opennn-5.0.5/opennn/weighted_squared_error.cpp

//   OpenNN: Open Neural Networks Library
//   www.opennn.net
//
//   W E I G T H E D   S Q U A R E D   E R R O R   C L A S S
//
//   Artificial Intelligence Techniques SL
//   artelnics@artelnics.com

#include "weighted_squared_error.h"

namespace OpenNN
{

/// Default constructor.
/// It creates a weighted squared error term not associated to any
/// neural network and not measured on any data set.
/// It also initializes all the rest of class members to their default values.

WeightedSquaredError::WeightedSquaredError() : LossIndex()
{
    set_default();
}


/// Neural network and data set constructor.
/// It creates a weighted squared error term object associated to a
/// neural network and measured on a data set.
/// It also initializes all the rest of class members to their default values.
/// @param new_neural_network_pointer Pointer to a neural network object.
/// @param new_data_set_pointer Pointer to a data set object.

WeightedSquaredError::WeightedSquaredError(NeuralNetwork* new_neural_network_pointer, DataSet* new_data_set_pointer)
    : LossIndex(new_neural_network_pointer, new_data_set_pointer)
{
    set_default();
}


/// Destructor.

WeightedSquaredError::~WeightedSquaredError()
{
}


/// Returns the weight of the positives.

type WeightedSquaredError::get_positives_weight() const
{
    return positives_weight;
}


/// Returns the weight of the negatives.

type WeightedSquaredError::get_negatives_weight() const
{
    return negatives_weight;
}


type WeightedSquaredError::get_normalizaton_coefficient() const
{
    return normalization_coefficient;
}


/// Set the default values for the object.

void WeightedSquaredError::set_default()
{
    if(has_data_set() && data_set_pointer->has_data())
    {
        set_weights();

        set_normalization_coefficient();
    }
    else
    {
        negatives_weight = -1.0;
        positives_weight = -1.0;

        normalization_coefficient = -1.0;
    }
}


/// Set a new weight for the positives values.
/// @param new_positives_weight New weight for the positives.

void WeightedSquaredError::set_positives_weight(const type& new_positives_weight)
{
    positives_weight = new_positives_weight;
}


/// Set a new weight for the negatives values.
/// @param new_negatives_weight New weight for the negatives.

void WeightedSquaredError::set_negatives_weight(const type& new_negatives_weight)
{
    negatives_weight = new_negatives_weight;
}


/// Set new weights for the positives and negatives values.
/// @param new_positives_weight New weight for the positives.
/// @param new_negatives_weight New weight for the negatives.

void WeightedSquaredError::set_weights(const type& new_positives_weight, const type& new_negatives_weight)
{
    positives_weight = new_positives_weight;
    negatives_weight = new_negatives_weight;
}


/// Calculates of the weights for the positives and negatives values with the data of the data set.

void WeightedSquaredError::set_weights()
{
    // Control sentence

#ifdef __OPENNN_DEBUG__

//    check();

#endif

    if(data_set_pointer->get_target_variables_number() == 0)
    {
        positives_weight = 1.0;
        negatives_weight = 1.0;
    }
    else if(data_set_pointer && data_set_pointer->get_target_columns()(0).type == DataSet::Binary)
    {
        const Tensor<Index, 1> target_distribution = data_set_pointer->calculate_target_distribution();

        const Index negatives = target_distribution[0];
        const Index positives = target_distribution[1];

        if(positives == 0 || negatives == 0)
        {
            positives_weight = 1.0;
            negatives_weight = 1.0;

            return;
        }

        negatives_weight = 1.0;
        positives_weight = static_cast<type>(negatives)/static_cast<type>(positives);
    }
    else
    {
        positives_weight = 1.0;
        negatives_weight = 1.0;
    }

}


/// Calculates of the normalization coefficient with the data of the data set.

void WeightedSquaredError::set_normalization_coefficient()
{
    // Control sentence

