一、TensorFlow基础使用
创建、启动图
1 #创建一个常量op 2 m1 = tf.constant([[3,3]]) 3 #创建一个常量op 4 m2 = tf.constant([[2],[3]]) 5 #创建一个矩阵乘法op，把m1和m2传入 6 product = tf.matmul(m1,m2) 7 print(product)
1 #定义一个会话，自动启动默认图 2 sess = tf.Session() 3 #调用sess的run方法来执行矩阵乘法op 4 #run(product)触发了图中3个op 5 result = sess.run(product) 6 print(result) 7 sess.close()
需要建立一个会话，才能得到两个常量的乘积
1 #上面这段启动session可以改写成如下形式，此时就不用手动的去调用sess.close去关闭session 2 with tf.Session() as sess: 3 #调用sess的run方法来执行矩阵乘法op 4 #run(product)触发了图中3个op 5 result = sess.run(product) 6 print(result)
变量
1 import tensorflow as tf
1 #定义变量 2 x = tf.Variable([1,2]) 3 a = tf.constant([3,3]) 4 #定义减法和加法的op 5 sub = tf.subtract(x,a) 6 add = tf.add(x,sub) 7 8 #使用变量前要先初始化 9 init = tf.global_variables_initializer() 10 11 #定义会话 12 with tf.Session() as sess: 13 #先运行初始化变量 14 sess.run(init) 15 16 print(sess.run(sub)) 17 print(sess.run(add))
1 #变量可以初始化，初始化的值为0，名字为counter 2 state = tf.Variable(0,name='counter') 3 4 new_value = tf.add(state,1) 5 6 #调用赋值操作，不能直接用等号赋值 7 update = tf.assign(state,new_value) 8 9 #因为state是变量，所以需要初始化 10 init = tf.global_variables_initializer() 11 12 #定义会话 13 with tf.Session() as sess: 14 #先运行初始化变量 15 sess.run(init) 16 #查看初始化的值 17 print(sess.run(state)) 18 19 for _ in range(5): 20 sess.run(update) 21 print(sess.run(state))
1 import tensorflow as tf
1 #Fetch： 可以同时运行多个op 2 input1 = tf.constant(3.0) 3 input2 = tf.constant(2.0) 4 input3 = tf.constant(5.0) 5 6 add = tf.add(input2,input3) 7 mul = tf.multiply(input1,add) 8 9 with tf.Session() as sess: 10 #fetch多个op时，只要用中括号将多个op括起来即可 11 result = sess.run([add,mul]) 12 print(result)
1 #Feed 2 #创建占位符 ：placeholder 3 input1 = tf.placeholder(tf.float32) 4 input2 = tf.placeholder(tf.float32) 5 6 #可以在运行output时，再将input1和input2的值传入，传入时采用字典的形式 7 output = tf.multiply(input1,input2) 8 with tf.Session() as sess: 9 #传入时采用字典的形式 10 print(sess.run(output,feed_dict={input1:[8.],input2:[3.]}))
tensorflow的简单使用
1 import tensorflow as tf 2 import numpy as np
1 #使用numpy生成100个随机点 2 x_data = np.random.rand(100) 3 y_data = x_data*0.1 + 0.2 4 5 #构建一个线性模型 6 b = tf.Variable(0.) 7 k = tf.Variable(0.) 8 y = x_data*k + b 9 10 #使用tensorflow训练出模型，即得到k和b的值 11 #定义二次代价函数 12 loss = tf.reduce_mean(tf.square(y_data - y)) 13 #使用梯度下降函数来进行训练的优化器，下面学习率是0.2 14 optimizer = tf.train.GradientDescentOptimizer(0.2) 15 #最小化代价函数 16 train = optimizer.minimize(loss) 17 18 19 #初始化变量 20 init = tf.global_variables_initializer() 21 with tf.Session() as sess: 22 #先运行初始化变量 23 sess.run(init) 24 25 for step in range(201): 26 sess.run(train) 27 if step%20 ==0 : 28 print(step,sess.run([k,b])) 29 1 30 2 31 3 32 4 33 5 34 6 35 7 36 8 37 9 38 10 39 11 40 12 41 13 42 14 43 15 44 16 45 17 46 18 47 19 48 20 49 21 50 22 51 23 52 24 53 25 54 26 55 27 56 28 57 1 58 2 59 3 60 4 61 5 62 6 63 7 64 8 65 9 66 10 67 11 68 12 69 13 70 14 71 15 72 16 73 17 74 18 75 19 76 20 77 21 78 22 79 23 80 24 81 25 82 26 83 27 84 28 85 0 [0.0517033, 0.099609137] 86 20 [0.1019343, 0.1989941] 87 40 [0.10121739, 0.19936697] 88 60 [0.10076618, 0.19960159] 89 80 [0.10048222, 0.19974925] 90 100 [0.1003035, 0.19984218] 91 120 [0.10019101, 0.19990067] 92 140 [0.10012019, 0.19993751] 93 160 [0.10007564, 0.19996066] 94 180 [0.10004761, 0.19997525] 95 200 [0.10002997, 0.19998442] 96 97 tensorflow线性回归以及分类的简单实用，softmax介绍 98 99 非线性回归 100 101 import tensorflow as tf 102 import numpy as np 103 import matplotlib.pyplot as plt 104 1 105 2 106 3 107 1 108 2 109 3 110 #使用numpy生成200个随机点 111 #生成范围[-0.5,0.5]的200个点，下面生成200行1列的数据 112 x_data = np.linspace(-0.5,0.5,200)[:,np.newaxis] 113 noise = np.random.normal(0,0.02,x_data.shape) 114 y_data = np.square(x_data) + noise 115 116 117 #定义两个placeholder 118 #下面[]里面表示行不确定，列只有一列 119 x = tf.placeholder(tf.float32,[None,1]) 120 y = tf.placeholder(tf.float32,[None,1]) 121 122 #定义神经网络中间层 123 #神经元定义为[1,10],是因为输入的神经元是1个，中间层的神经元定义了10个 124 Weight_L1 = tf.Variable(tf.random_normal([1,10])) 125 biases_L1 = tf.Variable(tf.zeros([1,10])) 126 Wx_plus_b_L1 = tf.matmul(x,Weight_L1) + biases_L1 127 L1 = tf.nn.tanh(Wx_plus_b_L1) 128 129 #定义神经网络输出层 130 Weight_L2 = tf.Variable(tf.random_normal([10,1])) 131 biases_L2 = tf.Variable(tf.zeros([1,1])) 132 Wx_plus_b_L2 = tf.matmul(L1,Weight_L2) + biases_L2 133 prediction = tf.nn.tanh(Wx_plus_b_L2) 134 135 136 #二次代价函数 137 loss = tf.reduce_mean(tf.square(y-prediction)) 138 139 140 #使用梯度下降法训练 141 train_step = tf.train.GradientDescentOptimizer(0.1).minimize(loss) 142 143 144 with tf.Session() as sess: 145 #变量初始化 146 sess.run(tf.global_variables_initializer()) 147 for _ in range(2000): 148 sess.run(train_step,feed_dict={x:x_data,y:y_data}) 149 150 #获得预测值 151 prediction_value = sess.run(prediction, feed_dict={x:x_data}) 152 153 #画图 154 plt.figure() 155 plt.scatter(x_data,y_data) 156 plt.plot(x_data,prediction_value,'r-',lw=5) 157 plt.show()
MNIST数据集分类简单版
