Deep Learning 33：读论文“Densely Connected Convolutional Networks”-------DenseNet 简单理解

一.读前说明

1.论文"Densely Connected Convolutional Networks"是现在为止效果最好的CNN架构,比Resnet还好,有必要学习一下它为什么效果这么好.

2.代码地址：https://github.com/liuzhuang13/DenseNet

3.这篇论文主要参考了Highway Networks，Residual Networks (ResNets)和GoogLeNet,所以在读本篇论文之前,有必要读一下这几篇论文,另外还可以看一下Very Deep Learning with Highway Networks

4.参考文献 :ResNet && DenseNet(原理篇), DenseNet模型

二.阅读笔记 

Abstract

最近的一些论文表明,如果卷积神经网络的各层到输入层和输出层的连接更短,那么该网络就大体上可以设计得更深、更准确、训练得更有效。本文基于此提出了“稠密卷积网络（DensNet），该网络每一层均以前馈的形式与其他任一层连接。因此，传统卷积网络有L层就只有L个连接，而DenseNet的任一层不仅与相邻层有连接，而且与它的随后的所有层都有直接连接，所以该网络有L(L+1)/2个直接连接。任意一层的输入都是其前面所有层的特征图，而该层自己的特征图是其随后所有层的输入。DenseNet有以下几个令人激动的优点：1.减轻了梯度消失问题；2.强化了特征传播；3.大幅度减少了参数数量。该网络结构在4个高竞争性的目标识别基准数据集上进行了评估，包括：CIFAR-10,CIFAR-100,SVHN,ImageNet。DenseNet在这些数据集上大部分都获得了巨大的提高，达到目前为止最高的识别准确率。

1.Introduction

在视觉识别中，CNN是一种强大的机器学习方法。尽管CNN在20年以前就被提出来，但是只是在最近几年，计算机硬件和网络结构的提高才使得真正的深层CNN的训练变成可能。最开始的LeNet5包含5层，VGG包含19层，只有去年的Highway Networks和ResNets才超过了100层这个关卡。

三.阅读感想：

翻译了一半，居然感觉完全不用翻译，真接看英文原文也能看懂，嗯对，这篇文章写得通俗易懂，根本不用像看那些什么hiton、begio、yanlecun之类大牛写的文章一样，直接一遍看过去，看得似懂非懂的。看这篇论文看完之后，感觉像吃了蜂蜜一样，看了还想看，连连最后实验结果分析和discuss也写得非常好，特别是discuss中那个图，该文创意非常棒，并且简单，最主要的是该文创意来源就是我最喜欢的那种，就是总结以前很多文章中效果好的原因，找出它们的共性，然后强化这个共性，从而得到更好的结果。

四.DenseNet结构：

.在CIFAR-10上用训练时的结构DenseNet-BC：

如果depth=40, growth_rate=12, bottleneck=True, reduction=0.5=1-compression，则每个denseblock里面的层数n_layers=((40-4)/3)//2=6.其中//2表示除以2后向下取整。
注：conv表示正常的2D卷积，CONV表示BN-ReLU-conv
结构如下：
input：(32,32,3)
conv(24,3,3),                                                         % 其中conv(24,3,3)=conv(filters=2*growth_rate=24,kernel_size=3,3)

#第1个dense block
CONV(48,1,1)-CONV(12,3,3)-merge(36)-           %  其中CONV(48,1,1)=CONV(filters=inter_channel = nb_filter*4=48,1,1),merge后nb_filter=24+12=36
CONV(48,1,1)-CONV(12,3,3)-merge(48)-           % 同上，merge后nb_filter=36+12=48
CONV(48,1,1)-CONV(12,3,3)-merge(60)-
CONV(48,1,1)-CONV(12,3,3)-merge(72)-
CONV(48,1,1)-CONV(12,3,3)-merge(84)-
CONV(48,1,1)-CONV(12,3,3)-merge(96)-          % 此时nb_filter每多一层就增加growth_rate=12个，这里1个dense block里有6层，故增加72个，所以nb_falter=24+72=96

#第1个Transition Layer
CONV(48,1,1)                                                      % nb_filter=nb_filter*compression=96*0.5=48
AveragePool(2,2,(2,2))                                         % pool_size=2,2  strides=(2,2)

