主要介绍了Python基于numpy灵活定义神经网络结构的方法,结合实例形式分析了神经网络结构的原理及Python
具体实现方法,涉及Python使用numpy扩展进行数学运算的相关操作技巧,需要的朋友可以参考下
本文实例讲述了Python基于numpy灵活定义神经网络结构的方法。分享给大家供大家参考,具体如下:
用numpy可以灵活定义神经网络结构,还可以应用numpy强大的矩阵运算功能!
一、用法
1). 定义一个三层神经网络:
'''示例一'''
bash
nn = NeuralNetworks([3,4,2]) # 定义神经网络
nn.fit(X,y) # 拟合
print(nn.predict(X)) #预测
说明:
输入层节点数目:3
隐藏层节点数目:4
输出层节点数目:2
2).定义一个五层神经网络:
bash
nn = NeuralNetworks([3,5,7,4,2]) # 定义神经网络
nn.fit(X,y) # 拟合
print(nn.predict(X)) #预测
说明:
输入层节点数目:3
隐藏层1节点数目:5
隐藏层2节点数目:7
隐藏层3节点数目:4
输出层节点数目:2
二、实现
要点: dtype=objec
bash
import numpy as np
class NeuralNetworks(object):
''''''
def __init__(self, n_layers=None, active_type=None, n_iter=10000, error=0.05, alpha=0.5, lamda=0.4):
'''搭建神经网络框架'''
# 各层节点数目 (向量)
self.n = np.array(n_layers) # 'n_layers必须为list类型,如:[3,4,2] 或 n_layers=[3,4,2]'
self.size = self.n.size # 层的总数
# 层 (向量)
self.z = np.empty(self.size, dtype=object) # 先占位(置空),dtype=object !如下皆然
self.a = np.empty(self.size, dtype=object)
self.data_a = np.empty(self.size, dtype=object)
# 偏置 (向量)
self.b = np.empty(self.size, dtype=object)
self.delta_b = np.empty(self.size, dtype=object)
# 权 (矩阵)
self.w = np.empty(self.size, dtype=object)
self.delta_w = np.empty(self.size, dtype=object)
# 填充
for i in range(self.size):
self.a[i] = np.zeros(self.n[i]) # 全零
self.z[i] = np.zeros(self.n[i]) # 全零
self.data_a[i] = np.zeros(self.n[i]) # 全零
if i < self.size - 1:
self.b[i] = np.ones(self.n[i+1]) # 全一
self.delta_b[i] = np.zeros(self.n[i+1]) # 全零
mu, sigma = 0, 0.1 # 均值、方差
self.w[i] = np.random.normal(mu, sigma, (self.n[i], self.n[i+1])) # # 正态分布随机化
self.delta_w[i] = np.zeros((self.n[i], self.n[i+1])) # 全零
完整的numpy反向传播代码如下
bash
import numpy as np
''' 参考:http://ufldl.stanford.edu/wiki/index.php/%E7%A5%9E%E7%BB%8F%E7%BD%91%E7%BB%9C
'''
class NeuralNetworks(object):
''''''
def __init__(self, n_layers=None, active_type=None, n_iter=10000, error=0.05, alpha=0.5, lamda=0.4):
'''搭建神经网络框架'''
self.n_iter = n_iter # 迭代次数
self.error = error # 允许最大误差
self.alpha = alpha # 学习速率
self.lamda = lamda # 衰减因子 # 此处故意拼写错误!
if n_layers is None:
raise '各层的节点数目必须设置!'