1 import tensorflow as tf 2 from tensorflow.examples.tutorials.mnist import input_data
1 #载入数据集 2 #one_hot是将数据改成0-1的格式 3 mnist = input_data.read_data_sets("MNIST_data",one_hot=True) 4 5 #每个批次的大小 6 batch_size = 100 7 8 #计算一共有多少个批次 9 n_batch = mnist.train.num_examples // batch_size 10 11 #定义两个placeholder,一个图片是28x28=784 12 x = tf.placeholder(tf.float32,[None,784]) 13 y = tf.placeholder(tf.float32,[None,10]) 14 15 #创建一个简单的神经网络 16 Weight_L1 = tf.Variable(tf.zeros([784,10])) 17 biases_L1 = tf.Variable(tf.zeros([10])) 18 Wx_plus_b_L1 = tf.matmul(x,Weight_L1) + biases_L1 19 prediction = tf.nn.softmax(Wx_plus_b_L1) 20 21 22 23 #二次代价函数 24 loss = tf.reduce_mean(tf.square(y-prediction)) 25 26 #使用梯度下降法训练 27 train_step = tf.train.GradientDescentOptimizer(0.2).minimize(loss) 28 29 #初始化变量 30 init = tf.global_variables_initializer() 31 32 #结果存在一个布尔型的列表中 33 correct_prediction = tf.equal(tf.argmax(y,1),tf.argmax(prediction,1)) #argmax返回一个张量中最大值存在的位置 34 35 #求准确率 36 accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32)) #将布尔型转换为float32, true->1 false ->0 37 38 with tf.Session() as sess: 39 sess.run(init) 40 41 for epoch in range(21): 42 for batch in range (n_batch): 43 #获取每个batch的图片 44 batch_xs,batch_ys = mnist.train.next_batch(batch_size) 45 sess.run(train_step,feed_dict={x:batch_xs,y:batch_ys}) 46 47 acc = sess.run(accuracy,feed_dict={x:mnist.test.images,y:mnist.test.labels}) 48 print("Iter " + str(epoch) + ",Testing Accuracy "+ str(acc))
1 #第三章作业，修改神经网络使得识别率达到95%以上 2 #载入数据集 3 #one_hot是将数据改成0-1的格式 4 mnist = input_data.read_data_sets("MNIST_data",one_hot=True) 5 6 #每个批次的大小 7 batch_size = 100 8 9 #计算一共有多少个批次 10 n_batch = mnist.train.num_examples // batch_size 11 12 #定义两个placeholder,一个图片是28x28=784 13 x = tf.placeholder(tf.float32,[None,784]) 14 y = tf.placeholder(tf.float32,[None,10]) 15 16 #创建一个简单的神经网络 17 Weight_L1 = tf.Variable(tf.zeros([784,20])) 18 biases_L1 = tf.Variable(tf.zeros([20])) 19 Wx_plus_b_L1 = tf.matmul(x,Weight_L1) + biases_L1 20 L1 = tf.nn.tanh(Wx_plus_b_L1) 21 22 23 #定义神经网络输出层 24 Weight_L2 = tf.Variable(tf.random_normal([20,10])) 25 biases_L2 = tf.Variable(tf.zeros([10])) 26 Wx_plus_b_L2 = tf.matmul(L1,Weight_L2) + biases_L2 27 prediction = tf.nn.softmax(Wx_plus_b_L2) 28 29 #二次代价函数 30 loss = tf.reduce_mean(tf.square(y-prediction)) 31 32 #使用梯度下降法训练 33 train_step = tf.train.GradientDescentOptimizer(0.2).minimize(loss) 34 35 #初始化变量 36 init = tf.global_variables_initializer() 37 38 #结果存在一个布尔型的列表中 39 correct_prediction = tf.equal(tf.argmax(y,1),tf.argmax(prediction,1)) #argmax返回一个张量中最大值存在的位置 40 41 #求准确率 42 accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32)) #将布尔型转换为float32, true->1 false ->0 43 44 with tf.Session() as sess: 45 sess.run(init) 46 47 for epoch in range(100): 48 for batch in range (n_batch): 49 #获取每个batch的图片 50 batch_xs,batch_ys = mnist.train.next_batch(batch_size) 51 sess.run(train_step,feed_dict={x:batch_xs,y:batch_ys}) 52 53 acc = sess.run(accuracy,feed_dict={x:mnist.test.images,y:mnist.test.labels}) 54 print("Iter " + str(epoch) + ",Testing Accuracy "+ str(acc))
二、交叉熵，过拟合，dropout以及tensorflow中各种优化器的介绍
交叉熵
1 import tensorflow as tf 2 from tensorflow.examples.tutorials.mnist import input_data
1 #将二次代价函数替换成交叉熵 2 # tf.nn.sigmod_cross_entropy_with_logits() 来表示跟sigmod搭配使用的交叉熵 3 # tf.nn.softmax_cross_entropy_with_logits() 来表示跟softmax搭配使用的交叉熵 4 #载入数据集 5 #one_hot是将数据改成0-1的格式 6 mnist = input_data.read_data_sets("MNIST_data",one_hot=True) 7 8 #每个批次的大小 9 batch_size = 100 10 11 #计算一共有多少个批次 12 n_batch = mnist.train.num_examples // batch_size 13 14 #定义两个placeholder,一个图片是28x28=784 15 x = tf.placeholder(tf.float32,[None,784]) 16 y = tf.placeholder(tf.float32,[None,10]) 17 18 #创建一个简单的神经网络 19 Weight_L1 = tf.Variable(tf.zeros([784,10])) 20 biases_L1 = tf.Variable(tf.zeros([10])) 21 Wx_plus_b_L1 = tf.matmul(x,Weight_L1) + biases_L1 22 prediction = tf.nn.softmax(Wx_plus_b_L1) 23 24 25 26 #二次代价函数 27 #loss = tf.reduce_mean(tf.square(y-prediction)) 28 29 #交叉熵 30 #首先看输入logits，它的shape是[batch_size, num_classes] ，一般来讲，就是神经网络最后一层的输入z。 31 #另外一个输入是labels，它的shape也是[batch_size, num_classes]，就是我们神经网络期望的输出。 32 loss=tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=prediction)) 33 #使用梯度下降法训练 34 train_step = tf.train.GradientDescentOptimizer(0.2).minimize(loss) 35 36 #初始化变量 37 init = tf.global_variables_initializer() 38 39 #结果存在一个布尔型的列表中 40 correct_prediction = tf.equal(tf.argmax(y,1),tf.argmax(prediction,1)) #argmax返回一个张量中最大值存在的位置 41 42 #求准确率 43 accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32)) #将布尔型转换为float32, true->1 false ->0 44 45 with tf.Session() as sess: 46 sess.run(init) 47 48 for epoch in range(21): 49 for batch in range (n_batch): 50 #获取每个batch的图片 51 batch_xs,batch_ys = mnist.train.next_batch(batch_size) 52 sess.run(train_step,feed_dict={x:batch_xs,y:batch_ys}) 53 54 acc = sess.run(accuracy,feed_dict={x:mnist.test.images,y:mnist.test.labels}) 55 print("Iter " + str(epoch) + ",Testing Accuracy "+ str(acc))