#ifdef __OPENNN_DEBUG__

    check();

#endif

    if(data_set_pointer->get_target_columns().size()==0)
    {
        normalization_coefficient = static_cast<type>(1);
    }
    else if(data_set_pointer && data_set_pointer->get_target_columns()(0).type == DataSet::Binary)
    {
        const Tensor<Index, 1> target_variables_indices = data_set_pointer->get_target_variables_indices();

        const Index negatives = data_set_pointer->calculate_used_negatives(target_variables_indices[0]);

        normalization_coefficient = negatives*negatives_weight*static_cast<type>(0.5);
    }
    else
    {
        normalization_coefficient = static_cast<type>(1);
    }
}


///
/// \brief set_data_set_pointer
/// \param new_data_set_pointer

void WeightedSquaredError::set_data_set_pointer(DataSet* new_data_set_pointer)
{
    data_set_pointer = new_data_set_pointer;

    set_weights();

    set_normalization_coefficient();
}


type WeightedSquaredError::weighted_sum_squared_error(const Tensor<type, 2> & x, const Tensor<type, 2> & y) const
{
#ifdef __OPENNN_DEBUG__

    const Index rows_number = x.dimension(0);
    const Index columns_number = x.dimension(1);

    const Index other_rows_number = y.dimension(0);

    if(other_rows_number != rows_number)
    {
        ostringstream buffer;

        buffer << "OpenNN Exception: Metrics functions.\n"
               << "double minkowski_error(const Matrix<double>&, const double&) method.\n"
               << "Other number of rows must be equal to this number of rows.\n";

        throw logic_error(buffer.str());
    }

    const Index other_columns_number = y.dimension(1);

    if(other_columns_number != columns_number)
    {
        ostringstream buffer;

        buffer << "OpenNN Exception: Metrics functions.\n"
               << "double minkowski_error(const Matrix<double>&, const double&) method.\n"
               << "Other number of columns must be equal to this number of columns.\n";

        throw logic_error(buffer.str());
    }

#endif

    const Tensor<bool, 2> if_sentence = y == y.constant(1);
    const Tensor<bool, 2> else_sentence = y == y.constant(0);

    Tensor<type, 2> f_1(x.dimension(0), x.dimension(1));

    Tensor<type, 2> f_2(x.dimension(0), x.dimension(1));

    Tensor<type, 2> f_3(x.dimension(0), x.dimension(1));

    f_1 = (x - y).square()*positives_weight;

    f_2 = (x - y).square()*negatives_weight;

    f_3 = x.constant(0);

    Tensor<type, 0> weighted_sum_squared_error = (if_sentence.select(f_1, else_sentence.select(f_2, f_3))).sum();

    return weighted_sum_squared_error(0);
}


void WeightedSquaredError::calculate_error(const DataSet::Batch& batch,
                     const NeuralNetwork::ForwardPropagation& forward_propagation,
                     LossIndex::BackPropagation& back_propagation) const
{
    const Index trainable_layers_number = neural_network_pointer->get_trainable_layers_number();

    const type error = weighted_sum_squared_error(forward_propagation.layers[trainable_layers_number-1].activations_2d,
                                                                 batch.targets_2d);

    const Index batch_samples_number = batch.samples_number;
    const Index total_samples_number = data_set_pointer->get_samples_number();

    back_propagation.error = error/((static_cast<type>(batch_samples_number)/static_cast<type>(total_samples_number))*normalization_coefficient);

    return;
}


void WeightedSquaredError::calculate_error_terms(const DataSet::Batch& batch,
                                                 const NeuralNetwork::ForwardPropagation& forward_propagation,
                                                 SecondOrderLoss& second_order_loss) const
{
    const Index trainable_layers_number = neural_network_pointer->get_trainable_layers_number();

    const Index batch_samples_number = batch.get_samples_number();
    const Index total_samples_number = data_set_pointer->get_samples_number();

    const Tensor<type, 2>& outputs = forward_propagation.layers(trainable_layers_number-1).activations_2d;
    const Tensor<type, 2>& targets = batch.targets_2d;

    const Eigen::array<int, 1> rows_sum = {Eigen::array<int, 1>({1})};

    const Tensor<bool, 2> if_sentence = outputs == outputs.constant(1);

