main.cpp

#include <iostream>
#include <fstream>
#include <Eigen/Dense>
#include <cstdlib>
#include "Layer.h"
#include "Trainset.h"
#include "misc.h"

#define PRINT(x) (cout<< #x << " = " << endl << x << endl)

using namespace Eigen;
using namespace std;

const int inputs = 2;
const int hidden = 10;
const int outputs = 1;

int main()
{

    Layer w1(inputs,hidden);
    Layer w2(hidden, outputs);

    VectorXd x(inputs);
    VectorXd y(hidden), dy(hidden), theta1(hidden);
    VectorXd z(outputs), dz(outputs), theta2(outputs), ans(outputs);

    Trainset train(2, 1, 100000);
    Trainset testcase(2,1,5000);
    
    int trainsize = train.getSize();
    for(int i = 0; i < trainsize; i++) {

        if (i % (trainsize / 100) == 0)
        {
            cout << "Training Process: " << (double) i / trainsize * 100 << '%' << " of " << trainsize << '\r';
            cout.flush();
        }

        x = train.getInput(i);

        theta1 = w1.calculate(x);
        y = sigmoid(theta1);

        theta2 = w2.calculate(y);
        z = sigmoid(theta2);

        ans = train.getOutput(i);
        
        dz = ans - z;
        
        MatrixXd dw1, dw2, diag_dtheta1, diag_dtheta2;
        VectorXd db1, db2, dtheta1, dtheta2;

        diag_dtheta1 = d_sigmoid(theta1).asDiagonal();
        diag_dtheta2 = d_sigmoid(theta2).asDiagonal();

        db2 = diag_dtheta2 * dz;
        dw2 = db2 * y.transpose();

        dy = w2.getMatrix().transpose() * diag_dtheta2 * dz;

        db1 = diag_dtheta1 * dy;
        dw1 = db1 * x.transpose();

        // PRINT(x);
        // PRINT(z);
        // PRINT(dz);
        double step = 0.1;

        w1.setMatrix(w1.getMatrix() + dw1 * step);
        w1.setOffset(w1.getOffset() + db1 * step);
        w2.setMatrix(w2.getMatrix() + dw2 * step);
        w2.setOffset(w2.getOffset() + db2 * step);

        y = sigmoid(w1.calculate(x));
        z = sigmoid(w2.calculate(y));

        // PRINT(z);
    }
    
    for (int i = 0; i < testcase.getSize(); i++)
    {
        x = testcase.getInput(i);
        y = sigmoid(w1.calculate(x));
        z = sigmoid(w2.calculate(y));
        testcase.setMyAns(z, i);
    }
    double accurate = testcase.calculatePrecision();
    cout << endl << endl << "----------" << endl;
    cout << "accurate: " << accurate * 100 << '%' << endl;
    cout << "----------" << endl << endl;

    fstream fout;
    fout.open("plot/dot.txt", ios::out);
    for (int i = 0; i < testcase.getSize(); i++)
    {
        x = testcase.getInput(i);
        z = testcase.getMyAns(i);
        for (int j = 0; j < x.rows(); j++)
        {
            fout << x[j] << ' ';
        }
        fout << z[0] << endl;
    }
    fout.close();
    // PRINT(w1.getMatrix());
    // PRINT(w1.getOffset());
    // PRINT(w2.getMatrix());
    // PRINT(w2.getOffset());
    return 0;
}