AlexNet

转自:https://blog.csdn.net/u012679707/article/details/80793916

【深度学习】AlexNet原理解析及实现

    Alex提出的alexnet网络结构模型,在imagenet2012图像分类challenge上赢得了冠军。

    要研究CNN类型DL网络模型在图像分类上的应用,就逃不开研究alexnet,这是CNN在图像分类上的经典模型。

一、Alexnet结构

alexNet为8层深度网络,其中5层卷积层和3层全连接层,不计LRN层和池化层。如下图所示:

    

                                                            图 Alexnet结构

详解各层训练参数的计算:

前五层:卷积层


后三层:全连接层

                    

整体计算图:

  

二、结构分析

        AlexNet每层的超参数如下图所示,其中输入尺寸为227*227,第一个卷积使用较大的核尺寸11*11,步长为4,有96个卷积核;紧接着一层LRN层;然后是最大池化层,核为3*3,步长为2。这之后的卷积层的核尺寸都比较小,5*5或3*3,并且步长为1,即扫描全图所有像素;而最大池化层依然为3*3,步长为2.

        我们可以发现,前几个卷积层的计算量很大,但参数量很小,只占Alexnet总参数的很小一部分。这就是卷积层的优点!通过较小的参数量来提取有效的特征。

        要注意,论文中指出,如果去掉任何一个卷积层,都会使网络的分类性能大幅下降。

            

三、AlexNet的新技术点

    AlexNet的新技术点(即大牛论文的contribution),如下:

(1)ReLU作为激活函数。

    ReLU为非饱和函数,论文中验证其效果在较深的网络超过了SIgmoid,成功解决了SIgmoid在网络较深时的梯度弥散问题

(2)Dropout避免模型过拟合

    在训练时使用Dropout随机忽略一部分神经元,以避免模型过拟合。在alexnet的最后几个全连接层中使用了Dropout。

(3)重叠的最大池化

    之前的CNN中普遍使用平均池化,而Alexnet全部使用最大池化,避免平均池化的模糊化效果。并且,池化的步长小于核尺寸,这样使得池化层的输出之间会有重叠和覆盖提升了特征的丰富性

(4)提出LRN层

    提出LRN层,对局部神经元的活动创建竞争机制,使得响应较大的值变得相对更大,并抑制其他反馈较小的神经元,增强了模型的泛化能力。

(5)GPU加速

(6)数据增强

    随机从256*256的原始图像中截取224*224大小的区域(以及水平翻转的镜像),相当于增强了(256-224)*(256-224)*2=2048倍的数据量。使用了数据增强后,减轻过拟合,提升泛化能力。避免因为原始数据量的大小使得参数众多的CNN陷入过拟合中。

四、AlexNet的搭建

    利用tensorflow实现ALexNet,环境为:win10+anaconda+python3+CPU(本人仅利用CPU,未使用GPU加速,所以最终模型训练速度较慢)。

    利用tensorboard可视化ALexNet结构为:

            

