图像分割库segmentation_models.pytorch

迷南。 2023-01-17 10:41 179阅读 0赞

segmentation\_models\_pytorch是一个基于PyTorch的图像分割神经网络

这个新集合由俄罗斯的程序员小哥Pavel Yakubovskiy一手打造。  
github地址：[https://github.com/qubvel/segmentation\_models.pytorch][https_github.com_qubvel_segmentation_models.pytorch]   
![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2hoaGhoaGhoaGh3d3d3d3d3d3d3_size_16_color_FFFFFF_t_70][]  
该库的主要功能是：

1.  高级API（只需两行即可创建神经网络）
2.  用于二分类和多类分割的9种模型架构（包括传奇的Unet）
3.  每种架构有104种可用的编码器
4.  所有编码器均具有预训练的权重，以实现更快更好的收敛

一、安装

PyPI version:

pip install segmentation-models-pytorch

*  Latest version from source:

pip install git+https://github.com/qubvel/segmentation_models.pytorch

二、使用

*  由于该库是基于PyTorch框架构建的，因此创建的细分模型只是一个PyTorch nn.Module，可以轻松地创建它：

import segmentation_models_pytorch as smp
    model = smp.Unet()

*  根据任务的不同，您可以通过选择具有更少或更多参数的主干并使用预训练的权重来初始化它来更改网络体系结构：

model = smp.Unet('resnet34', encoder_weights='imagenet')

*  更改模型中输出类的数量：

model = smp.Unet('resnet34', classes=3, activation='softmax')

*  所有模型均具有预训练的编码器，因此您必须按照权重预训练的相同方法准备数据：

from segmentation_models_pytorch.encoders import get_preprocessing_fn
    preprocess_input = get_preprocessing_fn('resnet18', pretrained='imagenet')

编码器列表

以下是SMP中支持的编码器列表。选择适当的编码器系列并单击以展开表，然后选择特定的编码器及其预先训练的权重（`encoder_name` and `encoder_weights` parameters）。

ResNet

<table> 
 <thead> 
  <tr> 
   <th>Encoder</th> 
   <th>Weights</th> 
   <th>Params, M</th> 
  </tr> 
 </thead> 
 <tbody> 
  <tr> 
   <td>resnet18</td> 
   <td>imagenet / ssl / swsl</td> 
   <td>11M</td> 
  </tr> 
  <tr> 
   <td>resnet34</td> 
   <td>imagenet</td> 
   <td>21M</td> 
  </tr> 
  <tr> 
   <td>resnet50</td> 
   <td>imagenet / ssl / swsl</td> 
   <td>23M</td> 
  </tr> 
  <tr> 
   <td>resnet101</td> 
   <td>imagenet</td> 
   <td>42M</td> 
  </tr> 
  <tr> 
   <td>resnet152</td> 
   <td>imagenet</td> 
   <td>58M</td> 
  </tr> 
 </tbody> 
</table>

ResNeXt

<table> 
 <thead> 
  <tr> 
   <th>Encoder</th> 
   <th>Weights</th> 
   <th>Params, M</th> 
  </tr> 
 </thead> 
 <tbody> 
  <tr> 
   <td>resnext50_32x4d</td> 
   <td>imagenet / ssl / swsl</td> 
   <td>22M</td> 
  </tr> 
  <tr> 
   <td>resnext101_32x4d</td> 
   <td>ssl / swsl</td> 
   <td>42M</td> 
  </tr> 
  <tr> 
   <td>resnext101_32x8d</td> 
   <td>imagenet / instagram / ssl / swsl</td> 
   <td>86M</td> 
  </tr> 
  <tr> 
   <td>resnext101_32x16d</td> 
   <td>instagram / ssl / swsl</td> 
   <td>191M</td> 
  </tr> 
  <tr> 
   <td>resnext101_32x32d</td> 
   <td>instagram</td> 
   <td>466M</td> 
  </tr> 
  <tr> 
   <td>resnext101_32x48d</td> 
   <td>instagram</td> 
   <td>826M</td> 
  </tr> 
 </tbody> 
</table>

