罗斯基白话：TensorFlow + 实战系列（五）实战MNIST

本文主要是介绍罗斯基白话：TensorFlow + 实战系列（五）实战MNIST，希望对大家解决编程问题提供一定的参考价值，需要的开发者们随着小编来一起学习吧！

白话TensorFlow +实战系列（五）
实战MNIST

这篇文章主要用全连接神经网络来实现MNIST手写数字识别的问题。首先介绍下MNIST数据集。

1）MNIST数据集

MNIST数据集是一个非常有名的手写数字识别数据集，它包含了60000张图片作为训练集，10000张图片为测试集，每张图为一个手写的0~9数字。如图：

其中每张图的大小均为28*28，这里大小指的的是像素。例如数字1所对应的像素矩阵为：

而我们要做的就是教会电脑识别每个手写数字。这个数据集非常经典，常作为学习神经网络的入门教材，一如每个程序员的第一个程序都是“helloword!”一样。

2）数据处理

数据集下载下来后有四个文件，分别为训练集图片，训练集答案，测试集图片，测试集答案。TensorFlow提供了一个类来处理MNIST数据，这个类会自动的将MNIST数据分为训练集，验证集与测试集，并且这些数据都是可以直接喂给神经网络作为输入用的。示例代码如下：

其中input_data.read_data_sets会自动将数据集进行处理，one_hot = True用独热方式表示，意思是每个数字由one_hot方式表，例如数字0 = [1,0,0,0,0,0,0,0,0,0]，1 = [0,1,0,0,0,0,0,0,0,0]。运行结果如下：

接下来就用一个全连接神经网络来识别数字。

3）全连接神经网络

首先定义超参数与参数，没啥好解释的，代码如下：

import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_databatch_size = 100
learning_rate = 0.8
trainig_step = 30000n_input = 784
n_hidden = 500
n_labels = 10

接着定义网络的结构，构建的网络只有一个隐藏层，隐藏层节点为500。代码如下：

def inference(x_input):with tf.variable_scope("hidden"):weights = tf.get_variable("weights", [n_input, n_hidden], initializer = tf.random_normal_initializer(stddev = 0.1))biases = tf.get_variable("biases", [n_hidden], initializer = tf.constant_initializer(0.0))hidden = tf.nn.relu(tf.matmul(x_input, weights) + biases)with tf.variable_scope("out"):weights  = tf.get_variable("weights", [n_hidden, n_labels], initializer = tf.random_normal_initializer(stddev = 0.1))biases = tf.get_variable("biases", [n_labels], initializer = tf.constant_initializer(0.0))output = tf.matmul(hidden, weights) + biasesreturn output

在输出层中，output并没有用到relu函数，因为在之后的softmax层中也是非线性激励，所以可以不用。

接着定义训练过程，代码如下：

def train(mnist):x = tf.placeholder("float", [None, n_input])y = tf.placeholder("float", [None, n_labels])pred = inference(x)#计算损失函数cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits = pred, labels = y))#定义优化器optimizer = tf.train.GradientDescentOptimizer(learning_rate = learning_rate).minimize(cross_entropy)#定义准确率计算correct_pred = tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1))accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))init = tf.global_variables_initializer()with tf.Session() as sess:sess.run(init)#定义验证集与测试集validate_data = {x: mnist.validation.images, y: mnist.validation.labels}test_data = {x: mnist.test.images, y: mnist.test.labels}for i in range(trainig_step):#xs,ys为每个batch_size的训练数据与对应的标签xs, ys = mnist.train.next_batch(batch_size)_, loss = sess.run([optimizer, cross_entropy], feed_dict={x: xs, y:ys})#每1000次训练打印一次损失值与验证准确率if i % 1000 == 0:validate_accuracy = sess.run(accuracy, feed_dict=validate_data)print("after %d training steps, the loss is %g, the validation accuracy is %g" % (i, loss, validate_accuracy))print("the training is finish!")#最终的测试准确率acc = sess.run(accuracy, feed_dict=test_data)print("the test accuarcy is:", acc)

其中每一步的函数作用可以参考我的第二篇博客：罗斯基白话：TensorFlow+实战系列（二）从零构建传统神经网络

里面有详细的解释。

完整代码如下：

import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_databatch_size = 100
learning_rate = 0.8
trainig_step = 30000n_input = 784
n_hidden = 500
n_labels = 10def inference(x_input):with tf.variable_scope("hidden"):weights = tf.get_variable("weights", [n_input, n_hidden], initializer = tf.random_normal_initializer(stddev = 0.1))biases = tf.get_variable("biases", [n_hidden], initializer = tf.constant_initializer(0.0))hidden = tf.nn.relu(tf.matmul(x_input, weights) + biases)with tf.variable_scope("out"):weights  = tf.get_variable("weights", [n_hidden, n_labels], initializer = tf.random_normal_initializer(stddev = 0.1))biases = tf.get_variable("biases", [n_labels], initializer = tf.constant_initializer(0.0))output = tf.matmul(hidden, weights) + biasesreturn outputdef train(mnist):x = tf.placeholder("float", [None, n_input])y = tf.placeholder("float", [None, n_labels])pred = inference(x)#计算损失函数cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits = pred, labels = y))#定义优化器optimizer = tf.train.GradientDescentOptimizer(learning_rate = learning_rate).minimize(cross_entropy)#定义准确率计算correct_pred = tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1))accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))init = tf.global_variables_initializer()with tf.Session() as sess:sess.run(init)#定义验证集与测试集validate_data = {x: mnist.validation.images, y: mnist.validation.labels}test_data = {x: mnist.test.images, y: mnist.test.labels}for i in range(trainig_step):#xs,ys为每个batch_size的训练数据与对应的标签xs, ys = mnist.train.next_batch(batch_size)_, loss = sess.run([optimizer, cross_entropy], feed_dict={x: xs, y:ys})#每1000次训练打印一次损失值与验证准确率if i % 1000 == 0:validate_accuracy = sess.run(accuracy, feed_dict=validate_data)print("after %d training steps, the loss is %g, the validation accuracy is %g" % (i, loss, validate_accuracy))print("the training is finish!")#最终的测试准确率acc = sess.run(accuracy, feed_dict=test_data)print("the test accuarcy is:", acc)def main(argv = None):mnist = input_data.read_data_sets("/tensorflow/mnst_data", one_hot=True)train(mnist)if __name__ == "__main__":tf.app.run()

最后执行的结果如图：