(1)首先看一下卷积层的搭建:带有LRN和池化层的卷积层

  1. with tf.name_scope('conv1') as scope:
  2. """
  3. images:227*227*3
  4. kernel: 11*11 *64
  5. stride:4*4
  6. padding:name
  7. #通过with tf.name_scope('conv1') as scope可以将scope内生成的Variable自动命名为conv1/xxx
  8. 便于区分不同卷积层的组建
  9. input: images[227*227*3]
  10. middle: conv1[55*55*96]
  11. output: pool1 [27*27*96]
  12. """
  13. kernel=tf.Variable(tf.truncated_normal([11,11,3,96],
  14. dtype=tf.float32,stddev=0.1),name="weights")
  15. conv=tf.nn.conv2d(images,kernel,[1,4,4,1],padding='SAME')
  16. biases=tf.Variable(tf.constant(0.0, shape=[96], dtype=tf.float32),
  17. trainable=True,name="biases")
  18. bias=tf.nn.bias_add(conv,biases) # w*x+b
  19. conv1=tf.nn.relu(bias,name=scope) # reLu
  20. print_architecture(conv1)
  21. parameters +=[kernel,biases]
  22. #添加LRN层和max_pool层
  23. """
  24. LRN会让前馈、反馈的速度大大降低(下降1/3),但最终效果不明显,所以只有ALEXNET用LRN,其他模型都放弃了
  25. """
  26. lrn1=tf.nn.lrn(conv1,depth_radius=4,bias=1,alpha=0.001/9,beta=0.75,name="lrn1")
  27. pool1=tf.nn.max_pool(lrn1,ksize=[1,3,3,1],strides=[1,2,2,1],
  28. padding="VALID",name="pool1")
  29. print_architecture(pool1)
(2)卷积层的搭建:不带有LRN和池化层的卷积层
  1. with tf.name_scope('conv3') as scope:
  2. """
  3. input: pool2[13*13*256]
  4. output: conv3 [13*13*384]
  5. """
  6. kernel = tf.Variable(tf.truncated_normal([3, 3, 256, 384],
  7. dtype=tf.float32, stddev=0.1), name="weights")
  8. conv = tf.nn.conv2d(pool2, kernel, [1, 1, 1, 1], padding='SAME')
  9. biases = tf.Variable(tf.constant(0.0, shape=[384], dtype=tf.float32),
  10. trainable=True, name="biases")
  11. bias = tf.nn.bias_add(conv, biases) # w*x+b
  12. conv3 = tf.nn.relu(bias, name=scope) # reLu
  13. parameters += [kernel, biases]
  14. print_architecture(conv3)

3)全连接层的搭建

  1. #全连接层6
  2. with tf.name_scope('fc6') as scope:
  3. """
  4. input:pool5 [6*6*256]
  5. output:fc6 [4096]
  6. """
  7. kernel = tf.Variable(tf.truncated_normal([6*6*256,4096],
  8. dtype=tf.float32, stddev=0.1), name="weights")
  9. biases = tf.Variable(tf.constant(0.0, shape=[4096], dtype=tf.float32),
  10. trainable=True, name="biases")
  11. # 输入数据变换
  12. flat = tf.reshape(pool5, [-1, 6*6*256] ) # 整形成m*n,列n为7*7*64
  13. # 进行全连接操作
  14. fc = tf.nn.relu(tf.matmul(flat, kernel) + biases,name='fc6')
  15. # 防止过拟合 nn.dropout
  16. fc6 = tf.nn.dropout(fc, keep_prob)
  17. parameters += [kernel, biases]
  18. print_architecture(fc6)

(4)训练测试:

    因未下载ImageNet数据集(太大),只是简单的测试了一下alexnet的性能。使用的是随机生成的图片来作为训练数据。

  1. def time_compute(session,target,info_string):
  2. num_step_burn_in=10 #预热轮数,头几轮迭代有显存加载、cache命中等问题可以因此跳过
  3. total_duration=0.0 #总时间
  4. total_duration_squared=0.0
  5. for i in range(num_batch+num_step_burn_in):
  6. start_time=time.time()
  7. _ = session.run(target)
  8. duration= time.time() -start_time
  9. if i>= num_step_burn_in:
  10. if i%10==0: #每迭代10次显示一次duration
  11. print("%s: step %d,duration=%.5f "% (datetime.now(),i-num_step_burn_in,duration))
  12. total_duration += duration
  13. total_duration_squared += duration *duration
  14. time_mean=total_duration /num_batch
  15. time_variance=total_duration_squared / num_batch - time_mean*time_mean
  16. time_stddev=math.sqrt(time_variance)
  17. #迭代完成,输出
  18. print("%s: %s across %d steps,%.3f +/- %.3f sec per batch "%
  19. (datetime.now(),info_string,num_batch,time_mean,time_stddev))
  20. def main():
  21. with tf.Graph().as_default():
  22. """仅使用随机图片数据 测试前馈和反馈计算的耗时"""
  23. image_size =224
  24. images=tf.Variable(tf.random_normal([batch_size,image_size,image_size,3],
  25. dtype=tf.float32,stddev=0.1 ) )
  26. fc8,parameters=inference(images)
  27. init=tf.global_variables_initializer()
  28. sess=tf.Session()
  29. sess.run(init)
  30. """
  31. AlexNet forward 计算的测评
  32. 传入的target:fc8(即最后一层的输出)
  33. 优化目标:loss
  34. 使用tf.gradients求相对于loss的所有模型参数的梯度
  35. AlexNet Backward 计算的测评
  36. target:grad
  37. """
  38. time_compute(sess,target=fc8,info_string="Forward")
  39. obj=tf.nn.l2_loss(fc8)
  40. grad=tf.gradients(obj,parameters)
  41. time_compute(sess,grad,"Forward-backward")

