TensorFlow学习笔记10：Inception V3

发表于 分类于 ， Valine： 本文字数： 25k

Inception V3结构

值得借鉴的设计CNN的思想和Trick：

  • Factorization into small convolutions很有效，可以降低参数量，减轻过拟合，增加网络非线性的表达能力。
  • 卷积网络从输入到输出，应该让图片尺寸逐渐减少，输出通道逐渐增加，即让空间结构简化，将空间信息转化为高阶抽象的特征信息。
  • Inception Module用多个分支提取不同抽象程度的高阶特征的思路很有效，可以丰富网络的表达能力。

TensorFlow实现

定义函数 inception_v3_arg_scope

函数 inception_v3_arg_scope 用来生成网络中经常用到的函数的默认参数，比如卷记的激活函数，权重初始化方式，标准化器等等。接下来嵌套一个slim.arg_scope，对卷积层生成函数slim.conv2d的几个人参数赋予默认值。最后返回定义好的scope。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
import tensorflow as tf
import tensorflow.contrib.slim as slim
#定义简单的函数产生截断的正态分布
trunc_normal = lambda stddev:tf.truncated_normal_initializer(0.0,stddev)

'''
定义函数 inception_v3_arg_scope 用来生成网络中经常用到的函数的默认参数
L2正则的Weight_decay默认值为0.0004
标准差stddev默认值为0.1
参数batch_norm_var_collection默认值为moving_vars
batch normalization参数字典：
    衰减系数decay为0.9997
    epsilon为0.001
    updates_collections为tf.GraphKeys.UPDATE_OPS
    字典variables_collections:
        beta和gamma设置为None
        moving_mean和moving_variance设置为batch_norm_var_collection

'''
def inception_v3_arg_scope(weight_decay=0.00004,stddev=0.1,
                        batch_norm_var_collection="moving_vars"):
    batch_norm_params = {
        "decay":0.9997,"epsilon":0.001,"updates_collections":tf.GraphKeys.UPDATE_OPS,
        "variables_collections":{
            "beta":None,"gamma":None,"moving_mean":[batch_norm_var_collection],
            "moving_variance":[batch_norm_var_collection]
        }
    }

    '''
    嵌套slim.arg_scope
    权重初始化器weight_initializer设置为trunc_normal(stddev)
    激活函数设置为ReLU
    标准化器设置为slim.batch_norm
    标准化器的参数设置为batch_norm_patams
    '''
    with slim.arg_scope([slim.conv2d,slim.fully_connected],
                        weights_regularizer=slim.l2_regularizer(weight_decay)):
        #对卷积层生成函数的几个参数赋予默认值
        with slim.arg_scope([slim.conv2d],
                            weights_regularizer = tf.truncated_normal_initializer(stddev=stddev),
                            activation_fc = tf.nn.relu,
                            normalizer_fc = slim.batch_norm,
                            normalizer_params = batch_norm_params) as scope:
            return scope

定义Inception V3的卷积部分

它可以生成Inception V3网络的卷积部分。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
'''
定义Inception V3的卷积部分
inputs为输入图片数据的tensor
scope为包含了函数默认参数的环境
用字典表end_points来保存某些关键节点供之后使用
'''
def inception_v3_base(inputs,scope=None):
    end_points = {}
    with tf.variable_scope(scope,"InceptionV3",[inputs]):
        #对slim.conv2d,slim.max_pool2d,slim.avg_pool2d三个函数的参数设置默认值
        with slim.arg_scope([slim.conv2d,slim.max_pool2d,slim.avg_pool2d],stride = 1,padding = "VALID"):                               
            #各参数分别为输入的tensor,输出的通道,卷积核尺寸,步长stride，padding模式
            net = slim.conv2d(inputs,num_outputs=32,kernel_size=[3,3],stride=2,scope="Conv2d_1a_3x3")
            net = slim.conv2d(net,num_outputs=32,kernel_size=[3,3],scope="Conv2d_2a_3x3")
            net = slim.conv2d(net,num_outputs=64,kernel_size=[3,3],padding="SAME",scope="Conv2d_2b_3x3")
            net = slim.max_pool2d(net,kernel_size=[3,3],stride=2,scope="MaxPool_3a_3x3")
            net = slim.conv2d(net,num_outputs=80,kernel_size=[1,1],scope="Conv2d_3b_1x1")
            net = slim.conv2d(net,num_outputs=192,kernel_size=[3,3],scope="Conv2d_4a_3x3")
            net = slim.max_pool2d(net,kernel_size=[3,3],stride=2,scope="MaxPool_5a_3x3")

定义第一个Inception模块

  • 第一个Inception Module
    第一个Inception Module名称为Mixed_5b。这个Inception Module中有4个分支：

