/*=========================================================================
 *
 *  Copyright UMC Utrecht and contributors
 *
 *  Licensed under the Apache License, Version 2.0 (the "License");
 *  you may not use this file except in compliance with the License.
 *  You may obtain a copy of the License at
 *
 *        http://www.apache.org/licenses/LICENSE-2.0.txt
 *
 *  Unless required by applicable law or agreed to in writing, software
 *  distributed under the License is distributed on an "AS IS" BASIS,
 *  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 *  See the License for the specific language governing permissions and
 *  limitations under the License.
 *
 *=========================================================================*/
#ifndef elxAdaptiveStochasticVarianceReducedGradient_hxx
#define elxAdaptiveStochasticVarianceReducedGradient_hxx

#include "elxAdaptiveStochasticVarianceReducedGradient.h"
#include <itkDeref.h>

#include <iomanip>
#include <string>
#include <vector>
#include <sstream>
#include <algorithm>
#include <utility>
#include "itkAdvancedImageToImageMetric.h"
#include "itkTimeProbe.h"
#include "itkCommand.h"
#include "itkEventObject.h"
#include "itkMacro.h"

#ifdef ELASTIX_USE_EIGEN
#  include <Eigen/Dense>
#  include <Eigen/Core>
#endif

#include "itkTimeProbesCollectorBase.h" // tmp


namespace elastix
{

/**
 * ********************** Constructor ***********************
 */

template <class TElastix>
AdaptiveStochasticVarianceReducedGradient<TElastix>::AdaptiveStochasticVarianceReducedGradient()
{
  this->m_MaximumNumberOfSamplingAttempts = 0;
  this->m_CurrentNumberOfSamplingAttempts = 0;
  this->m_PreviousErrorAtIteration = 0;
  this->m_AutomaticParameterEstimationDone = false;

  this->m_AutomaticParameterEstimation = false;
  this->m_MaximumStepLength = 1.0;

  this->m_NumberOfGradientMeasurements = 0;
  this->m_NumberOfJacobianMeasurements = 0;
  this->m_NumberOfSamplesForExactGradient = 100000;
  this->m_NumberOfSpatialSamples = 5000;
  this->m_NumberOfInnerLoopSamples = 10;
  this->m_SigmoidScaleFactor = 0.1;
  this->m_NoiseFactor = 0.8;

  this->m_NumberOfInnerIterations = 50;
  this->m_OutsideIterations = 10;

  this->m_RandomGenerator = RandomGeneratorType::GetInstance();
  this->m_AdvancedTransform = nullptr;

  this->m_UseNoiseCompensation = true;
  this->m_OriginalButSigmoidToDefault = false;

  // this->m_LearningRate = 1.0;
} // Constructor


/**
 * ***************** BeforeRegistration ***********************
 */

template <class TElastix>
void
AdaptiveStochasticVarianceReducedGradient<TElastix>::BeforeRegistration()
{
  /** Add the target cell "stepsize" to IterationInfo. */
  this->AddTargetCellToIterationInfo("2:Metric");
  this->AddTargetCellToIterationInfo("3a:Time");
  this->AddTargetCellToIterationInfo("3b:StepSize");
  this->AddTargetCellToIterationInfo("4:||Gradient||");

  /** Format the metric and stepsize as floats. */
  this->GetIterationInfoAt("2:Metric") << std::showpoint << std::fixed;
  this->GetIterationInfoAt("3a:Time") << std::showpoint << std::fixed;
  this->GetIterationInfoAt("3b:StepSize") << std::showpoint << std::fixed;
  this->GetIterationInfoAt("4:||Gradient||") << std::showpoint << std::fixed;

  this->m_SettingsVector.clear();

} // end BeforeRegistration()


/**
 * ***************** BeforeEachResolution ***********************
 */

template <class TElastix>
void
AdaptiveStochasticVarianceReducedGradient<TElastix>::BeforeEachResolution()
{
  /** Get the current resolution level. */
  unsigned int level = static_cast<unsigned int>(this->m_Registration->GetAsITKBaseType()->GetCurrentLevel());

  const unsigned int P = this->GetElastix()->GetElxTransformBase()->GetAsITKBaseType()->GetNumberOfParameters();

  const Configuration & configuration = itk::Deref(Superclass2::GetConfiguration());

  /** Set the maximumNumberOfInnerLoopIterations. */
  SizeValueType maximumNumberOfInnerLoopIterations = 50;
  configuration.ReadParameter(
    maximumNumberOfInnerLoopIterations, "MaximumNumberOfInnerLoopIterations", this->GetComponentLabel(), level, 0);
  this->m_NumberOfInnerIterations = maximumNumberOfInnerLoopIterations;

  /** Set the maximumNumberOfIterations. */
  SizeValueType maximumNumberOfIterations = 500;
  configuration.ReadParameter(
    maximumNumberOfIterations, "MaximumNumberOfIterations", this->GetComponentLabel(), level, 0);
  this->m_OutsideIterations = maximumNumberOfIterations;
  this->SetNumberOfIterations(this->m_OutsideIterations);

  /** Set the numberOfInnerLoopSamples. */
  SizeValueType numberOfInnerLoopSamples = 10;
  configuration.ReadParameter(
    numberOfInnerLoopSamples, "NumberOfInnerLoopSamples", this->GetComponentLabel(), level, 0);
  this->m_NumberOfInnerLoopSamples = numberOfInnerLoopSamples;

  /** Set the NumberOfSpatialSamples. */
  unsigned long numberOfSpatialSamples = 5000;
  configuration.ReadParameter(numberOfSpatialSamples, "NumberOfSpatialSamples", this->GetComponentLabel(), level, 0);
  this->m_NumberOfSpatialSamples = numberOfSpatialSamples;

  /** Set the gain parameter A. */
  double A = 20.0;
  configuration.ReadParameter(A, "SP_A", this->GetComponentLabel(), level, 0);
  this->SetParam_A(A);

