Inception V2 没有改进 Inception V1

如何解决Inception V2 没有改进 Inception V1

我正在学习 D2L 一书,GoogLeNet 章节中的一个练习是从 Rethinking the Inception Architecture for Computer Vision 论文中获取模型,所以我这样做并基本上重新创建了基于 Inception V1 架构的 Inception V2 架构在 D2L 书中提出,我唯一没有包括的是论文中描述的网格减少。我希望这会改善我的模型的结果,但如果有的话,它会使它们变得更糟。我正在使用关于 fashion_mnist 数据集的 GoogLeNet 章节中提供的 D2L 代码测试模型,并使用 Inception V1 获得了 0.908 的测试准确率,但是在我的 V2 实现中,我只得到了 0.885 的测试acc,这是我所能达到的最高值得到。我希望 V2 模型会更好,所以我真的不知道我哪里出错了,如果我在我的代码中犯了错误或什么,但是对 V2 架构或论文的任何见解都将不胜感激。

这是前 3 个 inception 块的代码(来自论文中的图 5)

class InceptionB1(nn.Module):
    # `c1`--`c4` are the number of output channels for each path
    def __init__(self,in_channels,c1,c2,c3,c4,**kwargs):
        super(InceptionB1,self).__init__(**kwargs)
        # Path 1 is a single 1 x 1 convolutional layer
        self.p1_1 = nn.Conv2d(in_channels,kernel_size=1)
        # Path 2 is a 1 x 1 convolutional layer followed by a 3 x 3
        # convolutional layer
        self.p2_1 = nn.Conv2d(in_channels,c2[0],kernel_size=1)
        self.p2_2 = nn.Conv2d(c2[0],c2[1],kernel_size=3,padding=1)
        # Path 3 is a 1 x 1 convolutional layer followed by 2 3x3 convs (factorization)
        # convolutional layer
        self.p3_1 = nn.Conv2d(in_channels,c3[0],kernel_size=1)
        self.p3_2 = nn.Conv2d(c3[0],c3[1],padding=1)
        # Replacing path3_2 with modified path using factorization through smaller convolutions
        self.p3_3 = nn.Conv2d(c3[1],padding = 1)
        # Path 4 is a 3 x 3 maximum pooling layer followed by a 1 x 1
        # convolutional layer
        self.p4_1 = nn.MaxPool2d(kernel_size=3,stride=1,padding=1)
        self.p4_2 = nn.Conv2d(in_channels,kernel_size=1)

    def forward(self,x):
        p1 = F.relu(self.p1_1(x))
        p2 = F.relu(self.p2_2(F.relu(self.p2_1(x))))
        p3 = F.relu(self.p3_3(self.p3_2(F.relu(self.p3_1(x)))))
        p4 = F.relu(self.p4_2(self.p4_1(x)))
        # Concatenate the outputs on the channel dimension
        return torch.cat((p1,p2,p3,p4),dim=1)

这是我对中间 5 个块的代码(论文中的图 6)

class InceptionB2(nn.Module):
    # `c1`--`c4` are the number of output channels for each path
    def __init__(self,**kwargs):
        super(InceptionB2,kernel_size=1)
        # Path 2 is a 1 x 1 convolutional layer followed by a 1 x 3 then 3x1 conv
        # convolutional layer
        self.p2_1 = nn.Conv2d(in_channels,kernel_size=(1,3),padding=1)
        self.p2_3 = nn.Conv2d(c2[1],kernel_size=(3,1),padding=0)
        # Path 3 is a 1 x 1 conv layer followed by a 1x3 conv then 3x1 then 1x3 then 3x1
        # convolutional layer
        self.p3_1 = nn.Conv2d(in_channels,padding=1)
        self.p3_3 = nn.Conv2d(c3[1],padding=0)
        self.p3_4 = nn.Conv2d(c3[1],padding=0)
        self.p3_5 = nn.Conv2d(c3[1],padding=1)
        # Path 4 is a 3 x 3 maximum pooling layer followed by a 1 x 1
        # convolutional layer
        self.p4_1 = nn.MaxPool2d(kernel_size=3,x):
        p1 = F.relu(self.p1_1(x))
        p2 = F.relu(self.p2_3(F.relu(self.p2_2(F.relu(self.p2_1(x))))))
        p3 = F.relu(self.p3_5(F.relu(self.p3_4(F.relu(self.p3_3(F.relu(self.p3_2(F.relu(self.p3_1(x))))))))))
        p4 = F.relu(self.p4_2(self.p4_1(x)))
        # Concatenate the outputs on the channel dimension
        return torch.cat((p1,dim=1)

这是我最后 2 个块的代码(论文中的图 7)

class InceptionB3(nn.Module):
    # `c1`--`c4` are the number of output channels for each path
    def __init__(self,**kwargs):
        super(InceptionB3,kernel_size=1)
        # Path 2 is a 1 x 1 convolutional layer followed by a 1x3 and 3x1
        # convolutional layer
        self.p2_1 = nn.Conv2d(in_channels,kernel_size=1)
        self.p2_2_1 = nn.Conv2d(c2[0],padding=0)
        self.p2_2_2 = nn.Conv2d(c2[0],padding=0)
        # Path 3 is a 1 x 1 convolutional layer followed by a 3x3 then 1x3 and 3x1
        # convolutional layer
        self.p3_1 = nn.Conv2d(in_channels,padding=1)
        self.p3_3_1 = nn.Conv2d(c3[1],padding=0)
        self.p3_3_2 = nn.Conv2d(c3[1],padding=0)
        # Path 4 is a 3 x 3 maximum pooling layer followed by a 1 x 1
        # convolutional layer
        self.p4_1 = nn.MaxPool2d(kernel_size=3,x):
        p1 = F.relu(self.p1_1(x))

