打通多个视觉任务的全能Backbone:HRNet
<meta http-equiv="Content-Type" content="text/html; charset=UTF-8"/><div>
<div>
<h2>
<img src="https://img1.tuicool.com/vee6RnZ.jpg!web" class="alignCenter" referrerpolicy="no-referrer"/>
</h2>
<h3>
<div>
<span> 磐创AI分享</span>
<strong>
<span> </span>
</strong>
<strong/>
</div>
</h3>
来源 | GiantPandaCV
编辑 | pprp
<h4>
<span>
<strong>【导读】</strong>
</span>
<span>HRNet是微软亚洲研究院的王井东老师领导的团队完成的，打通图像分类、图像分割、目标检测、人脸对齐、姿态识别、风格迁移、Image Inpainting、超分、optical flow、Depth estimation、边缘检测等网络结构。</span>
</h4>
王老师在ValseWebinar《物体和关键点检测》中亲自讲解了HRNet，讲解地非常透彻。以下文章主要参考了王老师在演讲中的解读，配合论文+代码部分，来为各位读者介绍这个全能的Backbone-HRNet。
引入
<figure>
<img src="https://img0.tuicool.com/feURFb2.png!web" class="alignCenter" referrerpolicy="no-referrer"/>
<figcaption>网络结构设计思路</figcaption>
</figure>
在人体姿态识别这类的任务中，需要生成一个高分辨率的heatmap来进行关键点检测。这就与一般的网络结构比如VGGNet的要求不同，因为VGGNet最终得到的feature map分辨率很低，损失了空间结构。
<figure>
<img src="https://img1.tuicool.com/rayemqu.png!web" class="alignCenter" referrerpolicy="no-referrer"/>
<figcaption>传统的解决思路</figcaption>
</figure>
获取高分辨率的方式大部分都是如上图所示，采用的是先降分辨率，然后再升分辨率的方法。U-Net、SegNet、DeconvNet、Hourglass本质上都是这种结构。
<figure>
<img src="https://img2.tuicool.com/i6Z3Qv3.png!web" class="alignCenter" referrerpolicy="no-referrer"/>
<figcaption>虽然看上去不同，但是本质是一致的</figcaption>
</figure>
<h4>核心</h4>
普通网络都是这种结构，不同分辨率之间是进行了串联
<figure>
<img src="https://img1.tuicool.com/EfiABju.png!web" class="alignCenter" referrerpolicy="no-referrer"/>
<figcaption>不断降分辨率</figcaption>
</figure>
王井东老师则是将不同分辨率的feature map进行并联：
<figure>
<img src="https://img0.tuicool.com/JRv6Jr2.png!web" class="alignCenter" referrerpolicy="no-referrer"/>
<figcaption>并联不同分辨率feature map</figcaption>
</figure>
在并联的基础上，添加不同分辨率feature map之间的交互(fusion)。
<figure>
<img src="https://img2.tuicool.com/j2IZZzq.png!web" class="alignCenter" referrerpolicy="no-referrer"/>
</figure>
具体fusion的方法如下图所示：
<figure>
<img src="https://img0.tuicool.com/Qv6VZjE.png!web" class="alignCenter" referrerpolicy="no-referrer"/>
</figure>
<ul>
<li>
<p class="no-text-indent">同分辨率的层直接复制。</p>
</li>
<li>
<p class="no-text-indent">需要升分辨率的使用bilinear upsample + 1x1卷积将channel数统一。</p>
</li>
<li>
<p class="no-text-indent">需要降分辨率的使用strided 3x3 卷积。</p>
</li>
<li>
<p class="no-text-indent">三个feature map融合的方式是相加。</p>
</li>
</ul>
至于为何要用strided 3x3卷积，这是因为卷积在降维的时候会出现信息损失，使用strided 3x3卷积是为了通过学习的方式，降低信息的损耗。所以这里没有用maxpool或者组合池化。
<figure>
<img src="https://img2.tuicool.com/buINryY.png!web" class="alignCenter" referrerpolicy="no-referrer"/>
<figcaption>HR示意图</figcaption>
</figure>
另外在读HRNet的时候会有一个问题，有四个分支的到底如何使用这几个分支呢？论文中也给出了几种方式作为最终的特征选择。
<figure>
<img src="https://img1.tuicool.com/EJ3U7vb.png!web" class="alignCenter" referrerpolicy="no-referrer"/>
<figcaption>三种特征融合方法</figcaption>
</figure>
(a)图展示的是HRNetV1的特征选择，只使用分辨率最高的特征图。
(b)图展示的是HRNetV2的特征选择，将所有分辨率的特征图(小的特征图进行upsample)进行concate，主要用于语义分割和面部关键点检测。
(c)图展示的是HRNetV2p的特征选择，在HRNetV2的基础上，使用了一个特征金字塔，主要用于目标检测网络。
再补充一个(d)图
<figure>
<img src="https://img2.tuicool.com/a6N3eyq.png!web" class="alignCenter" referrerpolicy="no-referrer"/>
<figcaption>HRNetV2分类网络后的特征选择</figcaption>
</figure>
(d)图展示的也是HRNetV2，采用上图的融合方式，主要用于训练分类网络。
-
将高低分辨率之间的链接由串联改为并联。
-
在整个网络结构中都保持了高分辨率的表征(最上边那个通路)。
-
在高低分辨率中引入了交互来提高模型性能。
<h4>效果</h4>
消融实验
<ol>
<li>
<p class="no-text-indent">对交互方法进行消融实验，证明了当前跨分辨率的融合的有效性。</p>
</li>
</ol>
<figure>
<img src="https://img2.tuicool.com/Y3QvA3R.png!web" class="alignCenter" referrerpolicy="no-referrer"/>
<figcaption>交互方法的消融实现</figcaption>
</figure>
<ol start="2">
<li>
<p class="no-text-indent">证明高分辨率feature map的表征能力</p>
</li>
</ol>
<figure>
<img src="https://img2.tuicool.com/7baMviM.png!web" class="alignCenter" referrerpolicy="no-referrer"/>
</figure>
1x代表不进行降维，2x代表分辨率变为原来一半，4x代表分辨率变为原来四分之一。W32、W48中的32、48代表卷积的宽度或者通道数。
姿态识别任务上的表现
<figure>
<img src="https://img2.tuicool.com/iamIz2U.png!web" class="alignCenter" referrerpolicy="no-referrer"/>
</figure>
以上的姿态识别采用的是top-down的方法。
<figure>
<img src="https://img2.tuicool.com/Qb2EfqB.png!web" class="alignCenter" referrerpolicy="no-referrer"/>
<figcaption>COCO验证集的结果</figcaption>
</figure>
在参数和计算量不增加的情况下，要比其他同类网络效果好很多。
<figure>
<img src="https://img0.tuicool.com/JR3UbiZ.png!web" class="alignCenter" referrerpolicy="no-referrer"/>
<figcaption>COCO测试集上的结果</figcaption>
</figure>
在19年2月28日时的PoseTrack Leaderboard，HRNet占领两个项目的第一名。
<figure>
<img src="https://img0.tuicool.com/Nf26vyZ.png!web" class="alignCenter" referrerpolicy="no-referrer"/>
<figcaption>PoseTrack Leaderboard</figcaption>
</figure>
<div>
<h4>语义分割任务中的表现</h4>
<p style="text-align:center">
<img src="https://img0.tuicool.com/RZFJvuJ.png!web" class="alignCenter" referrerpolicy="no-referrer"/>
</p>
</div>
<figure>
<div>CityScape验证集上的结果对比</div>
</figure>
<figure>
<img src="https://img1.tuicool.com/vUBv2qm.png!web" class="alignCenter" referrerpolicy="no-referrer"/>
<figcaption>Cityscapes测试集上的对比</figcaption>
</figure>
<h4>目标检测任务中的表现</h4>
<figure>
<img src="https://img1.tuicool.com/JBNf2un.png!web" class="alignCenter" referrerpolicy="no-referrer"/>
<figcaption>单模型单尺度模型对比</figcaption>
</figure>
<figure>
<img src="https://img0.tuicool.com/e2a26fF.png!web" class="alignCenter" referrerpolicy="no-referrer"/>
