recurrent_layer.h

//   OpenNN: Open Neural Networks Library
//   www.opennn.net
//
//   R E C U R R E N T   L A Y E R   C L A S S   H E A D E R
//
//   Artificial Intelligence Techniques SL
//   artelnics@artelnics.com
 
#ifndef RECURRENTLAYER_H
#define RECURRENTLAYER_H
 
// System includes
 
#include <cmath>
#include <cstdlib>
#include <fstream>
#include <iostream>
#include <string>
#include <sstream>
 
// OpenNN includes
 
#include "config.h"
#include "tensor_utilities.h"
#include "layer.h"
 
#include "probabilistic_layer.h"
#include "perceptron_layer.h"
 
namespace OpenNN
{
 
struct RecurrentLayerForwardPropagation;
struct RecurrentLayerBackPropagation;
 
 
#ifdef OPENNN_CUDA
        #include "../../opennn-cuda/opennn-cuda/struct_recurrent_layer_cuda.h"
#endif
 
 
class RecurrentLayer : public Layer
{
 
public:
 
 
    enum class ActivationFunction{Threshold, SymmetricThreshold, Logistic, HyperbolicTangent,
                            Linear, RectifiedLinear, ExponentialLinear,
                            ScaledExponentialLinear, SoftPlus, SoftSign, HardSigmoid};
 
   // Constructors
 
   explicit RecurrentLayer();
 
   explicit RecurrentLayer(const Index&, const Index&);
 
   // Destructor
   
   virtual ~RecurrentLayer();
 
   // Get methods
 
   bool is_empty() const;
 
   Index get_inputs_number() const;
   Index get_neurons_number() const;
 
   const Tensor<type, 1>& get_hidden_states() const;
 
   // Parameters
 
   Index get_timesteps() const;
 
   Tensor<type, 1> get_biases() const;
   const Tensor<type, 2>& get_input_weights() const;
   const Tensor<type, 2>& get_recurrent_weights() const;
 
   Index get_biases_number() const;
   Index get_input_weights_number() const;
   Index get_recurrent_weights_number() const;
 
   Index get_parameters_number() const;
   Tensor<type, 1> get_parameters() const;
 
   Tensor<type, 2> get_biases(const Tensor<type, 1>&) const;
   Tensor<type, 2> get_input_weights(const Tensor<type, 1>&) const;
   Tensor<type, 2> get_recurrent_weights(const Tensor<type, 1>&) const;
 
   // Activation functions
 
   const RecurrentLayer::ActivationFunction& get_activation_function() const;
 
   string write_activation_function() const;
 
   // Display messages
 
   const bool& get_display() const;
 
   // Set methods
 
   void set();
   void set(const Index&, const Index&);
   void set(const RecurrentLayer&);
 
   void set_default();
 
   // Architecture
 
   void set_inputs_number(const Index&);
   void set_neurons_number(const Index&);
   void set_input_shape(const Tensor<Index, 1>&);
 
   // Parameters
 
   void set_timesteps(const Index&);
 
   void set_biases(const Tensor<type, 1>&);
 
   void set_input_weights(const Tensor<type, 2>&);
 
   void set_recurrent_weights(const Tensor<type, 2>&);
 
   void set_parameters(const Tensor<type, 1>&, const Index& = 0);
 
   // Activation functions
 
   void set_activation_function(const ActivationFunction&);
   void set_activation_function(const string&);
 
   // Display messages
 
   void set_display(const bool&);
 
   // Parameters initialization methods
 
   void set_hidden_states_constant(const type&);
 
   void set_biases_constant(const type&);
 
   void set_input_weights_constant(const type&);
   void set_recurrent_weights_constant(const type&);
   void initialize_input_weights_Glorot(const type&, const type&);
 
   void set_parameters_constant(const type&);
 
   void set_parameters_random();
 
   // neuron layer combinations
 
   void calculate_combinations(const Tensor<type, 1>&,
                               const Tensor<type, 2>&,
                               const Tensor<type, 2>&,
                               const Tensor<type, 1>&,
                               Tensor<type, 1>&) const;
 
   void calculate_activations(const Tensor<type, 1>&,
                              Tensor<type, 1>&) const;
 
   void calculate_activations_derivatives(const Tensor<type, 1>&,
                                          Tensor<type, 1>&,
                                          Tensor<type, 1>&) const;
 
   void calculate_activations_derivatives(const Tensor<type, 2>&,
                                          Tensor<type, 2>&,
                                          Tensor<type, 2>&) const;
 
 
   // neuron layer outputs
 
   Tensor<type, 2> calculate_outputs(const Tensor<type, 2>&);
 
   void forward_propagate(const Tensor<type, 2>&, LayerForwardPropagation*);
 
   void forward_propagate(const Tensor<type, 2>&, const Tensor<type, 1>, LayerForwardPropagation*);
 
   void calculate_hidden_delta(LayerForwardPropagation*,
                               LayerBackPropagation*,
                               LayerBackPropagation*) const;
 
   void calculate_hidden_delta_perceptron(PerceptronLayerForwardPropagation*,
                                          PerceptronLayerBackPropagation*,
                                          RecurrentLayerBackPropagation*) const;
 
   void calculate_hidden_delta_probabilistic(ProbabilisticLayerForwardPropagation*,
                                             ProbabilisticLayerBackPropagation*,
                                             RecurrentLayerBackPropagation*) const;
 