        p2_1 = F.relu(self.p2_2_1(F.relu(self.p2_1(x))))
        p2_2 = F.relu(elf.p2_2_2(F.relu(self.p2_1(x))))
        p2 = torch.matmul(p2_1,p2_2)

        p3_1 = F.relu(self.p3_2(F.relu(self.p3_1(x))))
        p3_2_1 = F.relu(self.p3_3_1(p3_1))
        p3_2_2 = F.relu(sself.p3_3_2(p3_1))
        p3 = torch.matmul(p3_2_1,p3_2_2)

        p4 = F.relu(self.p4_2(self.p4_1(x)))
        # Concatenate the outputs on the channel dimension
        return torch.cat((p1,dim=1)

这是D2L章节的链接 https://d2l.ai/chapter_convolutional-modern/googlenet.html

This is the training graph for the v1 model

This is the training graph for the v2 model

版权声明:本文内容由互联网用户自发贡献,该文观点与技术仅代表作者本人。本站仅提供信息存储空间服务,不拥有所有权,不承担相关法律责任。如发现本站有涉嫌侵权/违法违规的内容, 请发送邮件至 dio@foxmail.com 举报,一经查实,本站将立刻删除。

相关推荐


依赖报错 idea导入项目后依赖报错,解决方案:https://blog.csdn.net/weixin_42420249/article/details/81191861 依赖版本报错:更换其他版本 无法下载依赖可参考:https://blog.csdn.net/weixin_42628809/a
错误1:代码生成器依赖和mybatis依赖冲突 启动项目时报错如下 2021-12-03 13:33:33.927 ERROR 7228 [ main] o.s.b.d.LoggingFailureAnalysisReporter : *************************** APPL
错误1:gradle项目控制台输出为乱码 # 解决方案:https://blog.csdn.net/weixin_43501566/article/details/112482302 # 在gradle-wrapper.properties 添加以下内容 org.gradle.jvmargs=-Df
错误还原:在查询的过程中,传入的workType为0时,该条件不起作用 <select id="xxx"> SELECT di.id, di.name, di.work_type, di.updated... <where> <if test=&qu
报错如下,gcc版本太低 ^ server.c:5346:31: 错误:‘struct redisServer’没有名为‘server_cpulist’的成员 redisSetCpuAffinity(server.server_cpulist); ^ server.c: 在函数‘hasActiveC
解决方案1 1、改项目中.idea/workspace.xml配置文件,增加dynamic.classpath参数 2、搜索PropertiesComponent,添加如下 <property name="dynamic.classpath" value="tru
删除根组件app.vue中的默认代码后报错:Module Error (from ./node_modules/eslint-loader/index.js): 解决方案:关闭ESlint代码检测,在项目根目录创建vue.config.js,在文件中添加 module.exports = { lin
查看spark默认的python版本 [root@master day27]# pyspark /home/software/spark-2.3.4-bin-hadoop2.7/conf/spark-env.sh: line 2: /usr/local/hadoop/bin/hadoop: No s
使用本地python环境可以成功执行 import pandas as pd import matplotlib.pyplot as plt # 设置字体 plt.rcParams['font.sans-serif'] = ['SimHei'] # 能正确显示负号 p
错误1:Request method ‘DELETE‘ not supported 错误还原:controller层有一个接口,访问该接口时报错:Request method ‘DELETE‘ not supported 错误原因:没有接收到前端传入的参数,修改为如下 参考 错误2:cannot r
错误1:启动docker镜像时报错:Error response from daemon: driver failed programming external connectivity on endpoint quirky_allen 解决方法:重启docker -> systemctl r
错误1:private field ‘xxx‘ is never assigned 按Altʾnter快捷键,选择第2项 参考:https://blog.csdn.net/shi_hong_fei_hei/article/details/88814070 错误2:启动时报错,不能找到主启动类 #
报错如下,通过源不能下载,最后警告pip需升级版本 Requirement already satisfied: pip in c:\users\ychen\appdata\local\programs\python\python310\lib\site-packages (22.0.4) Coll
错误1:maven打包报错 错误还原:使用maven打包项目时报错如下 [ERROR] Failed to execute goal org.apache.maven.plugins:maven-resources-plugin:3.2.0:resources (default-resources)
错误1:服务调用时报错 服务消费者模块assess通过openFeign调用服务提供者模块hires 如下为服务提供者模块hires的控制层接口 @RestController @RequestMapping("/hires") public class FeignControl
错误1:运行项目后报如下错误 解决方案 报错2:Failed to execute goal org.apache.maven.plugins:maven-compiler-plugin:3.8.1:compile (default-compile) on project sb 解决方案:在pom.
参考 错误原因 过滤器或拦截器在生效时,redisTemplate还没有注入 解决方案:在注入容器时就生效 @Component //项目运行时就注入Spring容器 public class RedisBean { @Resource private RedisTemplate<String
使用vite构建项目报错 C:\Users\ychen\work>npm init @vitejs/app @vitejs/create-app is deprecated, use npm init vite instead C:\Users\ychen\AppData\Local\npm-