1 #定义初始化和偏置
2 def weight_varibale(shape):
3 # 随机初始化权重和偏置，权重和偏置后面会跟着训练自动优化
4 #weight = tf.Variable(tf.compat.v1.random_normal([784, 10], mean=0.0, stddev=1.0), name='weight')
5 weight = tf.Variable(tf.compat.v1.random_normal(shape=shape, mean=0.0, stddev=1.0), name='weight')
6 return weight
7 def bias_varible(shape):
8 #bias = tf.Variable(tf.constant([0], shape=[10]))
9 bias = tf.Variable(tf.constant([0.0], shape=shape))
10 return bias
11 # 预测Nonew个样本的输出结果matrix [None,784]*[784*10]+[10]=[None,10]，即矩阵[None,784]样本的特征*权重[784，10]+偏置[10]=预测结果[None,10]
12
13
14 #卷积模型
15 def model():
16 '''
17 自定义总卷积模型
18 :return:
19 '''
20 # 1、建立数据的占位符 ，X[None,784] y_true [None,10]
21 with tf.compat.v1.variable_scope('data'):
22 x = tf.compat.v1.placeholder(tf.float32, [None, 784])
23 y_true = tf.compat.v1.placeholder(tf.int32, [None, 10])
24 #2、卷积层一,卷积、激活、池化
25 # 初始化权重和偏重--由于这三个层都要用到权重和偏重，唯一的区别的形状不同，故我们设计一下函数专门用于初始化权重和偏置
26 with tf.compat.v1.variable_scope('conv1'):
27 #因为卷积接收的input格式为[batch,heigth,width,channel]，即要对其进进行形状的改变。这里784开平方等于28，故单个样本从二维转换为四维为[None,784]--[None,28,28,1]
28 #假设卷积神经有32个filter,5*5,stribes=1,padding='SAME', 则输入[None,28,28,32]--输出为[None,28,28,32]
29 #假设池化2*2，shribes2,padding='SAME',则[None,28,28,32]-->[None,14,14,32]
30 w_conv1=weight_varibale([5,5,1,32]) #随机初始化权重，输入32个filter,5*5,shrides步长=1,32个样本。。。即权重的形状
31 b_conv1=bias_varible([32])
32 x_reshape=tf.reshape(x,[-1,28,28,1])#这里的负一，相当于None
33
34 #卷积从[none,28,28,1]转成了[None,28,28,32]再加上偏置
35 conv1=tf.compat.v1.nn.conv2d(x_reshape,w_conv1,strides=[1,1,1,1],padding='SAME')+b_conv1
36 #激活
37 x_relu=tf.compat.v1.nn.relu(conv1)
38 #池化层 2*2,strides 2 [none,28,28,32]->[None,14,14,32]
39 x_pool1=tf.compat.v1.nn.max_pool(x_relu,ksize=[1,2,2,1],strides=[1,2,2,1],padding='SAME')
40
41 with tf.compat.v1.variable_scope('conv2'):
42 #3、卷积层二 假设卷积神经有64个filter,5*5*32,stribes=1,padding='SAME', 则输入[None,14,14,32]--输出为[None,14,14,64]
43 #假设池化2 * 2，shribes2, padding = 'SAME', 则[None,14,14,64]-->[None, 7, 7, 64]
44 w_conv2 = weight_varibale([5, 5, 32, 64]) # 随机初始化权重，输入64个filter,5*5*32,shrides步长=1,32个样本。。。即权重的形状..这里的32代表输入的通道数，64代表输出通道数
45 b_conv2 = bias_varible([64])
46 # 卷积从[none,14,14,32]转成了[None,14,14，64]再加上偏置
47 conv2 = tf.compat.v1.nn.conv2d(x_pool1, w_conv2, strides=[1, 1, 1, 1], padding='SAME') + b_conv2
48 # 激活
49 x_relu2 = tf.compat.v1.nn.relu(conv2)
50 # 池化层 2*2,strides 2 [None,14,14，64]->[None,7,7，64]
51 x_pool2 = tf.compat.v1.nn.max_pool(x_relu2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
52
53 #4、全连接层[None,7,7，64]，变成二维数组[none,[7*7*64]*偏置[[7*7*64,10]+[10]=[None,10]
54 with tf.compat.v1.variable_scope('conv_full'):
55 w_fc=weight_varibale([7*7*64,10])
56 b_fc=bias_varible([10])
57 #修改形状态[None,7,7，64]--》[none,7*7*64]
58 x_fc_reshape=tf.reshape(x_pool2,[-1,7*7*64])
59 #进行矩阵运算得出每个样本的10个结果
60 y_predict=tf.matmul(x_fc_reshape,w_fc)+b_fc
61 return x,y_true, y_predict
62 def con_fc():
63 # 使用占位符时，tersorflow2.X以上会出现tf.placeholder() is not compatible with eager execution报错，需要加下面这段语，避免程序报此错误。
64 tf.compat.v1.disable_eager_execution()
65 #获取真实的数据
66 mnist = input_data.read_data_sets("./tmp/mnist/", one_hot=True)
67
68 #定义模型，得出输出--即卷积输出
69 x,y_true, y_predict=model()
70 #进行交叉熵损失计算
71 with tf.compat.v1.variable_scope('soft_cross'):
72 # 返回交叉熵的列表结果，对交叉熵求平均值
73 loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y_true, logits=y_predict))
74
75 # 梯度下降求出损失
76 with tf.compat.v1.variable_scope('optimizer'):
77 train_op = tf.compat.v1.train.GradientDescentOptimizer(0.0001).minimize(loss)
78 # 5、计算准确率,预测准确置为1
79 with tf.compat.v1.variable_scope('acc'):
80 # equal_list None个样本[1，0，1，1,.....]
81 equal_list = tf.equal(tf.argmax(y_true, 1), tf.argmax(y_predict, 1))
82 accuray = tf.reduce_mean(tf.cast(equal_list, tf.float32))
83
84
85 # 因为有变量，故要定义初始化变量的op
86 init_op = tf.compat.v1.global_variables_initializer()
87 # 开启回话去训练
88 with tf.compat.v1.Session() as sess:
89 # 初始化变量
90 sess.run(init_op)
91 for i in range(1000):
92 mnist_x, mnist_y = mnist.train.next_batch(50)
93 # feed_dict实时提供的数据 x训练集，y为真实的目标值
94 # 运行op训练
95 #train_op 一个张量
96 #feed_dict
97 sess.run(train_op, feed_dict={x: mnist_x, y_true: mnist_y})
98 print('训练第%d步，准确率为:%f' % (i, sess.run(accuray, feed_dict={x: mnist_x, y_true: mnist_y})))
99 return None
100
101
102 if __name__ == '__main__':
103 con_fc()
