Решил написать нейросеть XOR по примеру отсюда:
В итоге сеть правильно просчитывает ответ, а вот с обратным распространением не все так гладко. При тесте с одинаковыми входными значениями при вводе примера из статьи просчет нормальный, при вводе 1, 1 значение ответа сначала падают к нулю, затем стремятся к 0.5. При случайных числах на ввод как правило в какой-то момент значение ответа падает ниже 0.3 и в большинстве случаев идентично для любого ввода.
Буду очень признателен за любую помощью.
Классы нейронов:
import math
from classes import *
class Neuron:
def __init__(self, previousLayer, selfValue, weightsToNextLayer):
self.selfValue = selfValue
self.previousLayer = previousLayer
self.weightsToNextLayer = weightsToNextLayer
self.numberInLayer = None
self.sumFromNextLayer = 0
self.nextLayer = None
self.network = None
self.previousIterations = None
if self.weightsToNextLayer:
self.previousIterations = [0 for x in range(len(self.weightsToNextLayer))]
def counterVectors(self):
counter = 0
if self.previousLayer:
for x in range(self.previousLayer.getLengthOfLayer()):
counter+= self.previousLayer.getNeuronValueByNumber(x) * self.previousLayer.getNeuronWeightByNumber(x, self.numberInLayer)
return counter
def sigmoid(self):
vectorsResult = self.counterVectors()
return 1 / (1 + math.exp(-vectorsResult))
def selfValueDefine(self):
self.selfValue = self.sigmoid()
def getSelfValue(self):
return self.selfValue
def setNumberInLayer(self, number):
self.numberInLayer = number
def functionIn(self):
return (1 - self.selfValue) * self.selfValue
def gradient(self, number):
return self.selfValue * self.nextLayer.neuronsInLayer[number].delta()
def deltaWeight(self, number):
return self.network.epsilon * self.gradient(number) + self.network.alpha * self.previousIterations[number]
def changeWeights(self):
for x in range(len(self.weightsToNextLayer)):
self.weightsToNextLayer[x] += self.deltaWeight(x)
self.previousIterations[x] = self.deltaWeight(x)
def backpropagation(self):
self.changeWeights()
def sumToPreviousLayer(self):
if self.previousLayer:
for x in self.previousLayer:
x.sumFromNextLayer+= self.delta() * x.weightsToNextLayer[self.numberInLayer]
class HiddenNeuron(Neuron):
def delta(self):
sum = 0
for x in range(len(self.weightsToNextLayer)):
sum += self.weightsToNextLayer[x] * self.nextLayer.getNeuronValueByNumber(x)
return sum * self.functionIn()
class OutputNeuron(Neuron):
def __init__(self, previousLayer, selfValue):
super().__init__(previousLayer, selfValue, None)
self.idealResult = self.selfValue
self.selfValue = 0
def delta(self):
return (self.idealResult - self.selfValue) * self.functionIn()
def setIdeal(self, ideal):
self.idealResult = ideal
class InputNeuron(Neuron):
def __init__(self, selfValue, weightsToNextLayer):
super().__init__(None, selfValue, weightsToNextLayer)
def selfValueDefine(self):
return self.selfValue
def getSelfValue(self):
return self.selfValue
class BiasNeuron(Neuron):
def __init__(self, weightsToNextLayer):
super().__init__(None, 1, weightsToNextLayer)
Классы слоя и сети:
import math
from neuron_classes import *
class Layer:
def __init__(self, neuronsInLayer):
self.neuronsInLayer = neuronsInLayer
self.nextLayer = None
self.network = None
def initNeurons(self):
for x in self.neuronsInLayer:
x.nextLayer = self.nextLayer
x.network = self.network
def getNeuronValueByNumber(self, neuronNumber):
return self.neuronsInLayer[neuronNumber].selfValue
def getNeuronWeightByNumber(self, neuronNumber, weightNumber):
return self.neuronsInLayer[neuronNumber].weightsToNextLayer[weightNumber]
def getLengthOfLayer(self):
return len(self.neuronsInLayer)
def calculateNeurons(self):
counter = 0
for x in self.neuronsInLayer:
x.setNumberInLayer(counter)
x.selfValueDefine()
counter+= 1
def layerToString(self):
convertedLayer = ''
for x in self.neuronsInLayer:
convertedLayer += str(x.getSelfValue)
print(convertedLayer)
return convertedLayer