<figcaption>Mask R-CNN上结果</figcaption>
</figure>
分类任务上的表现
<figure>
<img src="https://img0.tuicool.com/nINfEry.png!web" class="alignCenter" referrerpolicy="no-referrer"/>
</figure>
ps: 王井东老师在这部分提到，分割的网络也需要使用分类的预训练模型，否则结果会差几个点。
<figure>
<img src="https://img2.tuicool.com/FN3Y3mV.png!web" class="alignCenter" referrerpolicy="no-referrer"/>
<figcaption>图像分类任务中和ResNet进行对比</figcaption>
</figure>
以上是HRNet和ResNet结果对比，同一个颜色的都是参数量大体一致的模型进行的对比，在参数量差不多甚至更少的情况下，HRNet能够比ResNet达到更好的效果。
HRNet( https://github.com/HRNet )工作量非常大，构建了六个库涉及语义分割、人体姿态检测、目标检测、图片分类、面部关键点检测、Mask R-CNN等库。全部内容如下图所示：
<figure>
<img src="https://img0.tuicool.com/ru6nErZ.png!web" class="alignCenter" referrerpolicy="no-referrer"/>
</figure>
笔者对HRNet代码构建非常感兴趣，所以以HRNet-Image-Classification库为例，来解析一下这部分代码。
先从简单的入手，BasicBlock
<figure>
<img src="https://img0.tuicool.com/AB3amym.png!web" class="alignCenter" referrerpolicy="no-referrer"/>
<figcaption>BasicBlock结构</figcaption>
</figure>
<pre class="prettyprint"><span style="line-height: 1.6 !important;"><span style="line-height:1.6 !important;">def</span> <span style="line-height:1.6 !important;">conv3x3</span><span style="line-height: 1.6 !important;">(in_planes, out_planes, stride=<span style="line-height:1.6 !important;">1</span>)</span>:</span><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">"""3x3 convolution with padding"""</span><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">return</span> nn.Conv2d(in_planes, out_planes, kernel_size=<span style="line-height:1.6 !important;">3</span>, stride=stride,<br style="line-height: 1.6 !important;"/> padding=<span style="line-height:1.6 !important;">1</span>, bias=<span style="line-height:1.6 !important;">False</span>)<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/><span style="line-height: 1.6 !important;"><span style="line-height:1.6 !important;">class</span> <span style="line-height:1.6 !important;">BasicBlock</span><span style="line-height: 1.6 !important;">(nn.Module)</span>:</span><br style="line-height: 1.6 !important;"/> expansion = <span style="line-height:1.6 !important;">1</span><br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height: 1.6 !important;"><span style="line-height:1.6 !important;">def</span> <span style="line-height:1.6 !important;">__init__</span><span style="line-height: 1.6 !important;">(self, inplanes, planes, stride=<span style="line-height:1.6 !important;">1</span>, downsample=None)</span>:</span><br style="line-height: 1.6 !important;"/> super(BasicBlock, self).__init__()<br style="line-height: 1.6 !important;"/> self.conv1 = conv3x3(inplanes, planes, stride)<br style="line-height: 1.6 !important;"/> self.bn1 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM)<br style="line-height: 1.6 !important;"/> self.relu = nn.ReLU(inplace=<span style="line-height:1.6 !important;">True</span>)<br style="line-height: 1.6 !important;"/> self.conv2 = conv3x3(planes, planes)<br style="line-height: 1.6 !important;"/> self.bn2 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM)<br style="line-height: 1.6 !important;"/> self.downsample = downsample<br style="line-height: 1.6 !important;"/> self.stride = stride<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height: 1.6 !important;"><span style="line-height:1.6 !important;">def</span> <span style="line-height:1.6 !important;">forward</span><span style="line-height: 1.6 !important;">(self, x)</span>:</span><br style="line-height: 1.6 !important;"/> residual = x<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> out = self.conv1(x)<br style="line-height: 1.6 !important;"/> out = self.bn1(out)<br style="line-height: 1.6 !important;"/> out = self.relu(out)<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> out = self.conv2(out)<br style="line-height: 1.6 !important;"/> out = self.bn2(out)<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">if</span> self.downsample <span style="line-height:1.6 !important;">is</span> <span style="line-height:1.6 !important;">not</span> <span style="line-height:1.6 !important;">None</span>:<br style="line-height: 1.6 !important;"/> residual = self.downsample(x)<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> out += residual<br style="line-height: 1.6 !important;"/> out = self.relu(out)<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">return</span> out<br style="line-height: 1.6 !important;"/></pre>
Bottleneck:
<figure>
<img src="https://img1.tuicool.com/367B3um.png!web" class="alignCenter" referrerpolicy="no-referrer"/>
<figcaption>Bottleneck结构图</figcaption>
</figure>
