每日Attention学习1——Parallel Aggregation Pyramid Pooling Module

模块出处

[CVPR 23] [link] [code] PIDNet: A Real-time Semantic Segmentation Network Inspired by PID Controllers

---

模块名称

Parallel Aggregation Pyramid Pooling Module (PAPPM)

---

模块作用

多尺度特征提取，更大感受野

---

模块结构

---

模块代码

import torch
import torch.nn as nn
import torch.nn.functional as F

class PAPPM(nn.Module):
    def __init__(self, inplanes, branch_planes, outplanes, BatchNorm=nn.BatchNorm2d):
        super(PAPPM, self).__init__()
        bn_mom = 0.1
        
        # Avg 5,2 + Conv
        self.scale1 = nn.Sequential(nn.AvgPool2d(kernel_size=5, stride=2, padding=2),
                                    BatchNorm(inplanes, momentum=bn_mom),
                                    nn.ReLU(inplace=True),
                                    nn.Conv2d(inplanes, branch_planes, kernel_size=1, bias=False),
                                    )
        # Avg 9,4 + Conv
        self.scale2 = nn.Sequential(nn.AvgPool2d(kernel_size=9, stride=4, padding=4),
                                    BatchNorm(inplanes, momentum=bn_mom),
                                    nn.ReLU(inplace=True),
                                    nn.Conv2d(inplanes, branch_planes, kernel_size=1, bias=False),
                                    )
        # Avg 17,8 + Conv
        self.scale3 = nn.Sequential(nn.AvgPool2d(kernel_size=17, stride=8, padding=8),
                                    BatchNorm(inplanes, momentum=bn_mom),
                                    nn.ReLU(inplace=True),
                                    nn.Conv2d(inplanes, branch_planes, kernel_size=1, bias=False),
                                    )
        # Avg Global + Conv
        self.scale4 = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)),
                                    BatchNorm(inplanes, momentum=bn_mom),
                                    nn.ReLU(inplace=True),
                                    nn.Conv2d(inplanes, branch_planes, kernel_size=1, bias=False),
                                    )
        # Conv
        self.scale0 = nn.Sequential(
                                    BatchNorm(inplanes, momentum=bn_mom),
                                    nn.ReLU(inplace=True),
                                    nn.Conv2d(inplanes, branch_planes, kernel_size=1, bias=False),
                                    )
        
        self.scale_process = nn.Sequential(
                                    BatchNorm(branch_planes*4, momentum=bn_mom),
                                    nn.ReLU(inplace=True),
                                    nn.Conv2d(branch_planes*4, branch_planes*4, kernel_size=3, 
                                    padding=1, groups=4, bias=False),
                                    )

      
        self.compression = nn.Sequential(
                                    BatchNorm(branch_planes * 5, momentum=bn_mom),
                                    nn.ReLU(inplace=True),
                                    nn.Conv2d(branch_planes * 5, outplanes, kernel_size=1, bias=False),
                                    )
        
        self.shortcut = nn.Sequential(
                                    BatchNorm(inplanes, momentum=bn_mom),
                                    nn.ReLU(inplace=True),
                                    nn.Conv2d(inplanes, outplanes, kernel_size=1, bias=False),
                                    )


    def forward(self, x):
        algc = False
        width = x.shape[-1]
        height = x.shape[-2]        
        scale_list = []

        x_ = self.scale0(x)
        scale_list.append(F.interpolate(self.scale1(x), size=[height, width],
                        mode='bilinear', align_corners=algc)+x_)
        scale_list.append(F.interpolate(self.scale2(x), size=[height, width],
                        mode='bilinear', align_corners=algc)+x_)
        scale_list.append(F.interpolate(self.scale3(x), size=[height, width],
                        mode='bilinear', align_corners=algc)+x_)
        scale_list.append(F.interpolate(self.scale4(x), size=[height, width],
                        mode='bilinear', align_corners=algc)+x_)
        
        scale_out = self.scale_process(torch.cat(scale_list, 1))
       
        out = self.compression(torch.cat([x_,scale_out], 1)) + self.shortcut(x)
        return out
    
if __name__ == '__main__':
    pappm = PAPPM(inplanes=1024, branch_planes=96, outplanes=256)
    x = torch.randn([3, 1024, 8, 8])
    out = pappm(x)
    print(out.shape)  # 3, 256, 8, 8

---

原文表述

为了更好地构建全局场景先验，PSPNet引入了金字塔池化模块（PPM），在卷积层之前串联多尺度池化映射，形成局部和全局上下文表示。文献[20]提出的深度聚合PPM（DAPPM）进一步提高了PPM的上下文嵌入能力，并显示出卓越的性能。然而，DAPPM的计算过程无法就其深度进行并行化，耗时较长，而且DAPPM每个尺度包含的通道过多，可能会超出轻量级模型的表示能力。因此，我们修改了DAPPM中的连接，使其可并行化，如图6所示，并将每个尺度的通道数从128个减少到96个。这种新的上下文采集模块被称为并行聚合PPM（PAPPM），并应用于PIDNet-M和PIDNet-S，以保证它们的速度。注意，对于更大的变体PIDNet-L，考虑到其深度，我们选择DAPPM。