建立一个带有全局平均池化层的卷积神经网络

一实例描述

使用全局平均池化层来代替传统的全连接层，使用了3个卷积层的同卷积操作，滤波器为5*5，每个卷积层后面都会跟个步长为2*2的池化层，滤波器为2*2。2层的卷积加池后是输出为10个通道卷积层，然后对这10个feature map进行全局平均池化，得到10个特征，再对这10个特征进行softmax计算，其结果来代表最终分类。

通过一个带有全局平均池化层的卷积神经网络对CIFAR数据集分类。

二代码

'''
导入头文件引入数据集
导入被切割后的24*24尺寸图片，每次取128个图片进行运算。
'''
import cifar10_input
import tensorflow as tf
import numpy as np
batch_size = 128
data_dir = '/tmp/cifar10_data/cifar-10-batches-bin'
print("begin")
images_train, labels_train = cifar10_input.inputs(eval_data = False,data_dir = data_dir, batch_size = batch_size)
images_test, labels_test = cifar10_input.inputs(eval_data = True, data_dir = data_dir, batch_size = batch_size)
print("begin data")
'''
定义网络结构
对于权重w的定义，统一使用函数truncated_normal生成标准差为0.1的随机数为其初始化。
对于权重b的定义，统一初始化为0.1。
'''
def weight_variable(shape):
  initial = tf.truncated_normal(shape, stddev=0.1)
  return tf.Variable(initial)
def bias_variable(shape):
  initial = tf.constant(0.1, shape=shape)
  return tf.Variable(initial)
'''
卷积操作的函数中，统一进行同卷积操作，即步长为1，padding='SAME'.
'''
def conv2d(x, W):
  return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')
'''
池化层有两个函数
一个是放在卷积后面，取最大值的方法，步长为2，padding='SAME',即将原尺寸的长和宽各除以2
另外一个是用来放在最后一层，取均值的方法，步长为最终生成的特征尺寸6*6（24*24经过两次池化变成了6*6），filter也为6*6.
倒数第二层没有最大池化的卷积层，因为共有10类，所有卷积输出的是10个通道，并使其全局平均池化为10个节点。
'''
def max_pool_2x2(x):
  return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],
                        strides=[1, 2, 2, 1], padding='SAME')  
                        
def avg_pool_6x6(x):
  return tf.nn.avg_pool(x, ksize=[1, 6, 6, 1],
                        strides=[1, 6, 6, 1], padding='SAME')
# tf Graph Input
x = tf.placeholder(tf.float32, [None, 24,24,3]) # cifar data image of shape 24*24*3
y = tf.placeholder(tf.float32, [None, 10]) # 0-9 数字=> 10 classes
W_conv1 = weight_variable([5, 5, 3, 64])
b_conv1 = bias_variable([64])
x_image = tf.reshape(x, [-1,24,24,3])
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
h_pool1 = max_pool_2x2(h_conv1)
W_conv2 = weight_variable([5, 5, 64, 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)
W_conv3 = weight_variable([5, 5, 64, 10])
b_conv3 = bias_variable([10])
h_conv3 = tf.nn.relu(conv2d(h_pool2, W_conv3) + b_conv3)
nt_hpool3=avg_pool_6x6(h_conv3)#10
nt_hpool3_flat = tf.reshape(nt_hpool3, [-1, 10])
y_conv=tf.nn.softmax(nt_hpool3_flat)
cross_entropy = -tf.reduce_sum(y*tf.log(y_conv))
'''
对于梯度优化算法，还是和多分类问题一样，使用AdamOptimizer函数，学习率为0.0001
'''
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(y,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
'''
运行session进行训练，启动session，迭代15000次数据集，这里记得要启动队列，同时读出来的label还要转成onehot编码
'''
sess = tf.Session()
sess.run(tf.global_variables_initializer())
tf.train.start_queue_runners(sess=sess)
for i in range(15000):#20000
  image_batch, label_batch = sess.run([images_train, labels_train])
  label_b = np.eye(10,dtype=float)[label_batch] #one hot
  
  train_step.run(feed_dict={x:image_batch, y: label_b},session=sess)
  
  if i%200 == 0:
    train_accuracy = accuracy.eval(feed_dict={
        x:image_batch, y: label_b},session=sess)
    print( "step %d, training accuracy %g"%(i, train_accuracy))
'''
评估结果
从测试数据集里将数据取出，放到模型里运行，查看模型的正确率。
'''
image_batch, label_batch = sess.run([images_test, labels_test])
label_b = np.eye(10,dtype=float)[label_batch]#one hot
print ("finished！ test accuracy %g"%accuracy.eval(feed_dict={
     x:image_batch, y: label_b},session=sess))