    1. 第一个分支有64输出通道的1×1卷积
    2. 第二个分支有48输出通道的1×1卷积，连接有64输出通道的5×5卷积
    3. 第三个分支有64输出通道的1×1卷积，再连续连接两个有96通道的3×3卷积

    4. 第四个分支有为3×3的平均池化，连接有32输出通道的1×1卷积

      最后使用tf.concat将四个分支的输出合并在一起，生成这个Inception Module的最终输出。4个分支的输出通道数之和为64+64+96+32=256，即最终输出的图片尺寸为35×35×256。

      1
      2
      3
      4
      5
      6
      7
      8
      9
      10
      11
      12
      13
      14
      15
      16
      17
      18
      #定义第一个Inception模块组
      with slim.arg_scope([slim.conv2d,slim.max_pool2d,slim.avg_pool2d],
                          stride = 1,padding = "SAME"):
          with tf.variable_scope("Mixed_5b"):
              with tf.variable_scope("Branch_0"):
                  batch_0 = slim.conv2d(net,num_outputs=64,kernel_size=[1,1],scope="Conv2d_0a_1x1")
              with tf.variable_scope("Branch_1"):
                  batch_1 = slim.conv2d(net,num_outputs=48,kernel_size=[1,1],scope="Conv2d_0a_1x1")
                  batch_1 = slim.conv2d(batch_1,num_outputs=64,kernel_size=[5,5],scope="Conv2d_0b_5x5")
              with tf.variable_scope("Branch_2"):
                  batch_2 = slim.conv2d(net,num_outputs=64,kernel_size=[1,1],scope="Conv2d_0a_1x1")
                  batch_2 = slim.conv2d(batch_2,num_outputs=96,kernel_size=[3,3],scope="Conv2d_0b_3x3")
                  batch_2 = slim.conv2d(batch_2,num_outputs=96,kernel_size=[3,3],scope="Conv2d_0c_3x3")
              with tf.variable_scope("Branch_3"):
                  batch_3 = slim.avg_pool2d(net,kernel_size=[3,3],scope="AvgPool_0a_3x3")
                  batch_3 = slim.conv2d(batch_3,num_outputs=32,kernel_size=[1,1],scope="Conv2d_0b_1x1")
       
              net = tf.concat([batch_0,batch_1,batch_2,batch_3],3)

  • 第二个Inception Module
    第二个Inception Module的名称为Mixed_5c。它同样也有四个分支，唯一不同的是第四个分支最后接的是64输出通道的1×1卷积。因此我们输出的tensor的最终尺寸为35×5×288。

    1
    2
    3
    4
    5
    6
    7
    8
    9
    10
    11
    12
    13
    14
    15
    16
    # 第二个Inception模块
    with tf.variable_scope("Mixed_5c"):
        with tf.variable_scope("Branch_0"):
            batch_0 = slim.conv2d(net, num_outputs=64, kernel_size=[1, 1], scope="Conv2d_0a_1x1")
        with tf.variable_scope("Branch_1"):
            batch_1 = slim.conv2d(net, num_outputs=48, kernel_size=[1, 1], scope="Conv2d_0b_1x1")
            batch_1 = slim.conv2d(batch_1, num_outputs=64, kernel_size=[5, 5], scope="Conv2d_0c_5x5")
        with tf.variable_scope("Branch_2"):
            batch_2 = slim.conv2d(net, num_outputs=64, kernel_size=[1, 1], scope="Conv2d_0a_1x1")
            batch_2 = slim.conv2d(batch_2, num_outputs=96, kernel_size=[3, 3], scope="Conv2d_0b_3x3")
            batch_2 = slim.conv2d(batch_2, num_outputs=96, kernel_size=[3, 3], scope="Conv2d_0c_3x3")
        with tf.variable_scope("Branch_3"):
            batch_3 = slim.avg_pool2d(net, kernel_size=[3, 3], scope="AvgPool_0a_3x3")
            batch_3 = slim.conv2d(batch_3, num_outputs=64, kernel_size=[1, 1], scope="Conv2d_0b_1x1")
        
        net = tf.concat([batch_0, batch_1, batch_2, batch_3], 3)

  • 第三个Inception Module
    第三个Inception Module的名称为Mixed_5d，和上一个Inception Module完全相同。

