itkCompositeTransform.h

/*=========================================================================
 *
 *  Copyright NumFOCUS
 *
 *  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
 *
 *         https://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 itkCompositeTransform_h
#define itkCompositeTransform_h

#include "itkMultiTransform.h"

#include <deque>

namespace itk
{

/** \class CompositeTransform
 * \brief This class contains a list of transforms and concatenates them by composition.
 *
 * This class concatenates transforms in \b reverse \b queue order by means of composition:
 *    \f$ T_0 o T_1 = T_0(T_1(x)) \f$
 * Transforms are stored in a container (queue), in the following order:
 *    \f$ T_0, T_1, ... , T_N-1 \f$
 * Transforms are added via a single method, AddTransform(). This adds the
 * transforms to the back of the queue. A single method for adding transforms
 * is meant to simplify the interface and prevent errors.
 * One use of the class is to optimize only a subset of included transforms.
 *
 * The sub transforms are the same dimensionality as this class.
 *
 * Example:
 * A user wants to optimize two Affine transforms together, then add a
 * Deformation Field (DF) transform, and optimize it separately.
 * They first add the two Affines, then run the optimization and both Affines
 * transform parameters are optimized. Next, they add the DF transform and call
 * SetOnlyMostRecentTransformToOptimizeOn, which clears the optimization flags
 * for both of the affine transforms, and leaves the flag set only for the DF
 * transform, since it was the last transform added. Now they run the
 * optimization and only the DF transform is optimized, but the affines are
 * included in the transformation during the optimization.
 *
 * Optimization Flags:
 * The m_TransformsToOptimize flags hold one flag for each transform in the
 * queue, designating if each transform is to be used for optimization. Note
 * that all transforms in the queue are applied in TransformPoint, regardless
 * of these flags states'. The methods GetParameters, SetParameters,
 * ComputeJacobianWithRespectToParameters, GetTransformCategory,
 * GetFixedParameters, and SetFixedParameters all query these
 * flags and include only those transforms whose corresponding flag is set.
 * Their input or output is a concatenated array of all transforms set for use
 * in optimization. The goal is to be able to optimize multiple transforms at
 * once, while leaving other transforms fixed. See the above example.
 *
 * Setting Optimization Flags:
 * A transform's optimization flag is set when it is added to the queue, and
 * remains set as other transforms are added. The methods
 * SetNthTransformToOptimize* and SetAllTransformToOptimize* are used to
 * set and clear flags arbitrarily. SetOnlyMostRecentTransformToOptimizeOn is
 * a convenience method for setting only the most recently added transform
 * for optimization, with the idea that this will be a common practice.
 *
 * Indexing:
 * The index values used in GetNthTransform and
 * SetNthTransformToOptimize* and SetAllTransformToOptimize* follow the
 * order in which transforms were added. Thus, the first transform added is at
 * index 0, the next at index 1, etc.
 *
 * Inverse:
 * The inverse transform is created by retrieving the inverse from each
 * sub transform and adding them to a composite transform in reverse order.
 * The m_TransformsToOptimizeFlags is copied in reverse for the inverse.
 *
 * \ingroup ITKTransform
 */
template <typename TParametersValueType = double, unsigned int VDimension = 3>
class ITK_TEMPLATE_EXPORT CompositeTransform : public MultiTransform<TParametersValueType, VDimension, VDimension>
{
public:
  ITK_DISALLOW_COPY_AND_MOVE(CompositeTransform);

  /** Standard class type aliases. */
  using Self = CompositeTransform;
  using Superclass = MultiTransform<TParametersValueType, VDimension, VDimension>;
  using Pointer = SmartPointer<Self>;
  using ConstPointer = SmartPointer<const Self>;

  /** \see LightObject::GetNameOfClass() */
  itkOverrideGetNameOfClassMacro(CompositeTransform);

  /** New macro for creation of through a Smart Pointer */
  itkNewMacro(Self);

