SqueezeNet

最新推荐文章于 2024-10-07 06:31:57 发布

学AI不秃头

最新推荐文章于 2024-10-07 06:31:57 发布

阅读量123

点赞数

分类专栏：深度学习文章标签：人工智能算法机器学习计算机视觉分类

本文链接：https://blog.csdn.net/a1691125058/article/details/133902214

版权

深度学习专栏收录该内容

4 篇文章 0 订阅

订阅专栏

简介

SqueezeNet是对标AlexNet，性能水平相当，但参数量大幅降低，比AlexNet少50x，模型大小小于0.5MB的一个轻量化模型。

为了减少参数量且保持准确率SqueezeNet提出里三个设计策略：

Strategy 1. Replace 3x3 filters with 1x1 filters.

把3x3的卷积换成1x1的卷积这样可以减少9倍的参数

Strategy 2. Decrease the number of input channels to 3x3 filters.

不仅减少了3x3卷积的数量而且还减少了卷积核的输入channel

Strategy 3. Downsample late in the network so that convolution layers have large activation maps.

下采样放在卷积层的后面，虽然今天的网络设计结构都是如此

The Fire Module

在这里插入图片描述

在Fire Module中作者提出使用三个超参数: $s_{1x1}$ 、 $e_{1x1}$ 、 $e_{3x3}$

这三个超参数分别是squeeze部分的1x1卷积核个数、expand部分中1x1卷积核的个数和

expand中3x3卷积核的个数。作者让 $s_{1x1}< (e_{1x1} + e_{1x1} )$ ,这样就达到减小了输入到3x3卷积层的channel数，是Stratege 2的体现。

网络架构

在这里插入图片描述

代码

Fire_block

在这里插入图片描述

we implement our expand layer with two separate convolution layers: a layer with 1x1 filters, and a layer with 3x3 filters. Then, we concatenate the outputs of these layers together in the channel dimension.

class Fire_block(nn.Module):
    def __init__(self,in_channel,S1x1,E1x1,E3x3):
        super(Fire_block, self).__init__()
        #Squeeze
        self.squeeze = nn.Conv2d(in_channel,S1x1,1)
        self.squeeze_activateion = nn.ReLU(inplace=True)
        #Expand
        self.expand1x1 = nn.Conv2d(S1x1,E1x1,1)
        self.expand1x1_activation = nn.ReLU(inplace=True)

        self.expand3x3 = nn.Conv2d(E1x1,E3x3,3,1,1)
        self.expand3x3_activation = nn.ReLU(inplace=True)
    def forward(self,x):
        x = self.squeeze(x)
        x = self.squeeze_activateion(x)
        #分成两个layer，再concat
        expand1x1 = self.expand1x1(x)
        expand1x1 = self.expand1x1_activation(expand1x1)

        expand3x3 = self.expand3x3(x)
        expand3x3 = self.expand3x3_activation(x)

        result = torch.cat([expand1x1,expand3x3],dim=1)
        return result

主网络

在这里插入图片描述

class SqueezeNet(nn.Module):
    def __init__(self,num_classes=5):
        super(SqueezeNet, self).__init__()

        self.features = nn.Sequential(
            nn.Conv2d(3, 96, 7, 2, 2),
            nn.MaxPool2d(3, 2),
            Fire_block(96, 16, 64, 64),
            Fire_block(128, 16, 64, 64),
            Fire_block(128, 32, 128, 128),
            nn.MaxPool2d(3, 2),
            Fire_block(256, 32, 128, 128),
            Fire_block(256, 48, 192, 192),
            Fire_block(384, 48, 192, 192),
            Fire_block(384, 64, 256, 256),
            nn.MaxPool2d(3, 2),
            Fire_block(512, 64, 256, 256),
        )

        self.classifer = nn.Sequential(
            nn.Conv2d(512,num_classes,1),
            nn.AdaptiveAvgPool2d(1),
            nn.Flatten(),
            nn.Softmax(dim=1)
        )
    def forward(self,x):
        x = self.features(x)
        x = self.classifer(x)
        return x