神经网络反向传播算法在Python中不起作用 [英] Neural network backpropagation algorithm not working in Python
本文介绍了神经网络反向传播算法在Python中不起作用的处理方法，对大家解决问题具有一定的参考价值，需要的朋友们下面随着小编来一起学习吧！
问题描述
按照示例，我正在用Python写一个神经网络.这里.鉴于神经网络在经过1万次训练后未能产生正确的值(误差范围内)，因此反向传播算法似乎无法正常工作.具体来说，在以下示例中，我正在训练它以计算正弦函数:
import numpy as np
class Neuralnet:
def __init__(self, neurons):
self.weights = []
self.inputs = []
self.outputs = []
self.errors = []
self.rate = .1
for layer in range(len(neurons)):
self.inputs.append(np.empty(neurons[layer]))
self.outputs.append(np.empty(neurons[layer]))
self.errors.append(np.empty(neurons[layer]))
for layer in range(len(neurons)-1):
self.weights.append(
np.random.normal(
scale=1/np.sqrt(neurons[layer]),
size=[neurons[layer], neurons[layer + 1]]
)
)
def feedforward(self, inputs):
self.inputs[0] = inputs
for layer in range(len(self.weights)):
self.outputs[layer] = np.tanh(self.inputs[layer])
self.inputs[layer + 1] = np.dot(self.weights[layer].T, self.outputs[layer])
self.outputs[-1] = np.tanh(self.inputs[-1])
def backpropagate(self, targets):
gradient = 1 - self.outputs[-1] * self.outputs[-1]
self.errors[-1] = gradient * (self.outputs[-1] - targets)
for layer in reversed(range(len(self.errors) - 1)):
gradient = 1 - self.outputs[layer] * self.outputs[layer]
self.errors[layer] = gradient * np.dot(self.weights[layer], self.errors[layer + 1])
for layer in range(len(self.weights)):
self.weights[layer] -= self.rate * np.outer(self.outputs[layer], self.errors[layer + 1])
def xor_example():
net = Neuralnet([2, 2, 1])
for step in range(100000):
net.feedforward([0, 0])
net.backpropagate([-1])
net.feedforward([0, 1])
net.backpropagate([1])
net.feedforward([1, 0])
net.backpropagate([1])
net.feedforward([1, 1])
net.backpropagate([-1])
net.feedforward([1, 1])
print(net.outputs[-1])
def identity_example():
net = Neuralnet([1, 3, 1])
for step in range(100000):
x = np.random.normal()
net.feedforward([x])
net.backpropagate([np.tanh(x)])
net.feedforward([-2])
print(net.outputs[-1])
def sine_example():
net = Neuralnet([1, 6, 1])
for step in range(100000):
x = np.random.normal()
net.feedforward([x])
net.backpropagate([np.tanh(np.sin(x))])
net.feedforward([3])
print(net.outputs[-1])
sine_example()
输出未能接近tanh(sin(3)) = 0.140190616.我怀疑有一个错误的索引或对齐方式，但Numpy并未引发任何此类错误.关于我哪里出问题的任何提示?
编辑:我忘了添加偏向神经元.这是更新的代码:
import numpy as np
class Neuralnet:
def __init__(self, neurons):
self.weights = []
self.outputs = []
self.inputs = []
self.errors = []
self.offsets = []
self.rate = .01
for layer in range(len(neurons)-1):
self.weights.append(
np.random.normal(
scale=1/np.sqrt(neurons[layer]),
size=[neurons[layer], neurons[layer + 1]]
)
)
self.outputs.append(np.empty(neurons[layer]))
self.inputs.append(np.empty(neurons[layer]))
self.errors.append(np.empty(neurons[layer]))
self.offsets.append(np.random.normal(scale=1/np.sqrt(neurons[layer]), size=neurons[layer + 1]))
self.inputs.append(np.empty(neurons[-1]))
self.errors.append(np.empty(neurons[-1]))
def feedforward(self, inputs):
self.inputs[0] = inputs
for layer in range(len(self.weights)):
self.outputs[layer] = np.tanh(self.inputs[layer])
self.inputs[layer + 1] = self.offsets[layer] + np.dot(self.weights[layer].T, self.outputs[layer])
def backpropagate(self, targets):
self.errors[-1] = self.inputs[-1] - targets
for layer in reversed(range(len(self.errors) - 1)):
gradient = 1 - self.outputs[layer] * self.outputs[layer]
self.errors[layer] = gradient * np.dot(self.weights[layer], self.errors[layer + 1])
for layer in range(len(self.weights)):
self.weights[layer] -= self.rate * np.outer(self.outputs[layer], self.errors[layer + 1])
self.offsets[layer] -= self.rate * self.errors[layer + 1]
def sine_example():
net = Neuralnet([1, 5, 1])
for step in range(10000):
x = np.random.uniform(-5, 5)
net.feedforward([x])
net.backpropagate([np.sin(x)])
net.feedforward([np.pi])
print(net.inputs[-1])
def xor_example():
net = Neuralnet([2, 2, 1])
for step in range(10000):
net.feedforward([0, 0])
net.backpropagate([-1])
net.feedforward([0, 1])
net.backpropagate([1])
net.feedforward([1, 0])
net.backpropagate([1])
net.feedforward([1, 1])
net.backpropagate([-1])
net.feedforward([1, 1])
print(net.outputs[-1])
def identity_example():
net = Neuralnet([1, 3, 1])
for step in range(10000):
x = np.random.normal()
net.feedforward([x])
net.backpropagate([x])
net.feedforward([-2])
print(net.outputs[-1])
identity_example()
解决方案
我认为您以错误的方式训练了NN.您有超过10000次迭代的循环，并在每个循环中提供了一个新样本.在这种情况下，NN将永远不会受到训练.