dropout
1 import tensorflow as tf 2 from tensorflow.examples.tutorials.mnist import input_data
1 #将二次代价函数替换成交叉熵 2 # tf.nn.sigmod_cross_entropy_with_logits() 来表示跟sigmod搭配使用的交叉熵 3 # tf.nn.softmax_cross_entropy_with_logits() 来表示跟softmax搭配使用的交叉熵 4 #载入数据集 5 #one_hot是将数据改成0-1的格式 6 mnist = input_data.read_data_sets("MNIST_data",one_hot=True) 7 8 #每个批次的大小 9 batch_size = 100 10 11 #计算一共有多少个批次 12 n_batch = mnist.train.num_examples // batch_size 13 14 #定义两个placeholder,一个图片是28x28=784 15 x = tf.placeholder(tf.float32,[None,784]) 16 y = tf.placeholder(tf.float32,[None,10]) 17 keep_prob = tf.placeholder(tf.float32) #dropout的参数，表示多少神经元被激活,1 表示所有神经元都是工作的 18 19 #创建一个简单的神经网络 20 Weight_L1 = tf.Variable(tf.truncated_normal([784,2000],stddev=0.1)) 21 biases_L1 = tf.Variable(tf.zeros([2000])+0.1) 22 Wx_plus_b_L1 = tf.nn.tanh(tf.matmul(x,Weight_L1) + biases_L1) 23 L1_drop = tf.nn.dropout(Wx_plus_b_L1,keep_prob) 24 25 26 Weight_L2 = tf.Variable(tf.truncated_normal([2000,2000],stddev=0.1)) 27 biases_L2 = tf.Variable(tf.zeros([2000])+0.1) 28 Wx_plus_b_L2 = tf.nn.tanh(tf.matmul(L1_drop,Weight_L2) + biases_L2) 29 L2_drop = tf.nn.dropout(Wx_plus_b_L2,keep_prob) 30 31 32 Weight_L3 = tf.Variable(tf.truncated_normal([2000,1000],stddev=0.1)) 33 biases_L3 = tf.Variable(tf.zeros([1000])+0.1) 34 Wx_plus_b_L3 = tf.nn.tanh(tf.matmul(L2_drop,Weight_L3) + biases_L3) 35 L3_drop = tf.nn.dropout(Wx_plus_b_L3,keep_prob) 36 37 Weight_L4 = tf.Variable(tf.truncated_normal([1000,10],stddev=0.1)) 38 biases_L4 = tf.Variable(tf.zeros([10])+0.1) 39 prediction = tf.nn.softmax(tf.matmul(L3_drop,Weight_L4) + biases_L4) 40 41 #二次代价函数 42 #loss = tf.reduce_mean(tf.square(y-prediction)) 43 44 #交叉熵 45 #首先看输入logits，它的shape是[batch_size, num_classes] ，一般来讲，就是神经网络最后一层的输入z。 46 #另外一个输入是labels，它的shape也是[batch_size, num_classes]，就是我们神经网络期望的输出。 47 loss=tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=prediction)) 48 #使用梯度下降法训练 49 train_step = tf.train.GradientDescentOptimizer(0.2).minimize(loss) 50 51 #初始化变量 52 init = tf.global_variables_initializer() 53 54 #结果存在一个布尔型的列表中 55 correct_prediction = tf.equal(tf.argmax(y,1),tf.argmax(prediction,1)) #argmax返回一个张量中最大值存在的位置 56 57 #求准确率 58 accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32)) #将布尔型转换为float32, true->1 false ->0 59 60 with tf.Session() as sess: 61 sess.run(init) 62 63 for epoch in range(21): 64 for batch in range (n_batch): 65 #获取每个batch的图片 66 batch_xs,batch_ys = mnist.train.next_batch(batch_size) 67 sess.run(train_step,feed_dict={x:batch_xs,y:batch_ys,keep_prob:1.0}) 68 69 test_acc = sess.run(accuracy,feed_dict={x:mnist.test.images,y:mnist.test.labels,keep_prob:1.0}) 70 train_acc = sess.run(accuracy,feed_dict={x:mnist.train.images,y:mnist.train.labels,keep_prob:1.0}) 71 72 print("Iter " + str(epoch) + ",Testing Accuracy "+ str(test_acc)+", training Accuracy "+str(train_acc))
1 import tensorflow as tf 2 from tensorflow.examples.tutorials.mnist import input_data
1 #将二次代价函数替换成交叉熵 2 # tf.nn.sigmod_cross_entropy_with_logits() 来表示跟sigmod搭配使用的交叉熵 3 # tf.nn.softmax_cross_entropy_with_logits() 来表示跟softmax搭配使用的交叉熵 4 #载入数据集 5 #one_hot是将数据改成0-1的格式 6 mnist = input_data.read_data_sets("MNIST_data",one_hot=True) 7 8 #每个批次的大小 9 batch_size = 100 10 11 #计算一共有多少个批次 12 n_batch = mnist.train.num_examples // batch_size 13 14 #定义两个placeholder,一个图片是28x28=784 15 x = tf.placeholder(tf.float32,[None,784]) 16 y = tf.placeholder(tf.float32,[None,10]) 17 18 #创建一个简单的神经网络 19 Weight_L1 = tf.Variable(tf.zeros([784,10])) 20 biases_L1 = tf.Variable(tf.zeros([10])) 21 Wx_plus_b_L1 = tf.matmul(x,Weight_L1) + biases_L1 22 prediction = tf.nn.softmax(Wx_plus_b_L1) 23 24 25 26 #二次代价函数 27 #loss = tf.reduce_mean(tf.square(y-prediction)) 28 29 #交叉熵 30 #首先看输入logits，它的shape是[batch_size, num_classes] ，一般来讲，就是神经网络最后一层的输入z。 31 #另外一个输入是labels，它的shape也是[batch_size, num_classes]，就是我们神经网络期望的输出。 32 loss=tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=prediction)) 33 34 #使用梯度下降法训练 35 #train_step = tf.train.GradientDescentOptimizer(0.2).minimize(loss) 36 #使用其他优化器 37 train_step = tf.train.AdamOptimizer(1e-2).minimize(loss) 38 39 #初始化变量 40 init = tf.global_variables_initializer() 41 42 #结果存在一个布尔型的列表中 43 correct_prediction = tf.equal(tf.argmax(y,1),tf.argmax(prediction,1)) #argmax返回一个张量中最大值存在的位置 44 45 #求准确率 46 accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32)) #将布尔型转换为float32, true->1 false ->0 47 48 with tf.Session() as sess: 49 sess.run(init) 50 51 for epoch in range(21): 52 for batch in range (n_batch): 53 #获取每个batch的图片 54 batch_xs,batch_ys = mnist.train.next_batch(batch_size) 55 sess.run(train_step,feed_dict={x:batch_xs,y:batch_ys}) 56 57 acc = sess.run(accuracy,feed_dict={x:mnist.test.images,y:mnist.test.labels}) 58 print("Iter " + str(epoch) + ",Testing Accuracy "+ str(acc))
