package neuralnetworkbase

import (
	rand "math/rand"
	"time"

	mat "gonum.org/v1/gonum/mat"
)

type NeuralNetwork struct {
	Count          int
	Sizes          []int
	Biases         []*mat.Dense
	Weights        []*mat.Dense
	A              []*mat.Dense
	Z              []*mat.Dense
	alpha          float64
	trainingCycles int
}

func (nn *NeuralNetwork) Result() *mat.Dense {
	return nn.A[nn.Count-1]
}

func NewNeuralNetwork(Sizes []int, nu float64, trainingCycles int) (nn *NeuralNetwork) {
	nn = &NeuralNetwork{}
	nn.Sizes = Sizes
	nn.Count = len(Sizes)
	nn.Weights = make([]*mat.Dense, nn.Count)
	nn.Biases = make([]*mat.Dense, nn.Count)
	nn.A = make([]*mat.Dense, nn.Count)
	nn.Z = make([]*mat.Dense, nn.Count)
	nn.alpha = nu / float64(nn.Sizes[0])
	nn.trainingCycles = trainingCycles

	for i := 1; i < nn.Count; i++ {
		nn.Weights[i] = generateRandomDense(nn.Sizes[i], nn.Sizes[i-1])
		nn.Biases[i] = generateRandomDense(nn.Sizes[i], 1)
	}
	return
}

func (nn *NeuralNetwork) Predict(aIn mat.Matrix) (maxIndex int, max float64) {
	nn.Forward(aIn)
	result := nn.Result()
	r, _ := result.Dims()
	max = 0.0
	maxIndex = 0
	for i := 0; i < r; i++ {
		if result.At(i, 0) > max {
			max = result.At(i, 0)
			maxIndex = i
		}
	}
	return
}

func (nn *NeuralNetwork) Train(dataSet, expect []*mat.Dense) {
	rand.Seed(time.Now().UnixNano())
	dataSetSize := len(dataSet)
	// randomIndex := rand.Int() % dataSetSize
	// fmt.Printf("Train: %v\n", randomIndex)
	for i := 0; i < nn.trainingCycles; i++ {
		for j := dataSetSize - 1; j >= 0; j -= 3 {
			if j < 0 {
				j = 0
			}
			nn.Backward(dataSet[j], expect[j])
		}
		// _, max := nn.Predict(dataSet[randomIndex])
		// if 1.0-max < 0.2 {
		// 	break
		// }
	}
}

func (nn *NeuralNetwork) SaveState(filename string) {

}

func (nn *NeuralNetwork) LoadState(filename string) {

}

func (nn *NeuralNetwork) Forward(aIn mat.Matrix) {
	nn.A[0] = mat.DenseCopyOf(aIn)

	for i := 1; i < nn.Count; i++ {
		nn.A[i] = mat.NewDense(nn.Sizes[i], 1, nil)
		aSrc := nn.A[i-1]
		aDst := nn.A[i]

		// r, c := nn.Weights[i].Dims()
		// fmt.Printf("r: %v,c: %v\n", r, c)

		// r, c = aSrc.Dims()
		// fmt.Printf("src r: %v,c: %v\n\n\n", r, c)

		aDst.Mul(nn.Weights[i], aSrc)
		aDst.Add(aDst, nn.Biases[i])
		nn.Z[i] = mat.DenseCopyOf(aDst)
		aDst.Apply(applySigmoid, aDst)
	}
}

func (nn *NeuralNetwork) Backward(aIn, aOut mat.Matrix) {
	nn.Forward(aIn)

	lastLayerNum := nn.Count - 1

	//Initial error
	err := &mat.Dense{}
	err.Sub(nn.Result(), aOut)

	sigmoidsPrime := &mat.Dense{}
	sigmoidsPrime.Apply(applySigmoidPrime, nn.Z[lastLayerNum])

	delta := &mat.Dense{}
	delta.MulElem(err, sigmoidsPrime)

	biases := mat.DenseCopyOf(delta)

	weights := &mat.Dense{}
	weights.Mul(delta, nn.A[lastLayerNum-1].T())

	newBiases := []*mat.Dense{makeBackGradien(biases, nn.Biases[lastLayerNum], nn.alpha)}
	newWeights := []*mat.Dense{makeBackGradien(weights, nn.Weights[lastLayerNum], nn.alpha)}

	err = delta
	for i := nn.Count - 2; i > 0; i-- {
		sigmoidsPrime := &mat.Dense{}
		sigmoidsPrime.Apply(applySigmoidPrime, nn.Z[i])

		delta := &mat.Dense{}
		wdelta := &mat.Dense{}
		wdelta.Mul(nn.Weights[i+1].T(), err)

		delta.MulElem(wdelta, sigmoidsPrime)
		err = delta

		biases := mat.DenseCopyOf(delta)

		weights := &mat.Dense{}
		weights.Mul(delta, nn.A[i-1].T())

		// Scale down
		newBiases = append([]*mat.Dense{makeBackGradien(biases, nn.Biases[i], nn.alpha)}, newBiases...)
		newWeights = append([]*mat.Dense{makeBackGradien(weights, nn.Weights[i], nn.alpha)}, newWeights...)
	}

	newBiases = append([]*mat.Dense{&mat.Dense{}}, newBiases...)
	newWeights = append([]*mat.Dense{&mat.Dense{}}, newWeights...)

	nn.Biases = newBiases
	nn.Weights = newWeights
}

func makeBackGradien(in mat.Matrix, actual mat.Matrix, alpha float64) *mat.Dense {
	scaled := &mat.Dense{}
	result := &mat.Dense{}

	scaled.Scale(alpha, in)
	result.Sub(actual, scaled)
	return result
}