编辑2:
新的训练集...
输入:
[
[0.0, 0.0],
[0.0, 1.0],
[0.0, 2.0],
[0.0, 3.0],
[0.0, 4.0],
[1.0, 0.0],
[1.0, 1.0],
[1.0, 2.0],
[1.0, 3.0],
[1.0, 4.0],
[2.0, 0.0],
[2.0, 1.0],
[2.0, 2.0],
[2.0, 3.0],
[2.0, 4.0],
[3.0, 0.0],
[3.0, 1.0],
[3.0, 2.0],
[3.0, 3.0],
[3.0, 4.0],
[4.0, 0.0],
[4.0, 1.0],
[4.0, 2.0],
[4.0, 3.0],
[4.0, 4.0]
]
输出:
[
[0.0],
[0.0],
[0.0],
[0.0],
[0.0],
[0.0],
[0.0],
[0.0],
[0.0],
[0.0],
[0.0],
[0.0],
[0.0],
[0.0],
[0.0],
[0.0],
[0.0],
[0.0],
[1.0],
[1.0],
[0.0],
[0.0],
[0.0],
[1.0],
[1.0]
]
编辑1:
我已经更新了我的代码。我修复了一些小问题,但是在网络学习后,所有输入组合的输出仍然相同。
这里解释了反向传播算法:反向传播算法
是的,这是一份作业,需要在一种简单的神经网络上实现一个简单的反向传播算法。
我选择使用Python作为语言,并选择了一个像这样的神经网络:
3层:1个输入层、1个隐藏层和1个输出层:
O O
O
O O
输入神经元上都有一个整数,输出神经元上有1或0。
这是我的完整实现(稍微有点长)。在下面,我将选择一些较短的相关代码片段,在其中我认为可能出现错误:
import os
import math
import Image
import random
from random import sample
#------------------------------ class definitions
class Weight:
def __init__(self, fromNeuron, toNeuron):
self.value = random.uniform(-0.5, 0.5)
self.fromNeuron = fromNeuron
self.toNeuron = toNeuron
fromNeuron.outputWeights.append(self)
toNeuron.inputWeights.append(self)
self.delta = 0.0 # delta value, this will accumulate and after each training cycle used to adjust the weight value
def calculateDelta(self, network):
self.delta += self.fromNeuron.value * self.toNeuron.error
class Neuron:
def __init__(self):
self.value = 0.0 # the output
self.idealValue = 0.0 # the ideal output
self.error = 0.0 # error between output and ideal output
self.inputWeights = []
self.outputWeights = []
def activate(self, network):
x = 0.0;
for weight in self.inputWeights:
x += weight.value * weight.fromNeuron.value
# sigmoid function
if x < -320:
self.value = 0
elif x > 320:
self.value = 1
else:
self.value = 1 / (1 + math.exp(-x))
class Layer:
def __init__(self, neurons):
self.neurons = neurons
def activate(self, network):
for neuron in self.neurons:
neuron.activate(network)
class Network:
def __init__(self, layers, learningRate):
self.layers = layers
self.learningRate = learningRate # the rate at which the network learns
self.weights = []
for hiddenNeuron in self.layers[1].neurons:
for inputNeuron in self.layers[0].neurons:
self.weights.append(Weight(inputNeuron, hiddenNeuron))
for outputNeuron in self.layers[2].neurons:
self.weights.append(Weight(hiddenNeuron, outputNeuron))
def setInputs(self, inputs):
self.layers[0].neurons[0].value = float(inputs[0])
self.layers[0].neurons[1].value = float(inputs[1])
def setExpectedOutputs(self, expectedOutputs):
self.layers[2].neurons[0].idealValue = expectedOutputs[0]
def calculateOutputs(self, expectedOutputs):
self.setExpectedOutputs(expectedOutputs)
self.layers[1].activate(self) # activation function for hidden layer
self.layers[2].activate(self) # activation function for output layer
def calculateOutputErrors(self):
for neuron in self.layers[2].neurons:
neuron.error = (neuron.idealValue - neuron.value) * neuron.value * (1 - neuron.value)
def calculateHiddenErrors(self):
for neuron in self.layers[1].neurons:
error = 0.0
for weight in neuron.outputWeights:
error += weight.toNeuron.error * weight.value
neuron.error = error * neuron.value * (1 - neuron.value)
def calculateDeltas(self):
for weight in self.weights:
weight.calculateDelta(self)
def train(self, inputs, expectedOutputs):
self.setInputs(inputs)
self.calculateOutputs(expectedOutputs)
self.calculateOutputErrors()
self.calculateHiddenErrors()
self.calculateDeltas()
def learn(self):
for weight in self.weights:
weight.value += self.learningRate * weight.delta
def calculateSingleOutput(self, inputs):
self.setInputs(inputs)
self.layers[1].activate(self)
self.layers[2].activate(self)
#return round(self.layers[2].neurons[0].value, 0)
return self.layers[2].neurons[0].value
#------------------------------ initialize objects etc
inputLayer = Layer([Neuron() for n in range(2)])
hiddenLayer = Layer([Neuron() for n in range(100)])
outputLayer = Layer([Neuron() for n in range(1)])
learningRate = 0.5
network = Network([inputLayer, hiddenLayer, outputLayer], learningRate)
# just for debugging, the real training set is much larger
trainingInputs = [
[0.0, 0.0],
[1.0, 0.0],
[2.0, 0.0],
[0.0, 1.0],
[1.0, 1.0],
[2.0, 1.0],
[0.0, 2.0],
[1.0, 2.0],
[2.0, 2.0]
]
trainingOutputs = [
[0.0],
[1.0],
[1.0],
[0.0],
[1.0],
[0.0],
[0.0],
[0.0],
[1.0]
]
#------------------------------ let's train
for i in range(500):
for j in range(len(trainingOutputs)):
network.train(trainingInputs[j], trainingOutputs[j])
network.learn()
#------------------------------ let's check
for pattern in trainingInputs:
print network.calculateSingleOutput(pattern)
现在的问题是,学习后网络似乎对所有输入组合都返回非常接近0.0的浮点数,即使那些应该接近1.0的组合也是如此。
我进行了100个周期的网络训练,在每个周期中执行以下操作:
对于训练集中的每组输入: - 设置网络输入 - 使用sigmoid函数计算输出 - 计算输出层中的误差 - 计算隐藏层中的误差 - 计算权重的增量
然后,我根据学习率和累积增量调整权重。
这是我的神经元激活函数:
def activationFunction(self, network):
"""
Calculate an activation function of a neuron which is a sum of all input weights * neurons where those weights start
"""
x = 0.0;
for weight in self.inputWeights:
x += weight.value * weight.getFromNeuron(network).value
# sigmoid function
self.value = 1 / (1 + math.exp(-x))
这是我计算增量的方法:
def calculateDelta(self, network):
self.delta += self.getFromNeuron(network).value * self.getToNeuron(network).error
这是我的算法的一般流程:
for i in range(numberOfIterations):
for k,expectedOutput in trainingSet.iteritems():
coordinates = k.split(",")
network.setInputs((float(coordinates[0]), float(coordinates[1])))
network.calculateOutputs([float(expectedOutput)])
network.calculateOutputErrors()
network.calculateHiddenErrors()
network.calculateDeltas()
oldWeights = network.weights
network.adjustWeights()
network.resetDeltas()
print "Iteration ", i
j = 0
for weight in network.weights:
print "Weight W", weight.i, weight.j, ": ", oldWeights[j].value, " ............ Adjusted value : ", weight.value
j += j
输出的最后两行是:
0.552785449458 # this should be close to 1
0.552785449458 # this should be close to 0
实际上,它返回所有输入组合的输出数字。
我有什么遗漏吗?
logging.log
语句来追踪边缘权重,并使用计算器逐步检查数字,看看它们的不一致之处所在。 - Katrielmath.exp
的原因。这导致另一个问题:当x变得太小或太大时,Python会引发异常,但这与观察到的虚假行为无关(只是一个普通的错误)。 - krisscalculateSingleOutput
中添加self.layers[2].runActivationFunctionForAllNeurons(self)
,它就可以工作了。但除了修复错误之外,收敛性比您编辑后的第一个版本要差,这很令人惊讶。我不知道是哪个更改导致了这种效果。 - kriss