    1
    2
    3
    4
    5
    6
    7
    8
    9
    10
    11
    12
    13
    14
    15
    16
    # 第三个Inception模块
    with tf.variable_scope("Mixed_5d"):
        with tf.variable_scope("Branch_0"):
            batch_0 = slim.conv2d(net, num_outputs=64, kernel_size=[1, 1], scope="Conv2d_0a_1x1")
        with tf.variable_scope("Branch_1"):
            batch_1 = slim.conv2d(net, num_outputs=48, kernel_size=[1, 1], scope="Conv2d_0b_1x1")
            batch_1 = slim.conv2d(batch_1, num_outputs=64, kernel_size=[5, 5], scope="Conv2d_0c_5x5")
        with tf.variable_scope("Branch_2"):
            batch_2 = slim.conv2d(net, num_outputs=64, kernel_size=[1, 1], scope="Conv2d_0a_1x1")
            batch_2 = slim.conv2d(batch_2, num_outputs=96, kernel_size=[3, 3], scope="Conv2d_0b_3x3")
            batch_2 = slim.conv2d(batch_2, num_outputs=96, kernel_size=[3, 3], scope="Conv2d_0c_3x3")
        with tf.variable_scope("Branch_3"):
            batch_3 = slim.avg_pool2d(net, kernel_size=[3, 3], scope="AvgPool_0a_3x3")
            batch_3 = slim.conv2d(batch_3, num_outputs=64, kernel_size=[1, 1], scope="Conv2d_0b_1x1")

        net = tf.concat([batch_0, batch_1, batch_2, batch_3], 3)

定义第二个Inception模块组

第二个Inception模块组是一个非常大的模块组，包含了5个Inception Module。其中第二个到第五个Inception Module的结构非常相似。

  • 第一个Inception Module
    第一个Inception Module的名称为Mixed_6a，包含三个分支：

    1. 第一个分支为384通道的3×3卷积，步长为2
    2. 第二个分支有三层，分别为64输出通道的1×1卷积，和两个96输出通道的3×3卷积，最后一层的步长为2。
    3. 第三个分支为3×3的池化层，步长为2
      最后用tf.concat将三个分支在输出通道上合并，最后的输出尺寸为17×17×(384+96+256)=17×17×768
      1
      2
      3
      4
      5
      6
      7
      8
      9
      10
      11
      12
      13
      14
      15
      # 定义第二个Inception模块组,第一个Inception模块
      with tf.variable_scope("Mixed_6a"):
          with tf.variable_scope("Branch_0"):
              batch_0 = slim.conv2d(net, num_outputs=384, kernel_size=[3,3],
                                      stride=2, padding="VALID",scope="Conv2d_1a_1x1")
          with tf.variable_scope("Branch_1"):
              batch_1 = slim.conv2d(net, num_outputs=64, kernel_size=[1, 1], scope="Conv2d_0a_1x1")
              batch_1 = slim.conv2d(batch_1, num_outputs=96, kernel_size=[3, 3], scope="Conv2d_0b_3x3")
              batch_1 = slim.conv2d(batch_1, num_outputs=96, kernel_size=[3, 3],
                                      stride=2, padding="VALID",scope="Conv2d_1a_1x1")
          with tf.variable_scope("Branch_2"):
              batch_2 = slim.max_pool2d(net,kernel_size=[3,3],stride=2,padding="VALID",
                                          scope="MaxPool_1a_3x3")

          net = tf.concat([batch_0, batch_1, batch_2], 3)

  • 第二个Inception Module
    名称为Mixed_6b，它有四个分支：

    1. 第一个分支为193输出通道的1×1卷积
    2. 第二个分支有三个卷积层，分别为128输出通道的1×1卷积，128输出通道的1×7卷积，以及192输出通道的7×1卷积，这里用到了Factorization into small convolutions思想，串联的1×7卷积和7×1卷积相当于合成一个7×7卷积。大大减少了参数，减轻了过拟合，同事多了一个激活函数增加了非线性特征变换。
    3. 第三个分支有五个卷积层，分别为128输出通道的1×1卷积，128输出通道的7×1卷积，128输出通道的1×1卷积，128输出通道的1×7卷积，128输出通道的7×1卷积和192输出通道的1×7卷积
    4. 第四个分支为3×3的平均池化层，再连接192输出通道的1×1卷积。
      最后四个分支合并，输出的tensor尺寸为17×17×(192+192+192+192)=17×17×768
      1
      2
      3
      4
      5
      6
      7
      8
      9
      10
      11
      12
      13
      14
      15
      16
      17
      18
      19
      # 定义第二个Inception模块组,第一个Inception模块
      with tf.variable_scope("Mixed_6b"):
          with tf.variable_scope("Branch_0"):
              batch_0 = slim.conv2d(net,num_outputs=192,kernel_size=[1,1],scope="Conv2d_0a_1x1")
          with tf.variable_scope("Branch_1"):
              batch_1 = slim.conv2d(net, num_outputs=128, kernel_size=[1, 1], scope="Conv2d_0a_1x1")
              batch_1 = slim.conv2d(batch_1, num_outputs=128, kernel_size=[1,7], scope="Conv2d_0b_1x7")
              batch_1 = slim.conv2d(batch_1, num_outputs=192, kernel_size=[7, 1],scope="Conv2d_0c_7x1")
          with tf.variable_scope("Branch_2"):
              batch_2 = slim.conv2d(net, num_outputs=128, kernel_size=[1, 1], scope="Conv2d_0a_1x1")
              batch_2 = slim.conv2d(batch_2, num_outputs=128, kernel_size=[7, 1], scope="Conv2d_0b_7x1")
              batch_2 = slim.conv2d(batch_2, num_outputs=128, kernel_size=[1, 7], scope="Conv2d_0c_1x7")
              batch_2 = slim.conv2d(batch_2, num_outputs=128, kernel_size=[7, 1], scope="Conv2d_0d_7x1")
              batch_2 = slim.conv2d(batch_2, num_outputs=192, kernel_size=[1, 7], scope="Conv2d_0e_1x7")
          with tf.variable_scope("Branch_3"):
              batch_3 = slim.avg_pool2d(net, kernel_size=[3, 3], scope="AvgPool_0a_3x3")
              batch_3 = slim.conv2d(batch_3, num_outputs=192, kernel_size=[1, 1], scope="Conv2d_0b_1x1")