#第2个dense block
CONV(48,1,1)-CONV(12,3,3)-merge(108)-        %  其中CONV(48,1,1)=CONV(filters=inter_channel = nb_filter*4=48,1,1),merge后nb_filter=96+12=108
CONV(48,1,1)-CONV(12,3,3)-merge(120)-
CONV(48,1,1)-CONV(12,3,3)-merge(132)-
CONV(48,1,1)-CONV(12,3,3)-merge(144)-
CONV(48,1,1)-CONV(12,3,3)-merge(156)-
CONV(48,1,1)-CONV(12,3,3)-merge(168)-       % 此时nb_filter每多一层就增加growth_rate=12个，这里1个dense block里有6层，故增加72个，所以nb_falter=96+72=168

#第2个Transition Layer
CONV(60,1,1)                                                      % nb_filter=nb_filter*compression=120*0.5=60
AveragePool(2,2,(2,2))                                         % pool_size=2,2  strides=(2,2)

#第3个dense block
CONV(48,1,1)-CONV(12,3,3)-merge(180)-       %  其中CONV(48,1,1)=CONV(filters=inter_channel = nb_filter*4=48,1,1)
CONV(48,1,1)-CONV(12,3,3)-merge(192)-
CONV(48,1,1)-CONV(12,3,3)-merge(204)-
CONV(48,1,1)-CONV(12,3,3)-merge(216)-
CONV(48,1,1)-CONV(12,3,3)-merge(228)-
CONV(48,1,1)-CONV(12,3,3)-merge(240)-      % 此时nb_filter每多一层就增加growth_rate=12个，这里1个dense block里有6层，故增加72个，所以nb_falter=168+72=240

