RNN循环神经网络

                 1
                #
                 导包，加载数据，定义变量
                 2
                import
                 tensorflow as tf

                 3
                #
                 流式计算图，构建第二次的时候会有内存，计算图形的时候初始化
                 4
                tf.reset_default_graph()

                 5
                import datetime # 打印时间 6import os  # 保存文件 7from tensorflow.examples.tutorials.mnist import input_data
 8 9import warnings
10 warnings.filterwarnings(‘ignore‘)
11# minst测试集12 mnist = input_data.read_data_sets(‘./‘, one_hot=True)
1314# 每次使用100条数据进行训练15 batch_size = 100
16# 图像向量17 width = 28
18 height = 28
19# LSTM隐藏神经元数量20 rnn_size = 256
21# 输出层one-hot向量长度的22 out_size = 10
2324def weight_variable(shape, w_alpha=0.01):
25     initial = w_alpha * tf.random_normal(shape)
26return tf.Variable(initial)
2728def bias_variable(shape, b_alpha=0.1):
29     initial = b_alpha * tf.random_normal(shape)
30return tf.Variable(initial)
3132# 权重及偏置33 w = weight_variable([rnn_size, out_size])
34 b = bias_variable([out_size])
3536# 将数据转化成RNN要求的数据37# 按照图片大小申请占位符38 X = tf.placeholder(tf.float32, [None, height, width])
39# 原排列[0,1,2]transpose为[1,0,2]代表前两维装置，如shape=(1,2,3)转为shape=(2,1,3)40# 这里的实际意义是把所有图像向量的相同行号向量转到一起，如x1的第一行与x2的第一行41 x = tf.transpose(X, [1, 0, 2])
42# reshape -1 代表自适应，这里按照图像每一列的长度为reshape后的列长度43 x = tf.reshape(x, [-1, width])
44# split默任在第一维即0 dimension进行分割，分割成height份，这里实际指把所有图片向量按对应行号进行重组45 x = tf.split(x, height)
4647# 构建神经网络48# LSTM49 lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(rnn_size)
50# 这里RNN会有与输入层相同数量的输出层，我们只需要最后一个输出51 outputs, status = tf.nn.static_rnn(lstm_cell, x, dtype=tf.float32)
5253 y_conv = tf.add(tf.matmul(outputs[-1], w), b)
54# 最小化损失优化55 Y = tf.placeholder(dtype=tf.float32,shape = [None,10])
56 loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=y_conv, labels=Y))
57 optimizer = tf.train.AdamOptimizer(0.01).minimize(loss)
58# 偏差 准确率59 correct = tf.equal(tf.argmax(y_conv, 1), tf.argmax(Y, 1))
60 accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
6162# 模型训练63# 启动会话.开始训练64 saver = tf.train.Saver()
65 session = tf.Session()
66session.run(tf.global_variables_initializer())
67 step = 0
68 acc_rate = 0.90
69while 1:
70     batch_x, batch_y = mnist.train.next_batch(batch_size)
71     batch_x = batch_x.reshape((batch_size, height, width))
72     session.run(optimizer, feed_dict={X:batch_x,Y:batch_y})
73# 每训练10次测试一次74if step % 10 == 0:
75         batch_x_test = mnist.test.images
76         batch_y_test = mnist.test.labels
77         batch_x_test = batch_x_test.reshape([-1, height, width])
78         acc = session.run(accuracy, feed_dict={X: batch_x_test, Y: batch_y_test})
79print(datetime.datetime.now().strftime(‘%c‘), ‘ step:‘, step, ‘ accuracy:‘, acc)
80# 偏差满足要求，保存模型81if acc >= acc_rate:
82             model_path = os.getcwd() + os.sep + str(acc_rate) + "mnist.model"83             saver.save(session, model_path, global_step=step)
84break85     step += 1
86 session.close()

                 1
                import
                 tensorflow as tf

                 2
                tf.reset_default_graph()

                 3
                from tensorflow.examples.tutorials.mnist import input_data
 4 5# 载入数据 6 mnist = input_data.read_data_sets("./", one_hot=True)
 7 8# 输入图片是28 9 n_input = 28
10 max_time = 28
11 lstm_size = 100  # 隐藏单元12 n_class = 10  # 10个分类13 batch_size = 50   # 每次50个样本14 n_batch_size = mnist.train.num_examples // batch_size    # 计算一共有多少批次1516# 这里None表示第一个维度可以是任意长度17# 创建占位符18 x = tf.placeholder(tf.float32,[None, 28*28])
19# 正确的标签20 y = tf.placeholder(tf.float32,[None, 10])
2122# 初始化权重 ,stddev为标准差 truncated_normal 裁剪过得正太分布，去掉特别大和特别小的值23 weight = tf.Variable(tf.truncated_normal([lstm_size, n_class], stddev=0.1))
24# 初始化偏置层25 biases = tf.Variable(tf.constant(0.1, shape=[n_class]))
2627# 定义RNN网络28def RNN(X, weights, biases):
29#  原始数据为[batch_size,28*28]30# input = [batch_size, max_time, n_input]31     input = tf.reshape(X,[-1, max_time, n_input])
32# 定义LSTM的基本单元33#     lstm_cell = tf.contrib.rnn.BasicLSTMCell(lstm_size)34     lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(lstm_size)
35# final_state[0] 是cell state36# final_state[1] 是hidden stat37     outputs, final_state = tf.nn.dynamic_rnn(lstm_cell, input, dtype=tf.float32)
38    display(final_state)
39     results = tf.nn.softmax(tf.matmul(final_state[1],weights)+biases)
40return results
4142# 计算RNN的返回结果43 prediction = RNN(x, weight, biases)
44# 损失函数45 loss = 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(loss)
48# 将结果存下来49 correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(prediction, 1))
50# 计算正确率51 accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
5253with tf.Session() as sess:
54    sess.run(tf.global_variables_initializer())
55for epoch in range(6):
56for batch in range(n_batch_size):
57# 取出下一批次数据58             batch_xs,batch_ys = mnist.train.next_batch(batch_size)
59             sess.run(train_step, feed_dict={x: batch_xs,y: batch_ys})
60if(batch%100==0):
61print(str(batch)+"/" + str(n_batch_size))
62         acc = sess.run(accuracy, feed_dict={x: mnist.test.images, y: mnist.test.labels})
63print("Iter" + str(epoch) + " ,Testing Accuracy = " + str(acc))