ResNeSt

<table> 
 <thead> 
  <tr> 
   <th>Encoder</th> 
   <th>Weights</th> 
   <th>Params, M</th> 
  </tr> 
 </thead> 
 <tbody> 
  <tr> 
   <td>timm-resnest14d</td> 
   <td>imagenet</td> 
   <td>8M</td> 
  </tr> 
  <tr> 
   <td>timm-resnest26d</td> 
   <td>imagenet</td> 
   <td>15M</td> 
  </tr> 
  <tr> 
   <td>timm-resnest50d</td> 
   <td>imagenet</td> 
   <td>25M</td> 
  </tr> 
  <tr> 
   <td>timm-resnest101e</td> 
   <td>imagenet</td> 
   <td>46M</td> 
  </tr> 
  <tr> 
   <td>timm-resnest200e</td> 
   <td>imagenet</td> 
   <td>68M</td> 
  </tr> 
  <tr> 
   <td>timm-resnest269e</td> 
   <td>imagenet</td> 
   <td>108M</td> 
  </tr> 
  <tr> 
   <td>timm-resnest50d_4s2x40d</td> 
   <td>imagenet</td> 
   <td>28M</td> 
  </tr> 
  <tr> 
   <td>timm-resnest50d_1s4x24d</td> 
   <td>imagenet</td> 
   <td>23M</td> 
  </tr> 
 </tbody> 
</table>

Res2Ne(X)t

<table> 
 <thead> 
  <tr> 
   <th>Encoder</th> 
   <th>Weights</th> 
   <th>Params, M</th> 
  </tr> 
 </thead> 
 <tbody> 
  <tr> 
   <td>timm-res2net50_26w_4s</td> 
   <td>imagenet</td> 
   <td>23M</td> 
  </tr> 
  <tr> 
   <td>timm-res2net101_26w_4s</td> 
   <td>imagenet</td> 
   <td>43M</td> 
  </tr> 
  <tr> 
   <td>timm-res2net50_26w_6s</td> 
   <td>imagenet</td> 
   <td>35M</td> 
  </tr> 
  <tr> 
   <td>timm-res2net50_26w_8s</td> 
   <td>imagenet</td> 
   <td>46M</td> 
  </tr> 
  <tr> 
   <td>timm-res2net50_48w_2s</td> 
   <td>imagenet</td> 
   <td>23M</td> 
  </tr> 
  <tr> 
   <td>timm-res2net50_14w_8s</td> 
   <td>imagenet</td> 
   <td>23M</td> 
  </tr> 
  <tr> 
   <td>timm-res2next50</td> 
   <td>imagenet</td> 
   <td>22M</td> 
  </tr> 
 </tbody> 
</table>

RegNet(x/y)