<pre class="prettyprint"><span style="line-height: 1.6 !important;"><span style="line-height:1.6 !important;">class</span> <span style="line-height:1.6 !important;">Bottleneck</span><span style="line-height: 1.6 !important;">(nn.Module)</span>:</span><br style="line-height: 1.6 !important;"/> expansion = <span style="line-height:1.6 !important;">4</span><br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height: 1.6 !important;"><span style="line-height:1.6 !important;">def</span> <span style="line-height:1.6 !important;">__init__</span><span style="line-height: 1.6 !important;">(self, inplanes, planes, stride=<span style="line-height:1.6 !important;">1</span>, downsample=None)</span>:</span><br style="line-height: 1.6 !important;"/> super(Bottleneck, self).__init__()<br style="line-height: 1.6 !important;"/> self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=<span style="line-height:1.6 !important;">1</span>, bias=<span style="line-height:1.6 !important;">False</span>)<br style="line-height: 1.6 !important;"/> self.bn1 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM)<br style="line-height: 1.6 !important;"/> self.conv2 = nn.Conv2d(planes, planes, kernel_size=<span style="line-height:1.6 !important;">3</span>, stride=stride,<br style="line-height: 1.6 !important;"/> padding=<span style="line-height:1.6 !important;">1</span>, bias=<span style="line-height:1.6 !important;">False</span>)<br style="line-height: 1.6 !important;"/> self.bn2 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM)<br style="line-height: 1.6 !important;"/> self.conv3 = nn.Conv2d(planes, planes * self.expansion, kernel_size=<span style="line-height:1.6 !important;">1</span>,<br style="line-height: 1.6 !important;"/> bias=<span style="line-height:1.6 !important;">False</span>)<br style="line-height: 1.6 !important;"/> self.bn3 = nn.BatchNorm2d(planes * self.expansion,<br style="line-height: 1.6 !important;"/> momentum=BN_MOMENTUM)<br style="line-height: 1.6 !important;"/> self.relu = nn.ReLU(inplace=<span style="line-height:1.6 !important;">True</span>)<br style="line-height: 1.6 !important;"/> self.downsample = downsample<br style="line-height: 1.6 !important;"/> self.stride = stride<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height: 1.6 !important;"><span style="line-height:1.6 !important;">def</span> <span style="line-height:1.6 !important;">forward</span><span style="line-height: 1.6 !important;">(self, x)</span>:</span><br style="line-height: 1.6 !important;"/> residual = x<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> out = self.conv1(x)<br style="line-height: 1.6 !important;"/> out = self.bn1(out)<br style="line-height: 1.6 !important;"/> out = self.relu(out)<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> out = self.conv2(out)<br style="line-height: 1.6 !important;"/> out = self.bn2(out)<br style="line-height: 1.6 !important;"/> out = self.relu(out)<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> out = self.conv3(out)<br style="line-height: 1.6 !important;"/> out = self.bn3(out)<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">if</span> self.downsample <span style="line-height:1.6 !important;">is</span> <span style="line-height:1.6 !important;">not</span> <span style="line-height:1.6 !important;">None</span>:<br style="line-height: 1.6 !important;"/> residual = self.downsample(x)<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> out += residual<br style="line-height: 1.6 !important;"/> out = self.relu(out)<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">return</span> out<br style="line-height: 1.6 !important;"/></pre>
HighResolutionModule,这是核心模块, 主要分为两个组件：branches和fuse layer。
<figure>
<img src="https://img1.tuicool.com/ZfeU3ej.jpg!web" class="alignCenter" referrerpolicy="no-referrer"/>
</figure>
<pre class="prettyprint"><span style="line-height: 1.6 !important;"><span style="line-height:1.6 !important;">class</span> <span style="line-height:1.6 !important;">HighResolutionModule</span><span style="line-height: 1.6 !important;">(nn.Module)</span>:</span><br style="line-height: 1.6 !important;"/> <span style="line-height: 1.6 !important;"><span style="line-height:1.6 !important;">def</span> <span style="line-height:1.6 !important;">__init__</span><span style="line-height: 1.6 !important;">(self, num_branches, blocks, num_blocks, num_inchannels,<br style="line-height: 1.6 !important;"/> num_channels, fuse_method, multi_scale_output=True)</span>:</span><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">'''<br style="line-height: 1.6 !important;"/> 调用：<br style="line-height: 1.6 !important;"/> # 调用高低分辨率交互模块， stage2 为例<br style="line-height: 1.6 !important;"/> HighResolutionModule(num_branches, # 2<br style="line-height: 1.6 !important;"/> block, # 'BASIC'<br style="line-height: 1.6 !important;"/> num_blocks, # [4, 4]<br style="line-height: 1.6 !important;"/> num_inchannels, # 上个stage的out channel<br style="line-height: 1.6 !important;"/> num_channels, # [32, 64]<br style="line-height: 1.6 !important;"/> fuse_method, # SUM<br style="line-height: 1.6 !important;"/> reset_multi_scale_output)<br style="line-height: 1.6 !important;"/> '''</span><br style="line-height: 1.6 !important;"/> super(HighResolutionModule, self).