tensorflow 卷积神经网络实现

来源:互联网 发布:算法初步例题 编辑:程序博客网 时间:2024/06/05 12:32

之前用deeplearning4j实现卷积神经网络,实际卷积比较好理解,这里刚刚好有一篇博客写的是比较详细的,我记录下,

卷积神经网络在tensorflow下实现情况,主要参考代码是:http://www.jeyzhang.com/tensorflow-learning-notes-2.html


代码如下:
  1. # -*- coding: utf-8 -*- import tensorflow as tf#导入input_data用于自动下载和安装MNIST数据集from tensorflow.examples.tutorials.mnist import input_datamnist = input_data.read_data_sets("MNIST_data/", one_hot=True)#创建一个交互式Sessionsess = tf.InteractiveSession()#创建两个占位符,x为输入网络的图像,y_为输入网络的图像类别x = tf.placeholder("float", shape=[None, 784])y_ = tf.placeholder("float", shape=[None, 10])#权重初始化函数def weight_variable(shape): #输出服从截尾正态分布的随机值 initial = tf.truncated_normal(shape, stddev=0.1) return tf.Variable(initial)#偏置初始化函数def bias_variable(shape): initial = tf.constant(0.1, shape=shape) return tf.Variable(initial)#创建卷积op#x 是一个4维张量,shape为[batch,height,width,channels]#卷积核移动步长为1。填充类型为SAME,可以不丢弃任何像素点def conv2d(x, W): return tf.nn.conv2d(x, W, strides=[1,1,1,1], padding="SAME")#创建池化op#采用最大池化,也就是取窗口中的最大值作为结果#x 是一个4维张量,shape为[batch,height,width,channels]#ksize表示pool窗口大小为2x2,也就是高2,宽2#strides,表示在height和width维度上的步长都为2def max_pool_2x2(x): return tf.nn.max_pool(x, ksize=[1,2,2,1], strides=[1,2,2,1], padding="SAME")#第1层,卷积层#初始化W为[5,5,1,32]的张量,表示卷积核大小为5*5,第一层网络的输入和输出神经元个数分别为1和32W_conv1 = weight_variable([5,5,1,32])#初始化b为[32],即输出大小b_conv1 = bias_variable([32])#把输入x(二维张量,shape为[batch, 784])变成4d的x_image,x_image的shape应该是[batch,28,28,1]#-1表示自动推测这个维度的sizex_image = tf.reshape(x, [-1,28,28,1])#把x_image和权重进行卷积,加上偏置项,然后应用ReLU激活函数,最后进行max_pooling#h_pool1的输出即为第一层网络输出,shape为[batch,14,14,1]h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)h_pool1 = max_pool_2x2(h_conv1)#第2层,卷积层#卷积核大小依然是5*5,这层的输入和输出神经元个数为32和64W_conv2 = weight_variable([5,5,32,64])b_conv2 = weight_variable([64])#h_pool2即为第二层网络输出,shape为[batch,7,7,1]h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)h_pool2 = max_pool_2x2(h_conv2)#第3层, 全连接层#这层是拥有1024个神经元的全连接层#W的第1维size为7*7*64,7*7是h_pool2输出的size,64是第2层输出神经元个数W_fc1 = weight_variable([7*7*64, 1024])b_fc1 = bias_variable([1024])#计算前需要把第2层的输出reshape成[batch, 7*7*64]的张量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)#Dropout层#为了减少过拟合,在输出层前加入dropoutkeep_prob = tf.placeholder("float")h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)#输出层#最后,添加一个softmax层#可以理解为另一个全连接层,只不过输出时使用softmax将网络输出值转换成了概率W_fc2 = weight_variable([1024, 10])b_fc2 = bias_variable([10])y_conv = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)#预测值和真实值之间的交叉墒cross_entropy = -tf.reduce_sum(y_ * tf.log(y_conv))#train op, 使用ADAM优化器来做梯度下降。学习率为0.0001train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)#评估模型,tf.argmax能给出某个tensor对象在某一维上数据最大值的索引。#因为标签是由0,1组成了one-hot vector,返回的索引就是数值为1的位置correct_predict = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))#计算正确预测项的比例,因为tf.equal返回的是布尔值,#使用tf.cast把布尔值转换成浮点数,然后用tf.reduce_mean求平均值accuracy = tf.reduce_mean(tf.cast(correct_predict, "float"))#初始化变量sess.run(tf.initialize_all_variables())#开始训练模型,循环20000次,每次随机从训练集中抓取50幅图像
    for i in range(2000):    batch = mnist.train.next_batch(50)    if i%100 == 0:        #每100次输出一次日志        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))    train_step.run(feed_dict={x:batch[0], y_:batch[1], keep_prob:0.5})print("test accuracy {0}".format( accuracy.eval(feed_dict={x:mnist.test.images, y_:mnist.test.labels, keep_prob:1.0})))
结果:

Extracting 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.gzstep 0, training accuracy 0.04step 1000, training accuracy 1step 2000, training accuracy 0.98step 3000, training accuracy 1