TensorFlow⼊门⽰例教程
本部分的代码⽬前都是基于GitHub⼤佬⾮常详细的TensorFlow的教程上，⾸先给出链接：本⼈对其中部分代码做了注释和中⽂翻译，会持续更新，⽬前包括： 1. 传统多层神经⽹络⽤语MNIST数据集分类（代码讲解，翻译）
1. 传统多层神经⽹络⽤语MNIST数据集分类（代码讲解，翻译）
1 \"\"\" Neural Network. 2
3 A 2-Hidden Layers Fully Connected Neural Network (a.k.a Multilayer Perceptron) 4 implementation with TensorFlow. This example is using the MNIST database 5 of handwritten digits (http://yann.lecun.com/exdb/mnist/). 6
7 Links:
8 [MNIST Dataset](http://yann.lecun.com/exdb/mnist/). 9
10 Author: Aymeric Damien
11 Project: https://github.com/aymericdamien/TensorFlow-Examples/ 12 \"\"\" 13
14 from __future__ import print_function 15
16 # Import MNIST data 17 # 导⼊mnist数据集
18 from tensorflow.examples.tutorials.mnist import input_data
19 mnist = input_data.read_data_sets(\"/tmp/data/\", one_hot=True) 20
21 # 导⼊tf
22 import tensorflow as tf 23
24 # Parameters 25 # 设定各种超参数
26 learning_rate = 0.1 # 学习率 27 num_steps = 500 # 训练500次
28 batch_size = 128 # 每批次取128个样本训练 29 display_step = 100 # 每训练100步显⽰⼀次 30
31 # Network Parameters 32 # 设定⽹络的超参数
33 n_hidden_1 = 256 # 1st layer number of neurons 34 n_hidden_2 = 256 # 2nd layer number of neurons
35 num_input = 784 # MNIST data input (img shape: 28*28) 36 num_classes = 10 # MNIST total classes (0-9 digits) 37
38 # tf Graph input
39 # tf图的输⼊，因为不知道到底输⼊⼤⼩是多少，因此设定占位符 40 X = tf.placeholder(\"float\", [None, num_input]) 41 Y = tf.placeholder(\"float\", [None, num_classes]) 42
43 # Store layers weight & bias 44 # 初始化w和b 45 weights = {
46 'h1': tf.Variable(tf.random_normal([num_input, n_hidden_1])), 47 'h2': tf.Variable(tf.random_normal([n_hidden_1, n_hidden_2])), 48 'out': tf.Variable(tf.random_normal([n_hidden_2, num_classes])) 49 }
50 biases = {
51 'b1': tf.Variable(tf.random_normal([n_hidden_1])), 52 'b2': tf.Variable(tf.random_normal([n_hidden_2])), 53 'out': tf.Variable(tf.random_normal([num_classes])) 54 } 55 56
57 # Create model 58 # 创建模型
59 def neural_net(x):
60 # Hidden fully connected layer with 256 neurons 61 # 隐藏层1，全连接了256个神经元
62 layer_1 = tf.add(tf.matmul(x, weights['h1']), biases['b1']) 63 # Hidden fully connected layer with 256 neurons # 隐藏层2，全连接了256个神经元
65 layer_2 = tf.add(tf.matmul(layer_1, weights['h2']), biases['b2']) 66 # Output fully connected layer with a neuron for each class 67 # 最后作为输出的全连接层，对每⼀分类连接⼀个神经元
68 out_layer = tf.matmul(layer_2, weights['out']) + biases['out'] 69 return out_layer 70
71 # Construct model 72 # 开启模型
73 # 输⼊数据X，得到得分向量logits 74 logits = neural_net(X)
75 # ⽤softmax分类器将得分向量转变成概率向量 76 prediction = tf.nn.softmax(logits) 77
78 # Define loss and optimizer 79 # 定义损失和优化器
80 # 交叉熵损失, 求均值得到---->loss_op
81 loss_op = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits( 82 logits=logits, labels=Y))
83 # 优化器使⽤的是Adam算法优化器
84 optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate) 85 # 最⼩化损失得到---->可以训练的train_op 86 train_op = optimizer.minimize(loss_op) 87
88 # Evaluate model # 评估模型
90 # tf.equal() 逐个元素进⾏判断，如果相等就是True，不相等，就是False。 91 correct_pred = tf.equal(tf.argmax(prediction, 1), tf.argmax(Y, 1)) 92 # tf.cast() 数据类型转换----> tf.reduce_mean() 再求均值 93 accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32)) 94
95 # Initialize the variables (i.e. assign their default value) 96 # 初始化这些变量（作⽤⽐如说，给他们分配随机默认值） 97 init = tf.global_variables_initializer() 98
99 # Start training
100 # 现在开始训练啦！
101 with tf.Session() as sess:102
103 # Run the initializer104 # 运⾏初始化器105 sess.run(init)106
107 for step in range(1, num_steps+1):
108 # 每批次128个训练，取出这128个对应的data：x；标签：y109 batch_x, batch_y = mnist.train.next_batch(batch_size)110 # Run optimization op (backprop)
111 # train_op是优化器得到的可以训练的op，通过反向传播优化模型112 sess.run(train_op, feed_dict={X: batch_x, Y: batch_y})113 # 每100步打印⼀次训练的成果
114 if step % display_step == 0 or step == 1:115 # Calculate batch loss and accuracy116 # 计算每批次的是损失和准确度
117 loss, acc = sess.run([loss_op, accuracy], feed_dict={X: batch_x,118 Y: batch_y})119 print(\"Step \" + str(step) + \ + \\120 \"{:.4f}\".format(loss) + \ + \\121 \"{:.3f}\".format(acc))122
123 print(\"Optimization Finished!\")124
125 # Calculate accuracy for MNIST test images126 # 看看在测试集上，我们的模型表现如何127 print(\"Testing Accuracy:\", \\
128 sess.run(accuracy, feed_dict={X: mnist.test.images,129 Y: mnist.test.labels}))