Relu-GlobalAveragePool-softmax

为验证以上的分析，用keras==1.2.0版本验证结果如下：

 Model created
 ____________________________________________________________________________________________________
 Layer (type) Output Shape Param # Connected to
 ====================================================================================================
 input_1 (InputLayer) (None, , , )
 ____________________________________________________________________________________________________
 initial_conv2D (Convolution2D) (None, , , )  input_1[][]
 ____________________________________________________________________________________________________
 batchnormalization_1 (BatchNorma (None, , , )  initial_conv2D[][]
 ____________________________________________________________________________________________________
 activation_1 (Activation) (None, , , )  batchnormalization_1[][]
 ____________________________________________________________________________________________________
 convolution2d_1 (Convolution2D) (None, , , )  activation_1[][]
 ____________________________________________________________________________________________________
 batchnormalization_2 (BatchNorma (None, , , )  convolution2d_1[][]
 ____________________________________________________________________________________________________
 activation_2 (Activation) (None, , , )  batchnormalization_2[][]
 ____________________________________________________________________________________________________
 convolution2d_2 (Convolution2D) (None, , , )  activation_2[][]
 ____________________________________________________________________________________________________
 merge_1 (Merge) (None, , , )  initial_conv2D[][]
 convolution2d_2[][]
 ____________________________________________________________________________________________________
 batchnormalization_3 (BatchNorma (None, , , )  merge_1[][]
 ____________________________________________________________________________________________________
 activation_3 (Activation) (None, , , )  batchnormalization_3[][]
 ____________________________________________________________________________________________________
 convolution2d_3 (Convolution2D) (None, , , )  activation_3[][]
 ____________________________________________________________________________________________________
 batchnormalization_4 (BatchNorma (None, , , )  convolution2d_3[][]
 ____________________________________________________________________________________________________
 activation_4 (Activation) (None, , , )  batchnormalization_4[][]
 ____________________________________________________________________________________________________
 convolution2d_4 (Convolution2D) (None, , , )  activation_4[][]
 ____________________________________________________________________________________________________
 merge_2 (Merge) (None, , , )  initial_conv2D[][]
 convolution2d_2[][]
 convolution2d_4[][]
 ____________________________________________________________________________________________________
 batchnormalization_5 (BatchNorma (None, , , )  merge_2[][]
 ____________________________________________________________________________________________________
 activation_5 (Activation) (None, , , )  batchnormalization_5[][]
 ____________________________________________________________________________________________________
 convolution2d_5 (Convolution2D) (None, , , )  activation_5[][]
 ____________________________________________________________________________________________________
 batchnormalization_6 (BatchNorma (None, , , )  convolution2d_5[][]
 ____________________________________________________________________________________________________
 activation_6 (Activation) (None, , , )  batchnormalization_6[][]
 ____________________________________________________________________________________________________
 convolution2d_6 (Convolution2D) (None, , , )  activation_6[][]
 ____________________________________________________________________________________________________
 merge_3 (Merge) (None, , , )  initial_conv2D[][]
 convolution2d_2[][]
 convolution2d_4[][]
 convolution2d_6[][]
 ____________________________________________________________________________________________________
 batchnormalization_7 (BatchNorma (None, , , )  merge_3[][]
 ____________________________________________________________________________________________________
 activation_7 (Activation) (None, , , )  batchnormalization_7[][]
 ____________________________________________________________________________________________________
 convolution2d_7 (Convolution2D) (None, , , )  activation_7[][]
 ____________________________________________________________________________________________________
 batchnormalization_8 (BatchNorma (None, , , )  convolution2d_7[][]
 ____________________________________________________________________________________________________
 activation_8 (Activation) (None, , , )  batchnormalization_8[][]
 ____________________________________________________________________________________________________
 convolution2d_8 (Convolution2D) (None, , , )  activation_8[][]
 ____________________________________________________________________________________________________
 merge_4 (Merge) (None, , , )  initial_conv2D[][]
 convolution2d_2[][]
 convolution2d_4[][]
 convolution2d_6[][]
 convolution2d_8[][]
 ____________________________________________________________________________________________________
 batchnormalization_9 (BatchNorma (None, , , )  merge_4[][]
 ____________________________________________________________________________________________________
 activation_9 (Activation) (None, , , )  batchnormalization_9[][]
 ____________________________________________________________________________________________________
 convolution2d_9 (Convolution2D) (None, , , )  activation_9[][]
 ____________________________________________________________________________________________________
 batchnormalization_10 (BatchNorm (None, , , )  convolution2d_9[][]
 ____________________________________________________________________________________________________
 activation_10 (Activation) (None, , , )  batchnormalization_10[][]
 ____________________________________________________________________________________________________
 convolution2d_10 (Convolution2D) (None, , , )  activation_10[][]
 ____________________________________________________________________________________________________
 merge_5 (Merge) (None, , , )  initial_conv2D[][]
 convolution2d_2[][]
 convolution2d_4[][]
 convolution2d_6[][]
 convolution2d_8[][]
 convolution2d_10[][]
 ____________________________________________________________________________________________________
 batchnormalization_11 (BatchNorm (None, , , )  merge_5[][]
 ____________________________________________________________________________________________________
 activation_11 (Activation) (None, , , )  batchnormalization_11[][]
 ____________________________________________________________________________________________________
 convolution2d_11 (Convolution2D) (None, , , )  activation_11[][]
 ____________________________________________________________________________________________________
 batchnormalization_12 (BatchNorm (None, , , )  convolution2d_11[][]
 ____________________________________________________________________________________________________
 activation_12 (Activation) (None, , , )  batchnormalization_12[][]
 ____________________________________________________________________________________________________
 convolution2d_12 (Convolution2D) (None, , , )  activation_12[][]
 ____________________________________________________________________________________________________
 merge_6 (Merge) (None, , , )  initial_conv2D[][]
 convolution2d_2[][]
 convolution2d_4[][]
 convolution2d_6[][]
 convolution2d_8[][]
 convolution2d_10[][]
 convolution2d_12[][]
 ____________________________________________________________________________________________________
 batchnormalization_13 (BatchNorm (None, , , )  merge_6[][]
 ____________________________________________________________________________________________________
 activation_13 (Activation) (None, , , )  batchnormalization_13[][]
 ____________________________________________________________________________________________________
 convolution2d_13 (Convolution2D) (None, , , )  activation_13[][]
