CNN Network Structure2

• 前言: ResNet
  
  
• 
  
  

0 More Details: HomeLink

1 Why ResNet

The deeper the network,
The more parameters,
The worse results we’ll get
WERIED!!!
WHY???
Too deep, accuracy saturated
Gradient vanished
How???

1.1 Structure: 2Qs

Structure

We can use conv/pool to reduce the size,
What about the shortcut?


How can we add 2 parts without same resolution?

1. Get rich set of primary features

2. Channel of input layer is less, big kernel doesn’t have to mean great of params

3. Computing reason:
   Res：7 * 7 * 3 * 64 *112 * 112 = 120M
   VGG: 3 * 3 * 3 * 64 * 224 * 224 +
   3 * 3 * 64 * 64 * 224 * 224 +
   3 * 3 * 64 * 128 * 112 *112 +
   3 * 3 * 128 * 128 * 56 * 56
   = x >>120M * 2

1.2 VggNet & Small kernel

Small kernel always better?

NO

Sacrificing more real details of the image

1.3 Structure: Revisit

50-layer : 49 Conv layer

扫描二维码关注公众号，回复： 12312617 查看本文章

conv1: 7 * 7

before 50-layer no 1 * 1

50-layer, num of channel: less less more…

how add between left and right ? add conv in left…

2 Structure: Advanced?

deeper

wider

3 Why better?

1 Solved gradient vanishing by using shortcut

2 Can be seen as assembled models