版权声明:本文为博主原创文章,欢迎转载,并请注明出处。联系方式:460356155@qq.com

全连接神经网络是深度学习的基础,理解它就可以掌握深度学习的核心概念:前向传播、反向误差传递、权重、学习率等。这里先用python创建模型,用minist作为数据集进行训练。

定义3层神经网络:输入层节点28*28(对应minist图片像素数)、隐藏层节点300、输出层节点10(对应0-9个数字)。

网络的激活函数采用sigmoid,网络权重的初始化采用正态分布。

完整代码如下:

  1 # -*- coding:utf-8 -*-
  2 
  3 u"""全连接神经网络训练学习MINIST"""
  4 
  5 __author__ = 'zhengbiqing 460356155@qq.com'
  6 
  7 
  8 import numpy
  9 import scipy.special
 10 import scipy.misc
 11 from PIL import Image
 12 import matplotlib.pyplot
 13 import pylab
 14 import datetime
 15 from random import shuffle
 16 
 17 
 18 #是否训练网络
 19 LEARN = True
 20 
 21 #是否保存网络
 22 SAVE_PARA = False
 23 
 24 #网络节点数
 25 INPUT = 784
 26 HIDDEN = 300
 27 OUTPUT = 10
 28 
 29 #学习率和训练次数
 30 LR = 0.05
 31 EPOCH = 10
 32 
 33 #训练数据集文件
 34 TRAIN_FILE = 'mnist_train.csv'
 35 TEST_FILE = 'mnist_test.csv'
 36 
 37 #网络保存文件名
 38 WEIGHT_IH = "minist_fc_wih.npy"
 39 WEIGHT_HO = "minist_fc_who.npy"
 40 
 41 
 42 #神经网络定义
 43 class NeuralNetwork:
 44     def __init__(self, inport_nodes, hidden_nodes, output_nodes, learnning_rate):
 45         #神经网络输入层、隐藏层、输出层节点数
 46         self.inodes = inport_nodes
 47         self.hnodes = hidden_nodes
 48         self.onodes = output_nodes
 49 
 50         #神经网络训练学习率
 51         self.learnning_rate = learnning_rate
 52 
 53         #用均值为0,标准方差为连接数的-0.5次方的正态分布初始化权重
 54         #权重矩阵行列分别为hidden * input、 output * hidden,和ih、ho相反
 55         self.wih = numpy.random.normal(0.0, pow(self.hnodes, -0.5), (self.hnodes, self.inodes))
 56         self.who = numpy.random.normal(0.0, pow(self.onodes, -0.5), (self.onodes, self.hnodes))
 57 
 58         #sigmoid函数为激活函数
 59         self.active_fun = lambda x: scipy.special.expit(x)        
 60 
 61     #设置神经网络权重,在加载已训练的权重时调用
 62     def set_weight(self, wih, who):
 63         self.wih = wih
 64         self.who = who
 65 
 66     #前向传播,根据输入得到输出
 67     def get_outputs(self, input_list):
 68         # 把list转换为N * 1的矩阵,ndmin=2二维,T转制
 69         inputs = numpy.array(input_list, ndmin=2).T
 70 
 71         # 隐藏层输入 = W dot X,矩阵乘法
 72         hidden_inputs = numpy.dot(self.wih, inputs)
 73         hidden_outputs = self.active_fun(hidden_inputs)
 74 
 75         final_inputs = numpy.dot(self.who, hidden_outputs)
 76         final_outputs = self.active_fun(final_inputs)
 77 
 78         return inputs, hidden_outputs, final_outputs
 79 
 80     #网络训练,误差计算,误差反向分配更新网络权重
 81     def train(self, input_list, target_list):
 82         inputs, hidden_outputs, final_outputs = self.get_outputs(input_list)
 83 
 84         targets = numpy.array(target_list, ndmin=2).T
 85 
 86         #误差计算
 87         output_errors = targets - final_outputs
 88         hidden_errors = numpy.dot(self.who.T, output_errors)
 89 
 90         #连接权重更新
 91         self.who += numpy.dot(self.learnning_rate * output_errors * final_outputs * (1 - final_outputs), hidden_outputs.T)
 92         self.wih += numpy.dot(self.learnning_rate * hidden_errors * hidden_outputs * (1 - hidden_outputs), inputs.T)
 93         
 94 
 95 #图像像素值变换
 96 def vals2input(vals):
 97     #[0,255]的图像像素值转换为i[0.01,1],以便sigmoid函数作非线性变换
 98     return (numpy.asfarray(vals) / 255.0 * 0.99) + 0.01
 99 
100 
101 '''
102 训练网络
103 train:是否训练网络,如果不训练则直接加载已训练得到的网络权重
104 epoch:训练次数
105 save:是否保存训练结果,即网络权重
106 '''
107 def net_train(train, epochs, save):
108     if train:
109         with open(TRAIN_FILE, 'r') as train_file:
110             train_list = train_file.readlines()
111 
112         for epoch in range(epochs):
113             #打乱训练数据
114             shuffle(train_list)
115 
116             for data in train_list:
117                 all_vals = data.split(',')
118                 #图像数据为0~255,转换到0.01~1区间,以便激活函数更有效
119                 inputs = vals2input(all_vals[1:])
120 
121                 #标签,正确的为0.99,其他为0.01
122                 targets = numpy.zeros(OUTPUT) + 0.01
123                 targets[int(all_vals[0])] = 0.99
124 
125                 net.train(inputs, targets)
126 
127             #每个epoch结束后用测试集检查识别准确度
128             net_test(epoch)
129             print('')
130 
131         if save:
132             #保存连接权重
133             numpy.save(WEIGHT_IH, net.wih)
134             numpy.save(WEIGHT_HO, net.who)
135     else:
136         #不训练直接加载已保存的权重
137         wih = numpy.load(WEIGHT_IH)
138         who = numpy.load(WEIGHT_HO)
139         net.set_weight(wih, who)
140 
141 
142 '''
143 用测试集检查准确率
144 '''
145 def net_test(epoch):
146     with open(TEST_FILE, 'r') as test_file:
147         test_list = test_file.readlines()
148 
149     ok = 0
150     errlist = [0] * 10
151 
152     for data in test_list:
153         all_vals = data.split(',')
154         inputs = vals2input(all_vals[1:])
155         _, _, net_out = net.get_outputs(inputs)
156 
157         max = numpy.argmax(net_out)
158         if max == int(all_vals[0]):
159             ok += 1
160         else:
161             # 识别错误统计,每个数字识别错误计数
162             # print('target:', all_vals[0], 'net_out:', max)
163             errlist[int(all_vals[0])] += 1
164 
165     print('EPOCH: {epoch} score: {score}'.format(epoch=epoch, score = ok / len(test_list) * 100))
166     print('error list: ', errlist, ' total: ', sum(errlist))
167 
168 
169 #变换图片的尺寸,保存变换后的图片
170 def resize_img(filein, fileout, width, height, type):
171     img = Image.open(filein)
172     out = img.resize((width, height), Image.ANTIALIAS)
173     out.save(fileout, type)
174 
175 
176 #用训练得到的网络识别一个图片文件
177 def img_test(img_file):
178     file_name_list = img_file.split('.')
179     file_name, file_type = file_name_list[0], file_name_list[1]
180     out_file = file_name + 'out' + '.' + file_type
181     resize_img(img_file, out_file, 28, 28, file_type)
182 
183     img_array = scipy.misc.imread(out_file, flatten=True)
184     img_data = 255.0 - img_array.reshape(784)
185     img_data = (img_data / 255.0 * 0.99) + 0.01
186 
187     _, _, net_out = net.get_outputs(img_data)
188     max = numpy.argmax(net_out)
189     print('pic recognized as: ', max)
190 
191 
192 #显示数据集某个索引对应的图片
193 def img_show(train, index):
194     file = TRAIN_FILE if train else TEST_FILE
195     with open(file, 'r') as test_file:
196         test_list = test_file.readlines()
197 
198     all_values = test_list[index].split(',')
199     print('number is: ', all_values[0])
200 
201     image_array = numpy.asfarray(all_values[1:]).reshape((28, 28))
202     matplotlib.pyplot.imshow(image_array, cmap='Greys', interpolation='None')
203     pylab.show()
204 
205 
206 start_time = datetime.datetime.now()
207 
208 net = NeuralNetwork(INPUT, HIDDEN, OUTPUT, LR)
209 net_train(LEARN, EPOCH, SAVE_PARA)
210 
211 if not LEARN:
212     net_test(0)
213 else:
214     print('MINIST FC Train:', INPUT, HIDDEN, OUTPUT, 'LR:', LR, 'EPOCH:', EPOCH)
215     print('train spend time: ', datetime.datetime.now() - start_time)
216 
217 #用画图软件创建图片文件,由得到的网络进行识别
218 # img_test('t9.png')
219 
220 #显示minist中的某个图片
221 # img_show(True, 1)