__init__()<br style="line-height: 1.6 !important;"/> self._check_branches(<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;font-style:italic;"># 检查分支数目是否合理</span><br style="line-height: 1.6 !important;"/> num_branches, blocks, num_blocks, num_inchannels, num_channels)<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> self.num_inchannels = num_inchannels<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;font-style:italic;"># 融合选用相加的方式</span><br style="line-height: 1.6 !important;"/> self.fuse_method = fuse_method<br style="line-height: 1.6 !important;"/> self.num_branches = num_branches<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> self.multi_scale_output = multi_scale_output<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;font-style:italic;"># 两个核心部分，一个是branches构建，一个是融合layers构建</span><br style="line-height: 1.6 !important;"/> self.branches = self._make_branches(<br style="line-height: 1.6 !important;"/> num_branches, blocks, num_blocks, num_channels)<br style="line-height: 1.6 !important;"/> self.fuse_layers = self._make_fuse_layers()<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> self.relu = nn.ReLU(<span style="line-height:1.6 !important;">False</span>)<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height: 1.6 !important;"><span style="line-height:1.6 !important;">def</span> <span style="line-height:1.6 !important;">_check_branches</span><span style="line-height: 1.6 !important;">(self, num_branches, blocks, num_blocks,<br style="line-height: 1.6 !important;"/> num_inchannels, num_channels)</span>:</span><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;font-style:italic;"># 分别检查参数是否符合要求,看models.py中的参数，blocks参数冗余了</span><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">if</span> num_branches != len(num_blocks):<br style="line-height: 1.6 !important;"/> error_msg = <span style="line-height:1.6 !important;">'NUM_BRANCHES({}) <> NUM_BLOCKS({})'</span>.format(<br style="line-height: 1.6 !important;"/> num_branches, len(num_blocks))<br style="line-height: 1.6 !important;"/> logger.error(error_msg)<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">raise</span> ValueError(error_msg)<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">if</span> num_branches != len(num_channels):<br style="line-height: 1.6 !important;"/> error_msg = <span style="line-height:1.6 !important;">'NUM_BRANCHES({}) <> NUM_CHANNELS({})'</span>.format(<br style="line-height: 1.6 !important;"/> num_branches, len(num_channels))<br style="line-height: 1.6 !important;"/> logger.error(error_msg)<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">raise</span> ValueError(error_msg)<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">if</span> num_branches != len(num_inchannels):<br style="line-height: 1.6 !important;"/> error_msg = <span style="line-height:1.6 !important;">'NUM_BRANCHES({}) <> NUM_INCHANNELS({})'</span>.format(<br style="line-height: 1.6 !important;"/> num_branches, len(num_inchannels))<br style="line-height: 1.6 !important;"/> logger.error(error_msg)<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">raise</span> ValueError(error_msg)<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height: 1.6 !important;"><span style="line-height:1.6 !important;">def</span> <span style="line-height:1.6 !important;">_make_one_branch</span><span style="line-height: 1.6 !important;">(self, branch_index, block, num_blocks, num_channels,<br style="line-height: 1.6 !important;"/> stride=<span style="line-height:1.6 !important;">1</span>)</span>:</span><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;font-style:italic;"># 构建一个分支，一个分支重复num_blocks个block</span><br style="line-height: 1.6 !important;"/> downsample = <span style="line-height:1.6 !important;">None</span><br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;font-style:italic;"># 这里判断，如果通道变大(分辨率变小)，则使用下采样</span><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">if</span> stride != <span style="line-height:1.6 !important;">1</span> <span style="line-height:1.6 !important;">or</span> \<br style="line-height: 1.6 !important;"/> self.num_inchannels[branch_index] != num_channels[branch_index] * block.expansion:<br style="line-height: 1.6 !important;"/> downsample = nn.Sequential(<br style="line-height: 1.6 !important;"/> nn.Conv2d(self.num_inchannels[branch_index],<br style="line-height: 1.6 !important;"/> num_channels[branch_index] * block.expansion,<br style="line-height: 1.6 !important;"/> kernel_size=<span style="line-height:1.6 !important;">1</span>, stride=stride, bias=<span style="line-height:1.6 !important;">False</span>),<br style="line-height: 1.6 !important;"/> nn.BatchNorm2d(num_channels[branch_index] * block.expansion,<br style="line-height: 1.6 !important;"/> momentum=BN_MOMENTUM),<br style="line-height: 1.6 !important;"/> )<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> layers = []<br style="line-height: 1.6 !important;"/> layers.append(block(self.num_inchannels[branch_index],<br style="line-height: 1.6 !important;"/> num_channels[branch_index], stride, downsample))<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> self.num_inchannels[branch_index] = \<br style="line-height: 1.6 !important;"/> num_channels[branch_index] * block.expansion<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">for</span> i <span style="line-height:1.6 !important;">in</span> range(<span style="line-height:1.6 !important;">1</span>, num_blocks[branch_index]):<br style="line-height: 1.6 !important;"/> layers.append(block(self.num_inchannels[branch_index],<br style="line-height: 1.6 !important;"/> num_channels[branch_index]))<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">return</span> nn.Sequential(*layers)<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height: 1.6 !important;"><span style="line-height:1.6 !important;">def</span> <span style="line-height:1.6 !important;">_make_branches</span><span style="line-height: 1.6 !important;">(self, num_branches, block, num_blocks, num_channels)</span>:</span><br style="line-height: 1.6 !important;"/> branches = []<br style="line-height: 1.6 !important;"/> <br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;font-style:italic;"># 通过循环构建多分支，每个分支属于不同的分辨率</span><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">for</span> i <span style="line-height:1.6 !important;">in</span> range(num_branches):<br style="line-height: 1.6 !important;"/> branches.append(<br style="line-height: 1.6 !important;"/> self._make_one_branch(i, block, num_blocks, num_channels))<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">return</span> nn.ModuleList(branches)<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height: 1.6 !important;"><span style="line-height:1.6 !important;">def</span> <span style="line-height:1.6 !important;">_make_fuse_layers</span><span style="line-height: 1.6 !important;">(self)</span>:</span><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">if</span> self.num_branches == <span style="line-height:1.6 !important;">1</span>:<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">return</span> <span style="line-height:1.6 !important;">None</span><br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> num_branches = self.num_branches <span style="line-height:1.6 !important;font-style:italic;"># 2</span><br style="line-height: 1.6 !important;"/> num_inchannels = self.num_inchannels<br style="line-height: 1.6 !important;"/> fuse_layers = []<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">for</span> i <span style="line-height:1.6 !important;">in</span> range(num_branches <span style="line-height:1.6 !important;">if</span> self.multi_scale_output <span style="line-height:1.6 !important;">else</span> <span style="line-height:1.6 !important;">1</span>):<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;font-style:italic;"># i代表枚举所有分支</span><br style="line-height: 1.6 !important;"/> fuse_layer = []<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">for</span> j <span style="line-height:1.6 !important;">in</span> range(num_branches):<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;font-style:italic;"># j代表处理的当前分支</span><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">if</span> j > i: <span style="line-height:1.6 !important;font-style:italic;"># 进行上采样，使用最近邻插值</span><br style="line-height: 1.6 !important;"/> fuse_layer.append(nn.Sequential(<br style="line-height: 1.6 !important;"/> nn.Conv2d(num_inchannels[j],<br style="line-height: 1.6 !important;"/> num_inchannels[i],<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">1</span>,<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">1</span>,<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">0</span>,<br style="line-height: 1.6 !important;"/> bias=<span style="line-height:1.6 !important;">False</span>),<br style="line-height: 1.6 !important;"/> nn.BatchNorm2d(num_inchannels[i],<br style="line-height: 1.6 !important;"/> momentum=BN_MOMENTUM),<br style="line-height: 1.6 !important;"/> nn.Upsample(scale_factor=<span style="line-height:1.6 !important;">2</span>**(j-i), mode=<span style="line-height:1.6 !important;">'nearest'</span>)))<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">elif</span> j == i:<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;font-style:italic;"># 本层不做处理</span><br style="line-height: 1.6 !important;"/> fuse_layer.append(<span style="line-height:1.6 !important;">None</span>)<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">else</span>:<br style="line-height: 1.6 !important;"/> conv3x3s = []<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;font-style:italic;"># 进行strided 3x3 conv下采样,如果跨两层，就使用两次strided 3x3 conv</span><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">for</span> k <span style="line-height:1.6 !important;">in</span> range(i-j):<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">if</span> k == i - j - <span style="line-height:1.6 !important;">1</span>:<br style="line-height: 1.6 !important;"/> num_outchannels_conv3x3 = num_inchannels[i]<br style="line-height: 1.6 !important;"/> conv3x3s.append(nn.Sequential(<br style="line-height: 1.6 !important;"/> nn.Conv2d(num_inchannels[j],<br style="line-height: 1.6 !important;"/> num_outchannels_conv3x3,<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">3</span>, <span style="line-height:1.6 !important;">2</span>, <span style="line-height:1.6 !important;">1</span>, bias=<span style="line-height:1.6 !important;">False</span>),<br style="line-height: 1.6 !important;"/> nn.BatchNorm2d(num_outchannels_conv3x3,<br style="line-height: 1.6 !important;"/> momentum=BN_MOMENTUM)))<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">else</span>:<br style="line-height: 1.6 !important;"/> num_outchannels_conv3x3 = num_inchannels[j]<br style="line-height: 1.6 !important;"/> conv3x3s.append(nn.Sequential(<br style="line-height: 1.6 !important;"/> nn.Conv2d(num_inchannels[j],<br style="line-height: 1.6 !important;"/> num_outchannels_conv3x3,<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">3</span>, <span style="line-height:1.6 !important;">2</span>, <span style="line-height:1.6 !important;">1</span>, bias=<span style="line-height:1.6 !important;">False</span>),<br style="line-height: 1.6 !important;"/> nn.BatchNorm2d(num_outchannels_conv3x3,<br style="line-height: 1.6 !important;"/> nn.ReLU(<span style="line-height:1.6 !important;">False</span>)))<br style="line-height: 1.6 !important;"/> fuse_layer.append(nn.Sequential(*conv3x3s))<br style="line-height: 1.6 !important;"/> fuse_layers.append(nn.ModuleList(fuse_layer))<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">return</span> nn.ModuleList(fuse_layers)<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height: 1.6 !important;"><span style="line-height:1.6 !important;">def</span> <span style="line-height:1.6 !important;">get_num_inchannels</span><span style="line-height: 1.6 !important;">(self)</span>:</span><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">return</span> self.num_inchannels<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height: 1.6 !important;"><span style="line-height:1.6 !important;">def</span> <span style="line-height:1.6 !important;">forward</span><span style="line-height: 1.6 !important;">(self, x)</span>:</span><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">if</span> self.num_branches == <span style="line-height:1.6 !important;">1</span>:<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">return</span> [self.branches[<span style="line-height:1.6 !important;">0</span>](x[<span style="line-height:1.6 !important;">0</span>])]<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">for</span> i <span style="line-height:1.6 !important;">in</span> range(self.num_branches):<br style="line-height: 1.6 !important;"/> x[i]=self.branches[i](x[i])<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> x_fuse=[]<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">for</span> i <span style="line-height:1.6 !important;">in</span> range(len(self.fuse_layers)):<br style="line-height: 1.6 !important;"/> y=x[<span style="line-height:1.6 !important;">0</span>] <span style="line-height:1.6 !important;">if</span> i == <span style="line-height:1.6 !important;">0</span> <span style="line-height:1.6 !important;">else</span> self.fuse_layers[i][<span style="line-height:1.6 !important;">0</span>](x[<span style="line-height:1.6 !important;">0</span>])<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">for</span> j <span style="line-height:1.6 !important;">in</span> range(<span style="line-height:1.6 !important;">1</span>, self.num_branches):<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">if</span> i == j:<br style="line-height: 1.6 !important;"/> y=y + x[j]<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">else</span>:<br style="line-height: 1.6 !important;"/> y=y + self.fuse_layers[i][j](x[j])<br style="line-height: 1.6 !important;"/> x_fuse.append(self.relu(y))<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;font-style:italic;"># 将fuse以后的多个分支结果保存到list中</span><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">return</span> x_fuse<br style="line-height: 1.6 !important;"/></pre>
models.py中保存的参数, 可以通过这些配置来改变模型的容量、分支个数、特征融合方法：
<pre class="prettyprint"><span style="line-height:1.6 !important;font-style:italic;"># high_resoluton_net related params for classification</span><br style="line-height: 1.6 !important;"/>POSE_HIGH_RESOLUTION_NET = CN()<br style="line-height: 1.6 !important;"/>POSE_HIGH_RESOLUTION_NET.PRETRAINED_LAYERS = [<span style="line-height:1.6 !important;">'*'</span>]<br style="line-height: 1.6 !important;"/>POSE_HIGH_RESOLUTION_NET.STEM_INPLANES = <span style="line-height:1.6 !important;">64</span><br style="line-height: 1.6 !important;"/>POSE_HIGH_RESOLUTION_NET.FINAL_CONV_KERNEL = <span style="line-height:1.6 !important;">1</span><br style="line-height: 1.6 !important;"/>POSE_HIGH_RESOLUTION_NET.WITH_HEAD = <span style="line-height:1.6 !important;">True</span><br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/>POSE_HIGH_RESOLUTION_NET.STAGE2 = CN()<br style="line-height: 1.6 !important;"/>POSE_HIGH_RESOLUTION_NET.STAGE2.NUM_MODULES = <span style="line-height:1.6 !important;">1</span><br style="line-height: 1.6 !important;"/>POSE_HIGH_RESOLUTION_NET.STAGE2.NUM_BRANCHES = <span style="line-height:1.6 !important;">2</span><br style="line-height: 1.6 !important;"/>POSE_HIGH_RESOLUTION_NET.STAGE2.NUM_BLOCKS = [<span style="line-height:1.6 !important;">4</span>, <span style="line-height:1.6 !important;">4</span>]<br style="line-height: 1.6 !important;"/>POSE_HIGH_RESOLUTION_NET.STAGE2.NUM_CHANNELS = [<span style="line-height:1.6 !important;">32</span>, <span