  /** Set the MaximumNumberOfSamplingAttempts. */
  SizeValueType maximumNumberOfSamplingAttempts = 0;
  configuration.ReadParameter(
    maximumNumberOfSamplingAttempts, "MaximumNumberOfSamplingAttempts", this->GetComponentLabel(), level, 0);
  this->SetMaximumNumberOfSamplingAttempts(maximumNumberOfSamplingAttempts);
  if (maximumNumberOfSamplingAttempts > 5)
  {
    log::warn(
      std::ostringstream{} << "\nWARNING: You have set MaximumNumberOfSamplingAttempts to "
                           << maximumNumberOfSamplingAttempts << ".\n"
                           << "  This functionality is known to cause problems (stack overflow) for large values.\n"
                           << "  If elastix stops or segfaults for no obvious reason, reduce this value.\n"
                           << "  You may select the RandomSparseMask image sampler to fix mask-related problems.\n");
  }

  /** Set/Get the initial time. Default: 0.0. Should be >=0. */
  double initialTime = 0.0;
  configuration.ReadParameter(initialTime, "SigmoidInitialTime", this->GetComponentLabel(), level, 0);
  this->SetInitialTime(initialTime);

  /** Set the maximum band size of the covariance matrix. */
  this->m_MaxBandCovSize = 192;
  configuration.ReadParameter(this->m_MaxBandCovSize, "MaxBandCovSize", this->GetComponentLabel(), level, 0);

  /** Set the number of random samples used to estimate the structure of the covariance matrix. */
  this->m_NumberOfBandStructureSamples = 10;
  configuration.ReadParameter(
    this->m_NumberOfBandStructureSamples, "NumberOfBandStructureSamples", this->GetComponentLabel(), level, 0);

  /** Set/Get whether the adaptive step size mechanism is desired. Default: true
   * NB: the setting is turned of in case of UseRandomSampleRegion=true.
   * Deprecated alias UseCruzAcceleration is also still supported.
   */
  bool useAdaptiveStepSizes = true;
  configuration.ReadParameter(useAdaptiveStepSizes, "UseCruzAcceleration", this->GetComponentLabel(), level, 0, false);
  configuration.ReadParameter(useAdaptiveStepSizes, "UseAdaptiveStepSizes", this->GetComponentLabel(), level, 0);
  this->SetUseAdaptiveStepSizes(useAdaptiveStepSizes);

  /** Set whether automatic gain estimation is required; default: true. */
  this->m_AutomaticParameterEstimation = true;
  configuration.ReadParameter(
    this->m_AutomaticParameterEstimation, "AutomaticParameterEstimation", this->GetComponentLabel(), level, 0);

  if (this->m_AutomaticParameterEstimation)
  {
    /** Set the maximum step length: the maximum displacement of a voxel in mm.
     * Compute default value: mean spacing of fixed and moving image.
     */
    const unsigned int fixdim = std::min((unsigned int)this->GetElastix()->FixedDimension, (unsigned int)2);
    const unsigned int movdim = std::min((unsigned int)this->GetElastix()->MovingDimension, (unsigned int)2);
    double             sum = 0.0;
    for (unsigned int d = 0; d < fixdim; ++d)
    {
      sum += this->GetElastix()->GetFixedImage()->GetSpacing()[d];
    }
    for (unsigned int d = 0; d < movdim; ++d)
    {
      sum += this->GetElastix()->GetMovingImage()->GetSpacing()[d];
    }
    this->m_MaximumStepLength = sum / static_cast<double>(fixdim + movdim);
    /** Read user setting */
    configuration.ReadParameter(this->m_MaximumStepLength, "MaximumStepLength", this->GetComponentLabel(), level, 0);

    /** Number of gradients N to estimate the average square magnitudes
     * of the exact gradient and the approximation error.
     * A value of 0 (default) means automatic estimation.
     */
    this->m_NumberOfGradientMeasurements = 0;
    configuration.ReadParameter(
      this->m_NumberOfGradientMeasurements, "NumberOfGradientMeasurements", this->GetComponentLabel(), level, 0);

    /** Set the number of Jacobian measurements M.
     * By default, if nothing specified by the user, M is determined as:
     * M = max( 1000, nrofparams );
     * This is a rather crude rule of thumb, which seems to work in practice.
     */
    this->m_NumberOfJacobianMeasurements = std::max(static_cast<unsigned int>(1000), static_cast<unsigned int>(P));
    configuration.ReadParameter(
      this->m_NumberOfJacobianMeasurements, "NumberOfJacobianMeasurements", this->GetComponentLabel(), level, 0);

    /** Set the number of image samples used to compute the 'exact' gradient.
     * By default, if nothing supplied by the user, 100000. This works in general.
     * If the image is smaller, the number of samples is automatically reduced later.
     */
    this->m_NumberOfSamplesForExactGradient = 100000;
    configuration.ReadParameter(
      this->m_NumberOfSamplesForExactGradient, "NumberOfSamplesForExactGradient", this->GetComponentLabel(), level, 0);

    /** Set/Get the scaling factor zeta of the sigmoid width. Large values
     * cause a more wide sigmoid. Default: 0.1. Should be >0.
     */
    double sigmoidScaleFactor = 0.1;
    configuration.ReadParameter(sigmoidScaleFactor, "SigmoidScaleFactor", this->GetComponentLabel(), level, 0);
    this->m_SigmoidScaleFactor = sigmoidScaleFactor;

  } // end if automatic parameter estimation
  else
  {
    /** If no automatic parameter estimation is used, a and alpha also need
     * to be specified.
     */
    double a = 400.0; // arbitrary guess
    double alpha = 0.602;
    configuration.ReadParameter(a, "SP_a", this->GetComponentLabel(), level, 0);
    configuration.ReadParameter(alpha, "SP_alpha", this->GetComponentLabel(), level, 0);
    this->SetParam_a(a);
    this->SetParam_alpha(alpha);

    /** Set/Get the maximum of the sigmoid. Should be >0. Default: 1.0. */
    double sigmoidMax = 1.0;
    configuration.ReadParameter(sigmoidMax, "SigmoidMax", this->GetComponentLabel(), level, 0);
    this->SetSigmoidMax(sigmoidMax);

    /** Set/Get the minimum of the sigmoid. Should be <0. Default: -0.8. */
    double sigmoidMin = -0.8;
    configuration.ReadParameter(sigmoidMin, "SigmoidMin", this->GetComponentLabel(), level, 0);
    this->SetSigmoidMin(sigmoidMin);

    /** Set/Get the scaling of the sigmoid width. Large values
     * cause a more wide sigmoid. Default: 1e-8. Should be >0.
     */
    double sigmoidScale = 1e-8;
    configuration.ReadParameter(sigmoidScale, "SigmoidScale", this->GetComponentLabel(), level, 0);
    this->SetSigmoidScale(sigmoidScale);

