DNN的BP算法Python简单实现
Posted
tags:
篇首语:本文由小常识网(cha138.com)小编为大家整理,主要介绍了DNN的BP算法Python简单实现相关的知识,希望对你有一定的参考价值。
BP算法是神经网络的基础,也是最重要的部分。由于误差反向传播的过程中,可能会出现梯度消失或者爆炸,所以需要调整损失函数。在LSTM中,通过sigmoid来实现三个门来解决记忆问题,用tensorflow实现的过程中,需要进行梯度修剪操作,以防止梯度爆炸。RNN的BPTT算法同样存在着这样的问题,所以步数超过5步以后,记忆效果大大下降。LSTM的效果能够支持到30多步数,太长了也不行。如果要求更长的记忆,或者考虑更多的上下文,可以把多个句子的LSTM输出组合起来作为另一个LSTM的输入。下面上传用Python实现的普通DNN的BP算法,激活为sigmoid.
字迹有些潦草,凑合用吧,习惯了手动绘图,个人习惯。后面的代码实现思路是最重要的:每个层有多个节点,层与层之间单向链接(前馈网络),因此数据结构可以设计为单向链表。实现的过程属于典型的递归,递归调用到最后一层后把每一层的back_weights反馈给上一层,直到推导结束。上传代码(未经过优化的代码):
测试代码:
import numpy as np
import NeuralNetWork as nw
if __name__ == ‘__main__‘:
print("test neural network")
data = np.array([[1, 0, 0, 0, 0, 0, 0, 0],
[0, 1, 0, 0, 0, 0, 0, 0],
[0, 0, 1, 0, 0, 0, 0, 0],
[0, 0, 0, 1, 0, 0, 0, 0],
[0, 0, 0, 0, 1, 0, 0, 0],
[0, 0, 0, 0, 0, 1, 0, 0],
[0, 0, 0, 0, 0, 0, 1, 0],
[0, 0, 0, 0, 0, 0, 0, 1]])
np.set_printoptions(precision=3, suppress=True)
for i in range(10):
network = nw.NeuralNetWork([8, 20, 8])
# 让输入数据与输出数据相等
network.fit(data, data, learning_rate=0.1, epochs=150)
print("\\n\\n", i, "result")
for item in data:
print(item, network.predict(item))
#NeuralNetWork.py
# encoding: utf-8 #NeuralNetWork.py import numpy as np; def logistic(inX): return 1 / (1+np.exp(-inX)) def logistic_derivative(x): return logistic(x) * (1 - logistic(x)) class Neuron: ‘‘‘ 构建神经元单元,每个单元都有如下属性:1.input;2.output;3.back_weight;4.deltas_item;5.weights. 每个神经元单元更新自己的weights,多个神经元构成layer,形成weights矩阵 ‘‘‘ def __init__(self,len_input): #输入的初始参数,随机取很小的值(<0.1) self.weights = np.random.random(len_input) * 0.1 #当前实例的输入 self.input = np.ones(len_input) #对下一层的输出值 self.output = 1.0 #误差项 self.deltas_item = 0.0 # 上一次权重增加的量,记录起来方便后面扩展时可考虑增加冲量 self.last_weight_add = 0 def calculate_output(self,x): #计算输出值 self.input = x; self.output = logistic(np.dot(self.weights,self.input)) return self.output def get_back_weight(self): #获取反馈差值 return self.weights * self.deltas_item def update_weight(self,target = 0,back_weight = 0,learning_rate=0.1,layer="OUTPUT"): #更新权重 if layer == "OUTPUT": self.deltas_item = (target - self.output) * logistic_derivative(self.input) elif layer == "HIDDEN": self.deltas_item = back_weight * logistic_derivative(self.input) delta_weight = self.input * self.deltas_item * learning_rate + 0.9 * self.last_weight_add #添加冲量 self.weights += delta_weight self.last_weight_add = delta_weight class NetLayer: ‘‘‘ 网络层封装,管理当前网络层的神经元列表 ‘‘‘ def __init__(self,len_node,in_count): ‘‘‘ :param len_node: 当前层的神经元数 :param in_count: 当前层的输入数 ‘‘‘ # 当前层的神经元列表 self.neurons = [Neuron(in_count) for _ in range(len_node)]; # 记录下一层的引用,方便递归操作 self.next_layer = None def calculate_output(self,inX): output = np.array([node.calculate_output(inX) for node in self.neurons]) if self.next_layer is not None: return self.next_layer.calculate_output(output) return output def get_back_weight(self): return sum([node.get_back_weight() for node in self.neurons]) def update_weight(self,learning_rate,target): layer = "OUTPUT" back_weight = np.zeros(len(self.neurons)) if self.next_layer is not None: back_weight = self.next_layer.update_weight(learning_rate,target) layer = "HIDDEN" for i,node in enumerate(self.neurons): target_item = 0 if len(target) <= i else target[i] node.update_weight(target = target_item,back_weight = back_weight[i],learning_rate=learning_rate,layer=layer) return self.get_back_weight() class NeuralNetWork: def __init__(self, layers): self.layers = [] self.construct_network(layers) pass def construct_network(self, layers): last_layer = None for i, layer in enumerate(layers): if i == 0: continue cur_layer = NetLayer(layer, layers[i - 1]) self.layers.append(cur_layer) if last_layer is not None: last_layer.next_layer = cur_layer last_layer = cur_layer def fit(self, x_train, y_train, learning_rate=0.1, epochs=100000, shuffle=False): ‘‘‘‘‘ 训练网络, 默认按顺序来训练 方法 1:按训练数据顺序来训练 方法 2: 随机选择测试 :param x_train: 输入数据 :param y_train: 输出数据 :param learning_rate: 学习率 :param epochs:权重更新次数 :param shuffle:随机取数据训练 ‘‘‘ indices = np.arange(len(x_train)) for _ in range(epochs): if shuffle: np.random.shuffle(indices) for i in indices: self.layers[0].calculate_output(x_train[i]) self.layers[0].update_weight(learning_rate, y_train[i]) pass def predict(self, x): return self.layers[0].calculate_output(x)
以上是关于DNN的BP算法Python简单实现的主要内容,如果未能解决你的问题,请参考以下文章