该模块融合了4种不同金字塔尺度的特征，第一行红色是最粗糙的特征–全局池化生成单个bin输出，后面三行是不同尺度的池化特征。
为了保证全局特征的权重，如果金字塔共有N个级别，则在每个级别后使用1×1 1×11×1的卷积将对于级别通道降为原本的1/N。再通过双线性插值获得未池化前的大小，最终concat到一起。
1 import torch
import torch.nn.functional as F
from torch import nn
from torchvision import models from utils import initialize_weights
from utils.misc import Conv2dDeformable
from .config import res101_path //金字塔模块,将从前面卷积结构提取的特征分别进行不同的池化操作,得到不同感受野以及全局语境信息(或者叫做不同层级的信息)
class _PyramidPoolingModule(nn.Module):
def __init__(self, in_dim, reduction_dim, setting):
super(_PyramidPoolingModule, self).__init__()
self.features = []
for s in setting: //对应不同的池化操作,单个bin,多个bin
self.features.append(nn.Sequential(
nn.AdaptiveAvgPool2d(s),
nn.Conv2d(in_dim, reduction_dim, kernel_size=1, bias=False),
nn.BatchNorm2d(reduction_dim, momentum=.95),
nn.ReLU(inplace=True)
))
self.features = nn.ModuleList(self.features) def forward(self, x):
x_size = x.size()
out = [x]
for f in self.features:
out.append(F.upsample(f(x), x_size[2:], mode='bilinear'))
out = torch.cat(out, 1)
return out //整个pspnet网络的结构
class PSPNet(nn.Module):
def __init__(self, num_classes, pretrained=True, use_aux=True):
super(PSPNet, self).__init__()
self.use_aux = use_aux
resnet = models.resnet101() //采用resnet101作为骨干模型,提取特征
if pretrained:
resnet.load_state_dict(torch.load(res101_path))
self.layer0 = nn.Sequential(resnet.conv1, resnet.bn1, resnet.relu, resnet.maxpool)
self.layer1, self.layer2, self.layer3, self.layer4 = resnet.layer1, resnet.layer2, resnet.layer3, resnet.layer4
//设置带洞卷积的参数(dilation),以及卷积的参数
for n, m in self.layer3.named_modules():
if 'conv2' in n:
m.dilation, m.padding, m.stride = (2, 2), (2, 2), (1, 1)
elif 'downsample.0' in n:
m.stride = (1, 1)
for n, m in self.layer4.named_modules():
if 'conv2' in n:
m.dilation, m.padding, m.stride = (4, 4), (4, 4), (1, 1)
elif 'downsample.0' in n:
m.stride = (1, 1)
//加入ppm模块,以及最后的连接层(卷积)
self.ppm = _PyramidPoolingModule(2048, 512, (1, 2, 3, 6))
self.final = nn.Sequential(
nn.Conv2d(4096, 512, kernel_size=3, padding=1, bias=False),
nn.BatchNorm2d(512, momentum=.95),
nn.ReLU(inplace=True),
nn.Dropout(0.1),
nn.Conv2d(512, num_classes, kernel_size=1)
) if use_aux:
self.aux_logits = nn.Conv2d(1024, num_classes, kernel_size=1)
initialize_weights(self.aux_logits)
# 初始化权重
initialize_weights(self.ppm, self.final) def forward(self, x):
x_size = x.size()
x = self.layer0(x)
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
if self.training and self.use_aux:
aux = self.aux_logits(x)
x = self.layer4(x)
x = self.ppm(x)
x = self.final(x)
if self.training and self.use_aux:
return F.upsample(x, x_size[2:], mode='bilinear'), F.upsample(aux, x_size[2:], mode='bilinear')
return F.upsample(x, x_size[2:], mode='bilinear')