  } // end else: no automatic parameter estimation

} // end BeforeEachResolution()


/**
 * ***************** AfterEachIteration *************************
 */

template <class TElastix>
void
AdaptiveStochasticVarianceReducedGradient<TElastix>::AfterEachIteration()
{
  /** Print some information. */
  this->GetIterationInfoAt("2:Metric") << this->GetValue();
  this->GetIterationInfoAt("3a:Time") << this->GetCurrentTime();
  this->GetIterationInfoAt("3b:StepSize") << this->GetLearningRate();
  bool asFastAsPossible = false;
  if (asFastAsPossible)
  {
    this->GetIterationInfoAt("4:||Gradient||") << "---";
  }
  else
  {
    this->GetIterationInfoAt("4:||Gradient||") << this->GetGradient().magnitude();
  }

  /** Select new spatial samples for the computation of the metric. */
  if (this->GetNewSamplesEveryIteration())
  {
    this->SelectNewSamples();
  }

} // end AfterEachIteration()


/**
 * ***************** AfterEachResolution *************************
 */

template <class TElastix>
void
AdaptiveStochasticVarianceReducedGradient<TElastix>::AfterEachResolution()
{
  /** Get the current resolution level. */
  unsigned int level = static_cast<unsigned int>(this->m_Registration->GetAsITKBaseType()->GetCurrentLevel());

  /**
   * enum StopConditionType{
   *   MaximumNumberOfIterations,
   *   MetricError,
   *   MinimumStepSize };
   */
  std::string stopcondition;

  switch (this->GetStopCondition())
  {
    case MetricError:
      stopcondition = "Error in metric";
      break;

    case MinimumStepSize:
      stopcondition = "The minimum step length has been reached";
      break;

    default:
      stopcondition = "Unknown";
      break;
  }

  /** Print the stopping condition. */
  log::info(std::ostringstream{} << "Stopping condition: " << stopcondition << ".");
  this->m_CurrentTime = 0.0;

  /** Store the used parameters, for later printing to screen. */
  SettingsType settings;
  settings.a = this->GetParam_a();
  settings.A = this->GetParam_A();
  settings.alpha = this->GetParam_alpha();
  settings.fmax = this->GetSigmoidMax();
  settings.fmin = this->GetSigmoidMin();
  settings.omega = this->GetSigmoidScale();
  this->m_SettingsVector.push_back(settings);

  /** Print settings that were used in this resolution. */
  SettingsVectorType tempSettingsVector;
  tempSettingsVector.push_back(settings);
  log::info(std::ostringstream{} << "Settings of " << this->elxGetClassName() << " in resolution " << level << ":");
  Superclass2::PrintSettingsVector(tempSettingsVector);

} // end AfterEachResolution()


/**
 * ******************* AfterRegistration ************************
 */

template <class TElastix>
void
AdaptiveStochasticVarianceReducedGradient<TElastix>::AfterRegistration()
{
  /** Print the best metric value. */

  double bestValue = this->GetValue();
  log::info(std::ostringstream{} << '\n'
                                 << "Final metric value  = " << bestValue << '\n'

                                 << "Settings of " << this->elxGetClassName() << " for all resolutions:");
  Superclass2::PrintSettingsVector(this->m_SettingsVector);

} // end AfterRegistration()


/**
 * ****************** StartOptimization *************************
 */

template <class TElastix>
void
AdaptiveStochasticVarianceReducedGradient<TElastix>::StartOptimization()
{
  /** Check if the entered scales are correct and != [ 1 1 1 ...]. */
  this->SetUseScales(false);
  const ScalesType & scales = this->GetScales();
  if (scales.GetSize() == this->GetInitialPosition().GetSize())
  {
    ScalesType unit_scales(scales.GetSize(), 1.0);
    if (scales != unit_scales)
    {
      /** only then: */
      this->SetUseScales(true);
    }
  }

  this->m_AutomaticParameterEstimationDone = false;

  itkDebugMacro("StartOptimization");

  this->m_CurrentIteration = 0;

  /** Get the number of parameters; checks also if a cost function has been set at all.
   * if not: an exception is thrown */
  this->GetScaledCostFunction()->GetNumberOfParameters();

  /** Initialize the scaledCostFunction with the currently set scales */
  this->InitializeScales();

  /** Set the current position as the scaled initial position */
  this->SetCurrentPosition(this->GetInitialPosition());

  this->ResumeOptimization();

} // end StartOptimization()

/**
 * ********************** AdvanceOneStep **********************
 */

template <class TElastix>
void
AdaptiveStochasticVarianceReducedGradient<TElastix>::AdvanceOneStep()
{
  itkDebugMacro("AdvancedOneStep");

  /** Get space dimension. */
  const unsigned int spaceDimension = this->GetScaledCostFunction()->GetNumberOfParameters();

  /** Get a reference to the previously allocated newPosition. */
  ParametersType & newPosition = this->m_ScaledCurrentPosition;

  /** Get a reference to the current position. */
  const ParametersType & currentPosition = this->GetScaledCurrentPosition();

  /** Update the new position. */
  const double learningRate = this->GetLearningRate();
  for (unsigned int j = 0; j < spaceDimension; ++j)
  {
    newPosition[j] = currentPosition[j] - learningRate * this->m_Gradient[j];
  }

  this->InvokeEvent(itk::IterationEvent());
}


/**
 * ********************** StopOptimization **********************
 */

template <class TElastix>
void
AdaptiveStochasticVarianceReducedGradient<TElastix>::StopOptimization()
{
  itkDebugMacro("StopOptimization");

  this->m_Stop = true;
  this->InvokeEvent(itk::EndEvent());
}


/**
 * ********************** ResumeOptimization **********************
 */

template <class TElastix>
void
AdaptiveStochasticVarianceReducedGradient<TElastix>::ResumeOptimization()
{
  itkDebugMacro("ResumeOptimization");

  this->m_Stop = false;

  InvokeEvent(itk::StartEvent());

  itk::TimeProbesCollectorBase timeCollector;

  timeCollector.Start("init");
  /** Set the NumberOfSpatialSamples. */
  SizeValueType numberOfSpatialSamples = this->m_NumberOfSpatialSamples;