继续优化网络，提高识别率
1 import tensorflow as tf 2 from tensorflow.examples.tutorials.mnist import input_data
1 #将二次代价函数替换成交叉熵 2 # tf.nn.sigmod_cross_entropy_with_logits() 来表示跟sigmod搭配使用的交叉熵 3 # tf.nn.softmax_cross_entropy_with_logits() 来表示跟softmax搭配使用的交叉熵 4 #载入数据集 5 #one_hot是将数据改成0-1的格式 6 mnist = input_data.read_data_sets("MNIST_data",one_hot=True) 7 8 #每个批次的大小 9 batch_size = 100 10 11 #计算一共有多少个批次 12 n_batch = mnist.train.num_examples // batch_size 13 14 #定义两个placeholder,一个图片是28x28=784 15 x = tf.placeholder(tf.float32,[None,784]) 16 y = tf.placeholder(tf.float32,[None,10]) 17 keep_prob = tf.placeholder(tf.float32) #dropout的参数，表示多少神经元被激活,1 表示所有神经元都是工作的 18 #学习率的变量 19 lr = tf.Variable(0.001,dtype = tf.float32) 20 21 #创建一个简单的神经网络 22 Weight_L1 = tf.Variable(tf.truncated_normal([784,500],stddev=0.1)) 23 biases_L1 = tf.Variable(tf.zeros([500])+0.1) 24 Wx_plus_b_L1 = tf.nn.tanh(tf.matmul(x,Weight_L1) + biases_L1) 25 L1_drop = tf.nn.dropout(Wx_plus_b_L1,keep_prob) 26 27 28 Weight_L2 = tf.Variable(tf.truncated_normal([500,300],stddev=0.1)) 29 biases_L2 = tf.Variable(tf.zeros([300])+0.1) 30 Wx_plus_b_L2 = tf.nn.tanh(tf.matmul(L1_drop,Weight_L2) + biases_L2) 31 L2_drop = tf.nn.dropout(Wx_plus_b_L2,keep_prob) 32 33 34 Weight_L3 = tf.Variable(tf.truncated_normal([300,10],stddev=0.1)) 35 biases_L3 = tf.Variable(tf.zeros([10])+0.1) 36 #Wx_plus_b_L3 = tf.nn.tanh(tf.matmul(L2_drop,Weight_L3) + biases_L3) 37 #L3_drop = tf.nn.dropout(Wx_plus_b_L3,keep_prob) 38 39 #Weight_L4 = tf.Variable(tf.truncated_normal([300,10],stddev=0.1)) 40 #biases_L4 = tf.Variable(tf.zeros([10])+0.1) 41 prediction = tf.nn.softmax(tf.matmul(L2_drop,Weight_L3) + biases_L3) 42 43 #二次代价函数 44 #loss = tf.reduce_mean(tf.square(y-prediction)) 45 46 #交叉熵 47 #首先看输入logits，它的shape是[batch_size, num_classes] ，一般来讲，就是神经网络最后一层的输入z。 48 #另外一个输入是labels，它的shape也是[batch_size, num_classes]，就是我们神经网络期望的输出。 49 loss=tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=prediction)) 50 #使用梯度下降法训练 51 #train_step = tf.train.GradientDescentOptimizer(0.2).minimize(loss) 52 #使用其他优化器 53 train_step = tf.train.AdamOptimizer(lr).minimize(loss) 54 55 56 #初始化变量 57 init = tf.global_variables_initializer() 58 59 #结果存在一个布尔型的列表中 60 correct_prediction = tf.equal(tf.argmax(y,1),tf.argmax(prediction,1)) #argmax返回一个张量中最大值存在的位置 61 62 #求准确率 63 accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32)) #将布尔型转换为float32, true->1 false ->0 64 65 with tf.Session() as sess: 66 sess.run(init) 67 68 for epoch in range(51): 69 sess.run(tf.assign(lr,0.001 * (0.95 ** epoch))) 70 for batch in range (n_batch): 71 #获取每个batch的图片 72 batch_xs,batch_ys = mnist.train.next_batch(batch_size) 73 sess.run(train_step,feed_dict={x:batch_xs,y:batch_ys,keep_prob:1.0}) 74 75 learning_rate = sess.run(lr) 76 77 test_acc = sess.run(accuracy,feed_dict={x:mnist.test.images,y:mnist.test.labels,keep_prob:1.0}) 78 train_acc = sess.run(accuracy,feed_dict={x:mnist.train.images,y:mnist.train.labels,keep_prob:1.0}) 79 80 print("Iter " + str(epoch) + ",Testing Accuracy "+ str(test_acc)+", training Accuracy "+str(train_acc))
三、tensorboard可视化
tensorboard网络结构
1 import tensorflow as tf 2 from tensorflow.examples.tutorials.mnist import input_data
1 #以3-2的程序为例介绍tensorboard的用法 2 #载入数据集 3 #one_hot是将数据改成0-1的格式 4 mnist = input_data.read_data_sets("MNIST_data",one_hot=True) 5 6 #每个批次的大小 7 batch_size = 100 8 9 #计算一共有多少个批次 10 n_batch = mnist.train.num_examples // batch_size 11 12 13 #*****要可视化网络结果，必须要用到命名空间*********** 14 with tf.name_scope('input'): 15 #定义两个placeholder,一个图片是28x28=784 16 x = tf.placeholder(tf.float32,[None,784],name='x-input') 17 y = tf.placeholder(tf.float32,[None,10],name='y-input') 18 19 #创建一个简单的神经网络 20 with tf.name_scope('layer'): 21 with tf.name_scope('wights'): 22 Weight_L1 = tf.Variable(tf.zeros([784,10]),name='W') 23 with tf.name_scope('biases'): 24 biases_L1 = tf.Variable(tf.zeros([10]),name='b') 25 with tf.name_scope('wx_plus_b'): 26 Wx_plus_b_L1 = tf.matmul(x,Weight_L1) + biases_L1 27 with tf.name_scope('softmax'): 28 prediction = tf.nn.softmax(Wx_plus_b_L1) 29 30 31 32 #二次代价函数 33 #loss = tf.reduce_mean(tf.square(y-prediction)) 34 with tf.name_scope('loss'): 35 loss=tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=prediction)) 36 #使用梯度下降法训练 37 with tf.name_scope('train'): 38 train_step = tf.train.GradientDescentOptimizer(0.2).minimize(loss) 39 40 #初始化变量 41 init = tf.global_variables_initializer() 42 43 #结果存在一个布尔型的列表中 44 with tf.name_scope('accuracy'): 45 with tf.name_scope('correct_prediction'): 46 correct_prediction = tf.equal(tf.argmax(y,1),tf.argmax(prediction,1)) #argmax返回一个张量中最大值存在的位置 47 with tf.name_scope('accuracy'): 48 #求准确率 49 accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32)) #将布尔型转换为float32, true->1 false ->0 50 51 with tf.Session() as sess: 52 sess.run(init) 53 writer = tf.summary.FileWriter('C:/Users/zgyxf183/Desktop/logs',sess.graph) 54 for epoch in range(1): 55 for batch in range (n_batch): 56 #获取每个batch的图片 57 batch_xs,batch_ys = mnist.train.next_batch(batch_size) 58 sess.run(train_step,feed_dict={x:batch_xs,y:batch_ys}) 59 60 acc = sess.run(accuracy,feed_dict={x:mnist.test.images,y:mnist.test.labels}) 61 print("Iter " + str(epoch) + ",Testing Accuracy "+ str(acc))
