CNN发展:Lenet、Alexnet、Network in Network(NiN)、VGG(16、19)、GoogleNet(Inception v1、v2、v3、v4)、ResNet(v1、v2;18、34、50、101和152)
(此外还有DenseNet、SqueezeNet、MobileNet、NASNet、SENet等;)
(tensorflow/tensorflow/contrib/slim/python/slim/nets有多种现成的网络;)
(tensorflow/models/tree/master/research/slim/nets有更多现成的网络;)
ImageNet论文:
ImageNet: A Large-Scale Hierarchical Image Database
网络演化:
https://www.cnblogs.com/52machinelearning/p/5821591.html
https://www.cnblogs.com/skyfsm/p/8451834.html
https://blog.youkuaiyun.com/dqcfkyqdxym3f8rb0/article/details/79314648
LeNet是用于手写体字符识别的非常高效的卷积神经网络;
(卷积层-池化层-卷积层-池化层-全连接层-全连接层-全连接层;)
LeNet相关论文:
Gradient-Based Learning Applied to Document Recognition
https://cuijiahua.com/blog/2018/01/dl_3.html
https://www.cnblogs.com/duanhx/articles/9655228.html
https://blog.youkuaiyun.com/lilong117194/article/details/79370831
AlexNet首次在CNN中成功应用了ReLU、Dropout、重叠最大池化(Overlapping maxPooling)、LRN(Local Response Normalization)、Data Augmentation(如PCA)、以及利用两台GPU(分组卷积)训练等Trick;
AlexNet将LeNet的思想发扬光大,把CNN的基本原理应用到了很深很宽的网络中;(5个卷积+3个全连接;输入图像大小224*224;)
AlexNet相关论文:
Imagenet classification with deep convolutional neural networks
https://www.cnblogs.com/alexanderkun/p/6917985.html
https://baike.baidu.com/item/AlexNet/22689612?fr=aladdin
https://www.cnblogs.com/alexanderkun/p/6918045.html
https://blog.youkuaiyun.com/luoluonuoyasuolong/article/details/81750190
https://www.jianshu.com/p/182073af7d95
https://blog.youkuaiyun.com/qq_24695385/article/details/80368618
VGG获得了2014年ILSVRC分类第二,定位第一;(当年分类第一是GoogLeNet)
VGG使用small filters(3*3conv),原因:
1. 3个3*3conv和1个7*7conv有效感受野一样,但是层数更深,非线性也越强
2. 参数更少,3*(3*3)< 7*7
VGG也使用了1*1卷积;
(1*1卷积本质是不同feature channel的线性叠加,不用考虑像素间的关系;可以实现通道数量的升维和降维;是性价比很高的聚合操作;最早出现在NiN论文)
VGG有19层的 Net-E 和16层的 Net-D ;前者比后者好
(实验表明LRN没有任何用处)
VGG相关论文:
Very deep convolutional networks for large-scale image recognition(vgg16、19)
https://www.sohu.com/a/241338315_787107
https://www.jianshu.com/p/68ac04943f9e !!!
https://www.cnblogs.com/bonelee/p/8921364.html
https://www.jianshu.com/p/e96b7a9b4229
GoogleNet没有全连接层,参数比Alexnet还少12倍,只有5百万个;总共22层,最大特色是Inception;Inception是全卷积的;
GoogleNet的Inception是基于Network in Network(NiN)改进的!
GoogleNet(v1)有多个loss(或softmax、输出);为啥这么设计?
(多个softmax,即辅助分类器,求平均来得到最后的结果;是一种正则化?)
(在实际测试的时候,这两个额外的softmax会被去掉?)
GoogleNet也能做目标检测?(只能作为例如Faster R-CNN的一部分?)
Inception v2将Inception v1的5*5卷积变为2个3*3卷积;(这个trick在vgg就出现过了吧!)
