教程 | 基于LSTM实现手写数字识别

• 2019 年 10 月 7 日
• 筆記

基于tensorflow，如何实现一个简单的循环神经网络，完成手写数字识别，附完整演示代码。

LSTM网络构建

01

基于tensorflow实现简单的LSTM网络，完成mnist手写数字数据集训练与识别。这个其中最重要的构建一个LSTM网络，tensorflow已经给我们提供相关的API， 我们只要使用相关API就可以轻松构建一个简单的LSTM网络。

首先定义输入与目标标签

# create RNN network  X = tf.placeholder(shape=[None, time_steps, num_features], dtype=tf.float32)  Y = tf.placeholder(shape=[None, 10], dtype=tf.float32)

其中

• None: 表示batchsize的大小或者数目
• time_steps: 网络把输出重新输入的次数
• num_features: 输入矩阵/神经元

构建LSTM单元

lstm_cell = rnn.BasicLSTMCell(num_hidden, forget_bias=1.0)  outputs, states = rnn.static_rnn(lstm_cell, x, dtype=tf.float32)

其中：

lstm_cell 表示 LSTM 的单元 num_hidden : 隐藏层节点数目 forget_bias: 遗忘门中要加上的增益偏置 outputs: 网络输出 states：状态

这样我们就构建好一个LSTM循环神经网络了，它的执行过程是很魔幻的。简直是神奇！以后再说。

代码程序执行与输出

02

完整的代码演示分为如下几个部分：

• 加载数据集
• 创建LSTM网络
• 训练网络
• 执行测试

import tensorflow as tf  from tensorflow.contrib import rnn  import numpy as np      from tensorflow.examples.tutorials.mnist import input_data  print(tf.__version__)  mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)    num_hidden = 128  time_steps = 28  num_features = 28  num_classes = 10  batch_size = 128    # create RNN network  X = tf.placeholder(shape=[None, time_steps, num_features], dtype=tf.float32)  Y = tf.placeholder(shape=[None, 10], dtype=tf.float32)    # Define weights  weights = {      'out': tf.Variable(tf.random_normal([num_hidden, num_classes]))  }  biases = {      'out': tf.Variable(tf.random_normal([num_classes]))  }      def rnn_network(x, weights, biases):      x = tf.unstack(x, time_steps, 1)      lstm_cell = rnn.BasicLSTMCell(num_hidden, forget_bias=1.0)      outputs, states = rnn.static_rnn(lstm_cell, x, dtype=tf.float32)      return tf.matmul(outputs[-1], weights['out']) + biases['out']      # 输入预测  logits = rnn_network(X, weights, biases)  prediction = tf.nn.softmax(logits)    # 定义损失函数与优化器  loss_op = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(      logits=logits, labels=Y))  optimizer = tf.train.AdamOptimizer()  train_op = optimizer.minimize(loss_op)    # 计算识别精度  correct_pred = tf.equal(tf.argmax(prediction, 1), tf.argmax(Y, 1))  accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))    # 开始训练  with tf.Session() as sess:      sess.run(tf.global_variables_initializer())      for step in range(1, 5001):          batch_x, batch_y = mnist.train.next_batch(batch_size)          # Reshape data to get 28 seq of 28 elements          batch_x = batch_x.reshape((batch_size, time_steps, num_features))          # Run optimization op (backprop)          sess.run(train_op, feed_dict={X: batch_x, Y: batch_y})          if step % 1000 == 0 or step == 1:              # Calculate batch loss and accuracy              loss, acc = sess.run([loss_op, accuracy], feed_dict={X: batch_x,                                                                   Y: batch_y})              print("Step " + str(step) + ", Loss= " +                     "{:.4f}".format(loss) + ", Training Accuracy= " +                     "{:.3f}".format(acc))        print("Optimization Finished!")        # 使用测试数据集测试训练号的模型， 测试128张手写数字图像      test_len = 128      test_data = mnist.test.images[:test_len].reshape((-1, time_steps, num_features))      test_label = mnist.test.labels[:test_len]      print("Testing Accuracy:",           sess.run(accuracy, feed_dict={X: test_data, Y: test_label}))

运行输出如下：