MoffatModelFittingTask.cpp

Go to the documentation of this file.

/*
 * MoffatModelTask.cpp
 *
 *  Created on: May 2, 2017
 *      Author: mschefer
 */
 
#include <iostream>
#include <tuple>
#include <vector>
#include <valarray>
#include <boost/any.hpp>
#include <mutex>
 
#include "AlexandriaKernel/memory_tools.h"
#include "SEImplementation/Plugin/MoffatModelFitting/MoffatModelFitting.h"
#include "SEImplementation/Plugin/MoffatModelFitting/MoffatModelFittingTask.h"
#include "ElementsKernel/PathSearch.h"
 
#include "SEImplementation/Image/ImageInterfaceTraits.h"
#include "SEImplementation/Image/ImagePsf.h"
 
#include "ModelFitting/Parameters/ManualParameter.h"
#include "ModelFitting/Parameters/DependentParameter.h"
 
#include "ModelFitting/Parameters/ExpSigmoidConverter.h"
#include "ModelFitting/Parameters/SigmoidConverter.h"
#include "ModelFitting/Parameters/NormalizedConverter.h"
#include "ModelFitting/Models/OnlySmooth.h"
#include "ModelFitting/Models/OldSharp.h"
#include "ModelFitting/Models/AutoSharp.h"
#include "ModelFitting/Models/CircularlySymmetricModelComponent.h"
#include "ModelFitting/Models/ScaledModelComponent.h"
#include "ModelFitting/Models/RotatedModelComponent.h"
#include "ModelFitting/Models/PointModel.h"
#include "ModelFitting/Models/ExtendedModel.h"
#include "ModelFitting/Models/FlattenedMoffatComponent.h"
 
#include "ModelFitting/Models/FrameModel.h"
#include "ModelFitting/Engine/ResidualEstimator.h"
#include "ModelFitting/Engine/LeastSquareEngineManager.h"
 
#include "ModelFitting/Engine/AsinhChiSquareComparator.h"
 
 
#include "ModelFitting/Engine/EngineParameterManager.h"
#include "ModelFitting/Engine/DataVsModelResiduals.h"
 
#include "SEImplementation/Plugin/DetectionFrameSourceStamp/DetectionFrameSourceStamp.h"
 
#include "SEImplementation/Plugin/PixelCentroid/PixelCentroid.h"
#include "SEImplementation/Property/PixelCoordinateList.h"
#include "SEImplementation/Plugin/PixelBoundaries/PixelBoundaries.h"
#include "SEImplementation/Plugin/ShapeParameters/ShapeParameters.h"
#include "SEImplementation/Plugin/PeakValue/PeakValue.h"
#include "SEImplementation/Plugin/IsophotalFlux/IsophotalFlux.h"
#include "SEImplementation/Plugin/DetectionFrameCoordinates/DetectionFrameCoordinates.h"
#include "SEImplementation/Plugin/DetectionFrameInfo/DetectionFrameInfo.h"
 
#include "SEImplementation/Image/VectorImageDataVsModelInputTraits.h"
 
#include "SEImplementation/Measurement/MultithreadedMeasurement.h"
 
#include "ModelFitting/Image/NullPsf.h"
 
 
namespace SourceXtractor {
 
using namespace ModelFitting;
using Euclid::make_unique;
 
namespace {
 
struct SourceModel {
  double m_size;
  std::shared_ptr<EngineParameter> dx, dy;
  std::shared_ptr<DependentParameter<std::shared_ptr<EngineParameter>>> x, y;
 
  double exp_i0_guess;
  std::shared_ptr<EngineParameter> moffat_i0, moffat_index, minkowski_exponent, flat_top_offset;
  std::shared_ptr<EngineParameter> moffat_x_scale, moffat_y_scale, moffat_rotation;
 
  SourceModel(double size, double x_guess, double y_guess, double pos_range,
      double exp_flux_guess, double exp_radius_guess, double exp_aspect_guess, double exp_rot_guess) :
 
    m_size(size),
    dx(std::make_shared<EngineParameter>(0, make_unique<SigmoidConverter>(-pos_range, pos_range))),
    dy(std::make_shared<EngineParameter>(0, make_unique<SigmoidConverter>(-pos_range, pos_range))),
 
    x(createDependentParameter([x_guess](double dx) { return dx + x_guess + 0.5; }, dx)),
    y(createDependentParameter([y_guess](double dy) { return dy + y_guess + 0.5; }, dy)),
 
