deeplab中 global average pooling(GAP) 原理
之前有疑惑deeplab中ASPP(Atrous Spatial Pyramid Pooling) 采用多尺度空洞卷积＋GAP并联提取多尺度语义信息时,GAP得到的结果是一维的,而空洞卷积得到的结果时二维的,不知道其中的细节是怎么操作从而实现并联的
查资料＋看源码
资料
https://sthalles.github.io/deep_segmentation_network/写到
为了增加全局信息,ASPP先对feature map 进行GAP,然后对GAP结果使用256个1*1 卷积,最后使用双线性差值
小实验
import torch.nn as nn
import torch
import torch.nn.functional as F
input = torch.randn(1,2,3,3)
print(input)
global_avg_pool = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)),
nn.Conv2d(2, 4, 1, stride=1, bias=False))
out=global_avg_pool(input)
print(out)
x5 = F.interpolate(out, size=(5,5), mode='bilinear', align_corners=True)
print(x5)
tensor([[[[-0.8494, 0.7672, -1.2737],
[ 0.4866, 0.9821, 1.9463],
[ 0.8208, -0.3884, 0.9766]],
[[ 0.8611, -0.8998, -0.1001],
[-0.8969, -0.7486, -0.7516],
[-1.3810, -0.2459, 0.3907]]]])
tensor([[[[ 0.3816]],
[[ 0.1271]],
[[ 0.1947]],
[[-0.0331]]]], grad_fn=)
tensor([[[[ 0.3816, 0.3816, 0.3816, 0.3816, 0.3816],
[ 0.3816, 0.3816, 0.3816, 0.3816, 0.3816],
[ 0.3816, 0.3816, 0.3816, 0.3816, 0.3816],
[ 0.3816, 0.3816, 0.3816, 0.3816, 0.3816],
[ 0.3816, 0.3816, 0.3816, 0.3816, 0.3816]],
[[ 0.1271, 0.1271, 0.1271, 0.1271, 0.1271],
[ 0.1271, 0.1271, 0.1271, 0.1271, 0.1271],
[ 0.1271, 0.1271, 0.1271, 0.1271, 0.1271],
[ 0.1271, 0.1271, 0.1271, 0.1271, 0.1271],
[ 0.1271, 0.1271, 0.1271, 0.1271, 0.1271]],
[[ 0.1947, 0.1947, 0.1947, 0.1947, 0.1947],
[ 0.1947, 0.1947, 0.1947, 0.1947, 0.1947],
[ 0.1947, 0.1947, 0.1947, 0.1947, 0.1947],
[ 0.1947, 0.1947, 0.1947, 0.1947, 0.1947],
[ 0.1947, 0.1947, 0.1947, 0.1947, 0.1947]],
[[-0.0331, -0.0331, -0.0331, -0.0331, -0.0331],
[-0.0331, -0.0331, -0.0331, -0.0331, -0.0331],
[-0.0331, -0.0331, -0.0331, -0.0331, -0.0331],
[-0.0331, -0.0331, -0.0331, -0.0331, -0.0331],
[-0.0331, -0.0331, -0.0331, -0.0331, -0.0331]]]],
grad_fn=)
源码
deeplab 代码
assp网络
inplanes = 2048
self.aspp1 = _ASPPModule(inplanes, 256, 1, padding=0, dilation=dilations[0], BatchNorm=BatchNorm)
self.aspp2 = _ASPPModule(inplanes, 256, 3, padding=dilations[1], dilation=dilations[1], BatchNorm=BatchNorm)
self.aspp3 = _ASPPModule(inplanes, 256, 3, padding=dilations[2], dilation=dilations[2], BatchNorm=BatchNorm)
self.aspp4 = _ASPPModule(inplanes, 256, 3, padding=dilations[3], dilation=dilations[3], BatchNorm=BatchNorm)
self.global_avg_pool = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)),
nn.Conv2d(inplanes, 256, 1, stride=1, bias=False),
BatchNorm(256),
nn.ReLU())
forward
x1 = self.aspp1(x)
x2 = self.aspp2(x)
x3 = self.aspp3(x)
x4 = self.aspp4(x)
x5 = self.global_avg_pool(x)
x5 = F.interpolate(x5, size=x4.size()[2:], mode='bilinear', align_corners=True)
x = torch.cat((x1, x2, x3, x4, x5), dim=1)