Python基于numpy灵活定义神经网络结构的方法
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''' 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:
'''示例二''' 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
如下实现方式为本人(@hhh5460)原创。 要点: dtype=object
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
下面完整代码是我学习斯坦福机器学习教程,完全自己敲出来的:
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
更多关于Python相关内容可查看本站专题:《Python数学运算技巧总结》、《Python数据结构与算法教程》、《Python函数使用技巧总结》、《Python字符串操作技巧汇总》及《Python入门与进阶经典教程》
希望本文所述对大家Python程序设计有所帮助。
相关推荐
cetrolchen 2020-08-14
georgesale 2020-06-27
hexianhao 2020-06-23
walegahaha 2020-06-15
皮皮虾大侠 2020-11-10
Site 2020-11-07
lwnylslwnyls 2020-11-06
86377811 2020-11-06
遇见人工智能 2020-11-03
xjtukuixing 2020-10-27
Evan 2020-10-26
xceman 2020-10-23
georgesale 2020-10-21
xceman 2020-10-15
WhiteHacker 2020-10-11
hexianhao 2020-10-06
大白机器人 2020-09-30
mpkno 2020-09-26
hiarxiaoliang 2020-09-23