#pragma once

#include <stdio.h>

template<class LAYER1_TYPE, class LAYER2_TYPE, class LAYER3_TYPE, class LAYER4_TYPE>
class NeuralNetFourLayers
{
	LAYER1_TYPE m_layer1;
	LAYER2_TYPE m_layer2;
	LAYER3_TYPE m_layer3;
	LAYER4_TYPE m_layer4;
	double m_lfLearningRate;
public:
	NeuralNetFourLayers(void);
	~NeuralNetFourLayers(void);

	void SetLearnRate(double lfLearnRate) { m_lfLearningRate = lfLearnRate; }

	bool Train(const double * const arrLfInputs, const double * const arrLfIntendedOutputs, double * arrLfScratch)
	{
		m_layer1.Eval(arrLfInputs);
		m_layer2.Eval(m_layer1.GetOutputVec());
		m_layer3.Eval(m_layer2.GetOutputVec());
		m_layer4.Eval(m_layer3.GetOutputVec());
		const double * arrLfActualOutputs = m_layer4.GetOutputVec();
		for(int i = 0; i < m_layer4.GetNumOutputs(); ++i)
		{
			arrLfScratch[i] = arrLfActualOutputs[i] - arrLfIntendedOutputs[i];
		}
		m_layer4.CalcErrorPartials(arrLfScratch);
		m_layer3.CalcErrorPartials(m_layer4.GetDerrDinput());
		m_layer2.CalcErrorPartials(m_layer3.GetDerrDinput());
		m_layer1.CalcErrorPartials(m_layer2.GetDerrDinput());
		m_layer4.DoBackpropStep(m_lfLearningRate, m_layer3.GetOutputVec());
		m_layer3.DoBackpropStep(m_lfLearningRate, m_layer2.GetOutputVec());
		m_layer2.DoBackpropStep(m_lfLearningRate, m_layer1.GetOutputVec());
		m_layer1.DoBackpropStep(m_lfLearningRate, arrLfInputs);
		return true;
	}

	const double * const Eval(const double * const arrLfInputs)
	{
		m_layer1.Eval(arrLfInputs);
		m_layer2.Eval(m_layer1.GetOutputVec());
		m_layer3.Eval(m_layer2.GetOutputVec());
		m_layer4.Eval(m_layer3.GetOutputVec());
		return m_layer4.GetOutputVec();
	}

	void Serialize(FILE * fiOut)
	{
		m_layer1.Serialize(fiOut);
		m_layer2.Serialize(fiOut);
		m_layer3.Serialize(fiOut);
		m_layer4.Serialize(fiOut);
	}

	void DeSerialize(FILE * fiIn)
	{
		m_layer1.DeSerialize(fiIn);
		m_layer2.DeSerialize(fiIn);
		m_layer3.DeSerialize(fiIn);
		m_layer4.DeSerialize(fiIn);
	}
};

template<class LAYER1_TYPE, class LAYER2_TYPE, class LAYER3_TYPE, class LAYER4_TYPE>
NeuralNetFourLayers<LAYER1_TYPE, LAYER2_TYPE, LAYER3_TYPE, LAYER4_TYPE>::NeuralNetFourLayers(void):m_lfLearningRate(0.1)
{
}

template<class LAYER1_TYPE, class LAYER2_TYPE, class LAYER3_TYPE, class LAYER4_TYPE>
NeuralNetFourLayers<LAYER1_TYPE, LAYER2_TYPE, LAYER3_TYPE, LAYER4_TYPE>::~NeuralNetFourLayers(void)
{
}