tensorboard网络运行
1 import tensorflow as tf 2 from tensorflow.examples.tutorials.mnist import input_data
1 #载入数据集 2 mnist = input_data.read_data_sets("MNIST_data",one_hot=True) 3 4 #每个批次的大小 5 batch_size = 100 6 #计算一共有多少个批次 7 n_batch = mnist.train.num_examples // batch_size 8 9 #参数概要 10 def variable_summaries(var): 11 with tf.name_scope('summaries'): 12 mean = tf.reduce_mean(var) 13 tf.summary.scalar('mean', mean)#平均值 14 with tf.name_scope('stddev'): 15 stddev = tf.sqrt(tf.reduce_mean(tf.square(var - mean))) 16 tf.summary.scalar('stddev', stddev)#标准差 17 tf.summary.scalar('max', tf.reduce_max(var))#最大值 18 tf.summary.scalar('min', tf.reduce_min(var))#最小值 19 tf.summary.histogram('histogram', var)#直方图 20 21 #命名空间 22 with tf.name_scope('input'): 23 #定义两个placeholder 24 x = tf.placeholder(tf.float32,[None,784],name='x-input') 25 y = tf.placeholder(tf.float32,[None,10],name='y-input') 26 27 with tf.name_scope('layer'): 28 #创建一个简单的神经网络 29 with tf.name_scope('wights'): 30 W = tf.Variable(tf.zeros([784,10]),name='W') 31 variable_summaries(W) 32 with tf.name_scope('biases'): 33 b = tf.Variable(tf.zeros([10]),name='b') 34 variable_summaries(b) 35 with tf.name_scope('wx_plus_b'): 36 wx_plus_b = tf.matmul(x,W) + b 37 with tf.name_scope('softmax'): 38 prediction = tf.nn.softmax(wx_plus_b) 39 40 #二次代价函数 41 # loss = tf.reduce_mean(tf.square(y-prediction)) 42 with tf.name_scope('loss'): 43 loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y,logits=prediction)) 44 tf.summary.scalar('loss',loss) 45 with tf.name_scope('train'): 46 #使用梯度下降法 47 train_step = tf.train.GradientDescentOptimizer(0.2).minimize(loss) 48 49 #初始化变量 50 init = tf.global_variables_initializer() 51 52 with tf.name_scope('accuracy'): 53 with tf.name_scope('correct_prediction'): 54 #结果存放在一个布尔型列表中 55 correct_prediction = tf.equal(tf.argmax(y,1),tf.argmax(prediction,1))#argmax返回一维张量中最大的值所在的位置 56 with tf.name_scope('accuracy'): 57 #求准确率 58 accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32)) 59 tf.summary.scalar('accuracy',accuracy) 60 61 #合并所有的summary 62 merged = tf.summary.merge_all() 63 64 with tf.Session() as sess: 65 sess.run(init) 66 writer = tf.summary.FileWriter('logs/',sess.graph) 67 for epoch in range(51): 68 for batch in range(n_batch): 69 batch_xs,batch_ys = mnist.train.next_batch(batch_size) 70 summary,_ = sess.run([merged,train_step],feed_dict={x:batch_xs,y:batch_ys}) 71 72 writer.add_summary(summary,epoch) 73 acc = sess.run(accuracy,feed_dict={x:mnist.test.images,y:mnist.test.labels}) 74 print("Iter " + str(epoch) + ",Testing Accuracy " + str(acc))
tensorboard可视化
1 import tensorflow as tf 2 from tensorflow.examples.tutorials.mnist import input_data 3 from tensorflow.contrib.tensorboard.plugins import projector
1 #载入数据集 2 mnist = input_data.read_data_sets("MNIST_data/",one_hot=True) 3 #运行次数 4 max_steps = 1001 5 #图片数量 6 image_num = 3000 7 #文件路径 8 DIR = "C:/Users/zgyxf183/Documents/jupyter/tensorFlow Learning/" 9 10 #定义会话 11 sess = tf.Session() 12 13 #载入图片 14 embedding = tf.Variable(tf.stack(mnist.test.images[:image_num]), trainable=False, name='embedding') 15 16 #参数概要 17 def variable_summaries(var): 18 with tf.name_scope('summaries'): 19 mean = tf.reduce_mean(var) 20 tf.summary.scalar('mean', mean)#平均值 21 with tf.name_scope('stddev'): 22 stddev = tf.sqrt(tf.reduce_mean(tf.square(var - mean))) 23 tf.summary.scalar('stddev', stddev)#标准差 24 tf.summary.scalar('max', tf.reduce_max(var))#最大值 25 tf.summary.scalar('min', tf.reduce_min(var))#最小值 26 tf.summary.histogram('histogram', var)#直方图 27 28 #命名空间 29 with tf.name_scope('input'): 30 #这里的none表示第一个维度可以是任意的长度 31 x = tf.placeholder(tf.float32,[None,784],name='x-input') 32 #正确的标签 33 y = tf.placeholder(tf.float32,[None,10],name='y-input') 34 35 #显示图片 36 with tf.name_scope('input_reshape'): 37 image_shaped_input = tf.reshape(x, [-1, 28, 28, 1]) 38 tf.summary.image('input', image_shaped_input, 10) 39 40 with tf.name_scope('layer'): 41 #创建一个简单神经网络 42 with tf.name_scope('weights'): 43 W = tf.Variable(tf.zeros([784,10]),name='W') 44 variable_summaries(W) 45 with tf.name_scope('biases'): 46 b = tf.Variable(tf.zeros([10]),name='b') 47 variable_summaries(b) 48 with tf.name_scope('wx_plus_b'): 49 wx_plus_b = tf.matmul(x,W) + b 50 with tf.name_scope('softmax'): 51 prediction = tf.nn.softmax(wx_plus_b) 52 53 with tf.name_scope('loss'): 54 #交叉熵代价函数 55 loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y,logits=prediction)) 