gctl_ai/lib/dnn/olayer_multientropy.cpp

/********************************************************
 *  ██████╗  ██████╗████████╗██╗
 * ██╔════╝ ██╔════╝╚══██╔══╝██║
 * ██║  ███╗██║        ██║   ██║
 * ██║   ██║██║        ██║   ██║
 * ╚██████╔╝╚██████╗   ██║   ███████╗
 *  ╚═════╝  ╚═════╝   ╚═╝   ╚══════╝
 * Geophysical Computational Tools & Library (GCTL)
 *
 * Copyright (c) 2022  Yi Zhang (yizhang-geo@zju.edu.cn)
 *
 * GCTL is distributed under a dual licensing scheme. 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 
 * of the License, or (at your option) any later version. You should have 
 * received a copy of the GNU Lesser General Public License along with this 
 * program. If not, see <http://www.gnu.org/licenses/>.
 * 
 * If the terms and conditions of the LGPL v.2. would prevent you from using 
 * the GCTL, please consider the option to obtain a commercial license for a 
 * fee. These licenses are offered by the GCTL's original author. As a rule, 
 * licenses are provided "as-is", unlimited in time for a one time fee. Please 
 * send corresponding requests to: yizhang-geo@zju.edu.cn. Please do not forget 
 * to include some description of your company and the realm of its activities. 
 * Also add information on how to contact you by electronic and paper mail.
 ******************************************************/

#include "olayer_multientropy.h"

gctl::multi_entropy::multi_entropy() {}

gctl::multi_entropy::~multi_entropy() {}

void gctl::multi_entropy::check_target_data(const matrix<double> &target)
{
    int nobs = target.col_size();
    int ncls = target.row_size();

    int one;
    for (size_t i = 0; i < nobs; i++)
    {
        one = 0;
        for (size_t j = 0; j < ncls; j++)
        {
            if (fabs(target[j][i] - 1.0) < 1e-6)
            {
                one++;
                continue;
            }
            
            if (fabs(target[j][i]) > 1e-6)
            {
                throw std::invalid_argument("[gctl::multi_entropy] Target data should only contain zero or one.");
            }
        }
        
        if (one != 1)
        {
            throw std::invalid_argument("[gctl::multi_entropy] Each column of target data should only contain one \"1\".");
        }
    }
    return;
}

void gctl::multi_entropy::check_target_data(const array<int> &target)
{
    int nobs = target.size();
    for (size_t i = 0; i < nobs; i++)
    {
        if (target[i] < 0)
        {
            throw std::invalid_argument("[gctl::multi_entropy] Target data must be non-negative.");
        }
    }
    return;
}

void gctl::multi_entropy::evaluation(const matrix<double> &prev_layer_data, const matrix<double> &target)
{
    int nobs = prev_layer_data.col_size();
    int ncls = prev_layer_data.row_size();

    if (target.col_size() != nobs || target.row_size() != ncls)
    {
        throw std::invalid_argument("[gctl::multi_entropy] Target data have incorrect dimension.");
    }

    // Compute the derivative of the input of this layer
    // L = -sum(log(phat) * y)
    // in = phat
    // d(L) / d(in) = -y / phat
    der_in_.resize(ncls, nobs, 0.0);

    int i, j;
#pragma omp parallel for private (i, j) schedule(guided)
    for (i = 0; i < ncls; i++)
    {
        for (j = 0; j < nobs; j++)
        {
            der_in_[i][j] = -1.0*target[i][j]/prev_layer_data[i][j];
        }
    }

    // L = -sum(log(phat) * y)
    // in = phat
    // d(L) / d(in) = -y / phat
    // m_din contains 0 if y = 0, and -1/phat if y = 1
    loss_ = 0.0;
//#pragma omp parallel for private (i, j) schedule(guided)
    for (i = 0; i < ncls; i++)
    {
        for (j = 0; j < nobs; j++)
        {
            loss_ -= std::log(prev_layer_data[i][j])*target[i][j];
        }
    }

    loss_ /= der_in_.col_size();
    return;
}

void gctl::multi_entropy::evaluation(const matrix<double> &prev_layer_data, const array<int> &target)
{
    // target is a vector of class labels that take values from [0, 1, ..., nclass - 1]
    // The i-th element of target is the class label for observation i
    int nobs = prev_layer_data.col_size();
    int ncls = prev_layer_data.row_size();

    if (target.size() != nobs)
    {
        throw std::invalid_argument("[gctl::multi_entropy] Target data have incorrect dimension.");
    }

    // Compute the derivative of the input of this layer
    // L = -log(phat[y])
    // in = phat
    // d(L) / d(in) = [0, 0, ..., -1/phat[y], 0, ..., 0]
    der_in_.resize(ncls, nobs);
    der_in_.assign_all(0.0);

    int i;
#pragma omp parallel for private (i) schedule(guided)
    for (i = 0; i < nobs; i++)
    {
        der_in_[target[i]][i] = -1.0/prev_layer_data[target[i]][i];
    }
    return;
}

double gctl::multi_entropy::loss_value() const
{
    return loss_;
}

std::string gctl::multi_entropy::get_output_name() const
{
    return "MultiClassEntropy";
}

gctl::olayer_type_e gctl::multi_entropy::get_output_type() const
{
    return MultiClassEntropy;
}