  SizeValueType spaceDimension = this->GetScaledCostFunction()->GetNumberOfParameters();

  this->m_Gradient.set_size(spaceDimension); // check this
  this->m_MeanGradient.set_size(spaceDimension);
  DerivativeType localCurrentGradient(spaceDimension);
  DerivativeType localPreviousGradient(spaceDimension);

  const unsigned int M = this->GetElastix()->GetNumberOfMetrics();

  std::vector<ImageRandomSamplerBasePointer>    samplerVec(M);
  std::vector<ImageRandomSamplerPointer>        randomSamplerVec(M);
  std::vector<ImageRandomSamplerPointer>        subRandomSamplerVec(M);
  std::vector<ImageRadomSampleContainerPointer> randomSampleContainer(M);
  /** set the number of samples. */
  //  unsigned int innerNumberOfSpatialSamples = 10;
  // unsigned int innerLoopIterations = 50;

  /** Print the elapsed time. */
  timeCollector.Stop("init");

  SizeValueType nrofsamplesCheck;
  while (!this->m_Stop)
  {
    // Check m_CurrentIteration right at the start of the loop, ensuring that
    // no step at all is performed when when m_NumberOfIterations is zero.
    if (this->m_CurrentIteration >= this->m_NumberOfIterations)
    {
      this->m_StopCondition = MaximumNumberOfIterations;
      this->StopOptimization();
      break;
    }

    /** The following code relies on the fact that all
     * components have been set up and that the initial
     * position has been set, so must be called in this
     * function. */
    if (this->GetAutomaticParameterEstimation() && !this->m_AutomaticParameterEstimationDone)
    {
      this->AutomaticParameterEstimation();
      // hack
      this->m_AutomaticParameterEstimationDone = true;
    } // endif

    for (unsigned int m = 0; m < M; ++m)
    {
      ImageSamplerBasePointer sampler = this->GetElastix()->GetElxMetricBase(m)->GetAdvancedMetricImageSampler();
      samplerVec[m] = dynamic_cast<ImageRandomSamplerBaseType *>(sampler.GetPointer());

      randomSamplerVec[m] = ImageRandomSamplerType::New();
      randomSamplerVec[m]->SetInput(samplerVec[m]->GetInput());
      randomSamplerVec[m]->SetInputImageRegion(samplerVec[m]->GetInputImageRegion());
      randomSamplerVec[m]->SetMask(samplerVec[m]->GetMask());
      randomSamplerVec[m]->SetNumberOfSamples(numberOfSpatialSamples);
      randomSamplerVec[m]->Update();
      this->GetElastix()->GetElxMetricBase(m)->SetAdvancedMetricImageSampler(randomSamplerVec[m]);
      randomSampleContainer[m] = randomSamplerVec[m]->GetOutput();
      nrofsamplesCheck = randomSampleContainer[m]->Size();
    }

    // this->SelectNewSamples();
    timeCollector.Start("copy");
    ParametersType previousPosition = this->GetScaledCurrentPosition();
    timeCollector.Stop("copy");

    timeCollector.Start("g1");
    this->GetRegistration()->GetAsITKBaseType()->GetModifiableMetric()->SetNumberOfWorkUnits(
      this->GetRegistration()->GetAsITKBaseType()->GetMetric()->GetThreader()->GetGlobalDefaultNumberOfThreads());
    this->GetScaledDerivativeWithExceptionHandling(previousPosition, this->m_MeanGradient);
    // this->GetScaledValueAndDerivative( this->GetScaledCurrentPosition() ,this->m_Value, this->m_PreviousGradient );
    timeCollector.Stop("g1");

    /** StopOptimization may have been called. */
    if (this->m_Stop)
    {
      break;
    }

    //
    // Inner loop!
    //
    for (unsigned int m = 0; m < M; ++m)
    {
      ImageSamplerBasePointer sampler = this->GetElastix()->GetElxMetricBase(m)->GetAdvancedMetricImageSampler();
      samplerVec[m] = dynamic_cast<ImageRandomSamplerBaseType *>(sampler.GetPointer());

      subRandomSamplerVec[m] = ImageRandomSamplerType::New();
      //       subRandomSamplerVec[ m ]->SetInput( randomSamplerVec[ m ] ->GetInput());
      //       subRandomSamplerVec[ m ]->SetInputImageRegion( randomSamplerVec[ m ]->
      //         GetInputImageRegion() );
      //       subRandomSamplerVec[ m ]->SetMask( randomSamplerVec[m]->GetMask() );
      subRandomSamplerVec[m]->SetInput(samplerVec[m]->GetInput());
      subRandomSamplerVec[m]->SetInputImageRegion(samplerVec[m]->GetInputImageRegion());
      subRandomSamplerVec[m]->SetMask(samplerVec[m]->GetMask());

      subRandomSamplerVec[m]->SetNumberOfSamples(this->m_NumberOfInnerLoopSamples);
      //      subRandomSamplerVec[ m ]-> SetRandomSampleContainer( randomSampleContainer[m], itp );
      subRandomSamplerVec[m]->Update();
      this->GetElastix()->GetElxMetricBase(m)->SetAdvancedMetricImageSampler(subRandomSamplerVec[m]);

      // temporary hack
      // multi-threading is killing performance when having only a few samples
      // (and large parameter vectors)
      if (this->m_NumberOfInnerLoopSamples < 300)
      {
        this->GetRegistration()->GetAsITKBaseType()->GetModifiableMetric()->SetNumberOfWorkUnits(1);
      }
    }

    const Configuration & configuration = itk::Deref(Superclass2::GetConfiguration());

    this->m_CurrentInnerIteration = 0;

    for (unsigned i = 0; i < this->m_NumberOfInnerIterations; ++i)
    {
      this->SelectNewSamples();
      timeCollector.Start("g23");
      this->GetScaledDerivativeWithExceptionHandling(previousPosition, localPreviousGradient);
      this->GetScaledValueAndDerivative(this->GetScaledCurrentPosition(), this->m_Value, localCurrentGradient);
      timeCollector.Stop("g23");

      timeCollector.Start("gvr");
      // this->m_Gradient = localCurrentGradient - localPreviousGradient + this->m_PreviousGradient;
      this->m_UseNoiseFactor = true;
      configuration.ReadParameter(this->m_UseNoiseFactor, "UseNoiseFactor", this->GetComponentLabel(), 0, 0);

