###### tags: `課程共筆`

# Squeeze Net

## 目錄
[TOC]

## 特色
- 研究壓縮CNN模型
- 模型小
- 訓練快
- 通訊成本低
- 容易佈署到edge端
- 盡可能不要太早把feature map壓縮得太小
- >Quote: Our intuition is that large activation maps (due to delayed downsampling) can lead to higher classification accuracy, with all else held equal.Indeed, K. He and H. Sun applied delayed downsampling to four different CNN architectures, and in each case delayed downsampling led to higher classification accuracy (He & Sun, 2015).


## 減少參數的方法
- 用1x1 conv取代3x3 conv
- Squeeze
- 用fire module把1x1和3x3的filter包起來，輸入會先經過1x1 conv壓縮，減少ch，然後送進3x3filter進行萃取，藉此減少參數量(1/9)
- Expand
- 3x3的conv會影響image shape，所以對他做了zero padding。藉此保證1x1與3x3的conv輸出shape相同。
> 

## Ref
> https://arxiv.org/pdf/1602.07360.pdf