深度神经网络Google Inception Net-V3结构图
前言
Google Inception Net在2014年的 ImageNet Large Scale Visual Recognition Competition (ILSVRC)中取得第一名,该网络以结构上的创新取胜,通过采用全局平均池化层取代全连接层,极大的降低了参数量,是非常实用的模型,一般称该网络模型为Inception V1。随后的Inception V2中,引入了Batch Normalization方法,加快了训练的收敛速度。在Inception V3模型中,通过将二维卷积层拆分成两个一维卷积层,不仅降低了参数数量,同时减轻了过拟合现象。
一、多少层?
Inception V3究竟有多少层呢?某书籍上说42层,某书籍上说46层。参考实现的源代码,仔细数一数,应该是47层。
![](https://upload-images.jianshu.io/upload_images/5211062-ddf211409ead150e.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/650/format/webp)
层次结构图.png
5(前面)+
3(block1_module1)+3(block1_module2)+3(block1_module3)+
3(block2_module1)+5(block2_module2)+5(block2_module3)+5(block2_module4)+5(block2_module5)+
4(block3_module1)+3(block3_module2)+3(block3_module3)
= 47层
Tips:上面的这张层次结构图出现在某些帖子和书籍中,根据实现的源码,标注的红色方框处应该是5个卷积层,而不是4个。
二、详细网络结构
详细的网络结构及其子网络结构如下。
![](https://upload-images.jianshu.io/upload_images/5211062-2e24e5da93968bb5.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1000/format/webp)
总体结构图 .png
![](https://upload-images.jianshu.io/upload_images/5211062-d91e49583550b0c5.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1000/format/webp)
block1_module1.png
![](https://upload-images.jianshu.io/upload_images/5211062-2230c40a32880b46.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1000/format/webp)
block1_module2.png
![](https://upload-images.jianshu.io/upload_images/5211062-8046e9a1baa4f72a.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1000/format/webp)
block1_module3.png
![](https://upload-images.jianshu.io/upload_images/5211062-a1127830401779d3.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1000/format/webp)
block2_module1.png
![](https://upload-images.jianshu.io/upload_images/5211062-bf6319a670ed4cbe.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1000/format/webp)
block2_module2.png
![](https://upload-images.jianshu.io/upload_images/5211062-c7b12cdfd5487a6c.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1000/format/webp)
block2_module3,4.png
![](https://upload-images.jianshu.io/upload_images/5211062-aaab45b4954d5221.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1000/format/webp)
block2_module5.png
![](https://upload-images.jianshu.io/upload_images/5211062-06627506ba6b09b0.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1000/format/webp)
block3_module1.png
![](https://upload-images.jianshu.io/upload_images/5211062-9a9c9ce21bbf7d8a.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1000/format/webp)
block3_module2.png
![](https://upload-images.jianshu.io/upload_images/5211062-7f8ec0d4f6a4bdd9.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1000/format/webp)
block3_module3.png
这么复杂精巧的网络结构是怎么设计出来的呢?是不断的进行数值实验吗?
还是靠爱... ...
![](https://upload-images.jianshu.io/upload_images/5211062-a30f66cd9915795c.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/495/format/webp)