tensorflow的基本用法(九)——定义卷积神经网络训练MNIST

文章作者：Tyan
博客：noahsnail.com  |  CSDN  |  简书

本文主要是使用tensorflow定义卷积神经网络来训练MNIST数据集。定义的神经网络结构为两个卷积层+两个连接层，每个卷积层包括卷积层、ReLU层和Pooling层。

#!/usr/bin/env python# _*_ coding: utf-8 _*_import tensorflow as tffrom tensorflow.examples.tutorials.mnist import input_data# 定义神经网络模型的评估部分def compute_accuracy(test_xs, test_ys):    # 使用全局变量prediction    global prediction    # 获得预测值y_pre    y_pre = sess.run(prediction, feed_dict = { xs: test_xs, keep_prob: 1})    # 判断预测值y和真实值y_中最大数的索引是否一致，y_pre的值为1-10概率    correct_prediction = tf.equal(tf.argmax(y_pre, 1), tf.argmax(test_ys, 1))    # 定义准确率的计算    accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))    # 计算准确率    result = sess.run(accuracy)    return result# 下载mnist数据mnist = input_data.read_data_sets('MNIST_data', one_hot=True)# 权重参数初始化def weight_variable(shape):    initial = tf.truncated_normal(shape, stddev = 0.1)    return tf.Variable(initial)# 偏置参数初始化def bias_variable(shape):    initial = tf.constant(0.1, shape = shape)    return tf.Variable(initial)# 定义卷积层def conv2d(x, W):    # stride的四个参数：[batch, height, width, channels], [batch_size, image_rows, image_cols, number_of_colors]    # height, width就是图像的高度和宽度，batch和channels在卷积层中通常设为1    return tf.nn.conv2d(x, W, strides = [1, 1, 1, 1], padding = 'SAME')def max_pool_2x2(x):    return tf.nn.max_pool(x, ksize = [1, 2, 2, 1], strides = [1, 2, 2, 1], padding = 'SAME')# 输入输出数据的placeholderxs = tf.placeholder(tf.float32, [None, 784])ys = tf.placeholder(tf.float32, [None, 10])# dropout的比例keep_prob = tf.placeholder(tf.float32)# 对数据进行重新排列，形成图像x_image = tf.reshape(xs, [-1, 28, 28, 1])print x_image.shape# 卷积层一# patch为5*5，in_size为1，即图像的厚度，如果是彩色，则为3，32是out_size，输出的大小W_conv1 = weight_variable([5, 5, 1, 32])b_conv1 = bias_variable([32])# ReLU操作，输出大小为28*28*32h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)# Pooling操作，输出大小为14*14*32h_pool1 = max_pool_2x2(h_conv1)# 卷积层二# patch为5*5，in_size为32，即图像的厚度，64是out_size，输出的大小W_conv2 = weight_variable([5, 5, 32, 64])b_conv2 = bias_variable([64])# ReLU操作，输出大小为14*14*64h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)# Pooling操作，输出大小为7*7*64h_pool2 = max_pool_2x2(h_conv2)# 全连接层一W_fc1 = weight_variable([7 * 7 * 64, 1024])b_fc1 = bias_variable([1024])# 输入数据变换h_pool2_flat = tf.reshape(h_pool2, [-1, 7 * 7 * 64])# 进行全连接操作h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)# 防止过拟合，dropouth_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)# 全连接层二W_fc2 = weight_variable([1024, 10])b_fc2 = bias_variable([10])# 预测prediction = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)# 计算losscross_entropy = tf.reduce_mean(-tf.reduce_sum(ys * tf.log(prediction), reduction_indices=[1])) # 神经网络训练train_step = tf.train.AdamOptimizer(0.0001).minimize(cross_entropy)# 定义Sessionsess = tf.Session()# 根据tensorflow版本选择初始化函数if int((tf.__version__).split('.')[1]) < 12 and int((tf.__version__).split('.')[0]) < 1:    init = tf.initialize_all_variables()else:    init = tf.global_variables_initializer()# 执行初始化sess.run(init)# 进行训练迭代for i in range(1000):    # 取出mnist数据集中的100个数据    batch_xs, batch_ys = mnist.train.next_batch(100)    # 执行训练过程并传入真实数据    sess.run(train_step, feed_dict={xs: batch_xs, ys: batch_ys, keep_prob: 0.5})    if i % 100 == 0:        print compute_accuracy(mnist.test.images, mnist.test.labels)

执行结果如下：

$ python practice4.pyExtracting MNIST_data/train-images-idx3-ubyte.gzExtracting MNIST_data/train-labels-idx1-ubyte.gzExtracting MNIST_data/t10k-images-idx3-ubyte.gzExtracting MNIST_data/t10k-labels-idx1-ubyte.gz0.08230.8750.92430.94270.95020.95730.95950.96230.9630.9687