style="line-height:1.6 !important;">64</span>]<br style="line-height: 1.6 !important;"/>POSE_HIGH_RESOLUTION_NET.STAGE2.BLOCK = <span style="line-height:1.6 !important;">'BASIC'</span><br style="line-height: 1.6 !important;"/>POSE_HIGH_RESOLUTION_NET.STAGE2.FUSE_METHOD = <span style="line-height:1.6 !important;">'SUM'</span><br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/>POSE_HIGH_RESOLUTION_NET.STAGE3 = CN()<br style="line-height: 1.6 !important;"/>POSE_HIGH_RESOLUTION_NET.STAGE3.NUM_MODULES = <span style="line-height:1.6 !important;">1</span><br style="line-height: 1.6 !important;"/>POSE_HIGH_RESOLUTION_NET.STAGE3.NUM_BRANCHES = <span style="line-height:1.6 !important;">3</span><br style="line-height: 1.6 !important;"/>POSE_HIGH_RESOLUTION_NET.STAGE3.NUM_BLOCKS = [<span style="line-height:1.6 !important;">4</span>, <span style="line-height:1.6 !important;">4</span>, <span style="line-height:1.6 !important;">4</span>]<br style="line-height: 1.6 !important;"/>POSE_HIGH_RESOLUTION_NET.STAGE3.NUM_CHANNELS = [<span style="line-height:1.6 !important;">32</span>, <span style="line-height:1.6 !important;">64</span>, <span style="line-height:1.6 !important;">128</span>]<br style="line-height: 1.6 !important;"/>POSE_HIGH_RESOLUTION_NET.STAGE3.BLOCK = <span style="line-height:1.6 !important;">'BASIC'</span><br style="line-height: 1.6 !important;"/>POSE_HIGH_RESOLUTION_NET.STAGE3.FUSE_METHOD = <span style="line-height:1.6 !important;">'SUM'</span><br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/>POSE_HIGH_RESOLUTION_NET.STAGE4 = CN()<br style="line-height: 1.6 !important;"/>POSE_HIGH_RESOLUTION_NET.STAGE4.NUM_MODULES = <span style="line-height:1.6 !important;">1</span><br style="line-height: 1.6 !important;"/>POSE_HIGH_RESOLUTION_NET.STAGE4.NUM_BRANCHES = <span style="line-height:1.6 !important;">4</span><br style="line-height: 1.6 !important;"/>POSE_HIGH_RESOLUTION_NET.STAGE4.NUM_BLOCKS = [<span style="line-height:1.6 !important;">4</span>, <span style="line-height:1.6 !important;">4</span>, <span style="line-height:1.6 !important;">4</span>, <span style="line-height:1.6 !important;">4</span>]<br style="line-height: 1.6 !important;"/>POSE_HIGH_RESOLUTION_NET.STAGE4.NUM_CHANNELS = [<span style="line-height:1.6 !important;">32</span>, <span style="line-height:1.6 !important;">64</span>, <span style="line-height:1.6 !important;">128</span>, <span style="line-height:1.6 !important;">256</span>]<br style="line-height: 1.6 !important;"/>POSE_HIGH_RESOLUTION_NET.STAGE4.BLOCK = <span style="line-height:1.6 !important;">'BASIC'</span><br style="line-height: 1.6 !important;"/>POSE_HIGH_RESOLUTION_NET.STAGE4.FUSE_METHOD = <span style="line-height:1.6 !important;">'SUM'</span><br style="line-height: 1.6 !important;"/></pre>
然后来看整个HRNet模型的构建, 由于整体代码量太大，这里仅仅来看forward函数。
<pre class="prettyprint"><span style="line-height: 1.6 !important;"><span style="line-height:1.6 !important;">def</span> <span style="line-height:1.6 !important;">forward</span><span style="line-height: 1.6 !important;">(self, x)</span>:</span><br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;font-style:italic;"># 使用两个strided 3x3conv进行快速降维</span><br style="line-height: 1.6 !important;"/> x=self.relu(self.bn1(self.conv1(x)))<br style="line-height: 1.6 !important;"/> x=self.relu(self.bn2(self.conv2(x)))<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;font-style:italic;"># 构建了一串BasicBlock构成的模块</span><br style="line-height: 1.6 !important;"/> x=self.layer1(x)<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;font-style:italic;"># 然后是多个stage，每个stage核心是调用HighResolutionModule模块</span><br style="line-height: 1.6 !important;"/> x_list=[]<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">for</span> i <span style="line-height:1.6 !important;">in</span> range(self.stage2_cfg[<span style="line-height:1.6 !important;">'NUM_BRANCHES'</span>]):<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">if</span> self.transition1[i] <span style="line-height:1.6 !important;">is</span> <span style="line-height:1.6 !important;">not</span> <span style="line-height:1.6 !important;">None</span>:<br style="line-height: 1.6 !important;"/> x_list.append(self.transition1[i](x))<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">else</span>:<br style="line-height: 1.6 !important;"/> x_list.append(x)<br style="line-height: 1.6 !important;"/> y_list=self.stage2(x_list)<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> x_list=[]<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">for</span> i <span style="line-height:1.6 !important;">in</span> range(self.stage3_cfg[<span style="line-height:1.6 !important;">'NUM_BRANCHES'</span>]):<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">if</span> self.transition2[i] <span style="line-height:1.6 !important;">is</span> <span style="line-height:1.6 !important;">not</span> <span style="line-height:1.6 !important;">None</span>:<br style="line-height: 1.6 !important;"/> x_list.append(self.transition2[i](y_list[<span style="line-height:1.6 !important;">-1</span>]))<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">else</span>:<br