(声明有误！请参见更新！)
您需要做的是生成大量真实样本 Y = sin(X) ，将其提供给您的网络 ONCE ，并反复遍历训练集，以最小化成本函数.要检查算法，您可能需要根据迭代次数绘制成本函数，并确保成本下降.
另一个重要点是权重的初始化.您的人数非常多，网络将花费大量时间进行收敛，尤其是在使用低费率时.最好在 [-eps .. eps] 的较小范围内均匀生成初始权重.
在我的代码中，我实现了两种不同的激活功能: sigmoid() 和 tanh() .您需要根据所选功能缩放输入:分别为[0 .. 1]和[-1 .. 1].
有些图像显示了成本函数以及sigmoid()和tanh()激活函数的最终预测:
您可以看到 sigmoid() 激活比tanh()效果更好.
与使用4层[1, 6, 4, 1]的较大网络相比，使用网络 [1, 6, 1] 时，我也得到了更好的预测.因此，NN的大小并不总是关键因素.这是对提到的4层网络的预测:
这是我的代码，带有一些注释.我试图在可能的地方使用您的符号.
import numpy as np
import math
import matplotlib.pyplot as plt
class Neuralnet:
def __init__(self, neurons, activation):
self.weights = []
self.inputs = []
self.outputs = []
self.errors = []
self.rate = 0.5
self.activation = activation #sigmoid or tanh
self.neurons = neurons
self.L = len(self.neurons) #number of layers
eps = 0.12; # range for uniform distribution -eps..+eps
for layer in range(len(neurons)-1):
self.weights.append(np.random.uniform(-eps,eps,size=(neurons[layer+1], neurons[layer]+1)))
###################################################################################################
def train(self, X, Y, iter_count):
m = X.shape[0];
for layer in range(self.L):
self.inputs.append(np.empty([m, self.neurons[layer]]))
self.errors.append(np.empty([m, self.neurons[layer]]))
if (layer < self.L -1):
self.outputs.append(np.empty([m, self.neurons[layer]+1]))
else:
self.outputs.append(np.empty([m, self.neurons[layer]]))
#accumulate the cost function
J_history = np.zeros([iter_count, 1])
for i in range(iter_count):
self.feedforward(X)
J = self.cost(Y, self.outputs[self.L-1])
J_history[i, 0] = J
self.backpropagate(Y)
#plot the cost function to check the descent
plt.plot(J_history)
plt.show()
###################################################################################################
def cost(self, Y, H):
J = np.sum(np.sum(np.power((Y - H), 2), axis=0))/(2*m)
return J
###################################################################################################
def feedforward(self, X):
m = X.shape[0];
self.outputs[0] = np.concatenate( (np.ones([m, 1]), X), axis=1)
for i in range(1, self.L):
self.inputs[i] = np.dot( self.outputs[i-1], self.weights[i-1].T )
if (self.activation == 'sigmoid'):
output_temp = self.sigmoid(self.inputs[i])
elif (self.activation == 'tanh'):
output_temp = np.tanh(self.inputs[i])
if (i < self.L - 1):
self.outputs[i] = np.concatenate( (np.ones([m, 1]), output_temp), axis=1)
else:
self.outputs[i] = output_temp
###################################################################################################
def backpropagate(self, Y):
self.errors[self.L-1] = self.outputs[self.L-1] - Y
for i in range(self.L - 2, 0, -1):
if (self.activation == 'sigmoid'):
self.errors[i] = np.dot( self.errors[i+1], self.weights[i][:, 1:] ) * self.sigmoid_prime(self.inputs[i])
elif (self.activation == 'tanh'):
self.errors[i] = np.dot( self.errors[i+1], self.weights[i][:, 1:] ) * (1 - self.outputs[i][:, 1:]*self.outputs[i][:, 1:])
for i in range(0, self.L-1):
grad = np.dot(self.errors[i+1].T, self.outputs[i]) / m
self.weights[i] = self.weights[i] - self.rate*grad
###################################################################################################
def sigmoid(self, z):
s = 1.0/(1.0 + np.exp(-z))
return s
###################################################################################################
def sigmoid_prime(self, z):
s = self.sigmoid(z)*(1 - self.sigmoid(z))
return s
###################################################################################################
def predict(self, X, weights):
m = X.shape[0];
self.inputs = []
self.outputs = []
self.weights = weights
for layer in range(self.L):
self.inputs.append(np.empty([m, self.neurons[layer]]))
if (layer < self.L -1):
self.outputs.append(np.empty([m, self.neurons[layer]+1]))
else:
self.outputs.append(np.empty([m, self.neurons[layer]]))
self.feedforward(X)
return self.outputs[self.L-1]
###################################################################################################
# MAIN PART
activation1 = 'sigmoid' # the input should be scaled into [ 0..1]
activation2 = 'tanh' # the input should be scaled into [-1..1]
activation = activation1
net = Neuralnet([1, 6, 1], activation) # structure of the NN and its activation function
##########################################################################################
