optimizer.cc

Go to the documentation of this file.

// -*- lsst-c++ -*-
/*
 * LSST Data Management System
 * Copyright 2008-2013 LSST Corporation.
 *
 * This product includes software developed by the
 * LSST Project (http://www.lsst.org/).
 *
 * This program is free software: you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation, either version 3 of the License, or
 * (at your option) any later version.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the LSST License Statement and
 * the GNU General Public License along with this program.  If not,
 * see <http://www.lsstcorp.org/LegalNotices/>.
 */

#include "pybind11/pybind11.h"

#include "ndarray/pybind11.h"

#include "lsst/pex/config/python.h"

#include "lsst/meas/modelfit/optimizer.h"
#include "lsst/meas/modelfit/Model.h"
#include "lsst/meas/modelfit/Likelihood.h"
#include "lsst/meas/modelfit/Prior.h"
#include "lsst/afw/table/Catalog.h"

namespace py = pybind11;
using namespace pybind11::literals;

namespace lsst {
namespace meas {
namespace modelfit {
namespace {

using PyOptimizerObjective = py::class_<OptimizerObjective, std::shared_ptr<OptimizerObjective>>;
using PyOptimizerControl = py::class_<OptimizerControl, std::shared_ptr<OptimizerControl>>;
using PyOptimizerHistoryRecorder =
        py::class_<OptimizerHistoryRecorder, std::shared_ptr<OptimizerHistoryRecorder>>;
using PyOptimizer = py::class_<Optimizer, std::shared_ptr<Optimizer>>;

static PyOptimizerObjective declareOptimizerObjective(py::module &mod) {
    PyOptimizerObjective cls(mod, "OptimizerObjective");
    // Class is abstract, so no constructor.
    cls.def_readonly("dataSize", &OptimizerObjective::dataSize);
    cls.def_readonly("parameterSize", &OptimizerObjective::parameterSize);
    cls.def_static("makeFromLikelihood", &OptimizerObjective::makeFromLikelihood, "likelihood"_a,
                   "prior"_a = nullptr);
    // class is abstract and not subclassable in Python, so we don't wrap the ctor
    cls.def("fillObjectiveValueGrid", &OptimizerObjective::fillObjectiveValueGrid, "parameters"_a,
            "output"_a);
    cls.def("computeResiduals", &OptimizerObjective::computeResiduals, "parameters"_a, "residuals"_a);
    cls.def("differentiateResiduals", &OptimizerObjective::differentiateResiduals, "parameters"_a,
            "derivatives"_a);
    cls.def("hasPrior", &OptimizerObjective::hasPrior);
    cls.def("computePrior", &OptimizerObjective::computePrior, "parameters"_a);
    cls.def("differentiatePrior", &OptimizerObjective::differentiatePrior, "parameters"_a, "gradient"_a,
            "hessian"_a);
    return cls;
}

static PyOptimizerControl declareOptimizerControl(py::module &mod) {
    PyOptimizerControl cls(mod, "OptimizerControl");
    LSST_DECLARE_CONTROL_FIELD(cls, OptimizerControl, noSR1Term);
    LSST_DECLARE_CONTROL_FIELD(cls, OptimizerControl, skipSR1UpdateThreshold);
    LSST_DECLARE_CONTROL_FIELD(cls, OptimizerControl, minTrustRadiusThreshold);
    LSST_DECLARE_CONTROL_FIELD(cls, OptimizerControl, gradientThreshold);
    LSST_DECLARE_CONTROL_FIELD(cls, OptimizerControl, numDiffRelStep);
    LSST_DECLARE_CONTROL_FIELD(cls, OptimizerControl, numDiffAbsStep);
    LSST_DECLARE_CONTROL_FIELD(cls, OptimizerControl, numDiffTrustRadiusStep);
    LSST_DECLARE_CONTROL_FIELD(cls, OptimizerControl, stepAcceptThreshold);
    LSST_DECLARE_CONTROL_FIELD(cls, OptimizerControl, trustRegionInitialSize);
    LSST_DECLARE_CONTROL_FIELD(cls, OptimizerControl, trustRegionGrowReductionRatio);
    LSST_DECLARE_CONTROL_FIELD(cls, OptimizerControl, trustRegionGrowStepFraction);
    LSST_DECLARE_CONTROL_FIELD(cls, OptimizerControl, trustRegionGrowFactor);
    LSST_DECLARE_CONTROL_FIELD(cls, OptimizerControl, trustRegionShrinkReductionRatio);
    LSST_DECLARE_CONTROL_FIELD(cls, OptimizerControl, trustRegionShrinkFactor);
    LSST_DECLARE_CONTROL_FIELD(cls, OptimizerControl, trustRegionSolverTolerance);
    LSST_DECLARE_CONTROL_FIELD(cls, OptimizerControl, maxInnerIterations);
    LSST_DECLARE_CONTROL_FIELD(cls, OptimizerControl, maxOuterIterations);
    LSST_DECLARE_CONTROL_FIELD(cls, OptimizerControl, doSaveIterations);
    cls.def(py::init<>());
    return cls;
}