56 tf.summary.scalar('loss',loss) 57 with tf.name_scope('train'): 58 #使用梯度下降法 59 train_step = tf.train.GradientDescentOptimizer(0.5).minimize(loss) 60 61 #初始化变量 62 sess.run(tf.global_variables_initializer()) 63 64 with tf.name_scope('accuracy'): 65 with tf.name_scope('correct_prediction'): 66 #结果存放在一个布尔型列表中 67 correct_prediction = tf.equal(tf.argmax(y,1),tf.argmax(prediction,1))#argmax返回一维张量中最大的值所在的位置 68 with tf.name_scope('accuracy'): 69 #求准确率 70 accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32))#把correct_prediction变为float32类型 71 tf.summary.scalar('accuracy',accuracy) 72 73 #产生metadata文件 74 if tf.gfile.Exists(DIR + 'projector/projector/metadata.tsv'): 75 tf.gfile.DeleteRecursively(DIR + 'projector/projector/metadata.tsv') 76 with open(DIR + 'projector/projector/metadata.tsv', 'w') as f: 77 labels = sess.run(tf.argmax(mnist.test.labels[:],1)) 78 for i in range(image_num): 79 f.write(str(labels[i]) + '\n') 80 81 #合并所有的summary 82 merged = tf.summary.merge_all() 83 84 85 projector_writer = tf.summary.FileWriter(DIR + 'projector/projector',sess.graph) 86 saver = tf.train.Saver() 87 config = projector.ProjectorConfig() 88 embed = config.embeddings.add() 89 embed.tensor_name = embedding.name 90 embed.metadata_path = DIR + 'projector/projector/metadata.tsv' 91 embed.sprite.image_path = DIR + 'projector/data/mnist_10k_sprite.png' 92 embed.sprite.single_image_dim.extend([28,28]) 93 projector.visualize_embeddings(projector_writer,config) 94 95 for i in range(max_steps): 96 #每个批次100个样本 97 batch_xs,batch_ys = mnist.train.next_batch(100) 98 run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE) 99 run_metadata = tf.RunMetadata() 100 summary,_ = sess.run([merged,train_step],feed_dict={x:batch_xs,y:batch_ys},options=run_options,run_metadata=run_metadata) 101 projector_writer.add_run_metadata(run_metadata, 'step%03d' % i) 102 projector_writer.add_summary(summary, i) 103 104 if i%100 == 0: 105 acc = sess.run(accuracy,feed_dict={x:mnist.test.images,y:mnist.test.labels}) 106 print ("Iter " + str(i) + ", Testing Accuracy= " + str(acc)) 107 108 saver.save(sess, DIR + 'projector/projector/a_model.ckpt', global_step=max_steps) 109 projector_writer.close() 110 sess.close()
四、卷积神经网络
卷积神经网络应用于MNIST数据集分类
1 import tensorflow as tf 2 from tensorflow.examples.tutorials.mnist import input_data
1 #将二次代价函数替换成交叉熵 2 # tf.nn.sigmod_cross_entropy_with_logits() 来表示跟sigmod搭配使用的交叉熵 3 # tf.nn.softmax_cross_entropy_with_logits() 来表示跟softmax搭配使用的交叉熵 4 #载入数据集 5 #one_hot是将数据改成0-1的格式 6 mnist = input_data.read_data_sets("MNIST_data",one_hot=True) 7 8 #每个批次的大小 9 batch_size = 100 10 11 #计算一共有多少个批次 12 n_batch = mnist.train.num_examples // batch_size 13 14 #初始化权值 15 def weight_variable(shape): 16 initial = tf.truncated_normal(shape,stddev=0.1)#生成一个截断的正态分布 17 return tf.Variable(initial) 18 19 #初始化偏移量 20 def bias_variable(shape): 21 initial = tf.constant(0.1,shape=shape) 22 return tf.Variable(initial) 23 24 25 #定义卷积层 26 def conv2d(x,W): 27 return tf.nn.conv2d(x,W,strides=[1,1,1,1],padding='SAME') 28 29 #定义池化层 30 def max_pool_2x2(x): 31 return tf.nn.max_pool(x,ksize=[1,2,2,1],strides=[1,2,2,1],padding='SAME') 32 33 34 35 #定义两个placeholder,一个图片是28x28=784 36 x = tf.placeholder(tf.float32,[None,784]) 37 y = tf.placeholder(tf.float32,[None,10]) 38 39 #改变x的格式转为4D的向量[batch, in_height, in_width, in_channels]` 40 #输入的数据是784行的一列数据，要转换成 28x28 41 x_image = tf.reshape(x,[-1,28,28,1]) 42 43 #初始化第一个卷积层的权值和偏置 44 W_conv1 = weight_variable([5,5,1,32])#5*5的采样窗口，32个卷积核从1个平面抽取特征 45 b_conv1 = bias_variable([32])#每一个卷积核一个偏置值 46 47 48 #把x_image和权值向量进行卷积，再加上偏置值，然后应用于relu激活函数 49 h_conv1 = tf.nn.relu(conv2d(x_image,W_conv1) + b_conv1) 50 h_pool1 = max_pool_2x2(h_conv1) 51 52 #初始化第二个卷积层的权值和偏置 53 W_conv2 = weight_variable([5,5,32,64])#5*5的采样窗口，64个卷积核从32个平面抽取特征 54 b_conv2 = bias_variable([64])#每一个卷积核一个偏置值 55 56 #把h_pool1和权值向量进行卷积，再加上偏置值，然后应用于relu激活函数 57 h_conv2 = tf.nn.relu(conv2d(h_pool1,W_conv2) + b_conv2) 58 h_pool2 = max_pool_2x2(h_conv2)#进行max-pooling 59 60 #28*28的图片第一次卷积后还是28*28，第一次池化后变为14*14 61 #第二次卷积后为14*14，第二次池化后变为了7*7 62 #进过上面操作后得到64张7*7的平面 63 64 65 66 #初始化第一个全连接层的权值 67 W_fc1 = weight_variable([7*7*64,1024])#上一层有7*7*64个神经元，全连接层有1024个神经元 68 b_fc1 = bias_variable([1024])#1024个节点 69 70 71 #把池化层2的输出扁平化为1维 72 h_pool2_flat = tf.reshape(h_pool2,[-1,7*7*64]) 73 74 #求第一个全连接层的输出 75 h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat,W_fc1) + b_fc1) 76 77 78 #keep_prob用来表示神经元的输出概率 79 keep_prob = tf.placeholder(tf.float32) 80 h_fc1_drop = tf.nn.dropout(h_fc1,keep_prob) 81 82 #初始化第二个全连接层 83 W_fc2 = weight_variable([1024,10]) 84 b_fc2 = bias_variable([10]) 85 86 #计算输出 87 prediction = tf.nn.softmax(tf.matmul(h_fc1_drop,W_fc2) + b_fc2) 88 89 #二次代价函数 90 #loss = tf.reduce_mean(tf.square(y-prediction)) 91 92 #交叉熵 93 #首先看输入logits，它的shape是[batch_size, num_classes] ，一般来讲，就是神经网络最后一层的输入z。 