def backpropagation(self):
for x in self.neuronsInLayer:
x.backpropagation()
class Network:
def __init__(self, layers, epsilon, alpha):
self.layers = layers
self.epsilon = epsilon
self.alpha = alpha
for x in range(len(layers) - 1):
layers[x].network = self
layers[x].nextLayer = layers[x + 1]
layers[x].initNeurons()
for x in range(len(layers) - 1):
layers[x].nextLayer = layers[x + 1]
self.forwardCounts = 0
def forwardPass(self):
for x in self.layers:
x.calculateNeurons()
self.forwardCounts+= 1
n = self.layers[len(self.layers) - 1]
return [self.layers[len(self.layers) - 1].getNeuronValueByNumber(x) for x in range(self.layers[len(self.layers) - 1].getLengthOfLayer())]
def backpropagation(self):
x = len(self.layers) - 1
while x > 0:
x -= 1
self.layers[x].backpropagation()
def setOutputIdealValues(self, outputIdealValues):
counter = 0
for x in self.layers[len(self.layers) - 1].neuronsInLayer:
x.idealValue = outputIdealValues[counter]
counter+= 1
def setInputs(self, inputs):
counter = 0
for x in self.layers[0].neuronsInLayer:
if type(x) != BiasNeuron:
x.selfValue = inputs[counter]
counter+= 1
def checkNetwork(self):
for y in range(len(self.layers)):
string = ''
for x in range(len(self.layers[y].neuronsInLayer)):
string += str(self.layers[y].neuronsInLayer[x].getSelfValue())
string += ' '
print(string)
Класс удобного доступа к сети:
from classes import *
from neuron_classes import *
class PackagedNetwork:
def __init__(self, inputNeuronsNumber, layersNumber, neuronsInLayer, outputNeuronsNumber, epsilon, alpha):
self.inputNeuronsNumber = inputNeuronsNumber
self.layersNumber = layersNumber
self.neuronsInLayer = neuronsInLayer
self.outputNeuronsNumber = outputNeuronsNumber
self.epsilon = epsilon
self.alpha = alpha
self.layers = [None for x in range(self.layersNumber)]
self.network = None
for y in range(layersNumber):
currentLayer = []
# input
if y == 0:
for x in range(inputNeuronsNumber):
currentLayer.append(InputNeuron(0, [1 * (i + 1) for i in range(neuronsInLayer)]))
# output
elif y == layersNumber - 1:
for x in range(outputNeuronsNumber):
currentLayer.append(OutputNeuron(self.layers[y - 1], 0))
# hidden
elif y == layersNumber - 2:
for x in range(neuronsInLayer):
currentLayer.append(HiddenNeuron(self.layers[y - 1], None, [1 * (i + 1) for i in range(self.outputNeuronsNumber)]))
else:
for x in range(neuronsInLayer):
currentLayer.append(HiddenNeuron(self.layers[y - 1], None, [1 * (i + 1) for i in range(self.neuronsInLayer)]))
layer = Layer(currentLayer)
self.layers[y] = layer
self.network = Network(self.layers, self.epsilon, self.alpha)
def forwardPass(self):
return self.network.forwardPass()
def backpropagation(self):
return self.network.backpropagation()
def setOutputIdealValues(self, outputIdealValues):
self.network.setOutputIdealValues(outputIdealValues)
def setInputs(self, inputs):
self.network.setInputs(inputs)
def checkNetwork(self):
self.network.checkNetwork()
Тест 1
from classes import *
from neuron_classes import *
import random
inputLayer = Layer([InputNeuron(1, [0.45, 0.78]), InputNeuron(0, [-0.12, 0.13])])
hiddenLayer = Layer([HiddenNeuron(inputLayer, None, [1.5]), HiddenNeuron(inputLayer, None, [-2.3])])
outputLayer = Layer([OutputNeuron(hiddenLayer, 1)])
network = Network([inputLayer, hiddenLayer, outputLayer], 0.7, 0.8)
for x in range(20):
print(network.forwardPass())
network.backpropagation()
Тест 2:
from package_classes import *
from neuron_classes import *
from classes import *
import random
network = PackagedNetwork(2, 3, 2, 1, 0.7, 0.3)
print('xor')
input1 = 0
input2 = 0
idealOutput = 0
for x in range(200):
input1 = random.randint(0, 1)
input2 = random.randint(0, 1)
idealOutput = int(input1 != input2)
network.setInputs([input1, input2])
network.setOutputIdealValues([idealOutput])
network.checkNetwork()
print(network.forwardPass(), input1, input2, idealOutput)
print()
network.backpropagation()
Простой