# TRAINING
m = 1000 #size of the training set
X = np.linspace(0, 4*math.pi, num = m).reshape(m, 1); # input training set
Y = np.sin(X) # target
kx = 0.1 # noise parameter
noise = (2.0*np.random.uniform(0, kx, m) - kx).reshape(m, 1)
Y = Y + noise # noisy target
# scaling of the target depending on the activation function
if (activation == 'sigmoid'):
Y_scaled = (Y/(1+kx) + 1)/2.0
elif (activation == 'tanh'):
Y_scaled = Y/(1+kx)
# number of the iteration for the training stage
iter_count = 20000
net.train(X, Y_scaled, iter_count) #training
# gained weights
trained_weights = net.weights
##########################################################################################
# PREDICTION
m_new = 40 #size of the prediction set
X_new = np.linspace(0, 4*math.pi, num = m_new).reshape(m_new, 1);
Y_new = net.predict(X_new, trained_weights) # prediction
#rescaling of the result
if (activation == 'sigmoid'):
Y_new = (2.0*Y_new - 1.0) * (1+kx)
elif (activation == 'tanh'):
Y_new = Y_new * (1+kx)
# visualization
plt.plot(X, Y)
plt.plot(X_new, Y_new, 'ro')
plt.show()
raw_input('press any key to exit')
更新
我想回顾一下您的代码中使用的培训方法的声明.实际上，每次迭代仅使用一个样本即可训练网络.在使用S型和tanh激活函数的在线培训中，我得到了有趣的结果:
使用Sigmoid进行在线培训(成本函数和预测)
使用Tanh进行在线培训(成本函数和预测)
可以看出，选择Sigmoid作为激活函数可以提供更好的性能.成本函数看起来不像离线培训时那样好，但是至少它会下降.
我在您的实现中绘制了成本函数，看起来也很生涩:
尝试使用S型甚至ReLU函数的代码是一个好主意.
这是更新的源代码.要在online和offline训练模式之间切换，只需更改method变量即可.
import numpy as np
import math
import matplotlib.pyplot as plt
class Neuralnet:
def __init__(self, neurons, activation):
self.weights = []
self.inputs = []
self.outputs = []
self.errors = []
self.rate = 0.2
self.activation = activation #sigmoid or tanh
self.neurons = neurons
self.L = len(self.neurons) #number of layers
eps = 0.12; #range for uniform distribution -eps..+eps
for layer in range(len(neurons)-1):
self.weights.append(np.random.uniform(-eps,eps,size=(neurons[layer+1], neurons[layer]+1)))
###################################################################################################
def train(self, X, Y, iter_count):
m = X.shape[0];
for layer in range(self.L):
self.inputs.append(np.empty([m, self.neurons[layer]]))
self.errors.append(np.empty([m, self.neurons[layer]]))
if (layer < self.L -1):
self.outputs.append(np.empty([m, self.neurons[layer]+1]))
else:
self.outputs.append(np.empty([m, self.neurons[layer]]))
#accumulate the cost function
J_history = np.zeros([iter_count, 1])
for i in range(iter_count):
self.feedforward(X)
J = self.cost(Y, self.outputs[self.L-1])
J_history[i, 0] = J
self.backpropagate(Y)
#plot the cost function to check the descent
#plt.plot(J_history)
#plt.show()
###################################################################################################
def cost(self, Y, H):
J = np.sum(np.sum(np.power((Y - H), 2), axis=0))/(2*m)
return J
###################################################################################################
def cost_online(self, min_x, max_x, iter_number):
h_arr = np.zeros([iter_number, 1])
y_arr = np.zeros([iter_number, 1])
for step in range(iter_number):
x = np.random.uniform(min_x, max_x, 1).reshape(1, 1)
self.feedforward(x)
h_arr[step, 0] = self.outputs[-1]
y_arr[step, 0] = np.sin(x)
J = np.sum(np.sum(np.power((y_arr - h_arr), 2), axis=0))/(2*iter_number)
return J
###################################################################################################
def feedforward(self, X):
m = X.shape[0];
self.outputs[0] = np.concatenate( (np.ones([m, 1]), X), axis=1)
for i in range(1, self.L):
self.inputs[i] = np.dot( self.outputs[i-1], self.weights[i-1].T )
if (self.activation == 'sigmoid'):
output_temp = self.sigmoid(self.inputs[i])
elif (self.activation == 'tanh'):
output_temp = np.tanh(self.inputs[i])
if (i < self.L - 1):
self.outputs[i] = np.concatenate( (np.ones([m, 1]), output_temp), axis=1)
else:
self.outputs[i] = output_temp
###################################################################################################
def backpropagate(self, Y):
self.errors[self.L-1] = self.outputs[self.L-1] - Y
for i in range(self.L - 2, 0, -1):
if (self.activation == 'sigmoid'):
self.errors[i] = np.dot( self.errors[i+1], self.weights[i][:, 1:] ) * self.sigmoid_prime(self.inputs[i])
elif (self.activation == 'tanh'):
self.errors[i] = np.dot( self.errors[i+1], self.weights[i][:, 1:] ) * (1 - self.outputs[i][:, 1:]*self.outputs[i][:, 1:])
for i in range(0, self.L-1):
grad = np.dot(self.errors[i+1].T, self.outputs[i]) / m
self.weights[i] = self.weights[i] - self.rate*grad
###################################################################################################
def sigmoid(self, z):
s = 1.0/(1.0 + np.exp(-z))
return s
###################################################################################################
def sigmoid_prime(self, z):
s = self.sigmoid(z)*(1 - self.sigmoid(z))
return s
###################################################################################################
def predict(self, X, weights):
m = X.shape[0];
self.inputs = []
self.outputs = []
self.weights = weights
for layer in range(self.L):
self.inputs.append(np.empty([m, self.neurons[layer]]))
if (layer < self.L -1):
self.outputs.append(np.empty([m, self.neurons[layer]+1]))
else:
self.outputs.append(np.empty([m, self.neurons[layer]]))
self.feedforward(X)
return self.outputs[self.L-1]
###################################################################################################
# MAIN PART
activation1 = 'sigmoid' #the input should be scaled into [0..1]
activation2 = 'tanh' #the input should be scaled into [-1..1]
activation = activation1
net = Neuralnet([1, 6, 1], activation) # structure of the NN and its activation function
method1 = 'online'
method2 = 'offline'
method = method1
kx = 0.1 #noise parameter
###################################################################################################
# TRAINING
if (method == 'offline'):
m = 1000 #size of the training set
X = np.linspace(0, 4*math.pi, num = m).reshape(m, 1); #input training set
Y = np.sin(X) #target
noise = (2.0*np.random.uniform(0, kx, m) - kx).reshape(m, 1)
Y = Y + noise #noisy target
#scaling of the target depending on the activation function
if (activation == 'sigmoid'):
Y_scaled = (Y/(1+kx) + 1)/2.0
elif (activation == 'tanh'):
Y_scaled = Y/(1+kx)
#number of the iteration for the training stage
iter_count = 20000
net.train(X, Y_scaled, iter_count) #training
elif (method == 'online'):
sampling_count = 100000 # number of samplings during the training stage
m = 1 #batch size
iter_count = sampling_count/m
for layer in range(net.L):
net.inputs.append(np.empty([m, net.neurons[layer]]))
net.errors.append(np.empty([m, net.neurons[layer]]))
if (layer < net.L -1):
net.outputs.append(np.empty([m, net.neurons[layer]+1]))
else:
net.outputs.append(np.empty([m, net.neurons[layer]]))
J_history = []
step_history = []
for i in range(iter_count):
X = np.random.uniform(0, 4*math.pi, m).reshape(m, 1)
Y = np.sin(X) #target
noise = (2.0*np.random.uniform(0, kx, m) - kx).reshape(m, 1)
Y = Y + noise #noisy target
#scaling of the target depending on the activation function
if (activation == 'sigmoid'):
Y_scaled = (Y/(1+kx) + 1)/2.0
elif (activation == 'tanh'):
Y_scaled = Y/(1+kx)
net.feedforward(X)
net.backpropagate(Y_scaled)
if (np.remainder(i, 1000) == 0):
J = net.cost_online(0, 4*math.pi, 1000)
J_history.append(J)
step_history.append(i)
plt.plot(step_history, J_history)
plt.title('Batch size ' + str(m) + ', rate ' + str(net.rate) + ', samples ' + str(sampling_count))
#plt.ylim([0, 0.1])
plt.show()
#gained weights
trained_weights = net.weights
##########################################################################################
# PREDICTION
m_new = 40 #size of the prediction set
X_new = np.linspace(0, 4*math.pi, num = m_new).reshape(m_new, 1);
Y_new = net.predict(X_new, trained_weights) #prediction
#rescaling of the result
if (activation == 'sigmoid'):
Y_new = (2.0*Y_new - 1.0) * (1+kx)
elif (activation == 'tanh'):
Y_new = Y_new * (1+kx)
#visualization
#fake sine curve to show the ideal signal
if (method == 'online'):
X = np.linspace(0, 4*math.pi, num = 100)
Y = np.sin(X)
plt.plot(X, Y)
plt.plot(X_new, Y_new, 'ro')
if (method == 'online'):
plt.title('Batch size ' + str(m) + ', rate ' + str(net.rate) + ', samples ' + str(sampling_count))
plt.ylim([-1.5, 1.5])
plt.show()
raw_input('press any key to exit')
现在我对您当前的代码有一些评论:
您的正弦函数如下:
def sine_example():
net = Neuralnet([1, 6, 1])
for step in range(100000):
x = np.random.normal()
net.feedforward([x])
net.backpropagate([np.tanh(np.sin(x))])
net.feedforward([3])
print(net.outputs[-1])
我不知道为什么在目标输入中使用tanh.如果您确实要使用正弦的正切值作为目标，则需要将其缩放为[-1..1]，因为tanh(sin(x))返回范围为[-0.76..0.76]的值.
下一件事是您的训练范围.您使用x = np.random.normal()生成样本.这是这种输入的分布:
之后，您希望网络预测3的正弦值，但是网络在训练阶段几乎从未见过此数字.我将使用更广泛范围内的均匀分布来生成样本.
I am writing a neural network in Python, following the example here. It seems that the backpropagation algorithm isn't working, given that the neural network fails to produce the right value (within a margin of error) after being trained 10 thousand times. Specifically, I am training it to compute the sine function in the following example:
import numpy as np
class Neuralnet:
def __init__(self, neurons):
self.weights = []
self.inputs = []
self.outputs = []
self.errors = []
self.rate = .1
for layer in range(len(neurons)):
self.inputs.append(np.empty(neurons[layer]))
self.outputs.append(np.empty(neurons[layer]))
self.errors.append(np.empty(neurons[layer]))
for layer in range(len(neurons)-1):
self.weights.append(
np.random.normal(
scale=1/np.sqrt(neurons[layer]),
size=[neurons[layer], neurons[layer + 1]]
)
)
def feedforward(self, inputs):
self.inputs[0] = inputs
for layer in range(len(self.weights)):
self.outputs[layer] = np.tanh(self.inputs[layer])
self.inputs[layer + 1] = np.dot(self.weights[layer].T, self.outputs[layer])
self.outputs[-1] = np.tanh(self.inputs[-1])
def backpropagate(self, targets):
gradient = 1 - self.outputs[-1] * self.outputs[-1]
self.errors[-1] = gradient * (self.outputs[-1] - targets)
for layer in reversed(range(len(self.errors) - 1)):
gradient = 1 - self.outputs[layer] * self.outputs[layer]
self.errors[layer] = gradient * np.dot(self.weights[layer], self.errors[layer + 1])
for layer in range(len(self.weights)):
self.weights[layer] -= self.rate * np.outer(self.outputs[layer], self.errors[layer + 1])
def xor_example():
net = Neuralnet([2, 2, 1])
for step in range(100000):
net.feedforward([0, 0])
net.backpropagate([-1])
net.feedforward([0, 1])
net.backpropagate([1])
net.feedforward([1, 0])
net.backpropagate([1])
net.feedforward([1, 1])
net.backpropagate([-1])
net.feedforward([1, 1])
print(net.outputs[-1])
def identity_example():
net = Neuralnet([1, 3, 1])
for step in range(100000):
x = np.random.normal()
net.feedforward([x])
net.backpropagate([np.tanh(x)])
net.feedforward([-2])
print(net.outputs[-1])
def sine_example():
net = Neuralnet([1, 6, 1])
for step in range(100000):
x = np.random.normal()
net.feedforward([x])
net.backpropagate([np.tanh(np.sin(x))])
net.feedforward([3])
print(net.outputs[-1])
sine_example()
The output fails to be close to tanh(sin(3)) = 0.140190616. I suspected a mistake involving wrong indices or alignment, but Numpy isn't raising any errors like these. Any tips on where I went wrong?
EDIT: I forgot to add the bias neurons. Here is the updated code:
import numpy as np
class Neuralnet:
def __init__(self, neurons):
self.weights = []
self.outputs = []
self.inputs = []
self.errors = []
self.offsets = []
self.rate = .01
for layer in range(len(neurons)-1):
self.weights.append(
np.random.normal(
scale=1/np.sqrt(neurons[layer]),
size=[neurons[layer], neurons[layer + 1]]
)
)
self.outputs.append(np.empty(neurons[layer]))
self.inputs.append(np.empty(neurons[layer]))
self.errors.append(np.empty(neurons[layer]))
self.offsets.append(np.random.normal(scale=1/np.sqrt(neurons[layer]), size=neurons[layer + 1]))
self.inputs.append(np.empty(neurons[-1]))
self.errors.append(np.empty(neurons[-1]))
def feedforward(self, inputs):
self.inputs[0] = inputs
for layer in range(len(self.weights)):
self.outputs[layer] = np.tanh(self.inputs[layer])
self.inputs[layer + 1] = self.offsets[layer] + np.dot(self.weights[layer].T, self.outputs[layer])
def backpropagate(self, targets):
self.errors[-1] = self.inputs[-1] - targets
for layer in reversed(range(len(self.errors) - 1)):
gradient = 1 - self.outputs[layer] * self.outputs[layer]
self.errors[layer] = gradient * np.dot(self.weights[layer], self.errors[layer + 1])
for layer in range(len(self.weights)):
self.weights[layer] -= self.rate * np.outer(self.outputs[layer], self.errors[layer + 1])
self.offsets[layer] -= self.rate * self.errors[layer + 1]
def sine_example():
net = Neuralnet([1, 5, 1])
for step in range(10000):
x = np.random.uniform(-5, 5)
net.feedforward([x])
net.backpropagate([np.sin(x)])
net.feedforward([np.pi])
print(net.inputs[-1])
def xor_example():
net = Neuralnet([2, 2, 1])
for step in range(10000):
net.feedforward([0, 0])
net.backpropagate([-1])
net.feedforward([0, 1])
net.backpropagate([1])
net.feedforward([1, 0])
net.backpropagate([1])
net.feedforward([1, 1])
net.backpropagate([-1])
net.feedforward([1, 1])
print(net.outputs[-1])
def identity_example():
net = Neuralnet([1, 3, 1])
for step in range(10000):
x = np.random.normal()
net.feedforward([x])
net.backpropagate([x])
net.feedforward([-2])
print(net.outputs[-1])
identity_example()
解决方案
I think you train the NN in the wrong way. You have a loop over 10000 iterations and feed a new sample in each cycle. The NN will never get trained in this case.
(the statement is wrong! See the update! )
What you need to do is to generate a large array of true samples Y = sin(X), give it to your network ONCE and iterate over the training set forwards and backwards, in order to minimize the cost function. To check the algorithm you may need to plot the cost function depending on the iteration number and make sure the cost goes down.
Another important point is the initialization of the weights. Your numbers are pretty large and the network will take a lot of time to converge, especially when using low rates. It's a good practice to generate the initial weights in some small range [-eps .. eps] uniformly.
In my code I implemented two different activation functions: sigmoid() and tanh(). You need to scale your inputs depending on the selected function: [0 .. 1] and [-1 .. 1] respectively.
Here are some images which show the cost function and the resulting predictions for sigmoid() and tanh() activation functions:
As you can see the sigmoid() activation gives a little bit better results, than the tanh().
Also I got much better predictions when using a network [1, 6, 1], compared to a bigger network with 4 layers [1, 6, 4, 1]. So the size of the NN is not always the crucial factor. Here is the prediction for the mentioned network with 4 layers:
Here is my code with some comments. I tried to use your notations where it was possible.
import numpy as np
import math
import matplotlib.pyplot as plt
class Neuralnet:
def __init__(self, neurons, activation):
self.weights = []
self.inputs = []
self.outputs = []
self.errors = []
self.rate = 0.5
self.activation = activation #sigmoid or tanh
self.neurons = neurons
self.L = len(self.neurons) #number of layers
eps = 0.12; # range for uniform distribution -eps..+eps
for layer in range(len(neurons)-1):
self.weights.append(np.random.uniform(-eps,eps,size=(neurons[layer+1], neurons[layer]+1)))
###################################################################################################
def train(self, X, Y, iter_count):
m = X.shape[0];
for layer in range(self.L):
self.inputs.append(np.empty([m, self.neurons[layer]]))
self.errors.append(np.empty([m, self.neurons[layer]]))
if (layer < self.L -1):
self.outputs.append(np.empty([m, self.neurons[layer]+1]))
else:
self.outputs.append(np.empty([m, self.neurons[layer]]))
#accumulate the cost function
J_history = np.zeros([iter_count, 1])
for i in range(iter_count):
self.feedforward(X)
J = self.cost(Y, self.outputs[self.L-1])
J_history[i, 0] = J
self.backpropagate(Y)
#plot the cost function to check the descent
plt.plot(J_history)
plt.show()
###################################################################################################
def cost(self, Y, H):
J = np.sum(np.sum(np.power((Y - H), 2), axis=0))/(2*m)
return J
###################################################################################################
def feedforward(self, X):
m = X.shape[0];
self.outputs[0] = np.concatenate( (np.ones([m, 1]), X), axis=1)
for i in range(1, self.L):
self.inputs[i] = np.dot( self.outputs[i-1], self.weights[i-1].T )
if (self.activation == 'sigmoid'):
output_temp = self.sigmoid(self.inputs[i])
elif (self.activation == 'tanh'):
output_temp = np.tanh(self.inputs[i])
if (i < self.L - 1):
self.outputs[i] = np.concatenate( (np.ones([m, 1]), output_temp), axis=1)
else:
self.outputs[i] = output_temp
###################################################################################################
def backpropagate(self, Y):
self.errors[self.L-1] = self.outputs[self.L-1] - Y
for i in range(self.L - 2, 0, -1):
if (self.activation == 'sigmoid'):
self.errors[i] = np.dot( self.errors[i+1], self.weights[i][:, 1:] ) * self.sigmoid_prime(self.inputs[i])
elif (self.activation == 'tanh'):
self.errors[i] = np.dot( self.errors[i+1], self.weights[i][:, 1:] ) * (1 - self.outputs[i][:, 1:]*self.outputs[i][:, 1:])
for i in range(0, self.L-1):
grad = np.dot(self.errors[i+1].T, self.outputs[i]) / m
self.weights[i] = self.weights[i] - self.rate*grad
###################################################################################################
def sigmoid(self, z):
s = 1.0/(1.0 + np.exp(-z))
return s
###################################################################################################
def sigmoid_prime(self, z):
s = self.sigmoid(z)*(1 - self.sigmoid(z))
return s
###################################################################################################
def predict(self, X, weights):
m = X.shape[0];
self.inputs = []
self.outputs = []
self.weights = weights
for layer in range(self.L):
self.inputs.append(np.empty([m, self.neurons[layer]]))
if (layer < self.L -1):
self.outputs.append(np.empty([m, self.neurons[layer]+1]))
else:
self.outputs.append(np.empty([m, self.neurons[layer]]))
self.feedforward(X)
return self.outputs[self.L-1]
###################################################################################################
# MAIN PART
activation1 = 'sigmoid' # the input should be scaled into [ 0..1]
activation2 = 'tanh' # the input should be scaled into [-1..1]
activation = activation1
net = Neuralnet([1, 6, 1], activation) # structure of the NN and its activation function
##########################################################################################
# TRAINING
m = 1000 #size of the training set
X = np.linspace(0, 4*math.pi, num = m).reshape(m, 1); # input training set
Y = np.sin(X) # target
kx = 0.1 # noise parameter
noise = (2.0*np.random.uniform(0, kx, m) - kx).reshape(m, 1)
Y = Y + noise # noisy target
# scaling of the target depending on the activation function
if (activation == 'sigmoid'):
Y_scaled = (Y/(1+kx) + 1)/2.0
elif (activation == 'tanh'):
Y_scaled = Y/(1+kx)
# number of the iteration for the training stage
iter_count = 20000
net.train(X, Y_scaled, iter_count) #training
# gained weights
trained_weights = net.weights
##########################################################################################
# PREDICTION
m_new = 40 #size of the prediction set
X_new = np.linspace(0, 4*math.pi, num = m_new).reshape(m_new, 1);
Y_new = net.predict(X_new, trained_weights) # prediction
#rescaling of the result
if (activation == 'sigmoid'):
Y_new = (2.0*Y_new - 1.0) * (1+kx)
elif (activation == 'tanh'):
Y_new = Y_new * (1+kx)
# visualization
plt.plot(X, Y)
plt.plot(X_new, Y_new, 'ro')
plt.show()
raw_input('press any key to exit')
UPDATE
I would like to take back the statement regarding the training method used in your code. The network can be indeed trained using only one sample per iteration. I got interesting results in online-training using both sigmoid and tanh activation functions:
Online-training using Sigmoid (cost function and prediction)
Online-training using Tanh (cost function and prediction)
As can be seen the choice of Sigmoid as activation function gives better performance. The cost function looks not that good as during the offline-training, but at least it tends to go down.
I plotted the cost function in your implementation, it looks pretty jerky as well:
May be it is a good idea to try your code with the sigmoid or even the ReLU function.
Here is the updated source code. To switch between online and offline training modes just change the method variable.
import numpy as np
import math
import matplotlib.pyplot as plt
class Neuralnet:
def __init__(self, neurons, activation):
self.weights = []
self.inputs = []
self.outputs = []
self.errors = []
self.rate = 0.2
self.activation = activation #sigmoid or tanh
self.neurons = neurons
self.L = len(self.neurons) #number of layers
eps = 0.12; #range for uniform distribution -eps..+eps
for layer in range(len(neurons)-1):
self.weights.append(np.random.uniform(-eps,eps,size=(neurons[layer+1], neurons[layer]+1)))
###################################################################################################
def train(self, X, Y, iter_count):
m = X.shape[0];
for layer in range(self.L):
self.inputs.append(np.empty([m, self.neurons[layer]]))
self.errors.append(np.empty([m, self.neurons[layer]]))
if (layer < self.L -1):
self.outputs.append(np.empty([m, self.neurons[layer]+1]))
else:
self.outputs.append(np.empty([m, self.neurons[layer]]))
#accumulate the cost function
J_history = np.zeros([iter_count, 1])
for i in range(iter_count):
self.feedforward(X)
J = self.cost(Y, self.outputs[self.L-1])
J_history[i, 0] = J
self.backpropagate(Y)
#plot the cost function to check the descent
#plt.plot(J_history)
#plt.show()
###################################################################################################
def cost(self, Y, H):
J = np.sum(np.sum(np.power((Y - H), 2), axis=0))/(2*m)
return J
###################################################################################################
def cost_online(self, min_x, max_x, iter_number):
h_arr = np.zeros([iter_number, 1])
y_arr = np.zeros([iter_number, 1])
for step in range(iter_number):
x = np.random.uniform(min_x, max_x, 1).reshape(1, 1)
self.feedforward(x)
h_arr[step, 0] = self.outputs[-1]
y_arr[step, 0] = np.sin(x)
J = np.sum(np.sum(np.power((y_arr - h_arr), 2), axis=0))/(2*iter_number)
return J
###################################################################################################
def feedforward(self, X):
m = X.shape[0];
self.outputs[0] = np.concatenate( (np.ones([m, 1]), X), axis=1)
for i in range(1, self.L):
self.inputs[i] = np.dot( self.outputs[i-1], self.weights[i-1].T )
if (self.activation == 'sigmoid'):
output_temp = self.sigmoid(self.inputs[i])
elif (self.activation == 'tanh'):
output_temp = np.tanh(self.inputs[i])
if (i < self.L - 1):
self.outputs[i] = np.concatenate( (np.ones([m, 1]), output_temp), axis=1)
else:
self.outputs[i] = output_temp
###################################################################################################
def backpropagate(self, Y):
self.errors[self.L-1] = self.outputs[self.L-1] - Y
for i in range(self.L - 2, 0, -1):
if (self.activation == 'sigmoid'):
self.errors[i] = np.dot( self.errors[i+1], self.weights[i][:, 1:] ) * self.sigmoid_prime(self.inputs[i])
elif (self.activation == 'tanh'):
self.errors[i] = np.dot( self.errors[i+1], self.weights[i][:, 1:] ) * (1 - self.outputs[i][:, 1:]*self.outputs[i][:, 1:])
for i in range(0, self.L-1):
grad = np.dot(self.errors[i+1].T, self.outputs[i]) / m
self.weights[i] = self.weights[i] - self.rate*grad
###################################################################################################
def sigmoid(self, z):
s = 1.0/(1.0 + np.exp(-z))
return s
###################################################################################################
def sigmoid_prime(self, z):
s = self.sigmoid(z)*(1 - self.sigmoid(z))
return s
###################################################################################################
def predict(self, X, weights):
m = X.shape[0];
self.inputs = []
self.outputs = []
self.weights = weights
for layer in range(self.L):
self.inputs.append(np.empty([m, self.neurons[layer]]))
if (layer < self.L -1):
self.outputs.append(np.empty([m, self.neurons[layer]+1]))
else:
self.outputs.append(np.empty([m, self.neurons[layer]]))
self.feedforward(X)
return self.outputs[self.L-1]
###################################################################################################
# MAIN PART
activation1 = 'sigmoid' #the input should be scaled into [0..1]
activation2 = 'tanh' #the input should be scaled into [-1..1]
activation = activation1
net = Neuralnet([1, 6, 1], activation) # structure of the NN and its activation function
method1 = 'online'
method2 = 'offline'
method = method1
kx = 0.1 #noise parameter
###################################################################################################
# TRAINING
if (method == 'offline'):
m = 1000 #size of the training set
X = np.linspace(0, 4*math.pi, num = m).reshape(m, 1); #input training set
Y = np.sin(X) #target
noise = (2.0*np.random.uniform(0, kx, m) - kx).reshape(m, 1)
Y = Y + noise #noisy target
#scaling of the target depending on the activation function
if (activation == 'sigmoid'):
Y_scaled = (Y/(1+kx) + 1)/2.0
elif (activation == 'tanh'):
Y_scaled = Y/(1+kx)
#number of the iteration for the training stage
iter_count = 20000
net.train(X, Y_scaled, iter_count) #training
elif (method == 'online'):
sampling_count = 100000 # number of samplings during the training stage
m = 1 #batch size
iter_count = sampling_count/m
for layer in range(net.L):
net.inputs.append(np.empty([m, net.neurons[layer]]))
net.errors.append(np.empty([m, net.neurons[layer]]))
if (layer < net.L -1):
net.outputs.append(np.empty([m, net.neurons[layer]+1]))
else:
net.outputs.append(np.empty([m, net.neurons[layer]]))
J_history = []
step_history = []
for i in range(iter_count):
X = np.random.uniform(0, 4*math.pi, m).reshape(m, 1)
Y = np.sin(X) #target
noise = (2.0*np.random.uniform(0, kx, m) - kx).reshape(m, 1)
Y = Y + noise #noisy target
#scaling of the target depending on the activation function
if (activation == 'sigmoid'):
Y_scaled = (Y/(1+kx) + 1)/2.0
elif (activation == 'tanh'):
Y_scaled = Y/(1+kx)
net.feedforward(X)
net.backpropagate(Y_scaled)
if (np.remainder(i, 1000) == 0):
J = net.cost_online(0, 4*math.pi, 1000)
J_history.append(J)
step_history.append(i)
plt.plot(step_history, J_history)
plt.title('Batch size ' + str(m) + ', rate ' + str(net.rate) + ', samples ' + str(sampling_count))
#plt.ylim([0, 0.1])
plt.show()
#gained weights
trained_weights = net.weights
##########################################################################################
# PREDICTION
m_new = 40 #size of the prediction set
X_new = np.linspace(0, 4*math.pi, num = m_new).reshape(m_new, 1);
Y_new = net.predict(X_new, trained_weights) #prediction
#rescaling of the result
if (activation == 'sigmoid'):
Y_new = (2.0*Y_new - 1.0) * (1+kx)
elif (activation == 'tanh'):
Y_new = Y_new * (1+kx)
#visualization
#fake sine curve to show the ideal signal
if (method == 'online'):
X = np.linspace(0, 4*math.pi, num = 100)
Y = np.sin(X)
plt.plot(X, Y)
plt.plot(X_new, Y_new, 'ro')
if (method == 'online'):
plt.title('Batch size ' + str(m) + ', rate ' + str(net.rate) + ', samples ' + str(sampling_count))
plt.ylim([-1.5, 1.5])
plt.show()
raw_input('press any key to exit')
Now I have some remarks to your current code:
Your sine function looks like this:
def sine_example():
net = Neuralnet([1, 6, 1])
for step in range(100000):
x = np.random.normal()
net.feedforward([x])
net.backpropagate([np.tanh(np.sin(x))])
net.feedforward([3])
print(net.outputs[-1])
I don't know why you use tanh in your target input. If you really want to use tanh of sine as target, you need to scale it to [-1..1], because tanh(sin(x)) returns values in range [-0.76..0.76].
The next thing is the range of your training set. You use x = np.random.normal() to generate the samples. Here is the distribution of such an input:
After it you want your network to predict the sine of 3, but the network has almost never seen this number during the training stage. I would use the uniform distribution in a wider range for sample generation instead.
这篇关于神经网络反向传播算法在Python中不起作用的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！
查看全文
