// Ann.cpp : Defines the entry point for the console application.
//

#include "stdafx.h"
# include <cmath>
#include <math.h>
#include <vector>
#include <iostream>
#include <iomanip>
#include <fstream>
#include <string>
using namespace std;
//Generating random number: either 0 or 1, uniform distribution, for XOR operation. Can remove later if using data from files.
int randint();
struct Sample
{
    double input[2];
    double output[2];
    string ToString() {
        string str;
        str = "input: " + to_string(input[0]) + " " + to_string(input[1]) + " output: " + to_string(output[0]) + " " + to_string(output[1])+"\n";
        return str;
    }
};

class Data
{
public:
    int getNumberOfInputs() { return inputs; }
    int getNumberOfOutputs() { return outputs; }

    double *getInput(int index)
    {
        double *input = data[index].input;
        return input;
    }

    double *getOutput(int index)
    {
        double *output = data[index].output;
        return output;
    }

    int getNumberOfSamples() { return samples; }

    void addSample(Sample sample)
    {
        data.push_back(sample);
        samples++;
        //cout << sample.ToString();
    }

    void setSizes(int input_size, int output_size)
    {
        inputs = input_size;
        outputs = output_size;
    }

protected:
    std::vector<Sample> data;
    int inputs;
    int outputs;
    int samples = 0;
};

class XOR : public Data
{
public:
    void generate(int n)
    {
        for (int i = 0; i < n; i++)
        {
            double input1 = randint();
            double input2 = randint();
            double output1 = input1 == input2;
            double output2 = input1 != input2;
            addSample({ { input1, input2 },{ output1, output2 } });
        }
    }

    XOR()
    {
        inputs = 2;
        outputs = 2;
        samples = 0;
    }
    void printInputs(int index)
    {
        cout << index << " index inputs: " << endl;
        for (int i = 0; i < inputs; i++)
        {
            cout << getInput(index)[i] << " ";
        }
        cout << endl;
    }

    void printOutputs(int index)
    {
        cout << index << " index outputs: " << endl;
        for (int i = 0; i < outputs; i++)
        {
            cout << fixed << setprecision(2) << data[index].output[i] << " ";
        }
        cout << endl;
    }
};

struct Topology
{
    int L;//sluoksniu sk
    std::vector<int> l;//kiekiai sluoksnyje
} topolygy;

int kiek = 0;//for statistics - ignore
double f(double x);
double f_deriv(double x);
double gL(double a, double z, double t);
double gl(int layer_id, int w_i, int w_j, double *a_arr, double *z_arr, double *t_arr, double *w_arr, int *s,int *sw, int L, int *l);
double w_gradient(int layer_id, int w_i, int w_j, double *a_arr, double *z_arr, double *t_arr, double *w_arr,int *s, int *sw, int L, int *l);
double delta_w(double grad, double dw);

const double ETA = 0.1;
const double ALPHA = 0.5;

class AnnBase {
public:
    virtual void prepare(Topology top) = 0;
    virtual void init(Topology top, double w_arr_1[]) = 0;
    virtual void train(double *a, double *b) = 0;
    virtual void feedForward(double *a, double *b) = 0;//a and b temp vars for now.
    virtual void destroy() = 0;
};

class AnnSerialDBL : public AnnBase {
public:
    void prepare(Topology top);
    void init(Topology top, double w_arr_1[]);
    void train(double *a, double *b);
    void feedForward(double *a,double *b);
    void destroy();
public:
    int sum1;//temp var to keep the length of z, so z could be reset for calcs. Can adapt code later, so this could be removed
    int L;
    int * l;
    int * s;
    double * a_arr;
    double * z_arr;
    int * W;
    int * sw;
    double * w_arr;
    double * dw_arr;
    double * t_arr;
};
int main()
{
    topolygy.L = 5;
    topolygy.l.push_back(2);
    topolygy.l.push_back(10);
    topolygy.l.push_back(10);
    topolygy.l.push_back(12);
    topolygy.l.push_back(2);
    AnnSerialDBL labas;
    labas.prepare(topolygy);
    labas.init(topolygy, NULL);

    for (int i = 0; i < topolygy.L; i++) {
    //  cout << ": " << labas.W[i];
    }
    XOR xo;
    xo.generate(10);
    for (int i = 0; i < xo.getNumberOfSamples(); i++) {
        labas.train(xo.getInput(i), xo.getOutput(i));
    }
    int sum = 0;
    for (int i = 0; i < labas.L; i++) {
        sum += topolygy.l.at(i) + 1;
    }
    for (int i = 0; i < sum; i++) {
    //  cout << labas.a_arr[i] << endl;
    }