          net = tf.concat([batch_0, batch_1, batch_2,batch_3], 3)

  • 第三个Inception Module
    名称为Mixed_6c，和前一个Inception Module非常相似，唯一不同的地方就是第二和第三分支中前几个卷积层的输出通道从128变成了160。

    1
    2
    3
    4
    5
    6
    7
    8
    9
    10
    11
    12
    13
    14
    15
    16
    17
    18
    19
    # 定义第二个Inception模块组,第三个Inception模块
    with tf.variable_scope("Mixed_6c"):
        with tf.variable_scope("Branch_0"):
            batch_0 = slim.conv2d(net,num_outputs=192,kernel_size=[1,1],scope="Conv2d_0a_1x1")
        with tf.variable_scope("Branch_1"):
            batch_1 = slim.conv2d(net, num_outputs=160, kernel_size=[1, 1], scope="Conv2d_0a_1x1")
            batch_1 = slim.conv2d(batch_1, num_outputs=160, kernel_size=[1,7], scope="Conv2d_0b_1x7")
            batch_1 = slim.conv2d(batch_1, num_outputs=160, kernel_size=[7, 1],scope="Conv2d_0c_7x1")
        with tf.variable_scope("Branch_2"):
            batch_2 = slim.conv2d(net, num_outputs=160, kernel_size=[1, 1], scope="Conv2d_0a_1x1")
            batch_2 = slim.conv2d(batch_2, num_outputs=160, kernel_size=[7, 1], scope="Conv2d_0b_7x1")
            batch_2 = slim.conv2d(batch_2, num_outputs=160, kernel_size=[1, 7], scope="Conv2d_0c_1x7")
            batch_2 = slim.conv2d(batch_2, num_outputs=160, kernel_size=[7, 1], scope="Conv2d_0d_7x1")
            batch_2 = slim.conv2d(batch_2, num_outputs=192, kernel_size=[1, 7], scope="Conv2d_0e_1x7")
        with tf.variable_scope("Branch_3"):
            batch_3 = slim.avg_pool2d(net, kernel_size=[3, 3], scope="AvgPool_0a_3x3")
            batch_3 = slim.conv2d(batch_3, num_outputs=192, kernel_size=[1, 1], scope="Conv2d_0b_1x1")

        net = tf.concat([batch_0, batch_1, batch_2,batch_3], 3)

  • 第四个Inception Module
    名称为Mixed_d，和Mixed_6c完全一致。

    1
    2
    3
    4
    5
    6
    7
    8
    9
    10
    11
    12
    13
    14
    15
    16
    17
    18
    19
    # 定义第二个Inception模块组,第四个Inception模块
    with tf.variable_scope("Mixed_6d"):
        with tf.variable_scope("Branch_0"):
            batch_0 = slim.conv2d(net,num_outputs=192,kernel_size=[1,1],scope="Conv2d_0a_1x1")
        with tf.variable_scope("Branch_1"):
            batch_1 = slim.conv2d(net, num_outputs=160, kernel_size=[1, 1], scope="Conv2d_0a_1x1")
            batch_1 = slim.conv2d(batch_1, num_outputs=160, kernel_size=[1,7], scope="Conv2d_0b_1x7")
            batch_1 = slim.conv2d(batch_1, num_outputs=160, kernel_size=[7, 1],scope="Conv2d_0c_7x1")
        with tf.variable_scope("Branch_2"):
            batch_2 = slim.conv2d(net, num_outputs=160, kernel_size=[1, 1], scope="Conv2d_0a_1x1")
            batch_2 = slim.conv2d(batch_2, num_outputs=160, kernel_size=[7, 1], scope="Conv2d_0b_7x1")
            batch_2 = slim.conv2d(batch_2, num_outputs=160, kernel_size=[1, 7], scope="Conv2d_0c_1x7")
            batch_2 = slim.conv2d(batch_2, num_outputs=160, kernel_size=[7, 1], scope="Conv2d_0d_7x1")
            batch_2 = slim.conv2d(batch_2, num_outputs=192, kernel_size=[1, 7], scope="Conv2d_0e_1x7")
        with tf.variable_scope("Branch_3"):
            batch_3 = slim.avg_pool2d(net, kernel_size=[3, 3], scope="AvgPool_0a_3x3")
            batch_3 = slim.conv2d(batch_3, num_outputs=192, kernel_size=[1, 1], scope="Conv2d_0b_1x1")

