TensorFlow使用RNN实现手写数字识别

 1 # https://www.cnblogs.com/felixwang2/p/9190664.html
 2 # TensorFlow(十二)：使用RNN实现手写数字识别
 3 
 4 import tensorflow as tf
 5 from tensorflow.examples.tutorials.mnist import input_data
 6 
 7 # 载入数据集
 8 mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
 9 
10 # 输入图片是28*28
11 n_inputs = 28  # 输入一行，一行有28个数据
12 max_time = 28  # 一共28行
13 lstm_size = 100  # 隐层单元
14 n_classes = 10  # 10个分类
15 batch_size = 50  # 每批次50个样本
16 n_batch = mnist.train.num_examples // batch_size  # 计算一共有多少个批次
17 
18 # 这里的none表示第一个维度可以是任意的长度
19 x = tf.placeholder(tf.float32, [None, 784])
20 # 正确的标签
21 y = tf.placeholder(tf.float32, [None, 10])
22 
23 # 初始化权值
24 weights = tf.Variable(tf.truncated_normal([lstm_size, n_classes], stddev=0.1))
25 # 初始化偏置值
26 biases = tf.Variable(tf.constant(0.1, shape=[n_classes]))
27 
28 
29 # 定义RNN网络
30 def RNN(X, weights, biases):
31     # inputs=[batch_size, max_time, n_inputs]
32     inputs = tf.reshape(X, [-1, max_time, n_inputs])
33     # 定义LSTM基本CELL
34     lstm_cell = tf.contrib.rnn.BasicLSTMCell(lstm_size)
35     # final_state[state,batch_size,cell.state_size]
36     # final_state[0]是cell state
37     # final_state[1]是hidden_state
38     # outputs: The RNN output 'Tensor'.
39     #  If time_major == False (default), this will be a `Tensor` shaped:
40     #       `[batch_size, max_time, cell.output_size]`.
41     #  If time_major == True, this will be a `Tensor` shaped:
42     #       `[max_time, batch_size, cell.output_size]`.
43     # final_state 记录的是最后一次的输出结果
44     # outputs 记录的是每一次的输出结果
45 
46     outputs, final_state = tf.nn.dynamic_rnn(lstm_cell, inputs, dtype=tf.float32)
47     results = tf.nn.softmax(tf.matmul(final_state[1], weights) + biases)
48     return results
49 
50 
51 # 计算RNN的返回结果
52 prediction = RNN(x, weights, biases)
53 # 损失函数
54 cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=prediction, labels=y))
55 # 使用AdamOptimizer进行优化
56 train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
57 # 结果存放在一个布尔型列表中
58 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
59 # 求准确率
60 accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))  # 把correct_prediction变为float32类型
61 # 初始化
62 init = tf.global_variables_initializer()
63 
64 gpu_options = tf.GPUOptions(allow_growth=True)
65 with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
66     sess.run(init)
67     for epoch in range(6):
68         for batch in range(n_batch):
69             batch_xs, batch_ys = mnist.train.next_batch(batch_size)
70             sess.run(train_step, feed_dict={x: batch_xs, y: batch_ys})
71 
72         acc = sess.run(accuracy, feed_dict={x: mnist.test.images, y: mnist.test.labels})
73         print("Iter " + str(epoch) + ", Testing Accuracy= " + str(acc))