recurrent_attention.hpp

Go to the documentation of this file.

#ifndef MLPACK_METHODS_ANN_LAYER_RECURRENT_ATTENTION_HPP
#define MLPACK_METHODS_ANN_LAYER_RECURRENT_ATTENTION_HPP

#include <mlpack/prereqs.hpp>
#include <boost/ptr_container/ptr_vector.hpp>

#include "../visitor/delta_visitor.hpp"
#include "../visitor/output_parameter_visitor.hpp"
#include "../visitor/reset_visitor.hpp"
#include "../visitor/weight_size_visitor.hpp"

#include "layer_types.hpp"
#include "add_merge.hpp"
#include "sequential.hpp"

namespace mlpack {
namespace ann  {

template <
    typename InputDataType = arma::mat,
    typename OutputDataType = arma::mat
>
class RecurrentAttention
{
 public:
  template<typename RNNModuleType, typename ActionModuleType>
  RecurrentAttention(const size_t outSize,
                     const RNNModuleType& rnn,
                     const ActionModuleType& action,
                     const size_t rho);

  template<typename eT>
  void Forward(arma::Mat<eT>&& input, arma::Mat<eT>&& output);

  template<typename eT>
  void Backward(const arma::Mat<eT>&& /* input */,
                arma::Mat<eT>&& gy,
                arma::Mat<eT>&& g);

  /*
   * Calculate the gradient using the output delta and the input activation.
   *
   * @param input The input parameter used for calculating the gradient.
   * @param error The calculated error.
   * @param gradient The calculated gradient.
   */
  template<typename eT>
  void Gradient(arma::Mat<eT>&& /* input */,
                arma::Mat<eT>&& /* error */,
                arma::Mat<eT>&& /* gradient */);

  std::vector<LayerTypes>& Model() { return network; }

  bool Deterministic() const { return deterministic; }
  bool& Deterministic() { return deterministic; }

  OutputDataType const& Parameters() const { return parameters; }
  OutputDataType& Parameters() { return parameters; }

  InputDataType const& InputParameter() const { return inputParameter; }
  InputDataType& InputParameter() { return inputParameter; }

  OutputDataType const& OutputParameter() const { return outputParameter; }
  OutputDataType& OutputParameter() { return outputParameter; }

  OutputDataType const& Delta() const { return delta; }
  OutputDataType& Delta() { return delta; }

  OutputDataType const& Gradient() const { return gradient; }
  OutputDataType& Gradient() { return gradient; }

  template<typename Archive>
  void Serialize(Archive& ar, const unsigned int /* version */);

 private:
  void IntermediateGradient()
  {
    intermediateGradient.zeros();

    // Gradient of the action module.
    if (backwardStep == (rho - 1))
    {
      boost::apply_visitor(GradientVisitor(std::move(initialInput),
          std::move(actionError)), actionModule);
    }
    else
    {
      boost::apply_visitor(GradientVisitor(std::move(boost::apply_visitor(
          outputParameterVisitor, actionModule)), std::move(actionError)),
          actionModule);
    }

    // Gradient of the recurrent module.
    boost::apply_visitor(GradientVisitor(std::move(boost::apply_visitor(
        outputParameterVisitor, rnnModule)), std::move(recurrentError)),
        rnnModule);

    attentionGradient += intermediateGradient;
  }

  size_t outSize;

  LayerTypes rnnModule;

  LayerTypes actionModule;

  size_t rho;

  size_t forwardStep;

  size_t backwardStep;

  bool deterministic;

  OutputDataType parameters;

  LayerTypes initialModule;

  LayerTypes recurrentModule;

  std::vector<LayerTypes> network;

  LayerTypes mergeModule;

  WeightSizeVisitor weightSizeVisitor;

  DeltaVisitor deltaVisitor;

  OutputParameterVisitor outputParameterVisitor;

  std::vector<arma::mat> feedbackOutputParameter;

  std::vector<arma::mat> moduleOutputParameter;

  OutputDataType delta;

  OutputDataType gradient;

  InputDataType inputParameter;

  OutputDataType outputParameter;

  arma::mat recurrentError;

  arma::mat actionError;

  arma::mat actionDelta;

  arma::mat rnnDelta;

  arma::mat initialInput;

  ResetVisitor resetVisitor;

  arma::mat attentionGradient;

  arma::mat intermediateGradient;
}; // class RecurrentAttention

} // namespace ann
} // namespace mlpack

// Include implementation.
#include "recurrent_attention_impl.hpp"

#endif