      if (this->m_UseNoiseFactor)
      {
        //         for( unsigned int j = 0; j < spaceDimension; ++j )
        //         {
        //           this->m_Gradient[j] = this->m_NoiseFactor * (localCurrentGradient[j] - localPreviousGradient[j]) +
        //           this->m_PreviousGradient[j];
        //         }
        for (unsigned int j = 0; j < spaceDimension; ++j)
        {
          this->m_Gradient[j] =
            this->m_NoiseFactor * (localCurrentGradient[j] - localPreviousGradient[j]) + this->m_MeanGradient[j];
        }
      }
      else
      {
        for (unsigned int j = 0; j < spaceDimension; ++j)
        {
          this->m_Gradient[j] = (localCurrentGradient[j] - localPreviousGradient[j]) + this->m_MeanGradient[j];
          // this->m_Gradient[j] = this->m_PreviousGradient[j];
        }
      }

      timeCollector.Stop("gvr");

      timeCollector.Start("step");
      this->SetLearningRate(this->Superclass1::Compute_a(this->Superclass1::GetCurrentTime()));
      this->AdvanceOneStep();
      timeCollector.Stop("step");

      this->Superclass1::UpdateCurrentTime();

      this->m_CurrentInnerIteration++;
      /** Preserve the previous position. */
    } // end inner forloop

    /** StopOptimization may have been called. */
    if (this->m_Stop)
    {
      break;
    }

    this->m_CurrentIteration++;

    /** Change the automatic estimation status to false. */
    this->m_AutomaticParameterEstimationDone = false;
    /** Reset the current time to initial time. */
    this->ResetCurrentTimeToInitialTime();

  } // end while

  timeCollector.Report(std::cout);

} // end ResumeOptimization()


/**
 * ****************** MetricErrorResponse *************************
 */

template <class TElastix>
void
AdaptiveStochasticVarianceReducedGradient<TElastix>::MetricErrorResponse(itk::ExceptionObject & err)
{
  if (this->GetCurrentIteration() != this->m_PreviousErrorAtIteration)
  {
    this->m_PreviousErrorAtIteration = this->GetCurrentIteration();
    this->m_CurrentNumberOfSamplingAttempts = 1;
  }
  else
  {
    this->m_CurrentNumberOfSamplingAttempts++;
  }

  if (this->m_CurrentNumberOfSamplingAttempts <= this->m_MaximumNumberOfSamplingAttempts)
  {
    this->SelectNewSamples();
    this->ResumeOptimization();
  }
  else
  {
    /** Stop optimisation and pass on exception. */
    this->Superclass1::MetricErrorResponse(err);
  }

} // end MetricErrorResponse()


/**
 * ******************* AutomaticParameterEstimation **********************
 */

template <class TElastix>
void
AdaptiveStochasticVarianceReducedGradient<TElastix>::AutomaticParameterEstimation()
{
  /** Setup timers. */
  itk::TimeProbe timer1;

  /** Total time. */
  timer1.Start();
  log::info(std::ostringstream{} << "Starting automatic parameter estimation for " << this->elxGetClassName()
                                 << " ...");

  const Configuration & configuration = itk::Deref(Superclass2::GetConfiguration());

  /** Decide which method is to be used. */
  std::string asgdParameterEstimationMethod = "Original";
  configuration.ReadParameter(
    asgdParameterEstimationMethod, "ASGDParameterEstimationMethod", this->GetComponentLabel(), 0, 0);

  /** Perform automatic optimizer parameter estimation by the desired method. */
  if (asgdParameterEstimationMethod == "Original")
  {
    /** Original ASGD estimation method. */
    this->m_OriginalButSigmoidToDefault = false;
    this->AutomaticParameterEstimationOriginal();
  }
  else if (asgdParameterEstimationMethod == "OriginalButSigmoidToDefault")
  {
    /** Original ASGD estimation method, but keeping the sigmoid parameters fixed. */
    this->m_OriginalButSigmoidToDefault = true;
    this->AutomaticParameterEstimationOriginal();
  }
  else if (asgdParameterEstimationMethod == "DisplacementDistribution")
  {
    /** Accelerated parameter estimation method. */
    this->AutomaticParameterEstimationUsingDisplacementDistribution();
  }

  /** Print the elapsed time. */
  timer1.Stop();
  log::info(std::ostringstream{} << "Automatic parameter estimation took "
                                 << Conversion::SecondsToDHMS(timer1.GetMean(), 6));

} // end AutomaticParameterEstimation()


/**
 * ******************* AutomaticParameterEstimationOriginal **********************
 */

template <class TElastix>
void
AdaptiveStochasticVarianceReducedGradient<TElastix>::AutomaticParameterEstimationOriginal()
{
  itk::TimeProbe timer2, timer3;

  /** Get the user input. */
  const double delta = this->GetMaximumStepLength();

  /** Compute the Jacobian terms. */
  double TrC = 0.0;
  double TrCC = 0.0;
  double maxJJ = 0.0;
  double maxJCJ = 0.0;

  /** Get current position to start the parameter estimation. */
  this->GetRegistration()->GetAsITKBaseType()->GetModifiableTransform()->SetParameters(this->GetCurrentPosition());

  /** Cast to advanced metric type. */
  using MetricType = typename ElastixType::MetricBaseType::AdvancedMetricType;
  MetricType * testPtr = dynamic_cast<MetricType *>(this->GetElastix()->GetElxMetricBase()->GetAsITKBaseType());
  if (!testPtr)
  {
    itkExceptionMacro("ERROR: AdaptiveStochasticVarianceReducedGradient expects the metric to be of type "
                      "AdvancedImageToImageMetric!");
  }

  /** Construct computeJacobianTerms to initialize the parameter estimation. */
  auto computeJacobianTerms = ComputeJacobianTermsType::New();
  computeJacobianTerms->SetFixedImage(testPtr->GetFixedImage());
  computeJacobianTerms->SetFixedImageRegion(testPtr->GetFixedImageRegion());
  computeJacobianTerms->SetFixedImageMask(testPtr->GetFixedImageMask());
  computeJacobianTerms->SetTransform(this->GetRegistration()->GetAsITKBaseType()->GetModifiableTransform());
  computeJacobianTerms->SetMaxBandCovSize(this->m_MaxBandCovSize);
  computeJacobianTerms->SetNumberOfBandStructureSamples(this->m_NumberOfBandStructureSamples);
  computeJacobianTerms->SetNumberOfJacobianMeasurements(this->m_NumberOfJacobianMeasurements);