<table> 
 <thead> 
  <tr> 
   <th>Encoder</th> 
   <th>Weights</th> 
   <th>Params, M</th> 
  </tr> 
 </thead> 
 <tbody> 
  <tr> 
   <td>timm-regnetx_002</td> 
   <td>imagenet</td> 
   <td>2M</td> 
  </tr> 
  <tr> 
   <td>timm-regnetx_004</td> 
   <td>imagenet</td> 
   <td>4M</td> 
  </tr> 
  <tr> 
   <td>timm-regnetx_006</td> 
   <td>imagenet</td> 
   <td>5M</td> 
  </tr> 
  <tr> 
   <td>timm-regnetx_008</td> 
   <td>imagenet</td> 
   <td>6M</td> 
  </tr> 
  <tr> 
   <td>timm-regnetx_016</td> 
   <td>imagenet</td> 
   <td>8M</td> 
  </tr> 
  <tr> 
   <td>timm-regnetx_032</td> 
   <td>imagenet</td> 
   <td>14M</td> 
  </tr> 
  <tr> 
   <td>timm-regnetx_040</td> 
   <td>imagenet</td> 
   <td>20M</td> 
  </tr> 
  <tr> 
   <td>timm-regnetx_064</td> 
   <td>imagenet</td> 
   <td>24M</td> 
  </tr> 
  <tr> 
   <td>timm-regnetx_080</td> 
   <td>imagenet</td> 
   <td>37M</td> 
  </tr> 
  <tr> 
   <td>timm-regnetx_120</td> 
   <td>imagenet</td> 
   <td>43M</td> 
  </tr> 
  <tr> 
   <td>timm-regnetx_160</td> 
   <td>imagenet</td> 
   <td>52M</td> 
  </tr> 
  <tr> 
   <td>timm-regnetx_320</td> 
   <td>imagenet</td> 
   <td>105M</td> 
  </tr> 
  <tr> 
   <td>timm-regnety_002</td> 
   <td>imagenet</td> 
   <td>2M</td> 
  </tr> 
  <tr> 
   <td>timm-regnety_004</td> 
   <td>imagenet</td> 
   <td>3M</td> 
  </tr> 
  <tr> 
   <td>timm-regnety_006</td> 
   <td>imagenet</td> 
   <td>5M</td> 
  </tr> 
  <tr> 
   <td>timm-regnety_008</td> 
   <td>imagenet</td> 
   <td>5M</td> 
  </tr> 
  <tr> 
   <td>timm-regnety_016</td> 
   <td>imagenet</td> 
   <td>10M</td> 
  </tr> 
  <tr> 
   <td>timm-regnety_032</td> 
   <td>imagenet</td> 
   <td>17M</td> 
  </tr> 
  <tr> 
   <td>timm-regnety_040</td> 
   <td>imagenet</td> 
   <td>19M</td> 
  </tr> 
  <tr> 
   <td>timm-regnety_064</td> 
   <td>imagenet</td> 
   <td>29M</td> 
  </tr> 
  <tr> 
   <td>timm-regnety_080</td> 
   <td>imagenet</td> 
   <td>37M</td> 
  </tr> 
  <tr> 
   <td>timm-regnety_120</td> 
   <td>imagenet</td> 
   <td>49M</td> 
  </tr> 
  <tr> 
   <td>timm-regnety_160</td> 
   <td>imagenet</td> 
   <td>80M</td> 
  </tr> 
  <tr> 
   <td>timm-regnety_320</td> 
   <td>imagenet</td> 
   <td>141M</td> 
  </tr> 
 </tbody> 
</table>

SE-Net

<table> 
 <thead> 
  <tr> 
   <th>Encoder</th> 
   <th>Weights</th> 
   <th>Params, M</th> 
  </tr> 
 </thead> 
 <tbody> 
  <tr> 
   <td>senet154</td> 
   <td>imagenet</td> 
   <td>113M</td> 
  </tr> 
  <tr> 
   <td>se_resnet50</td> 
   <td>imagenet</td> 
   <td>26M</td> 
  </tr> 
  <tr> 
   <td>se_resnet101</td> 
   <td>imagenet</td> 
   <td>47M</td> 
  </tr> 
  <tr> 
   <td>se_resnet152</td> 
   <td>imagenet</td> 
   <td>64M</td> 
  </tr> 
  <tr> 
   <td>se_resnext50_32x4d</td> 
   <td>imagenet</td> 
   <td>25M</td> 
  </tr> 
  <tr> 
   <td>se_resnext101_32x4d</td> 
   <td>imagenet</td> 
   <td>46M</td> 
  </tr> 
 </tbody> 
</table>

SK-ResNe(X)t

<table> 
 <thead> 
  <tr> 
   <th>Encoder</th> 
   <th>Weights</th> 
   <th>Params, M</th> 
  </tr> 
 </thead> 
 <tbody> 
  <tr> 
   <td>timm-skresnet18</td> 
   <td>imagenet</td> 
   <td>11M</td> 
  </tr> 
  <tr> 
   <td>timm-skresnet34</td> 
   <td>imagenet</td> 
   <td>21M</td> 
  </tr> 
  <tr> 
   <td>timm-skresnext50_32x4d</td> 
   <td>imagenet</td> 
   <td>25M</td> 
  </tr> 
 </tbody> 
</table>