94 #另外一个输入是labels，它的shape也是[batch_size, num_classes]，就是我们神经网络期望的输出。 95 loss=tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=prediction)) 96 #使用梯度下降法训练 97 #train_step = tf.train.GradientDescentOptimizer(0.2).minimize(loss) 98 #使用其他优化器 99 train_step = tf.train.AdamOptimizer(1e-4).minimize(loss) 100 101 102 #初始化变量 103 init = tf.global_variables_initializer() 104 105 #结果存在一个布尔型的列表中 106 correct_prediction = tf.equal(tf.argmax(y,1),tf.argmax(prediction,1)) #argmax返回一个张量中最大值存在的位置 107 108 #求准确率 109 accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32)) #将布尔型转换为float32, true->1 false ->0 110 111 with tf.Session() as sess: 112 sess.run(init) 113 114 for epoch in range(51): 115 116 for batch in range (n_batch): 117 #获取每个batch的图片 118 batch_xs,batch_ys = mnist.train.next_batch(batch_size) 119 sess.run(train_step,feed_dict={x:batch_xs,y:batch_ys,keep_prob:0.7}) 120 121 #learning_rate = sess.run(lr) 122 123 test_acc = sess.run(accuracy,feed_dict={x:mnist.test.images,y:mnist.test.labels,keep_prob:1.0}) 124 train_acc = sess.run(accuracy,feed_dict={x:mnist.train.images,y:mnist.train.labels,keep_prob:1.0}) 125 126 print("Iter " + str(epoch) + ",Testing Accuracy "+ str(test_acc)+", training Accuracy "+str(train_acc))
卷积神经网络可视化
1 import tensorflow as tf 2 from tensorflow.examples.tutorials.mnist import input_data
1 mnist = input_data.read_data_sets('MNIST_data',one_hot=True) 2 3 #每个批次的大小 4 batch_size = 100 5 #计算一共有多少个批次 6 n_batch = mnist.train.num_examples // batch_size 7 8 #参数概要 9 def variable_summaries(var): 10 with tf.name_scope('summaries'): 11 mean = tf.reduce_mean(var) 12 tf.summary.scalar('mean', mean)#平均值 13 with tf.name_scope('stddev'): 14 stddev = tf.sqrt(tf.reduce_mean(tf.square(var - mean))) 15 tf.summary.scalar('stddev', stddev)#标准差 16 tf.summary.scalar('max', tf.reduce_max(var))#最大值 17 tf.summary.scalar('min', tf.reduce_min(var))#最小值 18 tf.summary.histogram('histogram', var)#直方图 19 20 #初始化权值 21 def weight_variable(shape,name): 22 initial = tf.truncated_normal(shape,stddev=0.1)#生成一个截断的正态分布 23 return tf.Variable(initial,name=name) 24 25 #初始化偏置 26 def bias_variable(shape,name): 27 initial = tf.constant(0.1,shape=shape) 28 return tf.Variable(initial,name=name) 29 30 #卷积层 31 def conv2d(x,W): 32 #x input tensor of shape `[batch, in_height, in_width, in_channels]` 33 #W filter / kernel tensor of shape [filter_height, filter_width, in_channels, out_channels] 34 #`strides[0] = strides[3] = 1`. strides[1]代表x方向的步长，strides[2]代表y方向的步长 35 #padding: A `string` from: `"SAME", "VALID"` 36 return tf.nn.conv2d(x,W,strides=[1,1,1,1],padding='SAME') 37 38 #池化层 39 def max_pool_2x2(x): 40 #ksize [1,x,y,1] 41 return tf.nn.max_pool(x,ksize=[1,2,2,1],strides=[1,2,2,1],padding='SAME') 42 43 #命名空间 44 with tf.name_scope('input'): 45 #定义两个placeholder 46 x = tf.placeholder(tf.float32,[None,784],name='x-input') 47 y = tf.placeholder(tf.float32,[None,10],name='y-input') 48 with tf.name_scope('x_image'): 49 #改变x的格式转为4D的向量[batch, in_height, in_width, in_channels]` 50 x_image = tf.reshape(x,[-1,28,28,1],name='x_image') 51 52 53 with tf.name_scope('Conv1'): 54 #初始化第一个卷积层的权值和偏置 55 with tf.name_scope('W_conv1'): 56 W_conv1 = weight_variable([5,5,1,32],name='W_conv1')#5*5的采样窗口，32个卷积核从1个平面抽取特征 57 with tf.name_scope('b_conv1'): 58 b_conv1 = bias_variable([32],name='b_conv1')#每一个卷积核一个偏置值 59 60 #把x_image和权值向量进行卷积，再加上偏置值，然后应用于relu激活函数 61 with tf.name_scope('conv2d_1'): 62 conv2d_1 = conv2d(x_image,W_conv1) + b_conv1 63 with tf.name_scope('relu'): 64 h_conv1 = tf.nn.relu(conv2d_1) 65 with tf.name_scope('h_pool1'): 66 h_pool1 = max_pool_2x2(h_conv1)#进行max-pooling 67 68 with tf.name_scope('Conv2'): 69 #初始化第二个卷积层的权值和偏置 70 with tf.name_scope('W_conv2'): 71 W_conv2 = weight_variable([5,5,32,64],name='W_conv2')#5*5的采样窗口，64个卷积核从32个平面抽取特征 72 with tf.name_scope('b_conv2'): 73 b_conv2 = bias_variable([64],name='b_conv2')#每一个卷积核一个偏置值 74 75 #把h_pool1和权值向量进行卷积，再加上偏置值，然后应用于relu激活函数 76 with tf.name_scope('conv2d_2'): 77 conv2d_2 = conv2d(h_pool1,W_conv2) + b_conv2 78 with tf.name_scope('relu'): 79 h_conv2 = tf.nn.relu(conv2d_2) 80 with tf.name_scope('h_pool2'): 81 h_pool2 = max_pool_2x2(h_conv2)#进行max-pooling 82 83 #28*28的图片第一次卷积后还是28*28，第一次池化后变为14*14 84 #第二次卷积后为14*14，第二次池化后变为了7*7 85 #进过上面操作后得到64张7*7的平面 86 87 with tf.name_scope('fc1'): 88 #初始化第一个全连接层的权值 89 with tf.name_scope('W_fc1'): 90 W_fc1 = weight_variable([7*7*64,1024],name='W_fc1')#上一场有7*7*64个神经元，全连接层有1024个神经元 91 with tf.name_scope('b_fc1'): 92 b_fc1 = bias_variable([1024],name='b_fc1')#1024个节点 93 94 #把池化层2的输出扁平化为1维 95 with tf.name_scope('h_pool2_flat'): 96 h_pool2_flat = tf.reshape(h_pool2,[-1,7*7*64],name='h_pool2_flat') 97 #求第一个全连接层的输出 98 with tf.name_scope('wx_plus_b1'): 99 wx_plus_b1 = tf.matmul(h_pool2_flat,W_fc1) + b_fc1 100 with tf.name_scope('relu'): 101 h_fc1 = tf.nn.relu(wx_plus_b1) 