关于facebook复现Efficientnet与官方对比

古城微笑少年丶 2022-12-18 07:58 189阅读 0赞

## 0.链接 ##

<table> 
 <tbody> 
  <tr> 
   <td style="width:282px;">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;官方</td> 
   <td style="width:530px;">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;facebook&nbsp;</td> 
  </tr> 
  <tr> 
   <td style="width:282px;">&nbsp; &nbsp; &nbsp; <a href="https://github.com/qubvel/efficientnet" rel="nofollow">https://github.com/qubvel/efficientnet</a></td> 
   <td style="width:530px;">&nbsp; &nbsp; &nbsp; &nbsp; <a href="https://github.com/facebookresearch/pycls/blob/master/MODEL_ZOO.md" rel="nofollow">https://github.com/facebookresearch/pycls/blob/master/MODEL_ZOO.md</a></td> 
  </tr> 
 </tbody> 
</table>

## 1.框架 ##

<table style="width:404px;"> 
 <tbody> 
  <tr> 
   <td style="width:207px;">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; 官方</td> 
   <td style="width:196px;">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; facebook</td> 
  </tr> 
  <tr> 
   <td style="width:207px;">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; tenosorflow、keras</td> 
   <td style="width:196px;">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; pytorch</td> 
  </tr> 
 </tbody> 
</table>

## 2.环境对比 ##

<table style="width:486px;"> 
 <tbody> 
  <tr> 
   <td style="width:294px;">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; 官方</td> 
   <td style="width:189px;">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; facebook</td> 
  </tr> 
  <tr> 
   <td style="width:294px;"> <p>&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; keras_applications&gt;=1.0.7,&lt;=1.0.8</p> <p>&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; keras=2.2.4</p> </td> 
   <td style="width:189px;"> <p>&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; PyTorch 1.6</p> <p>&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; CUDA 10.0</p> <p>&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; cuDNN 7.6</p> </td> 
  </tr> 
 </tbody> 
</table>

# #

## 3.预训练模型错误率对比（error） ##

由于facebook的实现到B5，所以这里只对比到B5。数据都是Imagenet-1k。

<table style="width:308px;"> 
 <tbody> 
  <tr> 
   <td colspan="1" style="width:130px;">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; 模型</td> 
   <td colspan="2" style="width:175px;">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; error</td> 
  </tr> 
  <tr> 
   <td style="width:76px;">&nbsp; &nbsp;官方</td> 
   <td style="width:101px;">&nbsp; facebook</td> 
  </tr> 
  <tr> 
   <td style="width:130px;">&nbsp; &nbsp;EfficientNetB0</td> 
   <td style="width:76px;">&nbsp; &nbsp;22.8</td> 
   <td style="width:101px;">&nbsp; &nbsp; &nbsp; 24.9</td> 
  </tr> 
  <tr> 
   <td style="width:130px;">&nbsp; &nbsp;EfficientNetB1</td> 
   <td style="width:76px;">&nbsp; &nbsp;20.9</td> 
   <td style="width:101px;">&nbsp; &nbsp; &nbsp; 24.1</td> 
  </tr> 
  <tr> 
   <td style="width:130px;">&nbsp; &nbsp;EfficientNetB2</td> 
   <td style="width:76px;">&nbsp; &nbsp;19.8</td> 
   <td style="width:101px;">&nbsp; &nbsp; &nbsp; 23.5</td> 
  </tr> 
  <tr> 
   <td style="width:130px;">&nbsp; &nbsp;EfficientNetB3</td> 
   <td style="width:76px;">&nbsp; &nbsp;18.4</td> 
   <td style="width:101px;">&nbsp; &nbsp; &nbsp; 22.5</td> 
  </tr> 
  <tr> 
   <td style="width:130px;">&nbsp; &nbsp;EfficientNetB4</td> 
   <td style="width:76px;">&nbsp; &nbsp;17.0</td> 
   <td style="width:101px;">&nbsp; &nbsp; &nbsp; 21.4</td> 
  </tr> 
  <tr> 
   <td style="width:130px;">&nbsp; &nbsp;EfficientNetB5</td> 
   <td style="width:76px;">&nbsp; &nbsp;16.3</td> 
   <td style="width:101px;">&nbsp; &nbsp; &nbsp; 21.7</td> 
  </tr> 
 </tbody> 
</table>

## 参考 ##

1.官方EfficientNet

[2.facebook版本EfficientNet][2.facebook_EfficientNet]

[2.facebook_EfficientNet]: https://github.com/facebookresearch/pycls/blob/master/MODEL_ZOO.md