gctl_ai/examples/ex_mnist2.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 "../data/MNIST/mnist_database.h"
#include "../lib/dnn.h"

using namespace gctl;

int main(int argc, char const *argv[]) try
{
    mnist_database data("data/MNIST");

    matrix<double> train_obs(784, 60000), train_lab(10, 60000, 0.0);
    matrix<double> test_obs(784, 10000), test_lab(10, 10000, 0.0), predicts(10, 10000);

    const std::vector<std::vector<double> > &dt_obs = data.train_images();
    for (size_t i = 0; i < 60000; i++)
    {
        for (size_t j = 0; j < 784; j++)
        {
            train_obs[j][i] = dt_obs[i][j]/255.0;
        }
    }

    const std::vector<std::vector<double> > &dt_obs2 = data.test_images();
    for (size_t i = 0; i < 10000; i++)
    {
        for (size_t j = 0; j < 784; j++)
        {
            test_obs[j][i] = dt_obs2[i][j]/255.0;
        }
    }

    const std::vector<double> &dt_lab = data.train_labels();
    for (size_t i = 0; i < 60000; i++)
    {
        train_lab[dt_lab[i]][i] = 1.0;
    }

    const std::vector<double> &dt_lab2 = data.test_labels();
    for (size_t i = 0; i < 10000; i++)
    {
        test_lab[dt_lab2[i]][i] = 1.0;
    }

    dnn my_nn("Ex-MNIST");
    my_nn.add_hind_layer(1, 28, 28, 4, 4, 2, 2, Convolution, Same, PReLU);
    my_nn.add_hind_layer(169, 256, FullyConnected, PReLU);
    my_nn.add_hind_layer(256, 10, FullyConnected, SoftMax);
    my_nn.add_output_layer(MultiClassEntropy);
    my_nn.add_train_set(train_obs, train_lab, 5000);
    my_nn.init_network(0.0, 0.1);

    sgd_para my_para = my_nn.default_sgd_para();
    my_para.alpha = 0.01;
    my_para.epsilon = 1e-5;
    my_nn.train_network(my_para, gctl::ADAM);
    //lgd_para my_para = my_nn.default_lgd_para();
    //my_para.flight_times = 2000;
    //my_para.alpha = 0.1;
    //my_nn.train_network(my_para);

    my_nn.predict(test_obs, predicts);
    my_nn.save_network("data/saved_networks/mnist_m2");

    int wrong_predicts = 0;
    int test_id, pre_id;
    double test_scr, pre_scr;
    for (size_t j = 0; j < 10000; j++)
    {
        test_id = pre_id = 0;
        test_scr = pre_scr = 0;
        for (size_t i = 0; i < 10; i++)
        {
            if (test_lab[i][j] > test_scr) {test_scr = test_lab[i][j]; test_id = i;}
            if (predicts[i][j] > pre_scr) {pre_scr = predicts[i][j]; pre_id = i;}
        }

        if (test_id != pre_id)
        {
            wrong_predicts++;   
        }
    }

    std::cout << "Correct Rate = " << (10000 - wrong_predicts)/100.0 << "%\n";

    return 0;
}
catch (std::exception &e)
{
    GCTL_ShowWhatError(e.what(), GCTL_ERROR_ERROR, 0, 0, 0);
}