GradientDescent.cpp

/**
 * DESCRIPTION OF THE FILE
 *
 * @author Michal Kravčenko
 * @date 30.7.18 - 
 */

#include "GradientDescent.h"
#include "message.h"

namespace lib4neuro {
    GradientDescent::GradientDescent(double epsilon, size_t n_to_restart, int max_iters, size_t batch) {
        this->tolerance = epsilon;
        this->restart_frequency = n_to_restart;
        this->optimal_parameters = new std::vector<double>(0);
        this->maximum_niters = max_iters;
        this->batch = batch;
    }

    GradientDescent::~GradientDescent() {
        if (this->optimal_parameters) {
            delete this->optimal_parameters;
            this->optimal_parameters = nullptr;
        }
    }

    void GradientDescent::eval_step_size_mk(double &gamma,
                                            double beta,
                                            double &c,
                                            double grad_norm_prev,
                                            double grad_norm,
                                            double fi,
                                            double fim) {

        if (fi > fim) {
            c /= 1.0000005;
        } else if (fi < fim) {
            c *= 1.0000005;
        }

        gamma *= std::pow(c, 1.0 - 2.0 * beta) * std::pow(grad_norm_prev / grad_norm, 1.0 / c);

    }

    void GradientDescent::optimize(lib4neuro::ErrorFunction &ef, std::ofstream* ofs) {

        /* Copy data set max and min values, if it's normalized */
        if(ef.get_dataset()->is_normalized()) {
            ef.get_network_instance()->set_normalization_strategy_instance(
                    ef.get_dataset()->get_normalization_strategy());
        }

        COUT_INFO("Finding a solution via a Gradient Descent method with adaptive step-length..." << std::endl);

        if(ofs && ofs->is_open()) {
            *ofs << "Finding a solution via a Gradient Descent method with adaptive step-length..." << std::endl;
        }

        double grad_norm = this->tolerance * 10.0, gamma, sx, beta;
        double grad_norm_prev;
        size_t i;
        long long int iter_idx = this->maximum_niters;
        size_t iter_counter = 0;

        gamma = 1.0;
        double prev_val, val = 0.0, c = 1.25;

        size_t n_parameters = ef.get_dimension();


        std::vector<double> *gradient_current = new std::vector<double>(n_parameters);
        std::vector<double> *gradient_prev = new std::vector<double>(n_parameters);
        std::vector<double> *params_current = ef.get_parameters();
        std::vector<double> *params_prev = new std::vector<double>(n_parameters);
        std::vector<double> *ptr_mem;

//    std::vector<double> gradient_mem( n_parameters );
//    std::vector<double> parameters_analytical( n_parameters );


        std::fill(gradient_current->begin(), gradient_current->end(), 0.0);
        std::fill(gradient_prev->begin(), gradient_prev->end(), 0.0);
        while (grad_norm > this->tolerance && (iter_idx != 0)) {
            iter_idx--;
            iter_counter++;
            prev_val = val;
            grad_norm_prev = grad_norm;

            /* reset of the current gradient */
            std::fill(gradient_current->begin(), gradient_current->end(), 0.0);
//        std::fill(gradient_mem.begin(), gradient_mem.end(), 0.0);
            ef.calculate_error_gradient(*params_current, *gradient_current);
//        double error_analytical = this->calculate_gradient( ef.get_dataset()->get_data(), (size_t)2, params_current, gradient_current );

//        for(size_t k = 0; k < gradient_mem.size(); ++k){
//            printf("%f : %f\n", gradient_mem[ k ], gradient_current->at( k ));
//        }
//        printf("---------------------\n");

            grad_norm = 0.0;
            for (auto v: *gradient_current) {
                grad_norm += v * v;
            }
            grad_norm = std::sqrt(grad_norm);

            /* Update of the parameters */
            /* step length calculation */
            if (iter_counter < 10 || iter_counter % this->restart_frequency == 0) {
                /* fixed step length */
                gamma = 0.1 * this->tolerance;
            } else {
                /* angle between two consecutive gradients */
                sx = 0.0;
                for (i = 0; i < gradient_current->size(); ++i) {
                    sx += (gradient_current->at(i) * gradient_prev->at(i));
                }
                sx /= grad_norm * grad_norm_prev;
                beta = std::sqrt(std::acos(sx) / lib4neuro::PI);

                eval_step_size_mk(gamma, beta, c, grad_norm_prev, grad_norm, val, prev_val);
            }

            for (i = 0; i < gradient_current->size(); ++i) {
                (*params_prev)[i] = (*params_current)[i] - gamma * (*gradient_current)[i];
            }

            /* switcheroo */
            ptr_mem = gradient_prev;
            gradient_prev = gradient_current;
            gradient_current = ptr_mem;

            ptr_mem = params_prev;
            params_prev = params_current;
            params_current = ptr_mem;

            val = ef.eval(params_current);

            COUT_DEBUG(std::string("Iteration: ") << (unsigned int)(iter_counter)
                       << ". Step size: " << gamma
                       << ". C: " << c
                       << ". Gradient norm: " << grad_norm
                       << ". Total error: " << val
                       << "." << std::endl);

            WRITE_TO_OFS_DEBUG(ofs, "Iteration: " << (unsigned int)(iter_counter)
                                    << ". Step size: " << gamma
                                    << ". C: " << c
                                    << ". Gradient norm: " << grad_norm
                                    << ". Total error: " << val
                                    << "." << std::endl);


        }

        if(iter_idx == 0) {
            COUT_INFO("Maximum number of iterations (" << this->maximum_niters << ") was reached!" << std::endl);

            if(ofs && ofs->is_open()) {
                *ofs << "Maximum number of iterations (" << this->maximum_niters << ") was reached!" << std::endl;

            }

        } else {
            COUT_INFO("Gradient Descent method converged after "
                      << this->maximum_niters-iter_idx
                      << "iterations."
                      << std::endl);
#ifdef L4N_DEBUG
            if(ofs && ofs->is_open()) {
                *ofs << "Gradient Descent method converged after "
                      << this->maximum_niters-iter_idx
                      << "iterations."
                      << std::endl;
            }
#endif
        }

        *this->optimal_parameters = *params_current;

        delete gradient_current;
        delete gradient_prev;
        delete params_current;
        delete params_prev;
    }

    std::vector<double> *GradientDescent::get_parameters() {
        return this->optimal_parameters;
    }

}