cvpr2019 微软亚洲研究院的工作，主要思想是构建一个并行的多分辨率网络（有的应用只使用高分辨率特征，因此得名），这是一个系列的工作，包括分类，检测，分割等。终于来了一个非nas的网络。。。
项目地址 https://github.com/HRNet
知乎上的介绍 https://zhuanlan.zhihu.com/p/66848624
一篇讲的比较清楚的博客 https://ai-chen.github.io/%E8%AF%AD%E4%B9%89%E5%88%86%E5%89%B2%E8%AE%BA%E6%96%87%E9%98%85%E8%AF%BB/2019/06/20/HrNet-V2.html
虽然讲的是v2，其实跟v1几乎没有区别，分类模型只是差了一个head。
分类模型的论文https://arxiv.org/pdf/1904.04514.pdf
下面单独补充下分类的head(可参考下结构 https://github.com/HRNet/HRNet-Image-Classification/blob/master/figures/cls-head.png)
分为三个部分，incre_modules用于分别增加三个分支的通道数目
down_modules用于降低三个分支的通道数目以及分辨率，便于进行融合，输出一个1024维的特征
final_layer通过一个1*1输出2048维的特征，用于最后的类别输出（2048×num_cls）
def _make_head(self, pre_stage_channels): head_block = Bottleneck head_channels = [32, 64, 128, 256] # Increasing the #channels on each resolution # from C, 2C, 4C, 8C to 128, 256, 512, 1024 incre_modules = [] for i, channels in enumerate(pre_stage_channels): incre_module = self._make_layer(head_block, channels, head_channels[i], 1, stride=1) incre_modules.append(incre_module) incre_modules = nn.ModuleList(incre_modules) # downsampling modules downsamp_modules = [] for i in range(len(pre_stage_channels)-1): in_channels = head_channels[i] * head_block.expansion out_channels = head_channels[i+1] * head_block.expansion downsamp_module = nn.Sequential( nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=2, padding=1), nn.BatchNorm2d(out_channels, momentum=BN_MOMENTUM), nn.ReLU(inplace=True) ) downsamp_modules.append(downsamp_module) downsamp_modules = nn.ModuleList(downsamp_modules) final_layer = nn.Sequential( nn.Conv2d( in_channels=head_channels[3] * head_block.expansion, out_channels=2048, kernel_size=1, stride=1, padding=0 ), nn.BatchNorm2d(2048, momentum=BN_MOMENTUM), nn.ReLU(inplace=True) ) return incre_modules, downsamp_modules, final_layer