t N, class...EqFns, class...IneqFns>
Result<N> optimize(NLoptRAII &nl, Fn &&fn, const Input<N> &initvals,
const std::tuple<EqFns...> &equalities,
const std::tuple<IneqFns...> &inequalities)
{
Result<N> r;
OptData<Fn> data {fn, this, nl.ptr};
auto do_for_each_eq = [this, &nl](auto &arg) {
arg.self = this;
arg.opt_raw = nl.ptr;
using F = decltype(arg.fn);
nlopt_add_equality_constraint (nl.ptr, constrain_func<F, N>, &arg, ConstraintEps);
};
auto do_for_each_ineq = [this, &nl](auto &arg) {
arg.self = this;
arg.opt_raw = nl.ptr;
using F = decltype(arg.fn);
nlopt_add_inequality_constraint (nl.ptr, constrain_func<F, N>, &arg, ConstraintEps);
};
auto eq_data = wrap_tup<OptData>(equalities);
for_each_in_tuple(do_for_each_eq, eq_data);
auto ineq_data = wrap_tup<OptData>(inequalities);
for_each_in_tuple(do_for_each_ineq, ineq_data);
switch(m_dir) {
case OptDir::MIN:
nlopt_set_min_objective(nl.ptr, optfunc<Fn, N>, &data); break;
case OptDir::MAX:
nlopt_set_max_objective(nl.ptr, optfunc<Fn, N>, &data); break;
}
r.optimum = initvals;
r.resultcode = nlopt_optimize(nl.ptr, r.optimum.data(), &r.score);
return r;
}
public:
template<class Fn, size_t N, class...EqFns, class...IneqFns>
Result<N> optimize(Fn&& f,
const Input<N> &initvals,
const Bounds<N>& bounds,
const std::tuple<EqFns...> &equalities,
const std::tuple<IneqFns...> &inequalities)
{
if constexpr (IsAUGLAG<M>) {
NLoptRAII nl_wrap{NLOPT_AUGLAG, N};
set_up(nl_wrap, bounds, get_criteria());
NLoptRAII nl_glob{GetNLoptAlg_Global<M>, N};
set_up(nl_glob, bounds, get_criteria());
nlopt_set_local_optimizer(nl_wrap.ptr, nl_glob.ptr);
if constexpr (GetNLoptAlg_Local<M> < NLOPT_NUM_ALGORITHMS) {
NLoptRAII nl_loc{GetNLoptAlg_Local<M>, N};
set_up(nl_loc, bounds, m_loc_stopcr);
nlopt_set_local_optimizer(nl_glob.ptr, nl_loc.ptr);
return optimize(nl_wrap, std::forward<Fn>(f), initvals,
equalities, inequalities);
} else {
return optimize(nl_wrap, std::forward<Fn>(f), initvals,
equalities, inequalities);
}
} else {
NLoptRAII nl_glob{GetNLoptAlg_Global<M>, N};
set_up(nl_glob, bounds, get_criteria());
if constexpr (GetNLoptAlg_Local<M> < NLOPT_NUM_ALGORITHMS) {
NLoptRAII nl_loc{GetNLoptAlg_Local<M>, N};
set_up(nl_loc, bounds, m_loc_stopcr);
nlopt_set_local_optimizer(nl_glob.ptr, nl_loc.ptr);
return optimize(nl_glob, std::forward<Fn>(f), initvals,
equalities, inequalities);
} else {
return optimize(nl_glob, std::forward<Fn>(f), initvals,
equalities, inequalities);
}
}
assert(false);
return {};
}
explicit NLoptOpt(const StopCriteria &stopcr_glob = {})
: m_stopcr(stopcr_glob)
{}
void set_criteria(const StopCriteria &cr) { m_stopcr = cr; }
const StopCriteria &get_criteria() const noexcept { return m_stopcr; }
void set_loc_criteria(const StopCriteria &cr) { m_loc_stopcr = cr; }
const StopCriteria &get_loc_criteria() const noexcept { return m_loc_stopcr; }
void set_dir(OptDir dir) noexcept { m_dir = dir; }
void seed(long s) { nlopt_srand(s); }
};
template<class Alg> struct AlgFeatures_ {
static constexpr bool SupportsInequalities = false;
static constexpr bool SupportsEqualities = false;
};
} // namespace detail;
template<class Alg> constexpr bool SupportsEqualities =
detail::AlgFeatures_<remove_cvref_t<Alg>>::SupportsEqualities;
template<class Alg> constexpr bool SupportsInequalities =
detail::AlgFeatures_<remove_cvref_t<Alg>>::SupportsInequalities;
// Optimizers based on NLopt.