elif not isinstance(n_layers, list):
raise 'n_layers必须为list类型,如:[3,4,2] 或 n_layers=[3,4,2]'
# 节点数目 (向量)
self.n = np.array(n_layers)
self.size = self.n.size # 层的总数
# 层 (向量)
self.a = np.empty(self.size, dtype=object) # 先占位(置空),dtype=object !如下皆然
self.z = np.empty(self.size, dtype=object)
# 偏置 (向量)
self.b = np.empty(self.size, dtype=object)
self.delta_b = np.empty(self.size, dtype=object)
# 权 (矩阵)
self.w = np.empty(self.size, dtype=object)
self.delta_w = np.empty(self.size, dtype=object)
# 残差 (向量)
self.data_a = np.empty(self.size, dtype=object)
# 填充
for i in range(self.size):
self.a[i] = np.zeros(self.n[i]) # 全零
self.z[i] = np.zeros(self.n[i]) # 全零
self.data_a[i] = np.zeros(self.n[i]) # 全零
if i < self.size - 1:
self.b[i] = np.ones(self.n[i+1]) # 全一
self.delta_b[i] = np.zeros(self.n[i+1]) # 全零
mu, sigma = 0, 0.1 # 均值、方差
self.w[i] = np.random.normal(mu, sigma, (self.n[i], self.n[i+1])) # # 正态分布随机化
self.delta_w[i] = np.zeros((self.n[i], self.n[i+1])) # 全零
# 激活函数
self.active_functions = {
'sigmoid': self.sigmoid,
'tanh': self.tanh,
'radb': self.radb,
'line': self.line,
}
# 激活函数的导函数
self.derivative_functions = {
'sigmoid': self.sigmoid_d,
'tanh': self.tanh_d,
'radb': self.radb_d,
'line': self.line_d,
}
if active_type is None:
self.active_type = ['sigmoid'] * (self.size - 1) # 默认激活函数类型
else:
self.active_type = active_type
def sigmoid(self, z):
if np.max(z) > 600:
z[z.argmax()] = 600
return 1.0 / (1.0 + np.exp(-z))
def tanh(self, z):
return (np.exp(z) - np.exp(-z)) / (np.exp(z) + np.exp(-z))
def radb(self, z):
return np.exp(-z * z)
def line(self, z):
return z
def sigmoid_d(self, z):
return z * (1.0 - z)
def tanh_d(self, z):
return 1.0 - z * z
def radb_d(self, z):
return -2.0 * z * np.exp(-z * z)
def line_d(self, z):
return np.ones(z.size) # 全一
def forward(self, x):
'''正向传播(在线)'''
# 用样本 x 走一遍,刷新所有 z, a
self.a[0] = x
for i in range(self.size - 1):
self.z[i+1] = np.dot(self.a[i], self.w[i]) + self.b[i]
self.a[i+1] = self.active_functions[self.active_type[i]](self.z[i+1]) # 加了激活函数
def err(self, X, Y):
'''误差'''
last = self.size-1
err = 0.0
for x, y in zip(X, Y):
self.forward(x)
err += 0.5 * np.sum((self.a[last] - y)**2)
err /= X.shape[0]
err += sum([np.sum(w) for w in self.w[:last]**2])
return err
def backward(self, y):
'''反向传播(在线)'''
last = self.size - 1
# 用样本 y 走一遍,刷新所有delta_w, delta_b
self.data_a[last] = -(y - self.a[last]) * self.derivative_functions[self.active_type[last-1]](self.z[last]) # 加了激活函数的导函数
for i in range(last-1, 1, -1):
self.data_a[i] = np.dot(self.w[i], self.data_a[i+1]) * self.derivative_functions[self.active_type[i-1]](self.z[i]) # 加了激活函数的导函数
# 计算偏导
p_w = np.outer(self.a[i], self.data_a[i+1]) # 外积!感谢 numpy 的强大!
p_b = self.data_a[i+1]
# 更新 delta_w, delta_w
self.delta_w[i] = self.delta_w[i] + p_w
self.delta_b[i] = self.delta_b[i] + p_b
def update(self, n_samples):
'''更新权重参数'''
last = self.size - 1
for i in range(last):
self.w[i] -= self.alpha * ((1/n_samples) * self.delta_w[i] + self.lamda * self.w[i])
self.b[i] -= self.alpha * ((1/n_samples) * self.delta_b[i])
def fit(self, X, Y):
'''拟合'''
for i in range(self.n_iter):
# 用所有样本,依次
for x, y in zip(X, Y):
self.forward(x) # 前向,更新 a, z;
self.backward(y) # 后向,更新 delta_w, delta_b
# 然后,更新 w, b
self.update(len(X))
# 计算误差
err = self.err(X, Y)
if err < self.error:
break
# 整千次显示误差(否则太无聊!)
if i % 1000 == 0:
print('iter: {}, error: {}'.format(i, err))
def predict(self, X):
'''预测'''
last = self.size - 1
res = []
for x in X:
self.forward(x)
res.append(self.a[last])
return np.array(res)
if __name__ == '__main__':
nn = NeuralNetworks([2,3,4,3,1], n_iter=5000, alpha=0.4, lamda=0.3, error=0.06) # 定义神经网络
X = np.array([[0.,0.], # 准备数据
[0.,1.],
[1.,0.],
[1.,1.]])
y = np.array([0,1,1,0])
nn.fit(X,y) # 拟合
print(nn.predict(X)) # 预测