 ____________________________________________________________________________________________________
 averagepooling2d_1 (AveragePooli (None, , , )  convolution2d_13[][]
 ____________________________________________________________________________________________________
 batchnormalization_14 (BatchNorm (None, , , )  averagepooling2d_1[][]
 ____________________________________________________________________________________________________
 activation_14 (Activation) (None, , , )  batchnormalization_14[][]
 ____________________________________________________________________________________________________
 convolution2d_14 (Convolution2D) (None, , , )  activation_14[][]
 ____________________________________________________________________________________________________
 batchnormalization_15 (BatchNorm (None, , , )  convolution2d_14[][]
 ____________________________________________________________________________________________________
 activation_15 (Activation) (None, , , )  batchnormalization_15[][]
 ____________________________________________________________________________________________________
 convolution2d_15 (Convolution2D) (None, , , )  activation_15[][]
 ____________________________________________________________________________________________________
 merge_7 (Merge) (None, , , )  averagepooling2d_1[][]
 convolution2d_15[][]
 ____________________________________________________________________________________________________
 batchnormalization_16 (BatchNorm (None, , , )  merge_7[][]
 ____________________________________________________________________________________________________
 activation_16 (Activation) (None, , , )  batchnormalization_16[][]
 ____________________________________________________________________________________________________
 convolution2d_16 (Convolution2D) (None, , , )  activation_16[][]
 ____________________________________________________________________________________________________
 batchnormalization_17 (BatchNorm (None, , , )  convolution2d_16[][]
 ____________________________________________________________________________________________________
 activation_17 (Activation) (None, , , )  batchnormalization_17[][]
 ____________________________________________________________________________________________________
 convolution2d_17 (Convolution2D) (None, , , )  activation_17[][]
 ____________________________________________________________________________________________________
 merge_8 (Merge) (None, , , )  averagepooling2d_1[][]
 convolution2d_15[][]
 convolution2d_17[][]
 ____________________________________________________________________________________________________
 batchnormalization_18 (BatchNorm (None, , , )  merge_8[][]
 ____________________________________________________________________________________________________
 activation_18 (Activation) (None, , , )  batchnormalization_18[][]
 ____________________________________________________________________________________________________
 convolution2d_18 (Convolution2D) (None, , , )  activation_18[][]
 ____________________________________________________________________________________________________
 batchnormalization_19 (BatchNorm (None, , , )  convolution2d_18[][]
 ____________________________________________________________________________________________________
 activation_19 (Activation) (None, , , )  batchnormalization_19[][]
 ____________________________________________________________________________________________________
 convolution2d_19 (Convolution2D) (None, , , )  activation_19[][]
 ____________________________________________________________________________________________________
 merge_9 (Merge) (None, , , )  averagepooling2d_1[][]
 convolution2d_15[][]
 convolution2d_17[][]
 convolution2d_19[][]
 ____________________________________________________________________________________________________
 batchnormalization_20 (BatchNorm (None, , , )  merge_9[][]
 ____________________________________________________________________________________________________
 activation_20 (Activation) (None, , , )  batchnormalization_20[][]
 ____________________________________________________________________________________________________
 convolution2d_20 (Convolution2D) (None, , , )  activation_20[][]
 ____________________________________________________________________________________________________
 batchnormalization_21 (BatchNorm (None, , , )  convolution2d_20[][]
 ____________________________________________________________________________________________________
 activation_21 (Activation) (None, , , )  batchnormalization_21[][]
 ____________________________________________________________________________________________________
 convolution2d_21 (Convolution2D) (None, , , )  activation_21[][]
 ____________________________________________________________________________________________________
 merge_10 (Merge) (None, , , )  averagepooling2d_1[][]
 convolution2d_15[][]
 convolution2d_17[][]
 convolution2d_19[][]
 convolution2d_21[][]
 ____________________________________________________________________________________________________
 batchnormalization_22 (BatchNorm (None, , , )  merge_10[][]
 ____________________________________________________________________________________________________
 activation_22 (Activation) (None, , , )  batchnormalization_22[][]
 ____________________________________________________________________________________________________
 convolution2d_22 (Convolution2D) (None, , , )  activation_22[][]
 ____________________________________________________________________________________________________
 batchnormalization_23 (BatchNorm (None, , , )  convolution2d_22[][]
 ____________________________________________________________________________________________________
 activation_23 (Activation) (None, , , )  batchnormalization_23[][]
 ____________________________________________________________________________________________________
 convolution2d_23 (Convolution2D) (None, , , )  activation_23[][]
 ____________________________________________________________________________________________________
 merge_11 (Merge) (None, , , )  averagepooling2d_1[][]
 convolution2d_15[][]
 convolution2d_17[][]
 convolution2d_19[][]
 convolution2d_21[][]
 convolution2d_23[][]
 ____________________________________________________________________________________________________
 batchnormalization_24 (BatchNorm (None, , , )  merge_11[][]
 ____________________________________________________________________________________________________
 activation_24 (Activation) (None, , , )  batchnormalization_24[][]
 ____________________________________________________________________________________________________
 convolution2d_24 (Convolution2D) (None, , , )  activation_24[][]
 ____________________________________________________________________________________________________
 batchnormalization_25 (BatchNorm (None, , , )  convolution2d_24[][]
 ____________________________________________________________________________________________________
 activation_25 (Activation) (None, , , )  batchnormalization_25[][]
 ____________________________________________________________________________________________________
 convolution2d_25 (Convolution2D) (None, , , )  activation_25[][]
 ____________________________________________________________________________________________________
 merge_12 (Merge) (None, , , )  averagepooling2d_1[][]
 convolution2d_15[][]
 convolution2d_17[][]
 convolution2d_19[][]
 convolution2d_21[][]
 convolution2d_23[][]
 convolution2d_25[][]
 ____________________________________________________________________________________________________
 batchnormalization_26 (BatchNorm (None, , , )  merge_12[][]
 ____________________________________________________________________________________________________
 activation_26 (Activation) (None, , , )  batchnormalization_26[][]
 ____________________________________________________________________________________________________
 convolution2d_26 (Convolution2D) (None, , , )  activation_26[][]
 ____________________________________________________________________________________________________
 averagepooling2d_2 (AveragePooli (None, , , )  convolution2d_26[][]