  /** Check if use scales. */
  bool useScales = this->GetUseScales();
  if (useScales)
  {
    computeJacobianTerms->SetScales(this->m_ScaledCostFunction->GetScales());
    computeJacobianTerms->SetUseScales(true);
  }
  else
  {
    computeJacobianTerms->SetUseScales(false);
  }

  /** Compute the Jacobian terms. */
  log::info("  Computing JacobianTerms ...");
  timer2.Start();
  computeJacobianTerms->Compute(TrC, TrCC, maxJJ, maxJCJ);
  timer2.Stop();
  log::info(std::ostringstream{} << "  Computing the Jacobian terms took "
                                 << Conversion::SecondsToDHMS(timer2.GetMean(), 6));

  /** Determine number of gradient measurements such that
   * E + 2\sqrt(Var) < K E
   * with
   * E = E(1/N \sum_n g_n^T g_n) = sigma_1^2 TrC
   * Var = Var(1/N \sum_n g_n^T g_n) = 2 sigma_1^4 TrCC / N
   * K = 1.5
   * We enforce a minimum of 2.
   */
  timer3.Start();
  if (this->m_NumberOfGradientMeasurements == 0)
  {
    const double K = 1.5;
    if (TrCC > 1e-14 && TrC > 1e-14)
    {
      this->m_NumberOfGradientMeasurements =
        static_cast<unsigned int>(std::ceil(8.0 * TrCC / TrC / TrC / (K - 1) / (K - 1)));
    }
    else
    {
      this->m_NumberOfGradientMeasurements = 2;
    }
    this->m_NumberOfGradientMeasurements =
      std::max(static_cast<SizeValueType>(2), this->m_NumberOfGradientMeasurements);
    log::info(std::ostringstream{} << "  NumberOfGradientMeasurements to estimate sigma_i: "
                                   << this->m_NumberOfGradientMeasurements);
  }

  /** Measure square magnitude of exact gradient and approximation error. */
  const double sigma4factor = 1.0;
  double       sigma4 = 0.0;
  double       gg = 0.0;
  double       ee = 0.0;
  if (maxJJ > 1e-14)
  {
    sigma4 = sigma4factor * delta / std::sqrt(maxJJ);
  }
  this->SampleGradients(this->GetScaledCurrentPosition(), sigma4, gg, ee);
  timer3.Stop();
  log::info(std::ostringstream{} << "  Sampling the gradients took " << Conversion::SecondsToDHMS(timer3.GetMean(), 6));

  /** Determine parameter settings. */
  double sigma1 = 0.0;
  double sigma3 = 0.0;
  /** Estimate of sigma such that empirical norm^2 equals theoretical:
   * gg = 1/N sum_n g_n' g_n
   * sigma = gg / TrC
   */
  if (gg > 1e-14 && TrC > 1e-14)
  {
    sigma1 = std::sqrt(gg / TrC);
  }
  if (ee > 1e-14 && TrC > 1e-14)
  {
    sigma3 = std::sqrt(ee / TrC);
  }

  const double alpha = 1.0;
  const double A = this->GetParam_A();
  double       a_max = 0.0;
  if (sigma1 > 1e-14 && maxJCJ > 1e-14)
  {
    a_max = A * delta / sigma1 / std::sqrt(maxJCJ);
  }
  const double noisefactor = sigma1 * sigma1 / (sigma1 * sigma1 + sigma3 * sigma3 + 1e-14);
  const double a = a_max * noisefactor;

  const double omega = std::max(1e-14, this->m_SigmoidScaleFactor * sigma3 * sigma3 * std::sqrt(TrCC));
  const double fmax = 1.0;
  const double fmin = -0.99 + 0.98 * noisefactor;

  /** Set parameters in superclass. */
  this->SetParam_a(a);
  this->SetParam_alpha(alpha);

  /** Set parameters for original method. */
  if (!this->m_OriginalButSigmoidToDefault)
  {
    this->SetSigmoidMax(fmax);
    this->SetSigmoidMin(fmin);
    this->SetSigmoidScale(omega);
  }
} // end AutomaticParameterEstimationOriginal()


/**
 * *************** AutomaticParameterEstimationUsingDisplacementDistribution *****
 */

template <class TElastix>
void
AdaptiveStochasticVarianceReducedGradient<TElastix>::AutomaticParameterEstimationUsingDisplacementDistribution()
{
  itk::TimeProbe timer4, timer5;

  /** Get current position to start the parameter estimation. */
  this->GetRegistration()->GetAsITKBaseType()->GetModifiableTransform()->SetParameters(this->GetCurrentPosition());

  /** Get the user input. */
  const double delta = this->GetMaximumStepLength();
  double       maxJJ = 0;

  /** Cast to advanced metric type. */
  using MetricType = typename ElastixType::MetricBaseType::AdvancedMetricType;
  MetricType * testPtr = dynamic_cast<MetricType *>(this->GetElastix()->GetElxMetricBase()->GetAsITKBaseType());
  if (!testPtr)
  {
    itkExceptionMacro("ERROR: AdaptiveStochasticVarianceReducedGradient expects the metric to be of type "
                      "AdvancedImageToImageMetric!");
  }

  /** Construct computeJacobianTerms to initialize the parameter estimation. */
  typename ComputeDisplacementDistributionType::Pointer computeDisplacementDistribution =
    ComputeDisplacementDistributionType::New();
  computeDisplacementDistribution->SetFixedImage(testPtr->GetFixedImage());
  computeDisplacementDistribution->SetFixedImageRegion(testPtr->GetFixedImageRegion());
  computeDisplacementDistribution->SetFixedImageMask(testPtr->GetFixedImageMask());
  computeDisplacementDistribution->SetTransform(this->GetRegistration()->GetAsITKBaseType()->GetModifiableTransform());
  computeDisplacementDistribution->SetCostFunction(this->m_CostFunction);
  computeDisplacementDistribution->SetNumberOfJacobianMeasurements(this->m_NumberOfJacobianMeasurements);