Inception v2将Inception v1对称的conv分解为非对称的conv计算;(n*n卷积用一组1*n卷积和n*1卷积代替;)
Inception v2结合了两种下采样,pooling和Inception;
Inception v3和Inception v2区别只有一点,即v3在Aug loss中使用了BN进行regularization;
Inception v4研究了Inception模块与残差连接的结合;(得到Inception-ResNet-v1、Inception-ResNet-v2、Inception-v4三种网络;)
作者认为残差连接并不是深度网络所必须的(PS:ResNet的作者说残差连接是深度网络的标配);
GoogleNet相关论文:
Going deeper with convolutions(Inception v1)
Rethinking the Inception Architecture for Computer Vision(Inception v2和Inception v3)
Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning (Inception v4)
https://my.oschina.net/u/876354/blog/1637819
https://blog.youkuaiyun.com/sunbaigui/article/details/50807362
https://www.imooc.com/article/69524
https://blog.youkuaiyun.com/weixin_39953502/article/details/80966046
https://github.com/BVLC/caffe/blob/master/models/bvlc_googlenet/train_val.prototxt
https://blog.youkuaiyun.com/shuzfan/article/details/50738394
https://blog.youkuaiyun.com/loveliuzz/article/details/79135583
https://blog.youkuaiyun.com/stdcoutzyx/article/details/51052847
ResNet就是为了解决网络深度变深以后的性能退化和梯度消失问题;
ResNet核心思想是引入一个所谓的「恒等快捷连接」(identity shortcut connection),直接跳过一个或多个层;
(ResNet不是第一个利用快捷连接的模型,Highway Networks就引入了门控快捷连接;)
Resnet提供了两种选择方式,即identity mapping和residual mapping;
ResNet-152深度是VGG19的8倍,但是计算复杂度比后者低;
ResNet除了classification还能做detection、localization、segmentation?(只能作为例如Faster R-CNN的一部分)
ResNet V2和ResNet V1主要区别在于,作者通过研究ResNet残差学习单元的传播公式,发现前馈和反馈信号可以直接传输;
(捷径连接(shortcut connection)的非线性激活函数(如ReLU)替换为 Identity Mapping;)
(此外,ResNet V2在每一层都使用了Batch Normalization;)
ResNet相关论文:
Deep Residual Learning for Image Recognition(ResNet v1)
Identity Mappings in Deep Residual Networks(ResNet v2)
https://my.oschina.net/u/876354/blog/1622896
https://blog.youkuaiyun.com/lanran2/article/details/79057994 !!!
https://blog.youkuaiyun.com/u013181595/article/details/80990930
http://baijiahao.baidu.com/s?id=1598536455758606033&wfr=spider&for=pc
DenseNet将每一层与其他的层都链接起来;
(对于每一层网络,前面所有网络的特征图都被作为输入,同时其特征图也都被其他网络层作为输入所利用;)
(缓解梯度消失问题,加强特征传播,鼓励特征复用,极大的减少了参数量;同时内存占用非常高;)
(注意:dense connectivity仅仅是在一个dense block里的,不同dense block之间是没有dense connectivity的;)
DenseNet相关论文:
Densely Connected Convolutional Networks
https://www.imooc.com/article/36508
https://blog.youkuaiyun.com/sigai_csdn/article/details/82115254
https://blog.youkuaiyun.com/u012938704/article/details/53468483
SqueezeNet相关论文:
SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and< 0.5 MB model size
https://blog.youkuaiyun.com/csdnldp/article/details/78648543
https://blog.youkuaiyun.com/u011995719/article/details/78908755
http://www.imooc.com/article/264019
MobileNet相关论文:
Mobilenets: Efficient convolutional neural networks for mobile vision applications
https://blog.youkuaiyun.com/hongbin_xu/article/details/82957426
https://blog.youkuaiyun.com/u011974639/article/details/79199306
https://www.jianshu.com/p/854cb5857070
https://blog.youkuaiyun.com/c20081052/article/details/80703896
---------------------------------------------------------------
搭讪一下作者: