/// <summary>
/// Pomnoziti sve ulaze sa tezinama.
/// Sabrati sve rezultate mnozaca. Kao povratnu vrednost
/// vratiti vrednost aktivacione funkije za vrednost suma
/// svih mnozaca
/// </summary>
/// <param name="x">x is a vector of inputs</param>
/// <returns></returns>
public double forward(List <double> x)
{
List <double> inputs = new List <double>(x);
inputs.Add(1.0); //bias
List <double> forSum = new List <double>();
//TODO 6: Izracunati vrednost na izlazu vestackog neurona
for (int i = 0; i < inputs.Count; i++)
{
MultiplyNode mn = this.multiplyNodes[i];
double weight = mn.x[1];
List <double> newInputs = new List <double>();
newInputs.Add(inputs[i]);
newInputs.Add(weight);
forSum.Add(mn.forward(newInputs));
}
double summed = this.sumNode.forward(forSum);
double summed_act;
if (activation.Equals("sigmoid"))
{
summed_act = this.activation_node.forward(summed);
}
else
{
summed_act = this.activation_node2.forward(summed);
}
return(summed_act);
}
/// <summary>
/// </summary>
/// <param name="dz"></param>
/// <returns></returns>
public List <double> backward(List <double> dz)
{
List <double> dw = new List <double>();
List <double> dx = new List <double>();
double backward_signal = dz.Sum();
if (this.activation.Equals("sigmoid"))
{
backward_signal = this.activation_node.backward(backward_signal);
}
else
{
backward_signal = this.activation_node2.backward(backward_signal);
}
List <double> backWard = this.sumNode.backward(backward_signal);
for (int i = 0; i < this.multiplyNodes.Count; i++)
{
MultiplyNode mn = this.multiplyNodes[i];
List <double> retVals = mn.backward(backWard[i]);
dx.Add(retVals[0]);
dw.Add(retVals[1]);
}
//https://en.wikipedia.org/wiki/Gradient_descent
//df/dx1, df/dx2...
this.gradients.Add(dw);
return(dx);
}
/// <summary>
/// Greska se propagira u nazad kroz aktivacionu funkciju,
/// preko sabiraca, do svakog pojedinacnog mnozaca
/// </summary>
/// <param name="dz"></param>
/// <returns></returns>
public List <double> backward(List <double> dz)
{
//greska svakog ulaza
List <double> dw = new List <double>();
List <double> dx = new List <double>();
double backward_signal = dz.Sum();
//TODO 7: Izvrsiti propagaciju signala u nazad, prvo kroz aktivacionu funkciju,
// onda kroz sabirac pa kroz svaki pojedinacan mnozac
backward_signal = this.activation_node.backward(backward_signal);
//dobije se nazad lista gresaka po mnoziocima koje ulaze u sabirac
//z = x + y ===> dx = dz && dy = dz, te se dobije samo lista dz-ova onoliko puta koliko ima sabiraoca
List <double> sabiracListaGreska = this.sumNode.backward(backward_signal);
//za svaki sabiraoc (koji je ustv mnozac), izracunati gresku po promenljivoj i po tezini
for (int i = 0; i < sabiracListaGreska.Count; i++)
{
MultiplyNode mn = this.multiplyNodes[i];
List <double> parPromenljivaTezinaGreska = mn.backward(sabiracListaGreska[i]);
dx.Add(parPromenljivaTezinaGreska[0]);
dw.Add(parPromenljivaTezinaGreska[1]);
}
//https://en.wikipedia.org/wiki/Gradient_descent
//df/dx1, df/dx2...
this.gradients.Add(dw);
return(dx);
}
private static GaussianRandom grnd = new GaussianRandom(); //gausova raspodela za random vrednosti
/// <summary>
/// https://en.wikipedia.org/wiki/Artificial_neuron
/// ulaz1*tezina1
/// ulaz1*tezina2
/// bias(+1*tezina)
/// sve se sabere
/// provuce se kroz aktivacionu funkciju
/// </summary>
/// <param name="n_inputs"></param>
/// <param name="activation"> String naziv aktivacione funkcije </param>
public NeuronNode(int n_inputs, string activation)
{
this.n_inputs = n_inputs;
this.multiplyNodes = new List <MultiplyNode>(); //for inputs and weights
this.sumNode = new SumNode(); //for sum of inputs*weights
this.previous_deltas = new List <double>(); //vrednost delti u prosloj iteraciji
this.gradients = new List <List <double> >(); //lokalni gradijenti
MultiplyNode mulNode;
//init inputs
//collect inputs and corresponding weights
for (int i = 0; i < this.n_inputs; i++)
{
mulNode = new MultiplyNode();
double b = grnd.NextGaussian(0.3, 0.5);
mulNode.x = new List <double>()
{
1.0, b
};
this.multiplyNodes.Add(mulNode);
previous_deltas.Add(0.0);
}
//init bias node and weight
mulNode = new MultiplyNode();
double m = grnd.NextGaussian(0.0, 0.01);
mulNode.x = new List <double>()
{
1.0, m
};
this.multiplyNodes.Add(mulNode);
previous_deltas.Add(0.0);
// TODO: Ovde dodajem ako dodajem neku drugu aktivacionu funkciju
/*
* if else(activation.Equals("TangesHiperbolic").....
