[딥러닝] C++로 역전파(Back-propagation) 구현하기

출처: https://www.youtube.com/watch?v=XmK9f5IV8Uw#t=568.604741

#include <stdio.h>

#include <iostream>

#define MAX2(a,b) (a) > (b) ? (a) : (b)

class Neuron

{

public:

Neuron()

: w_(2.0),b_(1.0)

{}

Neuron(const double& w_input, const double& b_input)

: w_(w_input),b_(b_input)

{}

public: // attributes

double w_; // weight of one input

double b_; // bias

double input_, output_; //saved for back-prop

public: // behaviors

double getAct(const double& x)

{

// for linear or identity activation functions

return x;

// for ReLU activation functions

//return MAX2(0.0, x);

}

double feedForward(const double& _input)

{

input_ = _input;

// output y = f(\sigma) , 

// \sigma = w_ * input x + b

// for multiple inputs,

// \sigma = w0_ * x0_ + w1_ * x1_ + ... + b

const double sigma = w_ * input_ + b_;

output_ = getAct(sigma);

return output_;

}

double getActGrad(const double& x)

{

return 1.0;

}

void propBackward(const double& target)

{

const double alpha = 0.1; // learning rate

const double grad = (output_ - target) * getActGrad(output_);

w_ -= alpha * grad * input_; // last input_ came from d(wx+b)/dw = x

b_ -= alpha * grad * 1.0; // last 1.0 came from d(wx+b)/db = 1

}

void feedForwardAndPrint(const double& input)

{

printf("%f %f \n", input, feedForward(input));

}

};

void main()

{

// initialize my_neuron

Neuron my_neuron(2.0, 1.0);

for (int r = 0; r < 100; r++)

{

std::cout << "Training :: " << r << std::endl;

my_neuron.feedForwardAndPrint(1.0);

my_neuron.propBackward(4.0);

my_neuron.feedForwardAndPrint(1.0);

std::cout << "w =  " << my_neuron.w_ << " b =  " << my_neuron.b_ << std::endl;

}

}