DenseNet

<table> 
 <thead> 
  <tr> 
   <th>Encoder</th> 
   <th>Weights</th> 
   <th>Params, M</th> 
  </tr> 
 </thead> 
 <tbody> 
  <tr> 
   <td>densenet121</td> 
   <td>imagenet</td> 
   <td>6M</td> 
  </tr> 
  <tr> 
   <td>densenet169</td> 
   <td>imagenet</td> 
   <td>12M</td> 
  </tr> 
  <tr> 
   <td>densenet201</td> 
   <td>imagenet</td> 
   <td>18M</td> 
  </tr> 
  <tr> 
   <td>densenet161</td> 
   <td>imagenet</td> 
   <td>26M</td> 
  </tr> 
 </tbody> 
</table>

Inception

<table> 
 <thead> 
  <tr> 
   <th>Encoder</th> 
   <th>Weights</th> 
   <th>Params, M</th> 
  </tr> 
 </thead> 
 <tbody> 
  <tr> 
   <td>inceptionresnetv2</td> 
   <td>imagenet / imagenet+background</td> 
   <td>54M</td> 
  </tr> 
  <tr> 
   <td>inceptionv4</td> 
   <td>imagenet / imagenet+background</td> 
   <td>41M</td> 
  </tr> 
  <tr> 
   <td>xception</td> 
   <td>imagenet</td> 
   <td>22M</td> 
  </tr> 
 </tbody> 
</table>

EfficientNet

<table> 
 <thead> 
  <tr> 
   <th>Encoder</th> 
   <th>Weights</th> 
   <th>Params, M</th> 
  </tr> 
 </thead> 
 <tbody> 
  <tr> 
   <td>efficientnet-b0</td> 
   <td>imagenet</td> 
   <td>4M</td> 
  </tr> 
  <tr> 
   <td>efficientnet-b1</td> 
   <td>imagenet</td> 
   <td>6M</td> 
  </tr> 
  <tr> 
   <td>efficientnet-b2</td> 
   <td>imagenet</td> 
   <td>7M</td> 
  </tr> 
  <tr> 
   <td>efficientnet-b3</td> 
   <td>imagenet</td> 
   <td>10M</td> 
  </tr> 
  <tr> 
   <td>efficientnet-b4</td> 
   <td>imagenet</td> 
   <td>17M</td> 
  </tr> 
  <tr> 
   <td>efficientnet-b5</td> 
   <td>imagenet</td> 
   <td>28M</td> 
  </tr> 
  <tr> 
   <td>efficientnet-b6</td> 
   <td>imagenet</td> 
   <td>40M</td> 
  </tr> 
  <tr> 
   <td>efficientnet-b7</td> 
   <td>imagenet</td> 
   <td>63M</td> 
  </tr> 
  <tr> 
   <td>timm-efficientnet-b0</td> 
   <td>imagenet / advprop / noisy-student</td> 
   <td>4M</td> 
  </tr> 
  <tr> 
   <td>timm-efficientnet-b1</td> 
   <td>imagenet / advprop / noisy-student</td> 
   <td>6M</td> 
  </tr> 
  <tr> 
   <td>timm-efficientnet-b2</td> 
   <td>imagenet / advprop / noisy-student</td> 
   <td>7M</td> 
  </tr> 
  <tr> 
   <td>timm-efficientnet-b3</td> 
   <td>imagenet / advprop / noisy-student</td> 
   <td>10M</td> 
  </tr> 
  <tr> 
   <td>timm-efficientnet-b4</td> 
   <td>imagenet / advprop / noisy-student</td> 
   <td>17M</td> 
  </tr> 
  <tr> 
   <td>timm-efficientnet-b5</td> 
   <td>imagenet / advprop / noisy-student</td> 
   <td>28M</td> 
  </tr> 
  <tr> 
   <td>timm-efficientnet-b6</td> 
   <td>imagenet / advprop / noisy-student</td> 
   <td>40M</td> 
  </tr> 
  <tr> 
   <td>timm-efficientnet-b7</td> 
   <td>imagenet / advprop / noisy-student</td> 
   <td>63M</td> 
  </tr> 
  <tr> 
   <td>timm-efficientnet-b8</td> 
   <td>imagenet / advprop</td> 
   <td>84M</td> 
  </tr> 
  <tr> 
   <td>timm-efficientnet-l2</td> 
   <td>noisy-student</td> 
   <td>474M</td> 
  </tr> 
  <tr> 
   <td>timm-efficientnet-lite0</td> 
   <td>imagenet</td> 
   <td>4M</td> 
  </tr> 
  <tr> 
   <td>timm-efficientnet-lite1</td> 
   <td>imagenet</td> 
   <td>5M</td> 
  </tr> 
  <tr> 
   <td>timm-efficientnet-lite2</td> 
   <td>imagenet</td> 
   <td>6M</td> 
  </tr> 
  <tr> 
   <td>timm-efficientnet-lite3</td> 
   <td>imagenet</td> 
   <td>8M</td> 
  </tr> 
  <tr> 
   <td>timm-efficientnet-lite4</td> 
   <td>imagenet</td> 
   <td>13M</td> 
  </tr> 
 </tbody> 
</table>