   // Gradient
 
   void insert_gradient(LayerBackPropagation*, const Index& , Tensor<type, 1>&) const;
 
   void calculate_error_gradient(const Tensor<type, 2>&,
                                 LayerForwardPropagation*,
                                 LayerBackPropagation*) const;
 
   void calculate_biases_error_gradient(const Tensor<type, 2>&,
                                        RecurrentLayerForwardPropagation*,
                                        RecurrentLayerBackPropagation*) const;
 
   void calculate_input_weights_error_gradient(const Tensor<type, 2>&,
                                               RecurrentLayerForwardPropagation*,
                                               RecurrentLayerBackPropagation*) const;
 
   void calculate_recurrent_weights_error_gradient(const Tensor<type, 2>&,
                                                   RecurrentLayerForwardPropagation*,
                                                   RecurrentLayerBackPropagation*) const;
 
   // Expression methods
 
   string write_expression(const Tensor<string, 1>&, const Tensor<string, 1>&) const;
 
   string write_activation_function_expression() const;
 
   string write_expression_python() const;
   string write_combinations_python() const;
   string write_activations_python() const;
 
   // Serialization methods
 
   void from_XML(const tinyxml2::XMLDocument&);
 
   void write_XML(tinyxml2::XMLPrinter&) const;
 
protected:
 
   Index timesteps = 1;
 
 
   Tensor<type, 1> biases;
 
   Tensor<type, 2> input_weights;
 
 
   Tensor<type, 2> recurrent_weights;
 
 
   ActivationFunction activation_function = ActivationFunction::HyperbolicTangent;
 
   Tensor<type, 1> hidden_states;
 
 
   bool display = true;
 
#ifdef OPENNN_CUDA
    #include "../../opennn-cuda/opennn-cuda/recurrent_layer_cuda.h"
#else
};
#endif
 
struct RecurrentLayerForwardPropagation : LayerForwardPropagation
{
    explicit RecurrentLayerForwardPropagation() : LayerForwardPropagation()
    {
    }
 
    explicit RecurrentLayerForwardPropagation(const Index& new_batch_samples_number, Layer* new_layer_pointer) : LayerForwardPropagation()
    {
        set(new_batch_samples_number, new_layer_pointer);
    }
 
    void set(const Index& new_batch_samples_number, Layer* new_layer_pointer)
    {
        layer_pointer = new_layer_pointer;
 
        batch_samples_number = new_batch_samples_number;
 
        const Index neurons_number = layer_pointer->get_neurons_number();
        const Index inputs_number = layer_pointer->get_inputs_number();
 
        previous_activations.resize(neurons_number);
 
        current_inputs.resize(inputs_number);
        current_combinations.resize(neurons_number);
        current_activations_derivatives.resize(neurons_number);
 
        combinations.resize(batch_samples_number, neurons_number);
 
        activations.resize(batch_samples_number, neurons_number);
 
        activations_derivatives.resize(batch_samples_number, neurons_number);
    }
 
    void print() const
    {
    }
 
    Tensor<type, 1> previous_activations;
 
    Tensor<type, 1> current_inputs;
    Tensor<type, 1> current_combinations;
    Tensor<type, 1> current_activations_derivatives;
 
    Tensor<type, 2> combinations;
    Tensor<type, 2> activations;
    Tensor<type, 2> activations_derivatives;
};
 
 
struct RecurrentLayerBackPropagation : LayerBackPropagation
{
    explicit RecurrentLayerBackPropagation() : LayerBackPropagation()
    {
    }
 
    explicit RecurrentLayerBackPropagation(const Index& new_batch_samples_number, Layer* new_layer_pointer)
        : LayerBackPropagation()
    {
        set(new_batch_samples_number, new_layer_pointer);
    }
 
 
    void set(const Index& new_batch_samples_number, Layer* new_layer_pointer)
    {
        layer_pointer = new_layer_pointer;
 
        batch_samples_number = new_batch_samples_number;
 
        const Index neurons_number = layer_pointer->get_neurons_number();
        const Index inputs_number = layer_pointer->get_inputs_number();
 
        current_layer_deltas.resize(neurons_number);
 
        biases_derivatives.resize(neurons_number);
 
        input_weights_derivatives.resize(inputs_number * neurons_number);
 
        recurrent_weights_derivatives.resize(neurons_number * neurons_number);
 
        delta.resize(batch_samples_number, neurons_number);
 
        combinations_biases_derivatives.resize(neurons_number, neurons_number);
        combinations_weights_derivatives.resize(inputs_number*neurons_number, neurons_number);
        combinations_recurrent_weights_derivatives.resize(neurons_number*neurons_number, neurons_number);
    }
 
 
    void print() const
    {
 
    }
 
    Tensor<type, 1> current_layer_deltas;
 
    Tensor<type, 1> biases_derivatives;
 
    Tensor<type, 1> input_weights_derivatives;
 
    Tensor<type, 1> recurrent_weights_derivatives;
 
    Tensor<type, 2> combinations_biases_derivatives;
    Tensor<type, 2> combinations_weights_derivatives;
    Tensor<type, 2> combinations_recurrent_weights_derivatives;
 
    Tensor<type, 2> delta;
};
 
 
 
 
}
 
#endif
 
// OpenNN: Open Neural Networks Library.
// Copyright(C) 2005-2021 Artificial Intelligence Techniques, SL.
//
// This library is free software; you can redistribute it and/or
// modify it under the terms of the GNU Lesser General Public
// License as published by the Free Software Foundation; either
// version 2.1 of the License, or any later version.
//
// This library 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
// Lesser General Public License for more details.
 
// You should have received a copy of the GNU Lesser General Public
// License along with this library; if not, write to the Free Software
 
// Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA