网络主要由输入层(两个数构成),中间层(多个神经元构成的隐藏层),以及输出层构成;
import copy, numpy as np
np.random.seed(0)
# compute sigmoid nonlinearity #定义sigmoid函数
def sigmoid(x):
output = 1/(1+np.exp(-x))
return output
# convert output of sigmoid function to its derivative #计算sigmoid函数的倒数
def sigmoid_output_to_derivative(output):
return output*(1-output)
# training dataset generation
int2binary = {} #用于将输入的整数转为计算机可运行的二进制数用
binary_dim = 8 #定义了二进制数的长度=8
largest_number = pow(2,binary_dim) #二进制数最大能取的数就=256喽
binary = np.unpackbits(
np.array([range(largest_number)],dtype=np.uint8).T,axis=1)
for i in range(largest_number): #将二进制数与十进制数做个一一对应关系
int2binary[i] = binary[i]
# input variables
alpha = 0.1 #反向传播时参数w更新的速度
input_dim = 2 #输入数据的维度,程序是实现两个数相加的
hidden_dim = 16 #隐藏层神经元个数=16
output_dim = 1 #输出结果值是1维的
# initialize neural network weights #初始化神经网络的权重参数
synapse_0 = 2*np.random.random((input_dim,hidden_dim)) - 1 #输入至神经元的w0,维度为2X16,取值约束在[-1,1]间
synapse_1 = 2*np.random.random((hidden_dim,output_dim)) - 1 #神经元至输出层的权重w1,维度为16X1,取值约束在[-1,1]间
synapse_h = 2*np.random.random((hidden_dim,hidden_dim)) - 1 #神经元前一时刻状态至当前状态权重wh,维度为16X16,取值约束在[-1,1]间
synapse_0_update = np.zeros_like(synapse_0) #构造与w0相同维度的矩阵,并初始化为全0;
synapse_1_update = np.zeros_like(synapse_1)
synapse_h_update = np.zeros_like(synapse_h)
# training logic
for j in range(10000): #模型迭代次数,可自行更改
# generate a simple addition problem (a + b = c)
a_int = np.random.randint(largest_number/2) # int version #约束初始化的输入加数a的值不超过128
a = int2binary[a_int] # binary encoding #将加数a的转为对应二进制数
b_int = np.random.randint(largest_number/2) # int version
b = int2binary[b_int] # binary encoding
# true answer
c_int = a_int + b_int #真实和
c = int2binary[c_int]
# where we'll store our best guess (binary encoded)
d = np.zeros_like(c) #用于存储预测的和
overallError = 0 #打印显示误差
layer_2_deltas = list() #反向求导用
layer_1_values = list()
layer_1_values.append(np.zeros(hidden_dim)) #先对隐藏层前一时刻状态初始化为0
# moving along the positions in the binary encoding
for position in range(binary_dim): #前向传播;二进制求和,低位在右,高位在左
# generate input and output
X = np.array([[a[binary_dim - position - 1],b[binary_dim - position - 1]]]) #输入的a与b(二进制形式)
y = np.array([[c[binary_dim - position - 1]]]).T #真实label值
# hidden layer (input ~+ prev_hidden)
layer_1 = sigmoid(np.dot(X,synapse_0) + np.dot(layer_1_values[-1],synapse_h)) # X*w0+RNN前一时刻状态值*wh
# output layer (new binary representation)
layer_2 = sigmoid(np.dot(layer_1,synapse_1)) #layer_1*w1
# did we miss?... if so, by how much?
layer_2_error = y - layer_2 #求误差
layer_2_deltas.append((layer_2_error)*sigmoid_output_to_derivative(layer_2)) #代价函数
overallError += np.abs(layer_2_error[0]) #误差,打印显示用
# decode estimate so we can print it out
d[binary_dim - position - 1] = np.round(layer_2[0][0]) #预测的和
# store hidden layer so we can use it in the next timestep
layer_1_values.append(copy.deepcopy(layer_1)) #深拷贝,将RNN模块状态值存储,用于反向传播
future_layer_1_delta = np.zeros(hidden_dim)
for position in range(binary_dim): #反向传播,计算从左到右,即二进制高位到低位
X = np.array([[a[position],b[position]]])
layer_1 = layer_1_values[-position-1]
prev_layer_1 = layer_1_values[-position-2]
# error at output layer
layer_2_delta = layer_2_deltas[-position-1]
# error at hidden layer
layer_1_delta = (future_layer_1_delta.dot(synapse_h.T) + layer_2_delta.dot(synapse_1.T)) * sigmoid_output_to_derivative(layer_1)
# let's update all our weights so we can try again
synapse_1_update += np.atleast_2d(layer_1).T.dot(layer_2_delta) #对w1进行更新
synapse_h_update += np.atleast_2d(prev_layer_1).T.dot(layer_1_delta) #对wh进行更新
synapse_0_update += X.T.dot(layer_1_delta) #对w0进行更新
future_layer_1_delta = layer_1_delta
synapse_0 += synapse_0_update * alpha
synapse_1 += synapse_1_update * alpha
synapse_h += synapse_h_update * alpha
synapse_0_update *= 0
synapse_1_update *= 0
synapse_h_update *= 0
# print out progress
if(j % 1000 == 0): #每1000次打印结果
print ("Error:" + str(overallError))
print ("Pred:" + str(d))
print ("True:" + str(c))
out = 0
for index,x in enumerate(reversed(d)):
out += x*pow(2,index)
print (str(a_int) + " + " + str(b_int) + " = " + str(out))
print ("------------")
用anaconda带的prompt运行结果如下,随着迭代次数加大,运行结果正确率提高;
由于python中的对象之间赋值时是按引用传递的,如需拷贝对象,需用到标准库中的copy模块;
copy.copy是浅拷贝,只拷贝父对象,不会拷贝对象的内部的子对象;
copy.deepcopy深拷贝,拷贝对象及其子对象;
1.程序一开始需要import copy模块;
2.常见的非线性变化函数用的sigmoid,其定义在程序一开始做了定义;
3.sigmoid函数的求导大概就是: 令out=1/(1+exp(-x));则out对x求导得到out'=exp(-x)/(
1+exp(-x))^2=out*(1-out);
4.由于RNN模块有对前一状态的记忆,因此会对前一时刻的状态值做考虑;
5.
np.zeros_like 其作用是构建相同维度的矩阵,并赋值为0;
6.由于RNN模块中,初始时刻的前一时刻的状态值未知,赋值为0;
7.检查自己的python版本,print是否加()