卷积神经网络参考:https://www.cnblogs.com/chensheng-zhou/p/6380738.html
LeNet神经网络参考:https://my.oschina.net/u/876354/blog/1632862
LeNet是一个最典型的卷积神经网络,由卷积层、池化层、全连接层组成。其中卷积层与池化层配合,组成多个卷积组,逐层提取特征,最终通过若干个全连接层完成分类,其结构如下图。
MNIST 数据集来自美国国家标准与技术研究所, National Institute of Standards and Technology (NIST). 训练集 (training set) 由来自 250 个不同人手写的数字构成, 其中 50% 是高中学生, 50% 来自人口普查局 (the Census Bureau) 的工作人员. 测试集(test set) 也是同样比例的手写数字数据
手写字体识别
# coding:utf8
import os
import cv2
import numpy as np
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
sess = tf.InteractiveSession()
def getTrain():
trai n =[[] ,[]] # 指定训练集的格式,一维为输入数据,一维为其标签
# 读取所有训练图像,作为训练集
train_roo t ="mnist_train"
labels = os.listdir(train_root)
for label in labels:
imgpaths = os.listdir(os.path.join(train_root ,label))
for imgname in imgpaths:
img = cv2.imread(os.path.join(train_root ,label ,imgname) ,0)
array = np.array(img).flatten() # 将二维图像平铺为一维图像
arra y =MaxMinNormalization(array)
train[0].append(array)
label_ = [0 ,0 ,0 ,0 ,0 ,0 ,0 ,0 ,0 ,0]
label_[int(label)] = 1
train[1].append(label_)
train = shuff(train)
return train
def getTest():
tes t =[[] ,[]] # 指定训练集的格式,一维为输入数据,一维为其标签
# 读取所有训练图像,作为训练集
test_roo t ="mnist_test"
labels = os.listdir(test_root)
for label in labels:
imgpaths = os.listdir(os.path.join(test_root ,label))
for imgname in imgpaths:
img = cv2.imread(os.path.join(test_root ,label ,imgname) ,0)
array = np.array(img).flatten() # 将二维图像平铺为一维图像
arra y =MaxMinNormalization(array)
test[0].append(array)
label_ = [0 ,0 ,0 ,0 ,0 ,0 ,0 ,0 ,0 ,0]
label_[int(label)] = 1
test[1].append(label_)
test = shuff(test)
return test[0] ,test[1]
def shuff(data):
tem p =[]
for i in range(len(data[0])):
temp.append([data[0][i] ,data[1][i]])
import random
random.shuffle(temp)
dat a =[[] ,[]]
for tt in temp:
data[0].append(tt[0])
data[1].append(tt[1])
return data
count = 0
def getBatchNum(batch_size ,maxNum):
global count
if count = =0:
coun t =coun t +batch_size
return 0 ,min(batch_size ,maxNum)
else:
temp = count
coun t =coun t +batch_size
if min(count ,maxNum )= =maxNum:
coun t =0
return getBatchNum(batch_size ,maxNum)
return temp ,min(count ,maxNum)
def MaxMinNormalization(x):
x = (x - np.min(x)) / (np.max(x) - np.min(x))
return x
# 1、权重初始化,偏置初始化
def weight_variable(shape):
initial = tf.truncated_normal(shape ,stddev=0.1 ) # 正太分布的标准差设为0.1
return tf.Variable(initial)
def bias_variable(shape):
initial = tf.constant(0.1 ,shape=shape)
return tf.Variable(initial)
# 2、卷积层和池化层也是接下来要重复使用的,因此也为它们定义创建函数
def conv2d(x, w):
return tf.nn.conv2d(x, w, strides=[1 ,1 ,1 ,1] ,padding='SAME') # 保证输出和输入是同样大小
def max_pool_2x2(x):
return tf.nn.max_pool(x, ksize=[1 ,2 ,2 ,1], strides=[1 ,2 ,2 ,1] ,padding='SAME')
iterNum = 1000
batch_siz e =1024
print("load train dataset.")
trai n =getTrain()
print("load test dataset.")