        net = tf.concat([batch_0, batch_1, batch_2,batch_3], 3)

  • 第五个Inception Module
    名称为Mixed_6e，和前两个Inception Module也完全一致。这是第二个Inception模块组的最后一个Inception Module。我们将Mixed_6e存储于end_points中，作为Auxiliary Classifier辅助模型的分类。

    1
    2
    3
    4
    5
    6
    7
    8
    9
    10
    11
    12
    13
    14
    15
    16
    17
    18
    19
    20
    21
    # 定义第二个Inception模块组,第五个Inception模块
    with tf.variable_scope("Mixed_6e"):
        with tf.variable_scope("Branch_0"):
            batch_0 = slim.conv2d(net,num_outputs=192,kernel_size=[1,1],scope="Conv2d_0a_1x1")
        with tf.variable_scope("Branch_1"):
            batch_1 = slim.conv2d(net, num_outputs=160, kernel_size=[1, 1], scope="Conv2d_0a_1x1")
            batch_1 = slim.conv2d(batch_1, num_outputs=160, kernel_size=[1,7], scope="Conv2d_0b_1x7")
            batch_1 = slim.conv2d(batch_1, num_outputs=160, kernel_size=[7, 1],scope="Conv2d_0c_7x1")
        with tf.variable_scope("Branch_2"):
            batch_2 = slim.conv2d(net, num_outputs=160, kernel_size=[1, 1], scope="Conv2d_0a_1x1")
            batch_2 = slim.conv2d(batch_2, num_outputs=160, kernel_size=[7, 1], scope="Conv2d_0b_7x1")
            batch_2 = slim.conv2d(batch_2, num_outputs=160, kernel_size=[1, 7], scope="Conv2d_0c_1x7")
            batch_2 = slim.conv2d(batch_2, num_outputs=160, kernel_size=[7, 1], scope="Conv2d_0d_7x1")
            batch_2 = slim.conv2d(batch_2, num_outputs=192, kernel_size=[1, 7], scope="Conv2d_0e_1x7")
        with tf.variable_scope("Branch_3"):
            batch_3 = slim.avg_pool2d(net, kernel_size=[3, 3], scope="AvgPool_0a_3x3")
            batch_3 = slim.conv2d(batch_3, num_outputs=192, kernel_size=[1, 1], scope="Conv2d_0b_1x1")

        net = tf.concat([batch_0, batch_1, batch_2,batch_3], 3)
    #第二个模块组的最后一个Inception模块，将Mixed_6e存储于end_points中
    end_points["Mixed_6e"] = net

定义第三个Inception模块组

第三个Inception模块组包含了3个Inception Module。其中后两个Inception Module的结构非常相似。

  • 第一个Inception Module
    名称为Mixed_7a，包含了三个分支：

    1. 第一个分支为192输出通道的1×1卷积，再接320输出通道的3×3卷积，步长为2
    2. 第二个分支有四个卷积层，分别为192输出通道的1×1卷积，192输出通道的1×7卷积，192输出通道的7×1卷积和192输出通道的3×3卷积。最后一个卷积层步长为2，padding为VALID
    3. 第三个分支为一个3×3最大池化层，步长为2，padding为VALID。
      最后合并得到的tensor尺寸为为3×3×(320+192+768)=8×8×1280。
      1
      2
      3
      4
      5
      6
      7
      8
      9
      10
      11
      12
      13
      14
      15
      16
      17
      # 定义第三个Inception模块组,第一个Inception模块
      with tf.variable_scope("Mixed_7a"):
          with tf.variable_scope("Branch_0"):
              batch_0 = slim.conv2d(net, num_outputs=192, kernel_size=[1, 1], scope="Conv2d_0a_1x1")
              batch_0 = slim.conv2d(net, num_outputs=320, kernel_size=[3, 3],stride=2,
                                      padding="VALID",scope="Conv2d_1a_3x3")
          with tf.variable_scope("Branch_1"):
              batch_1 = slim.conv2d(net, num_outputs=192, kernel_size=[1, 1], scope="Conv2d_0a_1x1")
              batch_1 = slim.conv2d(batch_1, num_outputs=192, kernel_size=[1,7], scope="Conv2d_0b_1x7")
              batch_1 = slim.conv2d(batch_1, num_outputs=192, kernel_size=[7, 1],scope="Conv2d_0c_7x1")
              batch_1 = slim.conv2d(batch_1, num_outputs=192, kernel_size=[3, 3], stride=2,
                                      padding="VALID",scope="Conv2d_1a_3x3")
          with tf.variable_scope("Branch_2"):
              batch_2 = slim.max_pool2d(net, kernel_size=[3, 3], stride=2, padding="VALID",
                                          scope="MaxPool_1a_3x3")

          net = tf.concat([batch_0, batch_1, batch_2], 3)

  • 第二个Inception Module
    名称为Mixed_7b，它有四个分支：

    1. 第一个分支为320输出通道的1×1卷积
    2. 第二个分支先是一个384输出通道的1×1卷积，随后在分支内开两个分支，分别为384输出通道的1×3卷积和384输出通道的3×1卷积，然后用tf.concat合并，得到的tensor的尺寸为8×8×(384+384)=8×8×768
    3. 第三个分支先是48输出通道的1×1卷积，然后是384输出通道的3×3卷积，然后同样在分支内拆成两个分支，分别为384输出通道的1×3卷积和384输出通道的3×1卷积
    4. 第四个分支为一个3×3的平均池化层后接一个192输出通道的1×1卷积。
      最后合并得到的tensor的尺寸为8×8×(320+768+768+192)=8×8×2048
      1
      2
      3
      4
      5
      6
      7
      8
      9
      10
      11
      12
      13
      14
      15
      16
      17
      18
      19
      20
      # 定义第三个Inception模块组,第二个Inception模块
      with tf.variable_scope("Mixed_7b"):
          with tf.variable_scope("Branch_0"):
              batch_0 = slim.conv2d(net, num_outputs=320, kernel_size=[1, 1],scope="Conv2d_0a_1x1")
          with tf.variable_scope("Branch_1"):
              batch_1 = slim.conv2d(net, num_outputs=384, kernel_size=[1, 1], scope="Conv2d_0a_1x1")
              batch_1 = tf.concat([
                  slim.conv2d(batch_1,num_outputs=384,kernel_size=[1,3],scope="Conv2d_0b_1x3"),
                  slim.conv2d(batch_1,num_outputs=384,kernel_size=[3,1],scope="Conv2d_0b_3x1")],axis=3)
          with tf.variable_scope("Branch_2"):
              batch_2 = slim.conv2d(net,num_outputs=448,kernel_size=[1,1],scope="Conv2d_0a_1x1")
              batch_2 = slim.conv2d(batch_2,num_outputs=384,kernel_size=[3,3],scope="Conv2d_0b_3x3")
              batch_2 = tf.concat([
                  slim.conv2d(batch_2,num_outputs=384,kernel_size=[1,3],scope="Conv2d_0c_1x3"),
                  slim.conv2d(batch_2,num_outputs=384,kernel_size=[3,1],scope="Conv2d_0d_3x1")],axis=3)
          with tf.variable_scope("Branch_3"):
              batch_3 = slim.avg_pool2d(net,kernel_size=[3,3],scope="AvgPool_0a_3x3")
              batch_3 = slim.conv2d(batch_3,num_outputs=192,kernel_size=[1,1],scope="Conv2d_0b_1x1")

      net = tf.concat([batch_0, batch_1, batch_2,batch_3], 3)

  • 第三个Inception Module*
    名称为Mixed_7c，它和前一个Inception Module完全一致。最后我们返回这个Inception Module的结果，作为inception_v3_base函数的最终输出

    1
    2
    3
    4
    5
    6
    7
    8
    9
    10
    11
    12
    13
    14
    15
    16
    17
    18
    19
    20
    21
    22
    23