*/
if (activation.Equals("sigmoid"))
{
this.activation_node = new SigmoidNode();
}
else if (activation.Equals("tanh"))
{
this.activation_node = new TanhNode();
}
else
{
throw new NotImplementedException("Activation function is not supported");
}
}
/// <summary>
/// https://en.wikipedia.org/wiki/Artificial_neuron
/// input1*weight1
/// input1*weight2
/// bias(+1*weight)
/// </summary>
/// <param name="n_inputs"></param>
/// <param name="activation"></param>
public NeuronNode(int n_inputs, string activation)
{
this.n_inputs = n_inputs;
this.multiplyNodes = new List <MultiplyNode>(); //for inputs and weights
this.sumNode = new SumNode(); //for sum of inputs*weights
this.previous_deltas = new List <double>(); // values of las iteration
this.gradients = new List <List <double> >(); //local gradients
MultiplyNode mulNode;
//init inputs
//collect inputs and corresponding weights
for (int i = 0; i < this.n_inputs; i++)
{
mulNode = new MultiplyNode();
double b = grnd.NextGaussian(0.3, 0.5);
mulNode.x = new List <double>()
{
1.0, b
};
this.multiplyNodes.Add(mulNode);
previous_deltas.Add(0.0);
}
//init bias node and weight
mulNode = new MultiplyNode();
double m = grnd.NextGaussian(0.0, 0.01);
mulNode.x = new List <double>()
{
1.0, m
};
this.multiplyNodes.Add(mulNode);
previous_deltas.Add(0.0);
if (activation.Equals("sigmoid"))
{
this.activation_node = new SigmoidNode();
}
if (activation.Equals("relu"))
{
this.activation_node2 = new RelUNode();
}
else
{
throw new NotImplementedException("Activation function is not supported");
}
this.activation = activation;
}
/// <summary>
/// https://en.wikipedia.org/wiki/Artificial_neuron
/// ulaz1*tezina1
/// ulaz1*tezina2
/// bias(+1*tezina)
/// sve se sabere
/// provuce se kroz aktivacionu funkciju
/// </summary>
/// <param name="n_inputs"></param>
/// <param name="activation"></param>
public NeuronNode(int n_inputs, string activation)
{
this.n_inputs = n_inputs;
this.multiplyNodes = new List <MultiplyNode>(); //for inputs and weights
this.sumNode = new SumNode(); //for sum of inputs*weights
this.previous_deltas = new List <double>(); //vrednost delti u prosloj iteraciji
this.gradients = new List <List <double> >(); //lokalni gradijenti
MultiplyNode mulNode;
//init inputs
//collect inputs and corresponding weights
for (int i = 0; i < this.n_inputs; i++)
{
mulNode = new MultiplyNode();
mulNode.x = new List <double>()
{
1.0, (rnd.NextDouble() - 0.5) / 50.0
}; //izemdju -0.1 i 0.1
this.multiplyNodes.Add(mulNode);
previous_deltas.Add(0.0);
}
//init bias node and weight
mulNode = new MultiplyNode();
mulNode.x = new List <double>()
{
1.0, (rnd.NextDouble() - 0.5 / 50.0)
}; // -0.01 do 0.01
this.multiplyNodes.Add(mulNode);
previous_deltas.Add(0.0);
if (activation.Equals("sigmoid"))
{
this.activation_node = new SigmoidNode();
}
else
{
throw new NotImplementedException("Activation function is not supported");
}
}
/// <summary>
/// Pomnoziti sve ulaze sa tezinama.
/// Sabrati sve rezultate mnozaca. Kao povratnu vrednost
/// vratiti vrednost aktivacione funkije za vrednost suma
/// svih mnozaca
/// </summary>
/// <param name="x">x is a vector of inputs</param>
/// <returns></returns>
public double forward(List <double> x)
{
/*
* List<double> inputs = new List<double>(x);
* inputs.Add(1.0); //bias
*
* List<double> forSum = new List<double>();
* //TODO 6: Izracunati vrednost na izlazu vestackog neurona
* for (int i = 0; i < inputs.Count; i++)
* {
* double w_i = this.multiplyNodes[i].x[1]; //uzmi tezinu iz mnozackog cvora, drugi parametar
* MultiplyNode mn = new MultiplyNode();
*
*
* List<double> parTezinaInput = new List<double>();
*
*
* parTezinaInput.Add(w_i);
* parTezinaInput.Add(inputs[i]);
*
* //mora ovako jer mn.forward trazi parametre
*
*
* //pomnozi i dodaj u buduci sabirac
* forSum.Add(mn.forward(parTezinaInput));
*
* }
*
* double summed = 0.0;
* //SumNode sumNode = new SumNode();
* summed = sumNode.forward(forSum);
*
* // dobijena vrednost se propusti kroz aktivacionu funkciju
* double summed_act = this.activation_node.forward(summed);
* return summed_act;
*/
List <double> inputs = new List <double>(x);
inputs.Add(1.0); //bias
List <double> forSum = new List <double>();
//TODO 6: Izracunati vrednost na izlazu vestackog neurona
for (int i = 0; i < inputs.Count; i++)
{
MultiplyNode mn = this.multiplyNodes[i];
double weight = mn.x[1];
List <double> newInputs = new List <double>();
newInputs.Add(inputs[i]);
newInputs.Add(weight);
forSum.Add(mn.forward(newInputs));
}
double summed = this.sumNode.forward(forSum);
// dobijena vrednost se propusti kroz aktivacionu funkciju
double summed_act = this.activation_node.forward(summed);
return(summed_act);
}