102 103 #keep_prob用来表示神经元的输出概率 104 with tf.name_scope('keep_prob'): 105 keep_prob = tf.placeholder(tf.float32,name='keep_prob') 106 with tf.name_scope('h_fc1_drop'): 107 h_fc1_drop = tf.nn.dropout(h_fc1,keep_prob,name='h_fc1_drop') 108 109 with tf.name_scope('fc2'): 110 #初始化第二个全连接层 111 with tf.name_scope('W_fc2'): 112 W_fc2 = weight_variable([1024,10],name='W_fc2') 113 with tf.name_scope('b_fc2'): 114 b_fc2 = bias_variable([10],name='b_fc2') 115 with tf.name_scope('wx_plus_b2'): 116 wx_plus_b2 = tf.matmul(h_fc1_drop,W_fc2) + b_fc2 117 with tf.name_scope('softmax'): 118 #计算输出 119 prediction = tf.nn.softmax(wx_plus_b2) 120 121 #交叉熵代价函数 122 with tf.name_scope('cross_entropy'): 123 cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y,logits=prediction),name='cross_entropy') 124 tf.summary.scalar('cross_entropy',cross_entropy) 125 126 #使用AdamOptimizer进行优化 127 with tf.name_scope('train'): 128 train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy) 129 130 #求准确率 131 with tf.name_scope('accuracy'): 132 with tf.name_scope('correct_prediction'): 133 #结果存放在一个布尔列表中 134 correct_prediction = tf.equal(tf.argmax(prediction,1),tf.argmax(y,1))#argmax返回一维张量中最大的值所在的位置 135 with tf.name_scope('accuracy'): 136 #求准确率 137 accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32)) 138 tf.summary.scalar('accuracy',accuracy) 139 140 #合并所有的summary 141 merged = tf.summary.merge_all() 142 143 with tf.Session() as sess: 144 sess.run(tf.global_variables_initializer()) 145 train_writer = tf.summary.FileWriter('logs/train',sess.graph) 146 test_writer = tf.summary.FileWriter('logs/test',sess.graph) 147 for i in range(1001): 148 #训练模型 149 batch_xs,batch_ys = mnist.train.next_batch(batch_size) 150 sess.run(train_step,feed_dict={x:batch_xs,y:batch_ys,keep_prob:0.5}) 151 #记录训练集计算的参数 152 summary = sess.run(merged,feed_dict={x:batch_xs,y:batch_ys,keep_prob:1.0}) 153 train_writer.add_summary(summary,i) 154 #记录测试集计算的参数 155 batch_xs,batch_ys = mnist.test.next_batch(batch_size) 156 summary = sess.run(merged,feed_dict={x:batch_xs,y:batch_ys,keep_prob:1.0}) 157 test_writer.add_summary(summary,i) 158 159 if i%100==0: 160 test_acc = sess.run(accuracy,feed_dict={x:mnist.test.images,y:mnist.test.labels,keep_prob:1.0}) 161 train_acc = sess.run(accuracy,feed_dict={x:mnist.train.images[:10000],y:mnist.train.labels[:10000],keep_prob:1.0}) 162 print ("Iter " + str(i) + ", Testing Accuracy= " + str(test_acc) + ", Training Accuracy= " + str(train_acc))
LSTM
1 import tensorflow as tf 2 from tensorflow.examples.tutorials.mnist import input_data
1 #载入数据集 2 mnist = input_data.read_data_sets("MNIST_data/",one_hot=True) 3 4 # 输入图片是28*28 5 n_inputs = 28 #输入一行，一行有28个数据 6 max_time = 28 #一共28行 7 lstm_size = 100 #隐层单元 8 n_classes = 10 # 10个分类 9 batch_size = 50 #每批次50个样本 10 n_batch = mnist.train.num_examples // batch_size #计算一共有多少个批次 11 12 #这里的none表示第一个维度可以是任意的长度 13 x = tf.placeholder(tf.float32,[None,784]) 14 #正确的标签 15 y = tf.placeholder(tf.float32,[None,10]) 16 17 #初始化权值 18 weights = tf.Variable(tf.truncated_normal([lstm_size, n_classes], stddev=0.1)) 19 #初始化偏置值 20 biases = tf.Variable(tf.constant(0.1, shape=[n_classes])) 21 22 23 #定义RNN网络 24 def RNN(X,weights,biases): 25 # inputs=[batch_size, max_time, n_inputs] 26 inputs = tf.reshape(X,[-1,max_time,n_inputs]) 27 #定义LSTM基本CELL 28 lstm_cell = tf.contrib.rnn.core_rnn_cell.BasicLSTMCell(lstm_size) 29 # final_state[state, batch_size, cell.state_size] 30 # final_state[0]是cell state 31 # final_state[1]是hidden_state 32 # outputs: The RNN output `Tensor`. 33 # If time_major == False (default), this will be a `Tensor` shaped: 34 # `[batch_size, max_time, cell.output_size]`. 35 # If time_major == True, this will be a `Tensor` shaped: 36 # `[max_time, batch_size, cell.output_size]`. 37 outputs,final_state = tf.nn.dynamic_rnn(lstm_cell,inputs,dtype=tf.float32) 38 results = tf.nn.softmax(tf.matmul(final_state[1],weights) + biases) 39 return results 40 41 42 #计算RNN的返回结果 43 prediction= RNN(x, weights, biases) 44 #损失函数 45 cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=prediction,labels=y)) 46 #使用AdamOptimizer进行优化 47 train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy) 48 #结果存放在一个布尔型列表中 49 correct_prediction = tf.equal(tf.argmax(y,1),tf.argmax(prediction,1))#argmax返回一维张量中最大的值所在的位置 50 #求准确率 51 accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32))#把correct_prediction变为float32类型 52 #初始化 53 init = tf.global_variables_initializer() 54 55 with tf.Session() as sess: 56 sess.run(init) 57 for epoch in range(6): 58 for batch in range(n_batch): 59 batch_xs,batch_ys = mnist.train.next_batch(batch_size) 60 sess.run(train_step,feed_dict={x:batch_xs,y:batch_ys}) 61 62 acc = sess.run(accuracy,feed_dict={x:mnist.test.images,y:mnist.test.labels}) 63 print ("Iter " + str(epoch) + ", Testing Accuracy= " + str(acc))
