- 🍨 本文为🔗365天深度学习训练营 中的学习记录博客
- 🍖 原作者：K同学啊
📌 本周任务（自己改进一下）：
●1. 在DenseNet系列算法中插入SE-Net通道注意力机制，并完成乳腺癌数据集识别
●2. 改进思路是否可以迁移到其他地方呢
●3. 验证集accuracy比较
一、前言
SE-Net（Squeeze-and-Excitation Network），它是 ImageNet 2017 竞赛的冠军模型，由 WMW 团队提出。具有复杂度低，参数少和计算量小的优点。SE-Net 的设计思路简单，容易与现有的网络结构（如 Inception 和 ResNet）结合，增强这些网络的性能。传统的网络结构（如 Inception）主要关注空间维度的特征提升，而 SE-Net 则将重点放在特征通道之间的关系上。这意味着它关注的是不同特征通道的重要性，而不仅仅是空间位置。SE-Net 通过学习每个特征通道的重要性，来自动调整特征的权重。具体来说，它会提升那些对当前任务有用的特征，同时抑制那些不重要的特征。这一过程被称为“特征重标定”。SE 模块是 SE-Net 的核心组成部分，负责实现特征重标定的功能。虽然具体的 SE 模块图未显示，但通常它包括两个主要步骤：首先通过全局平均池化获取特征通道的全局信息，然后通过全连接层学习每个通道的重要性，最后根据这些重要性调整特征通道的输出。如下图所示：
对于给定的一个输入x，其特征通道数为C’，经过一系列操作变换后通道数变为C，SE-Net会进行如下操作
- Squeeze 操作： 通过全局平均池化将每个通道的特征压缩成一个单一的数值，从而得到一个全局空间信息的通道描述符。这一步可以视为对每个通道的特征进行“压缩”，从而总结出通道的全局信息。
- Excitation 操作： 采用一个全连接的神经网络，通常包含两层，第一层用来降维（减少模型复杂度和参数量），第二层用来恢复维度。这个过程通过 Sigmoid 函数输出每个通道的权重系数，从而实现对每个通道的“激励”。
- Scale 操作： 最后，通过将 Excitation 操作的输出（即通道的权重系数）与原始输入按元素相乘，实现了对特征的重新缩放。这种按权重调整通道输出的方法，增强了模型对有用特征的捕捉能力，同时抑制了不重要的特征。
二、SE模块的通用性
SE模块的一个优点在于他可以直接应用于现有的网络结构中，以Inception和ResNet为例，我们只需要在Inception模块和Residual模块后面加上SE模块即可。
三、SE模块的代码实现
class SELayer(nn.Module):
def __init__(self, channels, reduction= 16):
super(SELayer, self).__init__()
self.avg_pool = nn.AdaptiveAvgPool2d(1) #Squeeze操作
self.fc = nn.Sequential(
nn.Linear(channels, channels // reduction, bias=False),
nn.ReLU(inplace=True),
nn.Linear(channels // reduction, channels, bias=False),
nn.Sigmoid() #产生通道权重
)
def forward(self, x):
b, c, h, w = x.size()
y = self.avg_pool(x).view(b,c)
y = self.fc(y).view(b,c,1,1)
return x * y.expand_as(x)
四、SE模块插入到DenseNet网络中
我直接把网络结构放出来：
import torch
import torch.nn as nn
import torch.nn.functional as F
class SELayer(nn.Module):
def __init__(self, channels, reduction= 16):
super(SELayer, self).__init__()
self.avg_pool = nn.AdaptiveAvgPool2d(1) #Squeeze操作
self.fc = nn.Sequential(
nn.Linear(channels, channels // reduction, bias=False),
nn.ReLU(inplace=True),
nn.Linear(channels // reduction, channels, bias=False),
nn.Sigmoid() #产生通道权重
)
def forward(self, x):
b, c, h, w = x.size()
y = self.avg_pool(x).view(b,c)
y = self.fc(y).view(b,c,1,1)
return x * y.expand_as(x)
#定义卷积块
class ConvBlock(nn.Module):
def __init__(self, in_channels, growth_rate):
super(ConvBlock, self).__init__()
self.conv1 = nn.Sequential(
nn.BatchNorm2d(in_channels),
nn.ReLU(inplace=True),
nn.Conv2d(in_channels, 4 * growth_rate, kernel_size=1, bias=False),
nn.BatchNorm2d(4 * growth_rate),
nn.ReLU(inplace=True),
nn.Conv2d(4 * growth_rate, growth_rate, kernel_size=3, padding=1, bias=False)