  /** Sub transform type **/
  using TransformType = typename Superclass::TransformType;
  using typename Superclass::TransformTypePointer;
  /** InverseTransform type. */
  using typename Superclass::InverseTransformBasePointer;
  /** Scalar type. */
  using typename Superclass::ScalarType;
  /** Parameters type. */
  using typename Superclass::FixedParametersType;
  using typename Superclass::FixedParametersValueType;
  using typename Superclass::ParametersType;
  using typename Superclass::ParametersValueType;
  /** Derivative type */
  using typename Superclass::DerivativeType;
  /** Jacobian types. */
  using typename Superclass::JacobianType;
  using typename Superclass::JacobianPositionType;
  using typename Superclass::InverseJacobianPositionType;
  /** Transform category type. */
  using typename Superclass::TransformCategoryEnum;
  /** Standard coordinate point type for this class. */
  using typename Superclass::InputPointType;
  using typename Superclass::OutputPointType;
  /** Standard vector type for this class. */
  using typename Superclass::InputVectorType;
  using typename Superclass::OutputVectorType;
  /** Standard covariant vector type for this class */
  using typename Superclass::InputCovariantVectorType;
  using typename Superclass::OutputCovariantVectorType;
  /** Standard vnl_vector type for this class. */
  using typename Superclass::InputVnlVectorType;
  using typename Superclass::OutputVnlVectorType;
  /** Standard Vectorpixel type for this class */
  using typename Superclass::InputVectorPixelType;
  using typename Superclass::OutputVectorPixelType;
  /** Standard DiffusionTensor3D type alias for this class */
  using typename Superclass::InputDiffusionTensor3DType;
  using typename Superclass::OutputDiffusionTensor3DType;
  /** Standard SymmetricSecondRankTensor type alias for this class */
  using typename Superclass::InputSymmetricSecondRankTensorType;
  using typename Superclass::OutputSymmetricSecondRankTensorType;

  /** Transform queue type */
  using typename Superclass::TransformQueueType;

  /** The number of parameters defining this transform. */
  using typename Superclass::NumberOfParametersType;

  /** Optimization flags queue type */
  using TransformsToOptimizeFlagsType = std::deque<bool>;

  /** Dimension of the domain spaces. */
  static constexpr unsigned int InputDimension = VDimension;
  static constexpr unsigned int OutputDimension = VDimension;

  /** Active Transform state manipulation */

  virtual void
  SetNthTransformToOptimize(SizeValueType i, bool state)
  {
    this->m_TransformsToOptimizeFlags.at(i) = state;
    this->Modified();
  }

  virtual void
  SetNthTransformToOptimizeOn(SizeValueType i)
  {
    this->SetNthTransformToOptimize(i, true);
  }

  virtual void
  SetNthTransformToOptimizeOff(SizeValueType i)
  {
    this->SetNthTransformToOptimize(i, false);
  }

  virtual void
  SetAllTransformsToOptimize(bool state)
  {
    this->m_TransformsToOptimizeFlags.assign(this->m_TransformsToOptimizeFlags.size(), state);
    this->Modified();
  }

  virtual void
  SetAllTransformsToOptimizeOn()
  {
    this->SetAllTransformsToOptimize(true);
  }

  virtual void
  SetAllTransformsToOptimizeOff()
  {
    this->SetAllTransformsToOptimize(false);
  }

  /* With AddTransform() as the only way to add a transform, we
   * can have this method to easily allow user to optimize only
   * the transform added most recently. */
  virtual void
  SetOnlyMostRecentTransformToOptimizeOn()
  {
    this->SetAllTransformsToOptimize(false);
    this->SetNthTransformToOptimizeOn(this->GetNumberOfTransforms() - 1);
  }

  /* Get whether the Nth transform is set to be optimizied */
  /* NOTE: ambiguous function name here - are we getting if the Nth transform
      is set to be optimized, or the Nth of the transforms that are set to be
      optimized? */
  virtual bool
  GetNthTransformToOptimize(SizeValueType i) const
  {
    return this->m_TransformsToOptimizeFlags.at(i);
  }

  /** Access optimize flags */
  virtual const TransformsToOptimizeFlagsType &
  GetTransformsToOptimizeFlags() const
  {
    return this->m_TransformsToOptimizeFlags;
  }

  void
  ClearTransformQueue() override
  {
    Superclass::ClearTransformQueue();
    this->m_TransformsToOptimizeFlags.clear();
  }

