本博客将建立一个简单的卷积神经网络，可以把MNIST手写字符的识别准确率提高到99%。具体如下：

程序的开头是导入TensorFlow

import tensorflow as tfimport numpy as npimport os os.environ["CUDA_VISIBLE_DEVICES"] = "0"#指定GPU#从tensorflow.examples.tutorials.mnist引入模块from tensorflow.examples.tutorials.mnist import input_data

查看训练集、验证集、测试集的数据大小

#查看训练数据的大小print(mnist.train.images.shape) # (55000, 784)print(mnist.train.labels.shape) # (55000, 10)#查看验证数据的大小print(mnist.validation.images.shape) # (5000, 784)print(mnist.validation.labels.shape) # (5000, 10)#查看测试数据的大小print(mnist.test.images.shape) # (10000, 784)print(mnist.test.labels.shape) # (10000, 10)

#打印出第0幅图片的向量表示print(mnist.train.images[0, :])#打印出第0幅图片的标签print(mnist.train.labels[0, :])

#查看前20张训练图片的labelfor i in range(20):#得到one-hot表示，形如(0, 1, 0, 0, 0, 0, 0, 0, 0, 0)one_hot_label = mnist.train.labels[i, :]# 通过np.argmax我们可以直接获得原始的label# 因为只有1位为1，其他都是0label = np.argmax(one_hot_label)print('mnist_train_%d.jpg label: %d' % (i, label))

完整代码为

import tensorflow as tfimport numpy as npimport os os.environ["CUDA_VISIBLE_DEVICES"] = "0"#指定GPU#从tensorflow.examples.tutorials.mnist引入模块from tensorflow.examples.tutorials.mnist import input_data#从MNIST_data/中读取MNIST数据,当数据不存在时，会自动执行下载mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)#查看训练数据的大小print(mnist.train.images.shape) # (55000, 784)print(mnist.train.labels.shape) # (55000, 10)#查看验证数据的大小print(mnist.validation.images.shape) # (5000, 784)print(mnist.validation.labels.shape) # (5000, 10)#查看测试数据的大小print(mnist.test.images.shape) # (10000, 784)print(mnist.test.labels.shape) # (10000, 10)#打印出第0幅图片的向量表示print(mnist.train.images[0, :])#打印出第0幅图片的标签print(mnist.train.labels[0, :])#查看前20张训练图片的labelfor i in range(20):#得到one-hot表示，形如(0, 1, 0, 0, 0, 0, 0, 0, 0, 0)one_hot_label = mnist.train.labels[i, :]# 通过np.argmax我们可以直接获得原始的label# 因为只有1位为1，其他都是0label = np.argmax(one_hot_label)print('mnist_train_%d.jpg label: %d' % (i, label))#定义权重def weight_variable(shape):initial = tf.truncated_normal(shape, stddev=0.1)return tf.Variable(initial)#定义偏置def bias_variable(shape):initial = tf.constant(0.1, shape=shape)return tf.Variable(initial)#卷积层，步长为1def conv2d(x, W):return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')#最大池化层，步长为2def max_pool_2x2(x):return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],strides=[1, 2, 2, 1], padding='SAME')if __name__ == '__main__':#x为训练图像的占位符，输入图像是28*28=784x = tf.placeholder(tf.float32, [None, 784])#y_为训练图像标签的占位符，输出类别是10类y_ = tf.placeholder(tf.float32, [None, 10])#将单张图片从784维向量重新还原为28x28的矩阵图片x_image = tf.reshape(x, [-1, 28, 28, 1])with tf.variable_scope("conv1"):#第一层卷积层和池化层，卷积核大小为5*5，卷积核个数为32，W_conv1 = weight_variable([5, 5, 1, 32])b_conv1 = bias_variable([32])h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)h_pool1 = max_pool_2x2(h_conv1)with tf.variable_scope("conv2"):#第二层卷积层和池化层，卷积核大小为5*5，卷积核个数为64W_conv2 = weight_variable([5, 5, 32, 64])b_conv2 = bias_variable([64])h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)h_pool2 = max_pool_2x2(h_conv2)with tf.variable_scope("fc1"):#全连接层，输出为1024维的向量W_fc1 = weight_variable([7 * 7 * 64, 1024])b_fc1 = bias_variable([1024])h_pool2_flat = tf.reshape(h_pool2, [-1, 7 * 7 * 64])h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)#使用Dropout，keep_prob是一个占位符，训练时为0.5，测试时为1keep_prob = tf.placeholder(tf.float32)h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)#把1024维的向量转换成10维，对应10个类别W_fc2 = weight_variable([1024, 10])b_fc2 = bias_variable([10])y_conv = tf.matmul(h_fc1_drop, W_fc2) + b_fc2with tf.variable_scope("loss"):#不采用先Softmax再计算交叉熵的方法，而是直接用tf.nn.softmax_cross_entropy_with_logits直接计算loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=y_conv))with tf.variable_scope("optimizer"):#同样定义train_steptrain_step = tf.train.AdamOptimizer(1e-4).minimize(loss)with tf.variable_scope("acc"):#定义测试的准确率，tf.argmax（y，1）、tf.argmax（y_，1）的功能是取出数组中最大值的下标correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))#收集变量，单个数字值收集#记录loss，accuracy变化曲线tf.summary.scalar("losses",loss)tf.summary.scalar("acc",accuracy)#高纬度变量收集tf.summary.histogram('weightes',W_conv1)tf.summary.histogram('biases',b_conv1)#创建Session和变量初始化init_op = tf.global_variables_initializer()#定义一个合并变量的opmerged = tf.summary.merge_all()with tf.Session() as sess:#初始化所有变量sess.run(init_op)#建立events文件，然后写入filewriter = tf.summary.FileWriter('./tensorboard/',graph=sess.graph)#训练20000步for i in range(20000):batch = mnist.train.next_batch(64)train_step.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})#写入每步训练的值summary = sess.run(merged,feed_dict={x:batch[0], y_:batch[1], keep_prob: 0.5})filewriter.add_summary(summary,i)#每100步显示一次在验证集上的准确度if i % 100 == 0:#显示训练时的准确率train_accuracy = accuracy.eval(feed_dict={x: batch[0], y_: batch[1], keep_prob: 1.0})print("step %d, training accuracy %g" % (i, train_accuracy))#在验证集上的准确度print("validation accuracy %g" % accuracy.eval(feed_dict={x: mnist.validation.images, y_: mnist.validation.labels, keep_prob: 1.0}))#训练结束后报告在测试集上的准确度print("test accuracy %g" % accuracy.eval(feed_dict={x: mnist.test.images, y_: mnist.test.labels, keep_prob: 1.0}))

运行结果如下图所示

图

损失函数变化曲线

准确率变化曲线