MobileNet

<table> 
 <thead> 
  <tr> 
   <th>Encoder</th> 
   <th>Weights</th> 
   <th>Params, M</th> 
  </tr> 
 </thead> 
 <tbody> 
  <tr> 
   <td>mobilenet_v2</td> 
   <td>imagenet</td> 
   <td>2M</td> 
  </tr> 
 </tbody> 
</table>

DPN

<table> 
 <thead> 
  <tr> 
   <th>Encoder</th> 
   <th>Weights</th> 
   <th>Params, M</th> 
  </tr> 
 </thead> 
 <tbody> 
  <tr> 
   <td>dpn68</td> 
   <td>imagenet</td> 
   <td>11M</td> 
  </tr> 
  <tr> 
   <td>dpn68b</td> 
   <td>imagenet+5k</td> 
   <td>11M</td> 
  </tr> 
  <tr> 
   <td>dpn92</td> 
   <td>imagenet+5k</td> 
   <td>34M</td> 
  </tr> 
  <tr> 
   <td>dpn98</td> 
   <td>imagenet</td> 
   <td>58M</td> 
  </tr> 
  <tr> 
   <td>dpn107</td> 
   <td>imagenet+5k</td> 
   <td>84M</td> 
  </tr> 
  <tr> 
   <td>dpn131</td> 
   <td>imagenet</td> 
   <td>76M</td> 
  </tr> 
 </tbody> 
</table>

VGG

<table> 
 <thead> 
  <tr> 
   <th>Encoder</th> 
   <th>Weights</th> 
   <th>Params, M</th> 
  </tr> 
 </thead> 
 <tbody> 
  <tr> 
   <td>vgg11</td> 
   <td>imagenet</td> 
   <td>9M</td> 
  </tr> 
  <tr> 
   <td>vgg11_bn</td> 
   <td>imagenet</td> 
   <td>9M</td> 
  </tr> 
  <tr> 
   <td>vgg13</td> 
   <td>imagenet</td> 
   <td>9M</td> 
  </tr> 
  <tr> 
   <td>vgg13_bn</td> 
   <td>imagenet</td> 
   <td>9M</td> 
  </tr> 
  <tr> 
   <td>vgg16</td> 
   <td>imagenet</td> 
   <td>14M</td> 
  </tr> 
  <tr> 
   <td>vgg16_bn</td> 
   <td>imagenet</td> 
   <td>14M</td> 
  </tr> 
  <tr> 
   <td>vgg19</td> 
   <td>imagenet</td> 
   <td>20M</td> 
  </tr> 
  <tr> 
   <td>vgg19_bn</td> 
   <td>imagenet</td> 
   <td>20M</td> 
  </tr> 
 </tbody> 
</table>

[https_github.com_qubvel_segmentation_models.pytorch]: https://github.com/qubvel/segmentation_models.pytorch
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2hoaGhoaGhoaGh3d3d3d3d3d3d3_size_16_color_FFFFFF_t_70]: /images/20221021/ad7d8a26865c4d4fa5a6ce4632e45f7a.png