)
def forward(self, x):
x = self.conv1(x)
return x
class DenseBlock(nn.Module):
def __init__(self, num_layers, in_channels, growth_rate):
super(DenseBlock, self).__init__()
self.layers = nn.ModuleList()
for i in range(num_layers):
self.layers.append(ConvBlock(in_channels + i * growth_rate, growth_rate))
def forward(self, x):
for layer in self.layers:
new_features = layer(x)
x = torch.cat([x, new_features], dim=1)
return x
class TransitionBlock(nn.Module):
def __init__(self, in_channels, out_channels):
super(TransitionBlock, self).__init__()
self.bn = nn.BatchNorm2d(in_channels)
self.relu = nn.ReLU(inplace=True)
self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
self.avg_pool = nn.AvgPool2d(kernel_size=2, stride=2)
def forward(self, x):
x = self.bn(x)
x = self.relu(x)
x = self.conv(x)
x = self.avg_pool(x)
return x
# 构建DenseNet
class DenseNet(nn.Module):
def __init__(self, block_config, num_classes=1000, growth_rate=32):
super(DenseNet, self).__init__()
num_init_features = 64
self.features = nn.Sequential(
nn.Conv2d(3, num_init_features, kernel_size=7, stride=2, padding=3, bias=False),
nn.BatchNorm2d(num_init_features),
nn.ReLU(inplace=True),
nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
)
num_features = num_init_features
for i, num_blocks in enumerate(block_config):
block = DenseBlock(num_blocks, num_features, growth_rate)
self.features.add_module('denseblock{}'.format(i + 1), block)
num_features += num_blocks * growth_rate
if i != len(block_config) - 1:
trans = TransitionBlock(num_features, num_features // 2)
self.features.add_module('transition{}'.format(i + 1), trans)
num_features = num_features // 2
self.features.add_module('norm5', nn.BatchNorm2d(num_features))
self.classifier = nn.Linear(num_features, num_classes)
self.SE_layer = SELayer(num_features) # SE Layer should be initialized with the correct number of channels
def forward(self, x):
x = self.features(x)
x = self.SE_layer(x)
x = F.relu(x, inplace=True)
x = F.adaptive_avg_pool2d(x, (1, 1))
x = torch.flatten(x, 1)
x = self.classifier(x)
return x
# 定义不同的 DenseNet 配置
def DenseNet121(num_classes=3):
return DenseNet([6, 12, 24, 16], num_classes=num_classes)
def DenseNet169(num_classes=3):
return DenseNet([6, 12, 32, 32], num_classes=num_classes)
def DenseNet201(num_classes=3):
return DenseNet([6, 12, 48, 32], num_classes=num_classes)
import torchsummary
model1 = DenseNet121().cuda()
x = (3, 224, 224)
torchsummary.summary(model1,x)
model2 = DenseNet169().cuda()