    //Checking results
    for (int i = 0; i < 5;i++){
        double input[] = { randint()*1.0, randint()*1.0 };
        double output[] = { 0,0 };
        //cout << endl;
        labas.feedForward(input, output);

        cout << endl << "1: " << input[0] << " 2: " << input[1] << endl;
        cout << endl << "1: " << output[0] << " 2: " << output[1] << endl<<endl;

    }

    labas.destroy();
    int a;
    cout << kiek;
    cin >> a;

    return 0;
}

//returns random int, either 0 or 1
int randint() {
    double r = ((double)rand() / (RAND_MAX));
    int a = 0;
    if (r > 0.5) {
        a = 1;
    }
    else
    {
        a = 0;
    }
    return a;
}

void AnnSerialDBL::prepare(Topology top)
{
    l = new int[top.L];
    s = new int[top.L];

    int sum = 0;
    int mult = 0;
    for (int i = 0; i < top.L; i++) {
        sum += top.l.at(i) + 1;
    }
    sum1 = sum;
    for (int i = 0; i < top.L - 1; i++) {
        mult += (top.l.at(i) + 1)*top.l.at(i+1);
    }
    a_arr = new double[sum];
    z_arr = new double[sum];

    W = new int[top.L];
    sw = new int[top.L];

    w_arr = new double[mult];
    dw_arr = new double[mult];

    t_arr = new double[top.l.at(top.L - 1)];
}

void AnnSerialDBL::init(Topology top, double w_arr_1[] = NULL)
{
    L = top.L;
    //Neuronu kiekiai sluoksnyje
    for (int i = 0; i < top.L; i++) {
        l[i] = top.l.at(i) + 1;
    }

    //Sluoksniu pradzios indeksai
    for (int i = 0; i < top.L; i++) {
        s[i] = 0;
        for (int j = i; j > 0; j--) {
            s[i] += l[j - 1];
        }
    }

    //Bias neuronai
    for (int i = 0; i < top.L - 1; i++) {
        a_arr[s[i + 1] - 1] = 1;
    }

    //Svoriu kiekiai l-ame sluoksnyje
    for (int i = 0; i < top.L - 1; i++) {
        W[i] = l[i] * (l[i + 1] - 1);
        //cout << "Svoriu sk: " << W[i] << " Pradzios index: ";
        sw[i] = 0;
        if (i != 0) {
            for (int j = 0; j < i; j++) {
                sw[i] += W[j];
            }
        }
        if (w_arr_1 == NULL) {
            for (int j = 0; j < W[i]; j++) {
                w_arr[sw[i] + j] = (double)rand() / double(RAND_MAX);
                //cout << w_arr[sw[i] + j]<< endl;
                dw_arr[sw[i] + j] = 0;
            }
        }
        else {
            w_arr = w_arr_1; //ar reikia pokycius issisaugoti irgi?
        }

        //cout << sw[i] << " " << endl;
    }
}

void AnnSerialDBL::train(double *a,double *b)
{
    a_arr[0] = a[0];
    a_arr[1] = a[1];

    for (int j = 0; j < sum1; j++) {
        z_arr[j] = 0;
    }

    //FFEEED FORWARD
    for (int i = 0; i < L - 1; i++) {//per sluoksnius einu+
        for (int j = 0; j < l[i]; j++) { //kiek neuronu sluoksnyje+
            for (int k = 0; k < l[i + 1] - 1; k++) {//per sekancio sluoksnio z+
                z_arr[s[i + 1] + k] += w_arr[sw[i] + k + j*l[i + 1]] * a_arr[s[i] + j];
            //  cout << "w: "<< w_arr[sw[i] + k + j*l[i + 1]] << endl;
            //  cout << "a: " << a_arr[s[i] + j] << endl;
            //  cout << "z reiksmes: " << z_arr[s[i+1] + k] << endl;
                //cout << endl;
                kiek++;
            }
        }
        for (int k = 0; k < l[i + 1] - 1; k++) {//per sekancio sluoksnio z
            a_arr[s[i + 1] + k] = f(z_arr[s[i + 1] + k]);
            //cout << s[i + 1] + k << " a reiksmes: " << a_arr[s[i + 1] + k] << endl;
        }
    }

    t_arr[0] = b[0];
    t_arr[1] = b[1];

    //back propogation:
    for (int i = L-2; i >=0; i--) {//per sluoksnius
        for (int j = 0; j < l[i]; j++) {//per neuronus
            for (int k = 0; k < l[i + 1] - 1; k++) {//per kito sluoksnio neuronus
                dw_arr[sw[i] + k] = delta_w(w_gradient(i, j, k, a_arr, z_arr, t_arr, w_arr, s,sw, L, l), dw_arr[sw[i] + k]);
                w_arr[sw[i] + k] += dw_arr[sw[i] + k];

                //w_arr[sw[i] + k] = delta_w(w_gradient(i, j, k, a_arr, z_arr, t_arr, w_arr, s, sw, L, l), w_arr[sw[i] + k]);
                //dw_arr[sw[i] + k] += dw_arr[sw[i] + k];

            //  cout << w_arr[sw[i] + k] << " " << endl;
            }
        }
    }

}

void AnnSerialDBL::feedForward(double *a, double *b)
{
    a_arr[0] = a[0];
    a_arr[1] = a[1];
    for (int j = 0; j < sum1; j++) {
        z_arr[j] = 0;
    }
    //FFEEED FORWARD
    for (int i = 0; i < L - 1; i++) {//per sluoksnius einu+
        for (int j = 0; j < l[i]; j++) { //kiek neuronu sluoksnyje+
            for (int k = 0; k < l[i + 1] - 1; k++) {//per sekancio sluoksnio z+
                z_arr[s[i + 1] + k] += w_arr[sw[i] + k + j*l[i + 1]] * a_arr[s[i] + j];
                /*  cout << "w: "<< w_arr[sw[i] + k + j*l[i + 1]] << endl;
                cout << "a: " << a_arr[s[i] + j] << endl;
                cout << "z reiksmes: " << z_arr[s[i+1] + k] << endl;
                cout << endl;*/
            }
        }
        for (int k = 0; k < l[i + 1] - 1; k++) {//per sekancio sluoksnio z
            a_arr[s[i + 1] + k] = f(z_arr[s[i + 1] + k]);
            cout << s[i + 1] + k << " a reiksmes: " << a_arr[s[i + 1] + k] << endl;
        }
    }
    //cout << a_arr[s[L - 1]];
    //cout << a_arr[s[L - 1]+1];
    if (a_arr[s[L - 1]] < a_arr[s[L - 1] + 1]) {
        b[0] = 1;
        b[1] = 0;
    }
    else {
        b[0] = 0;
        b[1] = 1;
    }
    //data.output[0] = a_arr[s[L - 1]];
    //data.output[1] = a_arr[s[L - 1] + 1];
}

void AnnSerialDBL::destroy()
{
    delete l;
    l = NULL;
    delete s;
    s = NULL;

    delete a_arr;
    a_arr = NULL;
    delete z_arr;
    z_arr = NULL;

    delete W;
    W = NULL;
    delete sw;
    sw = NULL;

    delete w_arr;
    w_arr = NULL;
    delete dw_arr;
    dw_arr = NULL;

    delete t_arr;
    t_arr = NULL;
}

double f(double x) {
    double y = 1+exp(-x);
    if (y == 0) {
        cout << "Error 1";
    }
        if ((y-1) == 0) {
            //cout << "Error 2";
        }
    return 1 / y;
}


double f_deriv(double x) {
    double y = pow((1 + exp(-x)), 2);
    double z = exp(-x);
    if (y == 0) {
        cout << "Error 3";
    }
    if (z == 0) {
        //cout << "Error 4";
    }
    return exp(-x) / pow((1 + exp(-x)), 2);
}

double gL(double a, double z, double t) {
    double w = f_deriv(z) * (a - t);
    //cout << "z: " << z << " a: " << a << " t: " << t << endl;
    return w;
}

double gl(int layer_id, int w_i, int w_j, double *a_arr, double *z_arr, double *t_arr, double *w_arr, int *s,int *sw, int L, int *l) {
    double w = f_deriv(z_arr[s[layer_id] + w_j]);
    double sum = 0;
    for (int i = 0; i < l[layer_id + 1] - 1; i++) {
        if (layer_id + 2 == L) {
            sum += w_arr[sw[layer_id] + i] * gL(a_arr[s[layer_id + 1] + i], z_arr[s[layer_id + 1]+ i], t_arr[i]);
        }
        else {
            sum += w_arr[sw[layer_id] + w_j] * gl(layer_id + 1, w_i, i, a_arr, z_arr, t_arr,w_arr, s, sw, L, l);
        }
    }
    return w*sum;
}

double w_gradient(int layer_id, int w_i, int w_j, double *a_arr, double *z_arr, double *t_arr, double *w_arr,int *s, int *sw, int L, int *l) {
    double w = a_arr[s[layer_id] + w_i];
    if (layer_id + 2 == L) {
        w *= gL(a_arr[s[layer_id + 1] + w_j], z_arr[s[layer_id + 1] + w_j], t_arr[w_j]);
    }
    else {
        w *= gl(layer_id + 1, w_i, w_j, a_arr, z_arr, t_arr, w_arr, s, sw, L, l);
    }
    //cout << L << endl;
    //cout << w << " layer id:"<< layer_id <<endl;
    return w;
}

double delta_w(double grad, double dw) {
    return (-ETA)*grad + ALPHA*dw;
}