  /** Check if use scales. */
  if (this->GetUseScales())
  {
    computeDisplacementDistribution->SetUseScales(true);
    computeDisplacementDistribution->SetScales(this->m_ScaledCostFunction->GetScales());
    // this setting is not successful. Because copying the scales from ASGD to computeDisplacementDistribution is
    // failed.
  }
  else
  {
    computeDisplacementDistribution->SetUseScales(false);
  }

  const Configuration & configuration = itk::Deref(Superclass2::GetConfiguration());

  double      jacg = 0.0;
  std::string maximumDisplacementEstimationMethod = "2sigma";
  configuration.ReadParameter(
    maximumDisplacementEstimationMethod, "MaximumDisplacementEstimationMethod", this->GetComponentLabel(), 0, 0);

  /** Compute the Jacobian terms. */
  log::info("  Computing displacement distribution ...");
  timer4.Start();
  computeDisplacementDistribution->Compute(
    this->GetScaledCurrentPosition(), jacg, maxJJ, maximumDisplacementEstimationMethod);
  timer4.Stop();
  log::info(std::ostringstream{} << "  Computing the displacement distribution took "
                                 << Conversion::SecondsToDHMS(timer4.GetMean(), 6));

  /** Initial of the variables. */
  double       a = 0.0;
  const double A = this->GetParam_A();
  const double alpha = 1.0;

  this->m_UseNoiseCompensation = true;
  configuration.ReadParameter(this->m_UseNoiseCompensation, "NoiseCompensation", this->GetComponentLabel(), 0, 0);

  /** Use noise Compensation factor or not. */
  if (this->m_UseNoiseCompensation)
  {
    double sigma4 = 0.0;
    double gg = 0.0;
    double ee = 0.0;
    double sigma4factor = 1.0;

    /** Sample the grid and random sampler container to estimate the noise factor.*/
    if (this->m_NumberOfGradientMeasurements == 0)
    {
      this->m_NumberOfGradientMeasurements =
        std::max(static_cast<SizeValueType>(2), this->m_NumberOfGradientMeasurements);
      log::info(std::ostringstream{} << "  NumberOfGradientMeasurements to estimate sigma_i: "
                                     << this->m_NumberOfGradientMeasurements);
    }
    timer5.Start();
    if (maxJJ > 1e-14)
    {
      sigma4 = sigma4factor * delta / std::sqrt(maxJJ);
    }
    this->SampleGradients(this->GetScaledCurrentPosition(), sigma4, gg, ee);

    double noisefactor = gg / (gg + ee);
    this->m_NoiseFactor = noisefactor;
    a = delta * std::pow(A + 1.0, alpha) / jacg * noisefactor;
    timer5.Stop();
    log::info(std::ostringstream{} << "  Compute the noise compensation took "
                                   << Conversion::SecondsToDHMS(timer5.GetMean(), 6));
  }
  else
  {
    a = delta * std::pow(A + 1.0, alpha) / jacg;
  }

  /** Set parameters in superclass. */
  this->SetParam_a(a);
  this->SetParam_alpha(alpha);

} // end AutomaticParameterEstimationUsingDisplacementDistribution()


/**
 * ******************** SampleGradients **********************
 */

template <class TElastix>
void
AdaptiveStochasticVarianceReducedGradient<TElastix>::SampleGradients(const ParametersType & mu0,
                                                                     double                 perturbationSigma,
                                                                     double &               gg,
                                                                     double &               ee)
{
  /** Some shortcuts. */
  const unsigned int M = this->GetElastix()->GetNumberOfMetrics();

  /** Variables for sampler support. Each metric may have a sampler. */
  std::vector<bool>                                useRandomSampleRegionVec(M, false);
  std::vector<ImageRandomSamplerBasePointer>       randomSamplerVec(M);
  std::vector<ImageRandomCoordinateSamplerPointer> randomCoordinateSamplerVec(M);
  std::vector<ImageGridSamplerPointer>             gridSamplerVec(M);

  const Configuration & configuration = itk::Deref(Superclass2::GetConfiguration());

  /** If new samples every iteration, get each sampler, and check if it is
   * a kind of random sampler. If yes, prepare an additional grid sampler
   * for the exact gradients, and set the stochasticgradients flag to true.
   */
  bool stochasticgradients = false;
  if (this->GetNewSamplesEveryIteration())
  {
    for (unsigned int m = 0; m < M; ++m)
    {
      /** Get the sampler. */
      ImageSamplerBasePointer sampler = this->GetElastix()->GetElxMetricBase(m)->GetAdvancedMetricImageSampler();
      randomSamplerVec[m] = dynamic_cast<ImageRandomSamplerBaseType *>(sampler.GetPointer());
      randomCoordinateSamplerVec[m] = dynamic_cast<ImageRandomCoordinateSamplerType *>(sampler.GetPointer());

      if (randomSamplerVec[m].IsNotNull())
      {
        /** At least one of the metric has a random sampler. */
        stochasticgradients |= true;

        /** If the sampler is a randomCoordinateSampler set the UseRandomSampleRegion
         * property to false temporarily. It disturbs the parameter estimation.
         * At the end of this function the original setting is set back.
         * Also, the AdaptiveStepSize mechanism is turned off when any of the samplers
         * has UseRandomSampleRegion==true.
         * \todo Extend ASGD to really take into account random region sampling.
         */
        if (randomCoordinateSamplerVec[m].IsNotNull())
        {
          useRandomSampleRegionVec[m] = randomCoordinateSamplerVec[m]->GetUseRandomSampleRegion();
          if (useRandomSampleRegionVec[m])
          {
            if (this->GetUseAdaptiveStepSizes())
            {
              log::warn(
                "WARNING: UseAdaptiveStepSizes is turned off, because UseRandomSampleRegion is set to \"true\".");
              this->SetUseAdaptiveStepSizes(false);
            }
          }
          randomCoordinateSamplerVec[m]->SetUseRandomSampleRegion(false);

        } // end if random coordinate sampler

        /** Set up the grid sampler for the "exact" gradients.
         * Copy settings from the random sampler and update.
         */
        gridSamplerVec[m] = ImageGridSamplerType::New();
        gridSamplerVec[m]->SetInput(randomSamplerVec[m]->GetInput());
        gridSamplerVec[m]->SetInputImageRegion(randomSamplerVec[m]->GetInputImageRegion());
        gridSamplerVec[m]->SetMask(randomSamplerVec[m]->GetMask());