    // FIXME
    exp_i0_guess(exp_flux_guess / (M_PI * 2.0 * 0.346 * exp_radius_guess * exp_radius_guess * exp_aspect_guess)),
    moffat_i0(std::make_shared<EngineParameter>(exp_i0_guess, make_unique<ExpSigmoidConverter>(exp_i0_guess * .00001, exp_i0_guess * 1000))),
 
    moffat_index(std::make_shared<EngineParameter>(1, make_unique<ExpSigmoidConverter>(0.5, 8))),
    minkowski_exponent(std::make_shared<EngineParameter>(2, make_unique<ExpSigmoidConverter>(0.5, 10))),
    flat_top_offset(std::make_shared<EngineParameter>(1, make_unique<ExpSigmoidConverter>(0.000001, 10))),
 
    moffat_x_scale(std::make_shared<EngineParameter>(1, make_unique<ExpSigmoidConverter>(0.0001, 100.0))),
    moffat_y_scale(std::make_shared<EngineParameter>(1, make_unique<ExpSigmoidConverter>(0.0001, 100.0))),
    moffat_rotation(std::make_shared<EngineParameter>(-exp_rot_guess, make_unique<SigmoidConverter>(-2*M_PI, 2*M_PI)))
  {
  }
 
  void registerParameters(EngineParameterManager& manager) {
    manager.registerParameter(dx);
    manager.registerParameter(dy);
 
    manager.registerParameter(moffat_i0);
    manager.registerParameter(moffat_index);
    manager.registerParameter(minkowski_exponent);
    manager.registerParameter(flat_top_offset);
 
    manager.registerParameter(moffat_x_scale);
    manager.registerParameter(moffat_y_scale);
    manager.registerParameter(moffat_rotation);
  }
 
  void createModels(std::vector<std::shared_ptr<ModelFitting::ExtendedModel<ImageInterfaceTypePtr>>>& extended_models, std::vector<PointModel>& /*point_models*/) {
    // Moffat model
    {
      std::vector<std::unique_ptr<ModelComponent>> component_list {};
      auto moff = Euclid::make_unique<FlattenedMoffatComponent>(moffat_i0, moffat_index, minkowski_exponent, flat_top_offset);
      component_list.clear();
      component_list.emplace_back(std::move(moff));
      extended_models.emplace_back(std::make_shared<ModelFitting::ExtendedModel<ImageInterfaceTypePtr>>(
          std::move(component_list), moffat_x_scale, moffat_y_scale, moffat_rotation, m_size, m_size, x, y));
    }
  }
};
 
}
 
 
void MoffatModelFittingTask::computeProperties(SourceInterface& source) const {
  auto& source_stamp = source.getProperty<DetectionFrameSourceStamp>().getStamp();
  auto& variance_stamp = source.getProperty<DetectionFrameSourceStamp>().getVarianceStamp();
  auto& thresholded_stamp = source.getProperty<DetectionFrameSourceStamp>().getThresholdedStamp();
  auto& threshold_map_stamp = source.getProperty<DetectionFrameSourceStamp>().getThresholdMapStamp();
  PixelCoordinate stamp_top_left = source.getProperty<DetectionFrameSourceStamp>().getTopLeft();
 
  // Computes the minimum flux that a detection should have (min. detection threshold for every pixel)
  // This will be used instead of lower or negative fluxes that can happen for various reasons
  double min_flux = 0.;
  auto& pixel_coordinates = source.getProperty<PixelCoordinateList>().getCoordinateList();
  for (auto pixel : pixel_coordinates) {
    pixel -= stamp_top_left;
 
    min_flux += threshold_map_stamp.getValue(pixel);
  }
 
  double pixel_scale = 1;
 
  EngineParameterManager manager {};
  std::vector<ConstantModel> constant_models;
  std::vector<std::shared_ptr<ModelFitting::ExtendedModel<ImageInterfaceTypePtr>>> extended_models;
  std::vector<PointModel> point_models;
 
  auto& pixel_centroid = source.getProperty<PixelCentroid>();
  auto& shape_parameters = source.getProperty<ShapeParameters>();
  auto iso_flux = source.getProperty<IsophotalFlux>().getFlux();
 
  auto size = std::max<double>(source_stamp.getWidth(), source_stamp.getHeight());
 
  double radius_guess = shape_parameters.getEllipseA() / 2.0;
 
  double guess_x = pixel_centroid.getCentroidX() - stamp_top_left.m_x;
  double guess_y = pixel_centroid.getCentroidY() - stamp_top_left.m_y;
 
  double exp_flux_guess = std::max<double>(iso_flux, min_flux);
 
  double exp_reff_guess = radius_guess;
  double exp_aspect_guess = std::max<double>(shape_parameters.getEllipseB() / shape_parameters.getEllipseA(), 0.01);
  double exp_rot_guess = shape_parameters.getEllipseTheta();
 
  auto source_model = make_unique<SourceModel>(size, guess_x, guess_y, radius_guess * 2,
      exp_flux_guess, exp_reff_guess, exp_aspect_guess, exp_rot_guess);
 
  source_model->createModels(extended_models, point_models);
  source_model->registerParameters(manager);
 