        # 定义第三个Inception模块组,第三个Inception模块
        with tf.variable_scope("Mixed_7c"):
            with tf.variable_scope("Branch_0"):
                batch_0 = slim.conv2d(net, num_outputs=320, kernel_size=[1, 1],scope="Conv2d_0a_1x1")
            with tf.variable_scope("Branch_1"):
                batch_1 = slim.conv2d(net, num_outputs=384, kernel_size=[1, 1], scope="Conv2d_0a_1x1")
                batch_1 = tf.concat([
                    slim.conv2d(batch_1,num_outputs=384,kernel_size=[1,3],scope="Conv2d_0b_1x3"),
                    slim.conv2d(batch_1,num_outputs=384,kernel_size=[3,1],scope="Conv2d_0b_3x1")],axis=3)
            with tf.variable_scope("Branch_2"):
                batch_2 = slim.conv2d(net,num_outputs=448,kernel_size=[1,1],scope="Conv2d_0a_1x1")
                batch_2 = slim.conv2d(batch_2,num_outputs=384,kernel_size=[3,3],scope="Conv2d_0b_3x3")
                batch_2 = tf.concat([
                    slim.conv2d(batch_2,num_outputs=384,kernel_size=[1,3],scope="Conv2d_0c_1x3"),
                    slim.conv2d(batch_2,num_outputs=384,kernel_size=[3,1],scope="Conv2d_0d_3x1")],axis=3)
            with tf.variable_scope("Branch_3"):
                batch_3 = slim.avg_pool2d(net,kernel_size=[3,3],scope="AvgPool_0a_3x3")
                batch_3 = slim.conv2d(batch_3,num_outputs=192,kernel_size=[1,1],scope="Conv2d_0b_1x1")
     
        net = tf.concat([batch_0, batch_1, batch_2,batch_3], 3)
     
    return net,end_points

实现Inception_v3函数

实现Inception V3网络的最后一部分——全局平均池化，Softmax和Auxiliary Logits。

  • 函数Inception_v3的输入参数

    1
    2
    3
    4
    5
    6
    7
    8
    9
    10
    11
    12
    13
    14
    15
    16
    def inception_v3(inputs,num_classes=1000,is_training=True,droupot_keep_prob = 0.8,prediction_fn = slim.softmax,spatial_squeeze = True,reuse = None, scope="InceptionV3"): 
    """
    InceptionV3整个网络的构建
    param :
    inputs -- 输入tensor
    num_classes -- 最后分类数目
    is_training -- 是否是训练过程
    droupot_keep_prob -- dropout保留节点比例
    prediction_fn -- 最后分类函数，默认为softmax
    patial_squeeze -- 是否对输出去除维度为1的维度
    reuse -- 是否对网络和Variable重复使用
    scope -- 函数默认参数环境
    return:
    logits -- 最后输出结果
    end_points -- 包含辅助节点的重要节点字典表
    """

  • Auxiliary Logits部分的逻辑
    Auxiliary Logits作为辅助分类的节点，对分类结果预测有很大的帮助。通过end_points得到Mixed_6e后：

    • 连接一个5×5的平均池化，步长为3，padding设为VALID，这样输出的尺寸就从17×17×768变成5×5×768。
    • 接着连接一个128输出通道的1×1卷积和一个768输出通道的5×5卷积，这里权重初始化方式重设为标准差为0.01的正态分布，padding设置为VALID，输出尺寸变为1×1×768
    • 然后再接一个输出通道为num_classes的1×1卷积，不设激活函数和规范化函数权重初始方式重设为标准差为0.001的正态分布，这样输出就变成了1×1×1000
    • 最后使用tf.squeeze函数消除输出tensor中前两个为1的维度。将辅助分类节点的输出aux_logits储存到字典表end_points中。
      1
      2
      3
      4
      5
      6
      7
      8
      9
      10
      11
      12
      13
      14
      15
      16
      17
      aux_logits = end_points["Mixed_6e"]
      with tf.variable_scope("AuxLogits"):
          aux_logits = slim.avg_pool2d(aux_logits,kernel_size=[5,5],stride=3,
                                          padding="VALID",scope="Avgpool_1a_5x5")
          aux_logits = slim.conv2d(aux_logits,num_outputs=128,kernel_size=[1,1],scope="Conv2d_1b_1x1")
          aux_logits = slim.conv2d(aux_logits,num_outputs=768,kernel_size=[5,5],
                                      weights_initializer=trunc_normal(0.01),padding="VALID",
                                      scope="Conv2d_2a_5x5")
          aux_logits = slim.conv2d(aux_logits,num_outputs=num_classes,kernel_size=[1,1],
                                      activation_fn=None,normalizer_fn=None,
                                      weights_initializer=trunc_normal(0.001),scope="Conv2d_1b_1x1")
          #消除tensor中前两个维度为1的维度
          if spatial_squeeze:
              aux_logits = tf.squeeze(aux_logits,axis=[1,2],name="SpatialSqueeze")
          
          #将辅助节点分类的输出aux_logits存到end_points中
          end_points["AuxLogits"] = aux_logits