三运行结果

begin

begin data

step 0, training accuracy 0.09375

step 200, training accuracy 0.367188

step 400, training accuracy 0.367188

step 600, training accuracy 0.382812

step 800, training accuracy 0.460938

step 1000, training accuracy 0.492188

step 1200, training accuracy 0.578125

step 1400, training accuracy 0.53125

step 1600, training accuracy 0.515625

step 1800, training accuracy 0.445312

step 2000, training accuracy 0.5625

step 2200, training accuracy 0.570312

step 2400, training accuracy 0.515625

step 2600, training accuracy 0.554688

step 2800, training accuracy 0.515625

step 3000, training accuracy 0.539062

step 3200, training accuracy 0.5

step 3400, training accuracy 0.515625

step 3600, training accuracy 0.539062

step 3800, training accuracy 0.445312

step 4000, training accuracy 0.554688

step 4200, training accuracy 0.507812

step 4400, training accuracy 0.5625

step 4600, training accuracy 0.617188

step 4800, training accuracy 0.65625

step 5000, training accuracy 0.65625

step 5200, training accuracy 0.578125

step 5400, training accuracy 0.625

step 5600, training accuracy 0.484375

step 5800, training accuracy 0.59375

step 6000, training accuracy 0.578125

step 6200, training accuracy 0.6875

step 6400, training accuracy 0.539062

step 6600, training accuracy 0.609375

step 6800, training accuracy 0.523438

step 7000, training accuracy 0.640625

step 7200, training accuracy 0.578125

step 7400, training accuracy 0.640625

step 7600, training accuracy 0.625

step 7800, training accuracy 0.625

step 8000, training accuracy 0.617188

step 8200, training accuracy 0.570312

step 8400, training accuracy 0.648438

step 8600, training accuracy 0.601562

step 8800, training accuracy 0.617188

step 9000, training accuracy 0.609375

step 9200, training accuracy 0.65625

step 9400, training accuracy 0.625

step 9600, training accuracy 0.703125

step 9800, training accuracy 0.5625

step 10000, training accuracy 0.617188

step 10200, training accuracy 0.625

step 10400, training accuracy 0.703125

step 10600, training accuracy 0.625

step 10800, training accuracy 0.617188

step 11000, training accuracy 0.703125

step 11200, training accuracy 0.648438

step 11400, training accuracy 0.601562

step 11600, training accuracy 0.585938

step 11800, training accuracy 0.601562

step 12000, training accuracy 0.65625

step 12200, training accuracy 0.648438

step 12400, training accuracy 0.648438

step 12600, training accuracy 0.75

step 12800, training accuracy 0.585938

step 13000, training accuracy 0.703125

step 13200, training accuracy 0.609375

step 13400, training accuracy 0.671875

step 13600, training accuracy 0.703125

step 13800, training accuracy 0.648438

step 14000, training accuracy 0.671875

step 14200, training accuracy 0.640625

step 14400, training accuracy 0.632812

step 14600, training accuracy 0.679688

step 14800, training accuracy 0.625

finished！ test accuracy 0.640625

四运行说明

可以看出，识别效果得到了收敛，正确率在0.6左右，这个正确率不算很高，因为模型相对简单，只是用了两层的卷积操作。

实例中对于卷积核池化的使用也表明了一种习惯，一般在卷积过程中都会设为步长为1的same卷积，即大小不变，需要降维时则是全部通过池化来完成的。

建立一个带有全局平均池化层的卷积神经网络

猜你喜欢