  // Full frame model with all sources
  typedef std::shared_ptr<VectorImage<SourceXtractor::SeFloat>> VectorImageType;
  FrameModel<NullPsf<VectorImageType>, VectorImageType> frame_model {
    pixel_scale,
    (size_t) source_stamp.getWidth(), (size_t) source_stamp.getHeight(),
    std::move(constant_models),
    std::move(point_models),
    std::move(extended_models)
  };
 
  auto weight = VectorImage<SeFloat>::create(source_stamp.getWidth(), source_stamp.getHeight());
  std::fill(weight->getData().begin(), weight->getData().end(), 1);
 
  for (int y=0; y < source_stamp.getHeight(); y++) {
    for (int x=0; x < source_stamp.getWidth(); x++) {
      weight->at(x, y) = (thresholded_stamp.getValue(x, y) >= 0) ? 0 : 1;
    }
  }
 
  for (auto pixel : pixel_coordinates) {
    pixel -= stamp_top_left;
    weight->at(pixel.m_x, pixel.m_y) = 1;
  }
 
  const auto& detection_frame_info = source.getProperty<DetectionFrameInfo>();
  SeFloat gain = detection_frame_info.getGain();
  SeFloat saturation = detection_frame_info.getSaturation();
 
  for (int y=0; y < source_stamp.getHeight(); y++) {
    for (int x=0; x < source_stamp.getWidth(); x++) {
      auto back_var = variance_stamp.getValue(x, y);
      if (saturation > 0 && source_stamp.getValue(x, y) >= saturation) {
        weight->at(x, y) = 0;
      } else if (weight->at(x, y)>0) {
        if (gain > 0.0) {
          weight->at(x, y) = sqrt(1.0 / (back_var + source_stamp.getValue(x, y) / gain));
        } else {
          weight->at(x, y) = sqrt(1.0 / back_var); // infinite gain
        }
      }
    }
  }
 
  // FIXME we should be able to use the source_stamp Image interface directly
  auto image = VectorImage<SeFloat>::create(source_stamp);
 
  auto data_vs_model =
      createDataVsModelResiduals(image, std::move(frame_model), weight, AsinhChiSquareComparator{});
 
  ResidualEstimator res_estimator {};
  res_estimator.registerBlockProvider(move(data_vs_model));
 
  // Perform the minimization
  auto engine = LeastSquareEngineManager::create(m_least_squares_engine, m_max_iterations);
  auto solution = engine->solveProblem(manager, res_estimator);
  size_t iterations = solution.iteration_no;
 
  auto final_stamp = VectorImage<SeFloat>::create(source_stamp.getWidth(), source_stamp.getHeight());
 
  {
 
    // renders an image of the model for a single source with the final parameters
    std::vector<std::shared_ptr<ModelFitting::ExtendedModel<ImageInterfaceTypePtr>>> tmp_extended_models{};
    std::vector<PointModel> tmp_point_models{};
    std::vector<ConstantModel> tmp_constant_models{};
    source_model->createModels(tmp_extended_models, tmp_point_models);
    FrameModel<NullPsf<VectorImageType>, VectorImageType> frame_model_after{1,
                                                                            (size_t)source_stamp.getWidth(),
                                                                            (size_t)source_stamp.getHeight(),
                                                                            std::move(tmp_constant_models),
                                                                            std::move(tmp_point_models),
                                                                            std::move(tmp_extended_models)};
    auto                                                  final_image = frame_model_after.getImage();
 
    // integrates the flux for that source
    double total_flux = 0;
    for (int y = 0; y < source_stamp.getHeight(); y++) {
      for (int x = 0; x < source_stamp.getWidth(); x++) {
        PixelCoordinate pixel(x, y);
        pixel += stamp_top_left;
 
        // build final stamp
        final_stamp->setValue(x, y, final_stamp->getValue(x, y) + final_image->getValue(x, y));
 
        total_flux += final_image->getValue(x, y);
      }
    }
 
    auto coordinate_system = source.getProperty<DetectionFrameCoordinates>().getCoordinateSystem();
 
    SeFloat x = stamp_top_left.m_x + source_model->x->getValue() - 0.5f;
    SeFloat y = stamp_top_left.m_y + source_model->y->getValue() - 0.5f;
 
    source.setProperty<MoffatModelFitting>(
        x, y,
 
        source_model->moffat_i0->getValue(), source_model->moffat_index->getValue(),
        source_model->minkowski_exponent->getValue(), source_model->flat_top_offset->getValue(), source_model->m_size,
        source_model->moffat_x_scale->getValue(), source_model->moffat_y_scale->getValue(),
        source_model->moffat_rotation->getValue(),
 
        iterations);
  }
}
 
}