template<class M> class Optimizer<M, detail::NLoptOnly<M>> {
detail::NLoptOpt<M> m_opt;
public:
Optimizer& to_max() { m_opt.set_dir(detail::OptDir::MAX); return *this; }
Optimizer& to_min() { m_opt.set_dir(detail::OptDir::MIN); return *this; }
template<class Func, size_t N, class...EqFns, class...IneqFns>
Result<N> optimize(Func&& func,
const Input<N> &initvals,
const Bounds<N>& bounds,
const std::tuple<EqFns...> &eq_constraints = {},
const std::tuple<IneqFns...> &ineq_constraint = {})
{
static_assert(std::tuple_size_v<std::tuple<EqFns...>> == 0
|| SupportsEqualities<M>,
"Equality constraints are not supported.");
static_assert(std::tuple_size_v<std::tuple<IneqFns...>> == 0
|| SupportsInequalities<M>,
"Inequality constraints are not supported.");
return m_opt.optimize(std::forward<Func>(func), initvals, bounds,
eq_constraints,
ineq_constraint);
}
explicit Optimizer(StopCriteria stopcr = {}) : m_opt(stopcr) {}
Optimizer &set_criteria(const StopCriteria &cr)
{
m_opt.set_criteria(cr); return *this;
}
const StopCriteria &get_criteria() const { return m_opt.get_criteria(); }
void seed(long s) { m_opt.seed(s); }
void set_loc_criteria(const StopCriteria &cr) { m_opt.set_loc_criteria(cr); }
const StopCriteria &get_loc_criteria() const noexcept { return m_opt.get_loc_criteria(); }
};
// Predefinded NLopt algorithms
using AlgNLoptGenetic = detail::NLoptAlgComb<NLOPT_GN_ESCH>;
using AlgNLoptSubplex = detail::NLoptAlg<NLOPT_LN_SBPLX>;
using AlgNLoptSimplex = detail::NLoptAlg<NLOPT_LN_NELDERMEAD>;
using AlgNLoptCobyla = detail::NLoptAlg<NLOPT_LN_COBYLA>;
using AlgNLoptDIRECT = detail::NLoptAlg<NLOPT_GN_DIRECT>;
using AlgNLoptORIG_DIRECT = detail::NLoptAlg<NLOPT_GN_ORIG_DIRECT>;
using AlgNLoptISRES = detail::NLoptAlg<NLOPT_GN_ISRES>;
using AlgNLoptAGS = detail::NLoptAlg<NLOPT_GN_AGS>;
using AlgNLoptMLSL_Subplx = detail::NLoptAlgComb<NLOPT_GN_MLSL_LDS, NLOPT_LN_SBPLX>;
using AlgNLoptMLSL_Cobyla = detail::NLoptAlgComb<NLOPT_GN_MLSL, NLOPT_LN_COBYLA>;
using AlgNLoptGenetic_Subplx = detail::NLoptAlgComb<NLOPT_GN_ESCH, NLOPT_LN_SBPLX>;
// To craft auglag algorithms (constraint support through object function transformation)
using detail::NLoptAUGLAG;
namespace detail {
template<> struct AlgFeatures_<AlgNLoptCobyla> {
static constexpr bool SupportsInequalities = true;
static constexpr bool SupportsEqualities = true;
};
template<> struct AlgFeatures_<AlgNLoptISRES> {
static constexpr bool SupportsInequalities = true;
static constexpr bool SupportsEqualities = false;
};
template<> struct AlgFeatures_<AlgNLoptORIG_DIRECT> {
static constexpr bool SupportsInequalities = true;
static constexpr bool SupportsEqualities = false;
};
template<> struct AlgFeatures_<AlgNLoptAGS> {
static constexpr bool SupportsInequalities = true;
static constexpr bool SupportsEqualities = true;
};
template<class M> struct AlgFeatures_<NLoptAUGLAG<M>> {
static constexpr bool SupportsInequalities = true;
static constexpr bool SupportsEqualities = true;
};
} // namespace detail
}} // namespace Slic3r::opt
#endif // NLOPTOPTIMIZER_HPP