  /** Returns a boolean indicating whether it is possible or not to compute the
   * inverse of this current Transform. If it is possible, then the inverse of
   * the transform is returned in the inverseTransform variable passed by the user.
   * The inverse consists of the inverse of each sub-transform, in the \b reverse order
   * of the forward transforms. */
  bool
  GetInverse(Self * inverse) const;

  InverseTransformBasePointer
  GetInverseTransform() const override;

  /** Compute the position of point in the new space.
   *
   * Transforms are applied starting from the *back* of the
   * queue. That is, in reverse order of which they were added, in order
   * to work properly with ResampleFilter.
   *
   * Imagine a user wants to apply an Affine transform followed by a Deformation
   * Field (DF) transform. He adds the Affine, then the DF. Because the user
   * typically conceptualizes a transformation as being applied from the Moving
   * image to the Fixed image, this makes intuitive sense. But since the
   * ResampleFilter expects to transform from the Fixed image to the Moving
   * image, the transforms are applied in reverse order of addition, i.e. from
   * the back of the queue, and thus, DF then Affine.
   */
  OutputPointType
  TransformPoint(const InputPointType & inputPoint) const override;

  /**  Method to transform a vector. */
  using Superclass::TransformVector;
  OutputVectorType
  TransformVector(const InputVectorType &) const override;

  OutputVnlVectorType
  TransformVector(const InputVnlVectorType & inputVector) const override;

  OutputVectorPixelType
  TransformVector(const InputVectorPixelType & inputVector) const override;

  OutputVectorType
  TransformVector(const InputVectorType & inputVector, const InputPointType & inputPoint) const override;

  OutputVnlVectorType
  TransformVector(const InputVnlVectorType & inputVector, const InputPointType & inputPoint) const override;

  OutputVectorPixelType
  TransformVector(const InputVectorPixelType & inputVector, const InputPointType & inputPoint) const override;

  /**  Method to transform a CovariantVector. */
  using Superclass::TransformCovariantVector;
  OutputCovariantVectorType
  TransformCovariantVector(const InputCovariantVectorType &) const override;

  OutputVectorPixelType
  TransformCovariantVector(const InputVectorPixelType &) const override;

  OutputCovariantVectorType
  TransformCovariantVector(const InputCovariantVectorType & inputVector,
                           const InputPointType &           inputPoint) const override;

  OutputVectorPixelType
  TransformCovariantVector(const InputVectorPixelType & inputVector, const InputPointType & inputPoint) const override;

  /** Method to transform a DiffusionTensor3D */
  using Superclass::TransformDiffusionTensor3D;
  OutputDiffusionTensor3DType
  TransformDiffusionTensor3D(const InputDiffusionTensor3DType & inputTensor) const override;

  OutputVectorPixelType
  TransformDiffusionTensor3D(const InputVectorPixelType & inputTensor) const override;

  OutputDiffusionTensor3DType
  TransformDiffusionTensor3D(const InputDiffusionTensor3DType & inputTensor,
                             const InputPointType &             inputPoint) const override;

  OutputVectorPixelType
  TransformDiffusionTensor3D(const InputVectorPixelType & inputTensor,
                             const InputPointType &       inputPoint) const override;

  /** Method to transform a SymmetricSecondRankTensor */
  using Superclass::TransformSymmetricSecondRankTensor;
  OutputSymmetricSecondRankTensorType
  TransformSymmetricSecondRankTensor(const InputSymmetricSecondRankTensorType & inputTensor) const override;

  OutputVectorPixelType
  TransformSymmetricSecondRankTensor(const InputVectorPixelType & inputTensor) const override;

  OutputSymmetricSecondRankTensorType
  TransformSymmetricSecondRankTensor(const InputSymmetricSecondRankTensorType & inputTensor,
                                     const InputPointType &                     inputPoint) const override;

  OutputVectorPixelType
  TransformSymmetricSecondRankTensor(const InputVectorPixelType & inputTensor,
                                     const InputPointType &       inputPoint) const override;

  /** Special handling for composite transform. If all transforms
   * are linear, then return category Linear. Otherwise if all
   * transforms set to optimize are DisplacementFields, then
   * return DisplacementField category. */
  TransformCategoryEnum
  GetTransformCategory() const override;