 ____________________________________________________________________________________________________
 batchnormalization_27 (BatchNorm (None, , , )  averagepooling2d_2[][]
 ____________________________________________________________________________________________________
 activation_27 (Activation) (None, , , )  batchnormalization_27[][]
 ____________________________________________________________________________________________________
 convolution2d_27 (Convolution2D) (None, , , )  activation_27[][]
 ____________________________________________________________________________________________________
 batchnormalization_28 (BatchNorm (None, , , )  convolution2d_27[][]
 ____________________________________________________________________________________________________
 activation_28 (Activation) (None, , , )  batchnormalization_28[][]
 ____________________________________________________________________________________________________
 convolution2d_28 (Convolution2D) (None, , , )  activation_28[][]
 ____________________________________________________________________________________________________
 merge_13 (Merge) (None, , , )  averagepooling2d_2[][]
 convolution2d_28[][]
 ____________________________________________________________________________________________________
 batchnormalization_29 (BatchNorm (None, , , )  merge_13[][]
 ____________________________________________________________________________________________________
 activation_29 (Activation) (None, , , )  batchnormalization_29[][]
 ____________________________________________________________________________________________________
 convolution2d_29 (Convolution2D) (None, , , )  activation_29[][]
 ____________________________________________________________________________________________________
 batchnormalization_30 (BatchNorm (None, , , )  convolution2d_29[][]
 ____________________________________________________________________________________________________
 activation_30 (Activation) (None, , , )  batchnormalization_30[][]
 ____________________________________________________________________________________________________
 convolution2d_30 (Convolution2D) (None, , , )  activation_30[][]
 ____________________________________________________________________________________________________
 merge_14 (Merge) (None, , , )  averagepooling2d_2[][]
 convolution2d_28[][]
 convolution2d_30[][]
 ____________________________________________________________________________________________________
 batchnormalization_31 (BatchNorm (None, , , )  merge_14[][]
 ____________________________________________________________________________________________________
 activation_31 (Activation) (None, , , )  batchnormalization_31[][]
 ____________________________________________________________________________________________________
 convolution2d_31 (Convolution2D) (None, , , )  activation_31[][]
 ____________________________________________________________________________________________________
 batchnormalization_32 (BatchNorm (None, , , )  convolution2d_31[][]
 ____________________________________________________________________________________________________
 activation_32 (Activation) (None, , , )  batchnormalization_32[][]
 ____________________________________________________________________________________________________
 convolution2d_32 (Convolution2D) (None, , , )  activation_32[][]
 ____________________________________________________________________________________________________
 merge_15 (Merge) (None, , , )  averagepooling2d_2[][]
 convolution2d_28[][]
 convolution2d_30[][]
 convolution2d_32[][]
 ____________________________________________________________________________________________________
 batchnormalization_33 (BatchNorm (None, , , )  merge_15[][]
 ____________________________________________________________________________________________________
 activation_33 (Activation) (None, , , )  batchnormalization_33[][]
 ____________________________________________________________________________________________________
 convolution2d_33 (Convolution2D) (None, , , )  activation_33[][]
 ____________________________________________________________________________________________________
 batchnormalization_34 (BatchNorm (None, , , )  convolution2d_33[][]
 ____________________________________________________________________________________________________
 activation_34 (Activation) (None, , , )  batchnormalization_34[][]
 ____________________________________________________________________________________________________
 convolution2d_34 (Convolution2D) (None, , , )  activation_34[][]
 ____________________________________________________________________________________________________
 merge_16 (Merge) (None, , , )  averagepooling2d_2[][]
 convolution2d_28[][]
 convolution2d_30[][]
 convolution2d_32[][]
 convolution2d_34[][]
 ____________________________________________________________________________________________________
 batchnormalization_35 (BatchNorm (None, , , )  merge_16[][]
 ____________________________________________________________________________________________________
 activation_35 (Activation) (None, , , )  batchnormalization_35[][]
 ____________________________________________________________________________________________________
 convolution2d_35 (Convolution2D) (None, , , )  activation_35[][]
 ____________________________________________________________________________________________________
 batchnormalization_36 (BatchNorm (None, , , )  convolution2d_35[][]
 ____________________________________________________________________________________________________
 activation_36 (Activation) (None, , , )  batchnormalization_36[][]
 ____________________________________________________________________________________________________
 convolution2d_36 (Convolution2D) (None, , , )  activation_36[][]
 ____________________________________________________________________________________________________
 merge_17 (Merge) (None, , , )  averagepooling2d_2[][]
 convolution2d_28[][]
 convolution2d_30[][]
 convolution2d_32[][]
 convolution2d_34[][]
 convolution2d_36[][]
 ____________________________________________________________________________________________________
 batchnormalization_37 (BatchNorm (None, , , )  merge_17[][]
 ____________________________________________________________________________________________________
 activation_37 (Activation) (None, , , )  batchnormalization_37[][]
 ____________________________________________________________________________________________________
 convolution2d_37 (Convolution2D) (None, , , )  activation_37[][]
 ____________________________________________________________________________________________________
 batchnormalization_38 (BatchNorm (None, , , )  convolution2d_37[][]
 ____________________________________________________________________________________________________
 activation_38 (Activation) (None, , , )  batchnormalization_38[][]
 ____________________________________________________________________________________________________
 convolution2d_38 (Convolution2D) (None, , , )  activation_38[][]
 ____________________________________________________________________________________________________
 merge_18 (Merge) (None, , , )  averagepooling2d_2[][]
 convolution2d_28[][]
 convolution2d_30[][]
 convolution2d_32[][]
 convolution2d_34[][]
 convolution2d_36[][]
 convolution2d_38[][]
 ____________________________________________________________________________________________________
 batchnormalization_39 (BatchNorm (None, , , )  merge_18[][]
 ____________________________________________________________________________________________________
 activation_39 (Activation) (None, , , )  batchnormalization_39[][]
 ____________________________________________________________________________________________________
 globalaveragepooling2d_1 (Global (None, )  activation_39[][]
 ____________________________________________________________________________________________________
 dense_1 (Dense) (None, )  globalaveragepooling2d_1[][]
 ====================================================================================================
 Total params: ,
 Trainable params: ,
 Non-trainable params: ,
 ____________________________________________________________________________________________________
 Finished compiling
 Building model...