test0 ,test 1 =getTest()
# 3、参数
x = tf.placeholder(tf.float32, [None ,784], name="x-input")
y_ = tf.placeholder(tf.float32 ,[None ,10]) # 10列
# 4、第一层卷积,它由一个卷积接一个max pooling完成
w_conv1 = weight_variable([5 ,5 ,1 ,32])
b_conv1 = bias_variable([32]) # 每个输出通道都有一个对应的偏置量
x_image = tf.reshape(x ,[-1 ,28 ,28 ,1])
h_conv1 = tf.nn.relu(conv2d(x_image, w_conv1) + b_conv1) # 使用conv2d函数进行卷积操作,非线性处理
h_pool1 = max_pool_2x2(h_conv1) # 对卷积的输出结果进行池化操作
# 5、第二个和第一个一样,是为了构建一个更深的网络,把几个类似的堆叠起来
w_conv2 = weight_variable([5 ,5 ,32 ,64])
b_conv2 = bias_variable([64])
h_conv2 = tf.nn.relu(conv2d(h_pool1, w_conv2) + b_conv2 )# 输入的是第一层池化的结果
h_pool2 = max_pool_2x2(h_conv2)
# 6、密集连接层
# 把池化层输出的张量reshape(此函数可以重新调整矩阵的行、列、维数)成一些向量,加上偏置,然后对其使用Relu激活函数
w_fc1 = weight_variable([7 * 7 * 64, 1024])
b_fc1 = bias_variable([1024])
h_pool2_flat = tf.reshape(h_pool2, [-1 ,7 * 7 * 64])
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, w_fc1) + b_fc1)
# 7、使用dropout,防止过度拟合
keep_prob = tf.placeholder(tf.float32, name="keep_prob" )# placeholder是占位符
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
# 8、输出层,最后添加一个softmax层
w_fc2 = weight_variable([1024 ,10])
b_fc2 = bias_variable([10])
y_conv = tf.nn.softmax(tf.matmul(h_fc1_drop, w_fc2) + b_fc2, name="y-pred")
# 9、训练和评估模型
# 参数keep_prob控制dropout比例,然后每100次迭代输出一次日志
cross_entropy = tf.reduce_sum(-tf.reduce_sum(y_ * tf.log(y_conv) ,reduction_indices=[1]))
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
# 预测结果与真实值的一致性,这里产生的是一个bool型的向量
correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))
# 将bool型转换成float型,然后求平均值,即正确的比例
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
# 初始化所有变量,在2017年3月2号以后,用 tf.global_variables_initializer()替代tf.initialize_all_variables()
sess.run(tf.initialize_all_variables())
# 保存最后一个模型
saver = tf.train.Saver(max_to_keep=1)
for i in range(iterNum):
for j in range(int(len(train[1] ) /batch_size)):
imagesNu m =getBatchNum(batch_size ,len(train[1]))
batch = [train[0][imagesNum[0]:imagesNum[1]] ,train[1][imagesNum[0]:imagesNum[1]]]
train_step.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})
if i % 2 == 0:
train_accuracy = accuracy.eval(feed_dict={x: batch[0], y_: batch[1] ,keep_prob: 1.0})
print("Step %d ,training accuracy %g" % (i, train_accuracy))
print("test accuracy %f " % accuracy.eval(feed_dict={x: test0, y_ :test1, keep_prob: 1.0}))
# 保存模型于文件夹
saver.save(sess ,"save/model")
数据可视化
import tensorflow as tf
import numpy as np
import tkinter as tk
from tkinter import filedialog
from PIL import Image, ImageTk
from tkinter import filedialog
def creat_windows():
win = tk.Tk() # 创建窗口
sw = win.winfo_screenwidth()
sh = win.winfo_screenheight()
ww, wh = 400, 450
x, y = (s w -ww ) /2, (s h -wh ) /2
win.geometry("%dx%d+%d+%d " %(ww, wh, x, y- 40)) # 居中放置窗口
win.title('手写体识别') # 窗口命名
bg1_open = Image.open("timg.jpg").resize((300, 300))
bg1 = ImageTk.PhotoImage(bg1_open)
canvas = tk.Label(win, image=bg1)
canvas.pack()
var = tk.StringVar() # 创建变量文字
var.set('')
tk.Label(win, textvariable=var, bg='#C1FFC1', font=('宋体', 21), width=20, height=2).pack()
tk.Button(win, text='选择图片', width=20, height=2, bg='#FF8C00', command=lambda: main(var, canvas),
font=('圆体', 10)).pack()
win.mainloop()
def main(var, canvas):
file_path = filedialog.askopenfilename()
bg1_open = Image.open(file_path).resize((28, 28))
pic = np.array(bg1_open).reshape(784, )
bg1_resize = bg1_open.resize((300, 300))
bg1 = ImageTk.PhotoImage(bg1_resize)
canvas.configure(image=bg1)
canvas.image = bg1
init = tf.global_variables_initializer()
with tf.Session() as sess:
sess.run(init)
saver = tf.train.import_meta_graph('save/model.meta') # 载入模型结构
saver.restore(sess, 'save/model') # 载入模型参数
graph = tf.get_default_graph() # 加载计算图
x = graph.get_tensor_by_name("x-input:0") # 从模型中读取占位符变量
keep_prob = graph.get_tensor_by_name("keep_prob:0")
y_conv = graph.get_tensor_by_name("y-pred:0") # 关键的一句 从模型中读取占位符变量
prediction = tf.argmax(y_conv, 1)
predint = prediction.eval(feed_dict={x: [pic], keep_prob: 1.0}, session=sess) # feed_dict输入数据给placeholder占位符
answer = str(predint[0])
var.set("预测的结果是:" + answer)
if __name__ == "__main__":
creat_windows()
总结