  /** Get/Set Parameter functions work on the current list of transforms
      that are set to be optimized (active) using the
      'Set[Nth|All]TransformToOptimize' routines.
      The parameter data from each active transform is
      concatenated into a single ParametersType object.
      \note The sub-transforms are read in \b reverse queue order,
      so the returned array is ordered in the same way. That is,
      the last sub-transform to be added is returned first in the
      parameter array. This is the opposite of what's done in the
      parent MultiTransform class. */
  const ParametersType &
  GetParameters() const override;

  /* SetParameters only for transforms that are set to be optimized
   * See GetParameters() for parameter ordering. */
  void
  SetParameters(const ParametersType & inputParameters) override;

  /* GetFixedParameters only for transforms that are set to be optimized
   * See GetParameters() for parameter ordering. */
  const FixedParametersType &
  GetFixedParameters() const override;

  /* SetFixedParameters only for transforms that are set to be optimized.
   * See GetParameters() for parameter ordering. */
  void
  SetFixedParameters(const FixedParametersType & inputParameters) override;

  /* Get total number of parameters for transforms that are set to be
   * optimized */
  NumberOfParametersType
  GetNumberOfParameters() const override;

  /* Get total number of local parameters for transforms that are set
   * to be optimized */
  NumberOfParametersType
  GetNumberOfLocalParameters() const override;

  /* Get total number of fixed parameters for transforms that are set
   * to be optimized */
  NumberOfParametersType
  GetNumberOfFixedParameters() const override;

  /** Update the transform's parameters by the values in \c update.
   * See GetParameters() for parameter ordering. */
  void
  UpdateTransformParameters(const DerivativeType & update, ScalarType factor = 1.0) override;

  /**
   * Flatten the transform queue such that there are no nested composite transforms.
   */
  virtual void
  FlattenTransformQueue();

  /**
   * Compute the Jacobian with respect to the parameters for the composite
   * transform using Jacobian rule. See comments in the implementation.
   */
  void
  ComputeJacobianWithRespectToParameters(const InputPointType & p, JacobianType & outJacobian) const override;

  /**
   * Expanded interface to Compute the Jacobian with respect to the parameters for the composite
   * transform using Jacobian rule. This version takes in temporary
   * variables to avoid excessive constructions and memory allocations.
   * NOTE: outJacobian MUST be sized correctly prior to the call;
   * outJacobian's size should be [VDimension, this->GetNumberOfLocalParameters() ]
   * jacobianCache may be resized internally and will be reused between calls
   */
  void
  ComputeJacobianWithRespectToParametersCachedTemporaries(const InputPointType & p,
                                                          JacobianType &         outJacobian,
                                                          JacobianType &         cacheJacobian) const override;

protected:
  CompositeTransform() = default;
  ~CompositeTransform() override = default;
  void
  PrintSelf(std::ostream & os, Indent indent) const override;

  /** Clone the current transform */
  typename LightObject::Pointer
  InternalClone() const override;

  void
  PushFrontTransform(TransformTypePointer t) override
  {
    Superclass::PushFrontTransform(t);
    /* Add element to list of flags, and set true by default */
    this->m_TransformsToOptimizeFlags.push_front(true);
  }

  void
  PushBackTransform(TransformTypePointer t) override
  {
    Superclass::PushBackTransform(t);
    /* Add element to list of flags, and set true by default */
    this->m_TransformsToOptimizeFlags.push_back(true);
  }

  void
  PopFrontTransform() override
  {
    Superclass::PopFrontTransform();
    this->m_TransformsToOptimizeFlags.pop_front();
  }

  void
  PopBackTransform() override
  {
    Superclass::PopBackTransform();
    this->m_TransformsToOptimizeFlags.pop_back();
  }

  /** Get a list of transforms to optimize. Helper function. */
  const TransformQueueType &
  GetTransformsToOptimizeQueue() const;

  mutable TransformQueueType    m_TransformsToOptimizeQueue{};
  TransformsToOptimizeFlagsType m_TransformsToOptimizeFlags{};

private:
  mutable ModifiedTimeType m_PreviousTransformsToOptimizeUpdateTime{};
};

} // end namespace itk

#ifndef ITK_MANUAL_INSTANTIATION
#  include "itkCompositeTransform.hxx"
#endif

#endif // itkCompositeTransform_h