max pooling更注重重要的局部特征
average pooling更关注全局特征
两者concat可以丰富特征层
from torchvision.models import resnet18
import torch.nn as nn
import torch
class res18(nn.Module):
def __init__(self, num_classes):
super(res18, self).__init__()
self.base = resnet18(pretrained=False)
self.feature = nn.Sequential(
self.base.conv1,
self.base.bn1,
self.base.relu,
self.base.maxpool,
self.base.layer1,
self.base.layer2,
self.base.layer3,
self.base.layer4 #输出512通道
)
self.avg_pool = nn.AdaptiveAvgPool2d((1,1)) #自适应平均池化(batch, 512, 1, 1)
self.max_pool = nn.AdaptiveMaxPool2d((1,1)) #自适应最大池化(batch, 512, 1, 1)
self.reduce_layer = nn.Conv2d(1024, 512, 1) #1*1卷积进行降维
self.fc = nn.Sequential(
nn.Dropout(0.5),
nn.Linear(1024, num_classes) #fc层
)
def forward(self, x):
bs = x.shape[0] #batch size
x = self.feature(x) # 输出512通道
avgpool_x = self.avg_pool(x) #输出(batch, 512, 1, 1)
maxpool_x = self.max_pool(x) #输出(batch,512, 1, 1)
x = torch.cat([avgpool_x, maxpool_x], dim=1) #输出(batch, 1024, 1, 1)
x = self.reduce_layer(x).view(bs, -1) #输出[batch, 512])
logits = self.fc(x) #输出（batch，num_classes)
return logits
