Case Study

  Coursera, Deep Learning 4, Convolutional Neural Networks - week2-LMLPHP

(Note: 红色表示不重要)

LeNet-5 起初用来识别手写数字灰度图片

  Coursera, Deep Learning 4, Convolutional Neural Networks - week2-LMLPHP

AlexNet 输入的是227x227x3 的图片,输出1000 种类的结果

  Coursera, Deep Learning 4, Convolutional Neural Networks - week2-LMLPHP

VGG

VGG比AlexNet 结构更简单,filter 都是3x3的,max-pool 都是 2x2的.

  Coursera, Deep Learning 4, Convolutional Neural Networks - week2-LMLPHP

ResNets (Residual Network)

可用让很深的network 工作的很好. This really helps with the vanishing and exploding gradient problems.

  Coursera, Deep Learning 4, Convolutional Neural Networks - week2-LMLPHP

  Coursera, Deep Learning 4, Convolutional Neural Networks - week2-LMLPHP

为什么ResNet 会起作用呢?下图中所示如果vanishings时候 W =0, 设b也=0. a= a说明很容易保留

  Coursera, Deep Learning 4, Convolutional Neural Networks - week2-LMLPHP

  Coursera, Deep Learning 4, Convolutional Neural Networks - week2-LMLPHP

Networks in Networks and 1x1 Convolutions

1x1 convolutions 可以用来减少 channel数据,或者保持一样,甚至可以增大channel.

  Coursera, Deep Learning 4, Convolutional Neural Networks - week2-LMLPHP

  Coursera, Deep Learning 4, Convolutional Neural Networks - week2-LMLPHP

Inception network

就像大烩菜,把1x1, 3x3, 5x5, pooling 都揉到一起,就成了inception network.

  Coursera, Deep Learning 4, Convolutional Neural Networks - week2-LMLPHP

上图中有个问题是 computational cost 很高.

  Coursera, Deep Learning 4, Convolutional Neural Networks - week2-LMLPHP

这个问题可以用下面的方便解决。这个方法被证明不影响性能.

  Coursera, Deep Learning 4, Convolutional Neural Networks - week2-LMLPHP

下面是一个inception module

  Coursera, Deep Learning 4, Convolutional Neural Networks - week2-LMLPHP

一个incetption network 是有很多的inception module 组成的network. Inception 来自盗梦空间,和很深的网络有关联意思. 在整个网络中间部分分出来的branch 也是用来做predict的,在中间做预测一般是为了防止overfitting.

这个inception network 来自google的开发者,所以也叫 GoogLeNet, 后面的LeNet 是向 LeNet 的作者 Yann LeCun 致敬

  Coursera, Deep Learning 4, Convolutional Neural Networks - week2-LMLPHP

Practical advices for using ConvNets

在trainning data 少的情况下,可以用trasfer learning 的方法,在别人比较训练好的model 上修改后面的layer 来得到自己的model. 当然如果trainning set 够大,也可以自己从头到尾训练出自己的model.

  Coursera, Deep Learning 4, Convolutional Neural Networks - week2-LMLPHP

Data augmentation

  Coursera, Deep Learning 4, Convolutional Neural Networks - week2-LMLPHP

  Coursera, Deep Learning 4, Convolutional Neural Networks - week2-LMLPHP

  Coursera, Deep Learning 4, Convolutional Neural Networks - week2-LMLPHP

State of computer vision

  Coursera, Deep Learning 4, Convolutional Neural Networks - week2-LMLPHP

下面是一些tips针对benchmark/winning competitions, 但是实际工作中不常用.

  Coursera, Deep Learning 4, Convolutional Neural Networks - week2-LMLPHP

工作中常用的是下面的方法

  Coursera, Deep Learning 4, Convolutional Neural Networks - week2-LMLPHP

05-11 13:06