我应该做什么。我有一张黑白图像（100x100像素）：
一旦网络学习完毕，我希望它能根据权重再现这张图片，并尽可能接近原始图片。
这是我的反向传播实现代码：
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(10)])
outputLayer = Layer([Neuron() for n in range(1)])
learningRate = 0.4
network = Network([inputLayer, hiddenLayer, outputLayer], learningRate)
# let's get the training set
os.chdir("D:/stuff")
image = Image.open("backprop-input.gif")
pixels = image.load()
bbox = image.getbbox()
width = 5#bbox[2] # image width
height = 5#bbox[3] # image height
trainingInputs = []
trainingOutputs = []
b = w = 0
for x in range(0, width):
for y in range(0, height):
if (0, 0, 0, 255) == pixels[x, y]:
color = 0
b += 1
elif (255, 255, 255, 255) == pixels[x, y]:
color = 1
w += 1
trainingInputs.append([float(x), float(y)])
trainingOutputs.append([float(color)])
print "\nOriginal image ... Black:"+str(b)+" White:"+str(w)+"\n"
#------------------------------ let's train
for i in range(500):
for j in range(len(trainingOutputs)):
network.train(trainingInputs[j], trainingOutputs[j])
network.learn()
for w in network.weights:
w.delta = 0.0
#------------------------------ let's check
b = w = 0
for x in range(0, width):
for y in range(0, height):
out = network.calculateSingleOutput([float(x), float(y)])
if 0.0 == round(out):
color = (0, 0, 0, 255)
b += 1
elif 1.0 == round(out):
color = (255, 255, 255, 255)
w += 1
pixels[x, y] = color
#print out
print "\nAfter learning the network thinks ... Black:"+str(b)+" White:"+str(w)+"\n"
显然，我的实现出了一些问题。上面的代码返回：
原始图像...黑色：21 白色：4
在学习网络之后，它认为...
黑色：25 白色：0
如果我尝试使用更大的训练集（我只是为了测试目的从上面的图像中测试了25个像素），它会做同样的事情。在学习后，它会返回所有像素都应该是黑色的。
现在，如果我改用手动训练集，像这样：
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()
for w in network.weights:
w.delta = 0.0
#------------------------------ let's check
for inputs in trainingInputs:
print network.calculateSingleOutput(inputs)
输出示例如下：
0.0330125791296 # this should be 0, OK
0.953539182136 # this should be 1, OK
0.971854575477 # this should be 1, OK
0.00046146137467 # this should be 0, OK
0.896699762781 # this should be 1, OK
0.112909223162 # this should be 0, OK
0.00034058462280 # this should be 0, OK
0.0929886299643 # this should be 0, OK
0.940489647869 # this should be 1, OK
换句话说，网络猜对了所有像素（黑色和白色）。如果我使用实际图像中的像素而不是像上面那样硬编码训练集，为什么它会说所有像素都应该是黑色？
我尝试改变隐藏层中的神经元数量（最多100个神经元），但没有成功。
这是一项作业。
这也是我关于反向传播的上一个问题的延续。