(5)测试结果:

    结构输出   (注意,32是我设置的batch_size,即训练的图片数量为32)

                

    前向预测用时:


    后向训练(学习)用时:


    可以看出后向训练用时比前向推理用时长很多,大概是5倍。


【附录】完整代码

  1. # -*- coding:utf-8 -*-
  2. """
  3. @author:Lisa
  4. @file:alexNet.py
  5. @function:实现Alexnet深度模型
  6. @note:learn from《tensorflow实战》
  7. @time:2018/6/24 0024下午 5:26
  8. """
  9. import tensorflow as tf
  10. import time
  11. import math
  12. from datetime import datetime
  13. batch_size=32
  14. num_batch=100
  15. keep_prob=0.5
  16. def print_architecture(t):
  17. """print the architecture information of the network,include name and size"""
  18. print(t.op.name," ",t.get_shape().as_list())
  19. def inference(images):
  20. """ 构建网络 :5个conv+3个FC"""
  21. parameters=[] #储存参数
  22. with tf.name_scope('conv1') as scope:
  23. """
  24. images:227*227*3
  25. kernel: 11*11 *64
  26. stride:4*4
  27. padding:name
  28. #通过with tf.name_scope('conv1') as scope可以将scope内生成的Variable自动命名为conv1/xxx
  29. 便于区分不同卷积层的组建
  30. input: images[227*227*3]
  31. middle: conv1[55*55*96]
  32. output: pool1 [27*27*96]
  33. """
  34. kernel=tf.Variable(tf.truncated_normal([11,11,3,96],
  35. dtype=tf.float32,stddev=0.1),name="weights")
  36. conv=tf.nn.conv2d(images,kernel,[1,4,4,1],padding='SAME')
  37. biases=tf.Variable(tf.constant(0.0, shape=[96], dtype=tf.float32),
  38. trainable=True,name="biases")
  39. bias=tf.nn.bias_add(conv,biases) # w*x+b
  40. conv1=tf.nn.relu(bias,name=scope) # reLu
  41. print_architecture(conv1)
  42. parameters +=[kernel,biases]
  43. #添加LRN层和max_pool层
  44. """
  45. LRN会让前馈、反馈的速度大大降低(下降1/3),但最终效果不明显,所以只有ALEXNET用LRN,其他模型都放弃了
  46. """
  47. lrn1=tf.nn.lrn(conv1,depth_radius=4,bias=1,alpha=0.001/9,beta=0.75,name="lrn1")
  48. pool1=tf.nn.max_pool(lrn1,ksize=[1,3,3,1],strides=[1,2,2,1],
  49. padding="VALID",name="pool1")
  50. print_architecture(pool1)
  51. with tf.name_scope('conv2') as scope:
  52. """
  53. input: pool1[27*27*96]
  54. middle: conv2[27*27*256]
  55. output: pool2 [13*13*256]
  56. """
  57. kernel = tf.Variable(tf.truncated_normal([5, 5, 96, 256],
  58. dtype=tf.float32, stddev=0.1), name="weights")
  59. conv = tf.nn.conv2d(pool1, kernel, [1, 1, 1, 1], padding='SAME')
  60. biases = tf.Variable(tf.constant(0.0, shape=[256], dtype=tf.float32),
  61. trainable=True, name="biases")
  62. bias = tf.nn.bias_add(conv, biases) # w*x+b
  63. conv2 = tf.nn.relu(bias, name=scope) # reLu
  64. parameters += [kernel, biases]
  65. # 添加LRN层和max_pool层
  66. """
  67. LRN会让前馈、反馈的速度大大降低(下降1/3),但最终效果不明显,所以只有ALEXNET用LRN,其他模型都放弃了
  68. """
  69. lrn2 = tf.nn.lrn(conv2, depth_radius=4, bias=1, alpha=0.001 / 9, beta=0.75, name="lrn1")
  70. pool2 = tf.nn.max_pool(lrn2, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1],
  71. padding="VALID", name="pool2")
  72. print_architecture(pool2)
  73. with tf.name_scope('conv3') as scope:
  74. """