static PyOptimizerHistoryRecorder declareOptimizerHistoryRecorder(py::module &mod) {
    PyOptimizerHistoryRecorder cls(mod, "OptimizerHistoryRecorder");
    cls.def(py::init<afw::table::Schema &, std::shared_ptr<Model>, bool>(), "schema"_a, "model"_a,
            "doRecordDerivatives"_a);
    cls.def(py::init<afw::table::Schema const &>(), "schema"_a);
    cls.def("apply", &OptimizerHistoryRecorder::apply, "outerIterCount"_a, "innerIterCount"_a, "history"_a,
            "optimizer"_a);
    cls.def("unpackDerivatives",
            (void (OptimizerHistoryRecorder::*)(ndarray::Array<Scalar const, 1, 1> const &,
                                                ndarray::Array<Scalar, 1, 1> const &,
                                                ndarray::Array<Scalar, 2, 2> const &) const) &
                    OptimizerHistoryRecorder::unpackDerivatives,
            "nested"_a, "gradient"_a, "hessian"_a);
    cls.def("unpackDerivatives", (void (OptimizerHistoryRecorder::*)(
                                         afw::table::BaseRecord const &, ndarray::Array<Scalar, 1, 1> const &,
                                         ndarray::Array<Scalar, 2, 2> const &) const) &
                                         OptimizerHistoryRecorder::unpackDerivatives,
            "record"_a, "gradient"_a, "hessian"_a);
    // Other unpackDerivatives overloads do the same thing but with Eigen types,
    // which makes them redundant in Python where it's all just NumPy.
    cls.def("fillObjectiveModelGrid", &OptimizerHistoryRecorder::fillObjectiveModelGrid, "record"_a,
            "parameters"_a, "output"_a);
    cls.def_readonly("outer", &OptimizerHistoryRecorder::outer);
    cls.def_readonly("inner", &OptimizerHistoryRecorder::inner);
    cls.def_readonly("state", &OptimizerHistoryRecorder::state);
    cls.def_readonly("objective", &OptimizerHistoryRecorder::objective);
    cls.def_readonly("prior", &OptimizerHistoryRecorder::prior);
    cls.def_readonly("trust", &OptimizerHistoryRecorder::trust);
    cls.def_readonly("parameters", &OptimizerHistoryRecorder::parameters);
    cls.def_readonly("derivatives", &OptimizerHistoryRecorder::derivatives);
    return cls;
}

