Neuro/Net.cpp

#include "Net.h"

Net::Net(std::initializer_list<size_t> layerSizes)
{
	if (layerSizes.size() < 2)
	{
		throw std::exception("A net needs at least 2 layers");
	}

	for (size_t numNeurons : layerSizes)
	{
		push_back(Layer(numNeurons));
	}

	for (auto layerIt = begin(); layerIt != end() - 1; ++layerIt)
	{
		Layer &currentLayer = *layerIt;
		const Layer &nextLayer = *(layerIt + 1);

		currentLayer.addBiasNeuron();

		currentLayer.connectTo(nextLayer);
	}
}

void Net::feedForward(const std::vector<double> &inputValues)
{
	Layer &inputLayer = front();

	if (inputLayer.size() - 1 != inputValues.size())
	{
		throw std::exception("The number of input values has to match the input layer size");
	}

	inputLayer.setOutputValues(inputValues);

	for (auto layerIt = begin(); layerIt != end() - 1; ++layerIt)
	{
		const Layer &currentLayer = *layerIt;
		Layer &nextLayer = *(layerIt + 1);

		nextLayer.feedForward(currentLayer);
	}
}

std::vector<double> Net::getOutput()
{
	std::vector<double> result;

	const Layer &outputLayer = back();
	for (const Neuron &neuron : outputLayer)
	{
		result.push_back(neuron.getOutputValue());
	}

	return result;
}

void Net::backProp(const std::vector<double> &targetValues)
{
	Layer &outputLayer = back();

	if (targetValues.size() != outputLayer.size())
	{
		throw std::exception("The number of target values has to match the output layer size");
	}

	std::vector<double> resultValues = getOutput();
	size_t numResultValues = resultValues.size();

	// calculate rms error
	double rmsError = 0.0;

	for (unsigned int i = 0; i < numResultValues; ++i)
	{
		rmsError += std::pow(resultValues[i] - targetValues[i], 2);
	}

	rmsError = std::sqrt(rmsError / numResultValues);

	// calculate output neuron gradients 
	for (unsigned int i = 0; i < numResultValues; ++i)
	{
		outputLayer[i].calcOutputGradients(targetValues[i]);
	}

	// calculate hidden neuron gradients
	for (auto it = end() - 1; (it - 1) != begin(); --it)
	{
		Layer &hiddenLayer = *(it - 1);
		Layer &nextLayer = *it;

		for (Neuron &neuron : hiddenLayer)
		{
			neuron.calcHiddenGradients(nextLayer);
		}
	}

	// update the input weights
	for (auto it = end() - 1; it != begin(); --it)
	{
		Layer &currentLayer = *it;
		Layer &prevLayer = *(it - 1);

		currentLayer.updateInputWeights(prevLayer);
	}
}