  75. input: pool2[13*13*256]
  76. output: conv3 [13*13*384]
  77. """
  78. kernel = tf.Variable(tf.truncated_normal([3, 3, 256, 384],
  79. dtype=tf.float32, stddev=0.1), name="weights")
  80. conv = tf.nn.conv2d(pool2, kernel, [1, 1, 1, 1], padding='SAME')
  81. biases = tf.Variable(tf.constant(0.0, shape=[384], dtype=tf.float32),
  82. trainable=True, name="biases")
  83. bias = tf.nn.bias_add(conv, biases) # w*x+b
  84. conv3 = tf.nn.relu(bias, name=scope) # reLu
  85. parameters += [kernel, biases]
  86. print_architecture(conv3)
  87. with tf.name_scope('conv4') as scope:
  88. """
  89. input: conv3[13*13*384]
  90. output: conv4 [13*13*384]
  91. """
  92. kernel = tf.Variable(tf.truncated_normal([3, 3, 384, 384],
  93. dtype=tf.float32, stddev=0.1), name="weights")
  94. conv = tf.nn.conv2d(conv3, kernel, [1, 1, 1, 1], padding='SAME')
  95. biases = tf.Variable(tf.constant(0.0, shape=[384], dtype=tf.float32),
  96. trainable=True, name="biases")
  97. bias = tf.nn.bias_add(conv, biases) # w*x+b
  98. conv4 = tf.nn.relu(bias, name=scope) # reLu
  99. parameters += [kernel, biases]
  100. print_architecture(conv4)
  101. with tf.name_scope('conv5') as scope:
  102. """
  103. input: conv4[13*13*384]
  104. output: conv5 [6*6*256]
  105. """
  106. kernel = tf.Variable(tf.truncated_normal([3, 3, 384, 256],
  107. dtype=tf.float32, stddev=0.1), name="weights")
  108. conv = tf.nn.conv2d(conv4, kernel, [1, 1, 1, 1], padding='SAME')
  109. biases = tf.Variable(tf.constant(0.0, shape=[256], dtype=tf.float32),
  110. trainable=True, name="biases")
  111. bias = tf.nn.bias_add(conv, biases) # w*x+b
  112. conv5 = tf.nn.relu(bias, name=scope) # reLu
  113. pool5 = tf.nn.max_pool(conv5, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1],
  114. padding="VALID", name="pool5")
  115. parameters += [kernel, biases]
  116. print_architecture(pool5)
  117. #全连接层6
  118. with tf.name_scope('fc6') as scope:
  119. """
  120. input:pool5 [6*6*256]
  121. output:fc6 [4096]
  122. """
  123. kernel = tf.Variable(tf.truncated_normal([6*6*256,4096],
  124. dtype=tf.float32, stddev=0.1), name="weights")
  125. biases = tf.Variable(tf.constant(0.0, shape=[4096], dtype=tf.float32),
  126. trainable=True, name="biases")
  127. # 输入数据变换
  128. flat = tf.reshape(pool5, [-1, 6*6*256] ) # 整形成m*n,列n为7*7*64
  129. # 进行全连接操作
  130. fc = tf.nn.relu(tf.matmul(flat, kernel) + biases,name='fc6')
  131. # 防止过拟合 nn.dropout
  132. fc6 = tf.nn.dropout(fc, keep_prob)
  133. parameters += [kernel, biases]
  134. print_architecture(fc6)
  135. # 全连接层7
  136. with tf.name_scope('fc7') as scope:
  137. """
  138. input:fc6 [4096]
  139. output:fc7 [4096]
  140. """
  141. kernel = tf.Variable(tf.truncated_normal([4096, 4096],
  142. dtype=tf.float32, stddev=0.1), name="weights")
  143. biases = tf.Variable(tf.constant(0.0, shape=[4096], dtype=tf.float32),