static PyOptimizer declareOptimizer(py::module &mod) {
    PyOptimizer cls(mod, "Optimizer");
    // StateFlags enum is used as bitflag, so we wrap values as int class attributes.
    cls.attr("CONVERGED_GRADZERO") = py::cast(int(Optimizer::CONVERGED_GRADZERO));
    cls.attr("CONVERGED_TR_SMALL") = py::cast(int(Optimizer::CONVERGED_TR_SMALL));
    cls.attr("CONVERGED") = py::cast(int(Optimizer::CONVERGED));
    cls.attr("FAILED_MAX_INNER_ITERATIONS") = py::cast(int(Optimizer::FAILED_MAX_INNER_ITERATIONS));
    cls.attr("FAILED_MAX_OUTER_ITERATIONS") = py::cast(int(Optimizer::FAILED_MAX_OUTER_ITERATIONS));
    cls.attr("FAILED_MAX_ITERATIONS") = py::cast(int(Optimizer::FAILED_MAX_ITERATIONS));
    cls.attr("FAILED_EXCEPTION") = py::cast(int(Optimizer::FAILED_EXCEPTION));
    cls.attr("FAILED_NAN") = py::cast(int(Optimizer::FAILED_NAN));
    cls.attr("FAILED") = py::cast(int(Optimizer::FAILED));
    cls.attr("STATUS_STEP_REJECTED") = py::cast(int(Optimizer::STATUS_STEP_REJECTED));
    cls.attr("STATUS_STEP_ACCEPTED") = py::cast(int(Optimizer::STATUS_STEP_ACCEPTED));
    cls.attr("STATUS_STEP") = py::cast(int(Optimizer::STATUS_STEP));
    cls.attr("STATUS_TR_UNCHANGED") = py::cast(int(Optimizer::STATUS_TR_UNCHANGED));
    cls.attr("STATUS_TR_DECREASED") = py::cast(int(Optimizer::STATUS_TR_DECREASED));
    cls.attr("STATUS_TR_INCREASED") = py::cast(int(Optimizer::STATUS_TR_INCREASED));
    cls.attr("STATUS_TR") = py::cast(int(Optimizer::STATUS_TR));
    cls.attr("STATUS") = py::cast(int(Optimizer::STATUS));
    cls.def(py::init<std::shared_ptr<Optimizer::Objective const>, ndarray::Array<Scalar const, 1, 1> const &,
                     Optimizer::Control>(),
            "objective"_a, "parameters"_a, "ctrl"_a);
    cls.def("getObjective", &Optimizer::getObjective);
    cls.def("getControl", &Optimizer::getControl, py::return_value_policy::copy);
    cls.def("step", (bool (Optimizer::*)()) & Optimizer::step);
    cls.def("step", (bool (Optimizer::*)(Optimizer::HistoryRecorder const &, afw::table::BaseCatalog &)) &
                            Optimizer::step,
            "recorder"_a, "history"_a);
    cls.def("run", (int (Optimizer::*)()) & Optimizer::run);
    cls.def("run", (int (Optimizer::*)(Optimizer::HistoryRecorder const &, afw::table::BaseCatalog &)) &
                           Optimizer::run,
            "recorder"_a, "history"_a);
    cls.def("getState", &Optimizer::getState);
    cls.def("getObjectiveValue", &Optimizer::getObjectiveValue);
    cls.def("getParameters", &Optimizer::getParameters);
    cls.def("getResiduals", &Optimizer::getResiduals);
    cls.def("getGradient", &Optimizer::getGradient);
    cls.def("getHessian", &Optimizer::getHessian);
    cls.def("removeSR1Term", &Optimizer::removeSR1Term);
    return cls;
}

PYBIND11_MODULE(optimizer, mod) {
    py::module::import("lsst.meas.modelfit.model");
    py::module::import("lsst.meas.modelfit.likelihood");
    py::module::import("lsst.meas.modelfit.priors");

    auto clsObjective = declareOptimizerObjective(mod);
    auto clsControl = declareOptimizerControl(mod);
    auto clsHistoryRecorder = declareOptimizerHistoryRecorder(mod);
    auto cls = declareOptimizer(mod);
    cls.attr("Objective") = clsObjective;
    cls.attr("Control") = clsControl;
    cls.attr("HistoryRecorder") = clsHistoryRecorder;

    mod.def("solveTrustRegion", &solveTrustRegion, "x"_a, "F"_a, "g"_a, "r"_a, "tolerance"_a);
}

}
}
}
}  // namespace lsst::meas::modelfit::anonymous