  • 分类预测部分逻辑
    得到Mixed_7e即最后一个卷积层的输出后：

    • 接一个8×8全局平均池化，padding设置为VALID，tensor的尺寸就变成了1×1×2048
    • 然后接一个Dropout层，节点保留率为dropout_keep_prob
    • 接着连接一个输出通道为1000的1×1卷积，激活函数和规范化函数设为空
    • 然后用tf.squeeze去除输出tensor中维数为1的维度
    • 最后连接一个Softmax对结果进行分类预测，输出的结果存储到end_points中
      1
      2
      3
      4
      5
      6
      7
      8
      9
      10
      11
      12
      13
      14
      #正常分类预测
      with tf.variable_scope("Logits"):
          net = slim.avg_pool2d(net,kernel_size=[8,8],padding="VALID",
                                  scope="Avgpool_1a_8x8")
          net = slim.dropout(net,keep_prob=droupot_keep_prob,scope="Dropout_1b")
          end_points["Logits"] = net

          logits = slim.conv2d(net,num_outputs=num_classes,kernel_size=[1,1],activation_fn=None,
                                  normalizer_fn=None,scope="Conv2d_1c_1x1")
          if spatial_squeeze:
              logits = tf.squeeze(logits,axis=[1,2],name="SpatialSqueeze")

      end_points["Logits"] = logits
      end_points["Predictions"] = prediction_fn(logits,scope="Predictions")

inception_v3函数实现代码汇总：

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
def inception_v3(inputs,num_classes=1000,is_training=True,droupot_keep_prob = 0.8,
                 prediction_fn = slim.softmax,spatial_squeeze = True,reuse = None,scope="InceptionV3"):
    """
    InceptionV3整个网络的构建
    param :
    inputs -- 输入tensor
    num_classes -- 最后分类数目
    is_training -- 是否是训练过程
    droupot_keep_prob -- dropout保留节点比例
    prediction_fn -- 最后分类函数，默认为softmax
    patial_squeeze -- 是否对输出去除维度为1的维度
    reuse -- 是否对网络和Variable重复使用
    scope -- 函数默认参数环境
    return:
    logits -- 最后输出结果
    end_points -- 包含辅助节点的重要节点字典表
    """
    with tf.variable_scope(scope,"InceptionV3",[inputs,num_classes],
                           reuse=reuse) as scope:
        with slim.arg_scope([slim.batch_norm,slim.dropout],
                            is_training = is_training):
            net,end_points = inception_v3_base(inputs,scope=scope)     #前面定义的整个卷积网络部分
 
            #辅助分类节点部分
            with slim.arg_scope([slim.conv2d,slim.max_pool2d,slim.avg_pool2d],
                                stride = 1,padding = "SAME"):
                #通过end_points取到Mixed_6e
                aux_logits = end_points["Mixed_6e"]
                with tf.variable_scope("AuxLogits"):
                    aux_logits = slim.avg_pool2d(aux_logits,kernel_size=[5,5],stride=3,
                                                 padding="VALID",scope="Avgpool_1a_5x5")
                    aux_logits = slim.conv2d(aux_logits,num_outputs=128,kernel_size=[1,1],scope="Conv2d_1b_1x1")
                    aux_logits = slim.conv2d(aux_logits,num_outputs=768,kernel_size=[5,5],
                                             weights_initializer=trunc_normal(0.01),padding="VALID",
                                             scope="Conv2d_2a_5x5")
                    aux_logits = slim.conv2d(aux_logits,num_outputs=num_classes,kernel_size=[1,1],
                                             activation_fn=None,normalizer_fn=None,
                                             weights_initializer=trunc_normal(0.001),scope="Conv2d_1b_1x1")
                    #消除tensor中前两个维度为1的维度
                    if spatial_squeeze:
                        aux_logits = tf.squeeze(aux_logits,axis=[1,2],name="SpatialSqueeze")
 
                    end_points["AuxLogits"] = aux_logits    #将辅助节点分类的输出aux_logits存到end_points中
 
                #正常分类预测
                with tf.variable_scope("Logits"):
                    net = slim.avg_pool2d(net,kernel_size=[8,8],padding="VALID",
                                          scope="Avgpool_1a_8x8")
                    net = slim.dropout(net,keep_prob=droupot_keep_prob,scope="Dropout_1b")
                    end_points["Logits"] = net
 
                    logits = slim.conv2d(net,num_outputs=num_classes,kernel_size=[1,1],activation_fn=None,
                                         normalizer_fn=None,scope="Conv2d_1c_1x1")
                    if spatial_squeeze:
                        logits = tf.squeeze(logits,axis=[1,2],name="SpatialSqueeze")
                
                end_points["Logits"] = logits
                end_points["Predictions"] = prediction_fn(logits,scope="Predictions")
        
        return logits,end_points

0%