784-300-10简单的全连接神经网络训练结果准确率基本在97.7%左右,运行结果如下:

EPOCH: 0 score: 95.96000000000001
error list:  [13, 21, 31, 28, 51, 61, 33, 66, 44, 56]  total:  404

EPOCH: 1 score: 96.77
error list:  [15, 19, 27, 63, 37, 37, 21, 40, 18, 46]  total:  323

EPOCH: 2 score: 97.25
error list:  [9, 17, 26, 26, 24, 56, 21, 41, 22, 33]  total:  275

EPOCH: 3 score: 97.82
error list:  [9, 16, 21, 18, 20, 18, 22, 21, 31, 42]  total:  218

EPOCH: 4 score: 97.54
error list:  [12, 23, 17, 25, 15, 34, 19, 25, 22, 54]  total:  246

EPOCH: 5 score: 97.78999999999999
error list:  [10, 16, 20, 23, 21, 32, 18, 31, 26, 24]  total:  221

EPOCH: 6 score: 97.6
error list:  [9, 13, 26, 34, 27, 26, 20, 28, 22, 35]  total:  240

EPOCH: 7 score: 97.74000000000001
error list:  [12, 8, 26, 29, 27, 26, 25, 20, 27, 26]  total:  226

EPOCH: 8 score: 97.77
error list:  [7, 10, 27, 16, 29, 28, 23, 29, 26, 28]  total:  223

EPOCH: 9 score: 97.99
error list:  [11, 10, 32, 17, 18, 24, 14, 22, 21, 32]  total:  201

MINIST FC Train: 784 300 10 LR: 0.05 EPOCH: 10
train spend time:  0:05:54.137925

Process finished with exit code 0