五.疑问：

1.运行完keras实验之后发现，居然在每个CONV(48,1,1)-CONV(12,3,3)- 后面都有一个Merge，可是在代码中我并没有发现呀，哪里来的？肯定是我看漏了，可是它是从哪来的呢？

答：原来在dense_block的定义中有这样一句话看掉了：

     for i in range(nb_layers):
         x = conv_block(x, growth_rate, bottleneck, dropout_rate, weight_decay)
         feature_list.append(x)
         x = merge(feature_list, mode='concat', concat_axis=concat_axis)
         nb_filter += growth_rate

意思就是在每个这样一个模块后，都要进行Merge，即：就是把每一层的输出都串联在一起，从而组成一个新的tensor。

2.为什么每个denseblock里面的层数n_layers=((40-4)/3)//2=6.其中//2表示除以2后向下取整？即为什么是减4？

答：因为该结构中层，除了dense block 中有很多层外，还1个初始的卷积层、2个过渡层、以及1个最后分类输出层。注意：在该论文中，讲的结构深度depth为L，它并不包括输入层在内。

所以对本论文中的深度depth或L的定义如下：

a.初始的卷积conv，算作1层；

b.每个过渡层，算作1层；

c.每个dense block中的CONV(48,1,1)-CONV(12,3,3)模块，算作2层，即：1个CONV就算作1层；

d.最后的输出模块Relu-GlobalAveragePool-softmax，算作1层。

也可这么说：深度就是卷积层的层数加上1个softmax层。