  144. trainable=True, name="biases")
  145. # 进行全连接操作
  146. fc = tf.nn.relu(tf.matmul(fc6, kernel) + biases, name='fc7')
  147. # 防止过拟合 nn.dropout
  148. fc7 = tf.nn.dropout(fc, keep_prob)
  149. parameters += [kernel, biases]
  150. print_architecture(fc7)
  151. # 全连接层8
  152. with tf.name_scope('fc8') as scope:
  153. """
  154. input:fc7 [4096]
  155. output:fc8 [1000]
  156. """
  157. kernel = tf.Variable(tf.truncated_normal([4096, 1000],
  158. dtype=tf.float32, stddev=0.1), name="weights")
  159. biases = tf.Variable(tf.constant(0.0, shape=[1000], dtype=tf.float32),
  160. trainable=True, name="biases")
  161. # 进行全连接操作
  162. fc8 = tf.nn.xw_plus_b(fc7, kernel, biases, name='fc8')
  163. parameters += [kernel, biases]
  164. print_architecture(fc8)
  165. return fc8,parameters
  166. def time_compute(session,target,info_string):
  167. num_step_burn_in=10 #预热轮数,头几轮迭代有显存加载、cache命中等问题可以因此跳过
  168. total_duration=0.0 #总时间
  169. total_duration_squared=0.0
  170. for i in range(num_batch+num_step_burn_in):
  171. start_time=time.time()
  172. _ = session.run(target)
  173. duration= time.time() -start_time
  174. if i>= num_step_burn_in:
  175. if i%10==0: #每迭代10次显示一次duration
  176. print("%s: step %d,duration=%.5f "% (datetime.now(),i-num_step_burn_in,duration))
  177. total_duration += duration
  178. total_duration_squared += duration *duration
  179. time_mean=total_duration /num_batch
  180. time_variance=total_duration_squared / num_batch - time_mean*time_mean
  181. time_stddev=math.sqrt(time_variance)
  182. #迭代完成,输出
  183. print("%s: %s across %d steps,%.3f +/- %.3f sec per batch "%
  184. (datetime.now(),info_string,num_batch,time_mean,time_stddev))
  185. def main():
  186. with tf.Graph().as_default():
  187. """仅使用随机图片数据 测试前馈和反馈计算的耗时"""
  188. image_size =224
  189. images=tf.Variable(tf.random_normal([batch_size,image_size,image_size,3],
  190. dtype=tf.float32,stddev=0.1 ) )
  191. fc8,parameters=inference(images)
  192. init=tf.global_variables_initializer()
  193. sess=tf.Session()
  194. sess.run(init)
  195. """
  196. AlexNet forward 计算的测评
  197. 传入的target:fc8(即最后一层的输出)
  198. 优化目标:loss
  199. 使用tf.gradients求相对于loss的所有模型参数的梯度
  200. AlexNet Backward 计算的测评
  201. target:grad
  202. """
  203. time_compute(sess,target=fc8,info_string="Forward")
  204. obj=tf.nn.l2_loss(fc8)
  205. grad=tf.gradients(obj,parameters)
  206. time_compute(sess,grad,"Forward-backward")
  207. if __name__=="__main__":
  208. main()
------------------------------------------------------         END      ----------------------------------------------------------

参考:

《tensorflow实战》黄文坚(本文内容及代码大多源于此书,感谢!)

大牛论文《ImageNet Classification with Deep Convolutional Neural Networks Alex Krizhevsky

[caffe]深度学习之图像分类模型AlexNet解读  https://blog.csdn.net/sunbaigui/article/details/39938097(参数分析很详细)



原文地址:https://www.cnblogs.com/leebxo/p/10207320.html