        /** Decide which method is to be used. */
        std::string sampleGradientMethod = "OriginalLargeNumber";
        configuration.ReadParameter(sampleGradientMethod, "SampleGradientMethod", this->GetComponentLabel(), 0, 0);
        if (sampleGradientMethod == "OriginalLargeNumber")
        {
          gridSamplerVec[m]->SetNumberOfSamples(this->m_NumberOfSpatialSamples);
          gridSamplerVec[m]->Update();
        }
        else if (sampleGradientMethod == "UseNumberOfSpatialSamples")
        {
          gridSamplerVec[m]->SetNumberOfSamples(this->m_NumberOfSpatialSamples);
          gridSamplerVec[m]->Update();
          randomSamplerVec[m]->SetNumberOfSamples(this->m_NumberOfSpatialSamples);
          randomSamplerVec[m]->Update();
        }
        else if (sampleGradientMethod == "InnerLoopSamples")
        {
          gridSamplerVec[m]->SetNumberOfSamples(this->m_NumberOfSpatialSamples);
          gridSamplerVec[m]->Update();
          randomSamplerVec[m]->SetNumberOfSamples(this->m_NumberOfInnerLoopSamples);
          randomSamplerVec[m]->Update();
        }

      } // end if random sampler

    } // end for loop over metrics
  }   // end if NewSamplesEveryIteration.

  /** Prepare for progress printing. */
  const bool showProgressPercentage = configuration.RetrieveParameterValue(false, "ShowProgressPercentage", 0, false);
  const auto progressObserver = showProgressPercentage
                                  ? ProgressCommand::CreateAndSetUpdateFrequency(this->m_NumberOfGradientMeasurements)
                                  : nullptr;
  log::info("  Sampling gradients ...");

  /** Initialize some variables for storing gradients and their magnitudes. */
  const unsigned int P = this->GetElastix()->GetElxTransformBase()->GetAsITKBaseType()->GetNumberOfParameters();
  DerivativeType     approxgradient(P);
  DerivativeType     exactgradient(P);
  DerivativeType     diffgradient;
  double             exactgg = 0.0;
  double             diffgg = 0.0;

  /** Compute gg for some random parameters. */
  for (unsigned int i = 0; i < this->m_NumberOfGradientMeasurements; ++i)
  {
    if (progressObserver != nullptr)
    {
      /** Show progress 0-100% */
      progressObserver->UpdateAndPrintProgress(i);
    }
    /** Generate a perturbation, according to:
     *    \mu_i ~ N( \mu_0, perturbationsigma^2 I ).
     */
    ParametersType perturbedMu0 = mu0;
    this->AddRandomPerturbation(perturbedMu0, perturbationSigma);

    /** Compute contribution to exactgg and diffgg. */
    if (stochasticgradients)
    {
      /** Set grid sampler(s) and get exact derivative. */
      for (unsigned int m = 0; m < M; ++m)
      {
        if (gridSamplerVec[m].IsNotNull())
        {
          this->GetElastix()->GetElxMetricBase(m)->SetAdvancedMetricImageSampler(gridSamplerVec[m]);
        }
      }
      this->GetScaledDerivativeWithExceptionHandling(perturbedMu0, exactgradient);
      this->m_ExactGradient = exactgradient;

      /** Set random sampler(s), select new spatial samples and get approximate derivative. */
      for (unsigned int m = 0; m < M; ++m)
      {
        if (randomSamplerVec[m].IsNotNull())
        {
          this->GetElastix()->GetElxMetricBase(m)->SetAdvancedMetricImageSampler(randomSamplerVec[m]);
        }
      }
      this->SelectNewSamples();
      this->GetScaledDerivativeWithExceptionHandling(perturbedMu0, approxgradient);

      /** Compute error vector. */
      diffgradient = exactgradient - approxgradient;

      /** Compute g^T g and e^T e */
      exactgg += exactgradient.squared_magnitude();
      diffgg += diffgradient.squared_magnitude();
    }
    else // no stochastic gradients
    {
      /** Get exact gradient. */
      this->GetScaledDerivativeWithExceptionHandling(perturbedMu0, exactgradient);

      /** Compute g^T g. NB: diffgg=0. */
      exactgg += exactgradient.squared_magnitude();
    } // end else: no stochastic gradients

  } // end for loop over gradient measurements
  if (progressObserver != nullptr)
  {
    progressObserver->PrintProgress(1.0);
  }
  /** Compute means. */
  exactgg /= this->m_NumberOfGradientMeasurements;
  diffgg /= this->m_NumberOfGradientMeasurements;

  /** For output: gg and ee.
   * gg and ee will be divided by Pd, but actually need to be divided by
   * the rank, in case of maximum likelihood. In case of no maximum likelihood,
   * the rank equals Pd.
   */
  gg = exactgg;
  ee = diffgg;

  /** Set back useRandomSampleRegion flag to what it was. */
  for (unsigned int m = 0; m < M; ++m)
  {
    if (randomCoordinateSamplerVec[m].IsNotNull())
    {
      randomCoordinateSamplerVec[m]->SetUseRandomSampleRegion(useRandomSampleRegionVec[m]);
    }
  }

} // end SampleGradients()


/**
 * *************** GetScaledDerivativeWithExceptionHandling ***************
 */

template <class TElastix>
void
AdaptiveStochasticVarianceReducedGradient<TElastix>::GetScaledDerivativeWithExceptionHandling(
  const ParametersType & parameters,
  DerivativeType &       derivative)
{
  double dummyvalue = 0;
  try
  {
    this->GetScaledValueAndDerivative(parameters, dummyvalue, derivative);
  }
  catch (const itk::ExceptionObject &)
  {
    this->m_StopCondition = MetricError;
    this->StopOptimization();
    throw;
  }

} // end GetScaledDerivativeWithExceptionHandling()


/**
 * *************** AddRandomPerturbation ***************
 */

template <class TElastix>
void
AdaptiveStochasticVarianceReducedGradient<TElastix>::AddRandomPerturbation(ParametersType & parameters, double sigma)
{
  /** Add delta ~ sigma * N(0,I) to the input parameters. */
  for (unsigned int p = 0; p < parameters.GetSize(); ++p)
  {
    parameters[p] += sigma * this->m_RandomGenerator->GetNormalVariate(0.0, 1.0);
  }

} // end AddRandomPerturbation()

} // end namespace elastix

#endif // end #ifndef elxAdaptiveStochasticVarianceReducedGradient_hxx