    Tensor<type, 2> f_1(outputs.dimension(0), outputs.dimension(1));

    Tensor<type, 2> f_2(outputs.dimension(0), outputs.dimension(1));

    f_1 = ((outputs - targets))*positives_weight;

    f_2 = ((outputs - targets))*negatives_weight;

    second_order_loss.error_terms = ((if_sentence.select(f_1, f_2)).sum(rows_sum).square()).sqrt();

    Tensor<type, 0> error;
    error.device(*thread_pool_device) = second_order_loss.error_terms.contract(second_order_loss.error_terms, AT_B);

    const type coefficient = ((static_cast<type>(batch_samples_number)/static_cast<type>(total_samples_number))*normalization_coefficient);

    second_order_loss.error = error()/coefficient;
}


// Gradient methods

void WeightedSquaredError::calculate_output_gradient(const DataSet::Batch& batch,
                               const NeuralNetwork::ForwardPropagation& forward_propagation,
                               BackPropagation& back_propagation) const
{
     #ifdef __OPENNN_DEBUG__

     check();

     #endif


     const Index trainable_layers_number = neural_network_pointer->get_trainable_layers_number();

     const Tensor<type, 2>& outputs = forward_propagation.layers(trainable_layers_number-1).activations_2d;
     const Tensor<type, 2>& targets = batch.targets_2d;

     const Index batch_samples_number = batch.targets_2d.size();
     const Index total_samples_number = data_set_pointer->get_samples_number();

     const type coefficient = static_cast<type>(2.0)/((static_cast<type>(batch_samples_number)/static_cast<type>(total_samples_number))*normalization_coefficient);

//     cout << "+ w " << positives_weight <<endl;
//     cout << "- w " << negatives_weight <<endl;
//     cout << "batch samples " << batch_samples_number <<endl;
//     cout << "total samples " << total_samples_number<<endl;
//     cout << "output " << outputs<<endl;
//     cout << "target " << targets<<endl;
//     cout << "norm  " << normalization_coefficient<<endl;
//     cout << "error " << back_propagation.error<<endl;


     const Tensor<bool, 2> if_sentence = targets == targets.constant(1);
     const Tensor<bool, 2> else_sentence = targets == targets.constant(0);

     Tensor<type, 2> f_1(outputs.dimension(0), outputs.dimension(1));

     Tensor<type, 2> f_2(outputs.dimension(0), outputs.dimension(1));

     Tensor<type, 2> f_3(outputs.dimension(0), outputs.dimension(1));

     f_1 = coefficient*(outputs - targets)*positives_weight;

     f_2 = coefficient*(outputs - targets)*negatives_weight;

     f_3 = outputs.constant(0);

//     cout << f_2;
//     back_propagation.output_gradient = (if_sentence.select(f_1, else_sentence.select(f_2, f_3)));
     back_propagation.output_gradient.device(*thread_pool_device) = (if_sentence.select(f_1, else_sentence.select(f_2, f_3)));

//     cout<<back_propagation.output_gradient<<endl;

//     system("pause");

}


void WeightedSquaredError::calculate_Jacobian_gradient(const DataSet::Batch& batch,
                                    LossIndex::SecondOrderLoss& second_order_loss) const
{
#ifdef __OPENNN_DEBUG__

    check();

#endif

    const Index batch_samples_number = batch.get_samples_number();
    const Index total_samples_number = data_set_pointer->get_samples_number();

    const type coefficient = 2/((static_cast<type>(batch_samples_number)/static_cast<type>(total_samples_number))*normalization_coefficient);

    second_order_loss.gradient.device(*thread_pool_device) = second_order_loss.error_terms_Jacobian.contract(second_order_loss.error_terms, AT_B);

    second_order_loss.gradient.device(*thread_pool_device) = coefficient*second_order_loss.gradient;
}

// Hessian method

void WeightedSquaredError::calculate_hessian_approximation(const DataSet::Batch& batch,
                                                           LossIndex::SecondOrderLoss& second_order_loss) const
{
#ifdef __OPENNN_DEBUG__

    check();

#endif

    const Index batch_samples_number = batch.get_samples_number();
    const Index total_samples_number = data_set_pointer->get_samples_number();

    const type coefficient = 2/((static_cast<type>(batch_samples_number)/static_cast<type>(total_samples_number))*normalization_coefficient);

    second_order_loss.hessian.device(*thread_pool_device) = second_order_loss.error_terms_Jacobian.contract(second_order_loss.error_terms_Jacobian, AT_B);

    second_order_loss.hessian.device(*thread_pool_device) = coefficient*second_order_loss.hessian;
}


/// Returns a string with the name of the weighted squared error loss type, "WEIGHTED_SQUARED_ERROR".

string WeightedSquaredError::get_error_type() const
{
    return "WEIGHTED_SQUARED_ERROR";
}


/// Returns a string with the name of the weighted squared error loss type in text format.

string WeightedSquaredError::get_error_type_text() const
{
    return "Weighted squared error";
}


/// Serializes the cross entropy error object into a XML document of the TinyXML library without keep the DOM tree in memory.
/// See the OpenNN manual for more information about the format of this document.
/// @param file_stream

void WeightedSquaredError::write_XML(tinyxml2::XMLPrinter& file_stream) const
{
    ostringstream buffer;

    // Error type

    file_stream.OpenElement("WeightedSquaredError");

    // Positives weight

    file_stream.OpenElement("PositivesWeight");

    buffer.str("");
    buffer << positives_weight;

    file_stream.PushText(buffer.str().c_str());

    file_stream.CloseElement();

    // Negatives weight

    file_stream.OpenElement("NegativesWeight");

    buffer.str("");
    buffer << negatives_weight;

    file_stream.PushText(buffer.str().c_str());

    file_stream.CloseElement();

    // Close error

    file_stream.CloseElement();
}


/// Loads a weighted squared error object from a XML document.
/// @param document Pointer to a TinyXML document with the object data.

void WeightedSquaredError::from_XML(const tinyxml2::XMLDocument& document)
{
    const tinyxml2::XMLElement* root_element = document.FirstChildElement("WeightedSquaredError");

    if(!root_element)
    {
        ostringstream buffer;

        buffer << "OpenNN Exception: WeightedSquaredError class.\n"
               << "void from_XML(const tinyxml2::XMLDocument&) method.\n"
               << "Weighted squared element is nullptr.\n";

        throw logic_error(buffer.str());
    }

    // Positives weight

    const tinyxml2::XMLElement* positives_weight_element = root_element->FirstChildElement("PositivesWeight");

    if(positives_weight_element)
    {
        const string string = positives_weight_element->GetText();

        try
        {
            set_positives_weight(static_cast<type>(atof(string.c_str())));
        }
        catch(const logic_error& e)
        {
            cerr << e.what() << endl;
        }
    }

    // Negatives weight

    const tinyxml2::XMLElement* negatives_weight_element = root_element->FirstChildElement("NegativesWeight");

    if(negatives_weight_element)
    {
        const string string = negatives_weight_element->GetText();

        try
        {
            set_negatives_weight(static_cast<type>(atof(string.c_str())));
        }
        catch(const logic_error& e)
        {
            cerr << e.what() << endl;
        }
    }
}

}

// OpenNN: Open Neural Networks Library.
// Copyright(C) 2005-2020 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