style="line-height: 1.6 !important;"/> x_list.append(y_list[i])<br style="line-height: 1.6 !important;"/> y_list=self.stage3(x_list)<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> x_list=[]<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">for</span> i <span style="line-height:1.6 !important;">in</span> range(self.stage4_cfg[<span style="line-height:1.6 !important;">'NUM_BRANCHES'</span>]):<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">if</span> self.transition3[i] <span style="line-height:1.6 !important;">is</span> <span style="line-height:1.6 !important;">not</span> <span style="line-height:1.6 !important;">None</span>:<br style="line-height: 1.6 !important;"/> x_list.append(self.transition3[i](y_list[<span style="line-height:1.6 !important;">-1</span>]))<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">else</span>:<br style="line-height: 1.6 !important;"/> x_list.append(y_list[i])<br style="line-height: 1.6 !important;"/> y_list=self.stage4(x_list)<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;font-style:italic;"># 添加分类头，上文中有显示，在分类问题中添加这种头</span><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;font-style:italic;"># 在其他问题中换用不同的头</span><br style="line-height: 1.6 !important;"/> y=self.incre_modules[<span style="line-height:1.6 !important;">0</span>](y_list[<span style="line-height:1.6 !important;">0</span>])<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">for</span> i <span style="line-height:1.6 !important;">in</span> range(len(self.downsamp_modules)):<br style="line-height: 1.6 !important;"/> y=self.incre_modules[i+<span style="line-height:1.6 !important;">1</span>](y_list[i+<span style="line-height:1.6 !important;">1</span>]) + \<br style="line-height: 1.6 !important;"/> self.downsamp_modules[i](y)<br style="line-height: 1.6 !important;"/> y=self.final_layer(y)<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">if</span> torch._C._get_tracing_state():<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;font-style:italic;"># 在不写C代码的情况下执行forward，直接用python版本</span><br style="line-height: 1.6 !important;"/> y=y.flatten(start_dim=<span style="line-height:1.6 !important;">2</span>).mean(dim=<span style="line-height:1.6 !important;">2</span>)<br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">else</span>:<br style="line-height: 1.6 !important;"/> y=F.avg_pool2d(y, kernel_size=y.size()<br style="line-height: 1.6 !important;"/> [<span style="line-height:1.6 !important;">2</span>:]).view(y.size(<span style="line-height:1.6 !important;">0</span>), <span style="line-height:1.6 !important;">-1</span>)<br style="line-height: 1.6 !important;"/> y=self.classifier(y)<br style="line-height: 1.6 !important;"/><br style="line-height: 1.6 !important;"/> <span style="line-height:1.6 !important;">return</span> y<br style="line-height: 1.6 !important;"/></pre>
<h4>总结</h4>
HRNet核心方法是：在模型的整个过程中，保存高分辨率表征的同时使用让不同分辨率的feature map进行特征交互。
HRNet在非常多的CV领域有广泛的应用，比如ICCV2019的东北虎关键点识别比赛中，HRNet就起到了一定的作用。并且在分类部分的实验证明了在同等参数量的情况下，可以取代ResNet进行分类。
senet是hrnet的一个特例，hrnet不仅有通道注意力，同时也有空间注意力 -- akkaze-郑安坤
<figure>
<img src="https://img0.tuicool.com/jE7vuyB.png!web" class="alignCenter" referrerpolicy="no-referrer"/>
<figcaption>SELayer核心实现</figcaption>
</figure>
SELayer首先通过一个全局平均池化得到一个一维向量，然后通过两个全连接层，将信息进行压缩和扩展，通过sigmoid以后得到每个通道的权值，然后用这个权值与原来的feature map相乘，进行信息上的优化。
<figure>
<img src="https://img2.tuicool.com/j6ziErV.png!web" class="alignCenter" referrerpolicy="no-referrer"/>
<figcaption>HRNet一个结构</figcaption>
</figure>
可以看到上图用红色箭头串起来的是不是和SELayer很相似。为什么说SENet是HRNet的一个特例，但从这个结构来讲，可以这么看：
<ul>
<li>
<p class="no-text-indent">SENet没有像HRNet这样分辨率变为原来的一半，分辨率直接变为1x1，比较极端。变为1x1向量以后，SENet中使用了两个全连接网络来学习通道的特征分布；但是在HRNet中，使用了几个卷积(Residual block)来学习特征。</p>
</li>
<li>
<p class="no-text-indent">SENet在主干部分(高分辨率分支)没有安排卷积进行特征的学习；HRNet在主干部分(高分辨率分支)安排了几个卷积(Residual block)来学习特征。</p>
</li>
<li>
<p class="no-text-indent">特征融合部分SENet和HRNet区分比较大，SENet使用的对应通道相乘的方法，HRNet则使用的是相加。之所以说SENet是通道注意力机制是因为通过全局平均池化后没有了空间特征，只剩通道的特征；HRNet则可以看作同时保留了空间特征和通道特征，所以说HRNet不仅有通道注意力，同时也有空间注意力。</p>
</li>
</ul>
HRNet团队有10人之多，构建了分类、分割、检测、关键点检测等库，工作量非常大，而且做了很多扎实的实验证明了这种思路的有效性。所以是否可以认为HRNet属于SENet之后又一个更优的backbone呢？还需要自己实践中使用这种想法和思路来验证。
<h4>参考</h4>
https://arxiv.org/pdf/1908.07919
https://www.bilibili.com/video/BV1WJ41197dh?t=508
https://github.com/HRNet
</div>
- End -
<p style="text-align:center">
<img src="https://img0.tuicool.com/6Fj63a6.jpg!web" class="alignCenter" referrerpolicy="no-referrer"/>
</p>
<mpcps frameborder="0"/>
<span>京东满100-50优惠活动等你来参与！</span>
<span>✄------------------------------------------------</span>
<span>看到这里，说明你喜欢这篇文章，请点击「</span>
<span>在看</span>
<span>」或顺手「</span>
<span>转发</span>
<span>」「</span>
<span>点赞</span>
<span>」。</span>
<span>欢迎微信搜索「</span>
<span>panchuangxx</span>
<span>」，添加小编</span>
<span>磐小小仙</span>
<span>微信，每日朋友圈更新一篇高质量推文（无广告），为您提供更多精彩内容。</span>
<span>▼ </span>
<span>▼</span>
<span> </span>
<span> 扫描二维码添加小编</span>
<span> ▼</span>
<span>
<span> </span>
<p style="text-align:center">
<img src="https://img0.tuicool.com/ym2Yny6.jpg!web" class="alignCenter" referrerpolicy="no-referrer"/>
</p>