model3 = DenseNet201().cuda()
代码输出：
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [-1, 64, 112, 112] 9,408
BatchNorm2d-2 [-1, 64, 112, 112] 128
ReLU-3 [-1, 64, 112, 112] 0
MaxPool2d-4 [-1, 64, 56, 56] 0
BatchNorm2d-5 [-1, 64, 56, 56] 128
ReLU-6 [-1, 64, 56, 56] 0
Conv2d-7 [-1, 128, 56, 56] 8,192
BatchNorm2d-8 [-1, 128, 56, 56] 256
ReLU-9 [-1, 128, 56, 56] 0
Conv2d-10 [-1, 32, 56, 56] 36,864
ConvBlock-11 [-1, 32, 56, 56] 0
BatchNorm2d-12 [-1, 96, 56, 56] 192
ReLU-13 [-1, 96, 56, 56] 0
Conv2d-14 [-1, 128, 56, 56] 12,288
BatchNorm2d-15 [-1, 128, 56, 56] 256
ReLU-16 [-1, 128, 56, 56] 0
Conv2d-17 [-1, 32, 56, 56] 36,864
ConvBlock-18 [-1, 32, 56, 56] 0
BatchNorm2d-19 [-1, 128, 56, 56] 256
ReLU-20 [-1, 128, 56, 56] 0
Conv2d-21 [-1, 128, 56, 56] 16,384
BatchNorm2d-22 [-1, 128, 56, 56] 256
ReLU-23 [-1, 128, 56, 56] 0
Conv2d-24 [-1, 32, 56, 56] 36,864
ConvBlock-25 [-1, 32, 56, 56] 0
BatchNorm2d-26 [-1, 160, 56, 56] 320
ReLU-27 [-1, 160, 56, 56] 0
Conv2d-28 [-1, 128, 56, 56] 20,480
BatchNorm2d-29 [-1, 128, 56, 56] 256
ReLU-30 [-1, 128, 56, 56] 0
Conv2d-31 [-1, 32, 56, 56] 36,864
ConvBlock-32 [-1, 32, 56, 56] 0
BatchNorm2d-33 [-1, 192, 56, 56] 384
ReLU-34 [-1, 192, 56, 56] 0
Conv2d-35 [-1, 128, 56, 56] 24,576
BatchNorm2d-36 [-1, 128, 56, 56] 256
ReLU-37 [-1, 128, 56, 56] 0
Conv2d-38 [-1, 32, 56, 56] 36,864
ConvBlock-39 [-1, 32, 56, 56] 0
BatchNorm2d-40 [-1, 224, 56, 56] 448
ReLU-41 [-1, 224, 56, 56] 0
Conv2d-42 [-1, 128, 56, 56] 28,672
BatchNorm2d-43 [-1, 128, 56, 56] 256
ReLU-44 [-1, 128, 56, 56] 0
Conv2d-45 [-1, 32, 56, 56] 36,864
ConvBlock-46 [-1, 32, 56, 56] 0
DenseBlock-47 [-1, 256, 56, 56] 0
BatchNorm2d-48 [-1, 256, 56, 56] 512
ReLU-49 [-1, 256, 56, 56] 0
Conv2d-50 [-1, 128, 56, 56] 32,768
AvgPool2d-51 [-1, 128, 28, 28] 0
TransitionBlock-52 [-1, 128, 28, 28] 0
BatchNorm2d-53 [-1, 128, 28, 28] 256
ReLU-54 [-1, 128, 28, 28] 0
Conv2d-55 [-1, 128, 28, 28] 16,384
BatchNorm2d-56 [-1, 128, 28, 28] 256
ReLU-57 [-1, 128, 28, 28] 0
Conv2d-58 [-1, 32, 28, 28] 36,864
ConvBlock-59 [-1, 32, 28, 28] 0
BatchNorm2d-60 [-1, 160, 28, 28] 320
ReLU-61 [-1, 160, 28, 28] 0
Conv2d-62 [-1, 128, 28, 28] 20,480
BatchNorm2d-63 [-1, 128, 28, 28] 256
ReLU-64 [-1, 128, 28, 28] 0
Conv2d-65 [-1, 32, 28, 28] 36,864
ConvBlock-66 [-1, 32, 28, 28] 0
BatchNorm2d-67 [-1, 192, 28, 28] 384
ReLU-68 [-1, 192, 28, 28] 0
Conv2d-69 [-1, 128, 28, 28] 24,576
BatchNorm2d-70 [-1, 128, 28, 28] 256
ReLU-71 [-1, 128, 28, 28] 0
Conv2d-72 [-1, 32, 28
最低0.47元/天 解锁文章
扫一扫
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
