【图像分类】【深度学习】【Pytorch版本】GoogLeNet(InceptionV4)模型算法详解

【图像分类】【深度学习】【Pytorch版本】GoogLeNet(InceptionV4)模型算法详解

文章目录

  • 【图像分类】【深度学习】【Pytorch版本】GoogLeNet(InceptionV4)模型算法详解
  • 前言
  • GoogLeNet(InceptionV4)讲解
    • Stem结构
    • Inception-A结构
    • Inception- B结构
    • Inception-C结构
    • redution-A结构
    • redution-B结构
    • GoogLeNet(InceptionV4)模型结构
  • GoogLeNet(InceptionV4) Pytorch代码
  • 完整代码
  • 总结


前言

GoogLeNet(InceptionV4)是由谷歌的Szegedy, Christian等人在《Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning【AAAI-2017】》【论文地址】一文中提出的改进模型,InceptionV4保留了此前的Inception模块的核心思想基础上进行了改进和优化,InceptionV4的所有模块都采用了统一的设计原则,即采用Inception模块作为基本单元,通过堆叠纯Inception基本单元来实现复杂的网络结构。

因为InceptionV4、Inception-Resnet-v1和Inception-Resnet-v2同出自一篇论文,大部分读者对InceptionV4存在误解,认为它是Inception模块与残差学习的结合,其实InceptionV4没有使用残差学习的思想,它基本延续了Inception v2/v3的结构,只有Inception-Resnet-v1和Inception-Resnet-v2才是Inception模块与残差学习的结合产物。


GoogLeNet(InceptionV4)讲解

InceptionV4的三种基础Inception结构与InceptionV3【参考】中使用的结构基本一样,但InceptionV4引入了一些新的模块形状及其间的连接设计,在网络的早期阶段引入了“Stem”模块,用于快速降低特征图的分辨率,从而减少后续Inception模块的计算量。

Stem结构

stem结构实际上是替代了此前的Inception系列网络中Inception结构组之前的网络层,Stem中借鉴了InceptionV3中使用的并行结构、不对称卷积核结构,并使用1*1的卷积核用来降维和增加非线性,可以在保证信息损失足够小的情况下,使得计算量降低。

所有卷积中没有标记为V表示填充方式为"SAME Padding",输入和输出维度一致;标记为V表示填充方式为"VALID Padding",输出维度视具体情况而定。

Inception-A结构

对应InceptionV3中的结构Ⅰ。

Inception- B结构

对应InceptionV3中的结构Ⅱ,只是1×3卷积和3×1卷积变成了1×7卷积和7×1卷积。

Inception-C结构

对应InceptionV3中的结构Ⅲ,只是3×3卷积变成了1×3卷积和3×1卷积的串联结构。

redution-A结构

对应InceptionV3中的特殊结构。

k和l表示卷积个数,不同网络结构的redution-A结构k和l是不同的,Inception-ResNet会在其他博文中介绍。

redution-B结构

采用并行、不对称卷积和1*1的卷积来降低计算量。

GoogLeNet(InceptionV4)模型结构

下图是原论文给出的关于 GoogLeNet(InceptionV4)模型结构的详细示意图:

GoogLeNet(InceptionV4)在图像分类中分为两部分:backbone部分: 主要由InceptionV4模块、Stem模块和池化层(汇聚层)组成,分类器部分:由全连接层组成。
InceptionV4三种Inception模块的个数分别为4、7、3个,而InceptionV3中则为3、5、2个,因此InceptionV4的层次更深、结构更复杂,feature map更多。为了降低计算量,在Inception-A和Inception-B后面分别添加了Reduction-A和Reduction-B的结构,用来降低计算量。


GoogLeNet(InceptionV4) Pytorch代码

卷积层组: 卷积层+BN层+激活函数

# 卷积组: Conv2d+BN+ReLU
class BasicConv2d(nn.Module):def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0):super(BasicConv2d, self).__init__()self.conv = nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding)self.bn = nn.BatchNorm2d(out_channels)self.relu = nn.ReLU(inplace=True)def forward(self, x):x = self.conv(x)x = self.bn(x)x = self.relu(x)return x

Stem模块: 卷积层组+池化层

# Stem:BasicConv2d+MaxPool2d
class Stem(nn.Module):def __init__(self, in_channels, out_channels):super(Stem, self).__init__()# conv3*3(32 stride2 valid)self.conv1 = BasicConv2d(in_channels, 32, kernel_size=3, stride=2)# conv3*3(32 valid)self.conv2 = BasicConv2d(32, 32, kernel_size=3)# conv3*3(64)self.conv3 = BasicConv2d(32, 64, kernel_size=3, padding=1)# maxpool3*3(stride2 valid) & conv3*3(96 stride2 valid)self.maxpool4 = nn.MaxPool2d(kernel_size=3, stride=2)self.conv4 = BasicConv2d(64, 96, kernel_size=3, stride=2)# conv1*1(64)+conv3*3(96 valid)self.conv5_1_1 = BasicConv2d(160, 64, kernel_size=1)self.conv5_1_2 = BasicConv2d(64, 96, kernel_size=3)# conv1*1(64)+conv7*1(64)+conv1*7(64)+conv3*3(96 valid)self.conv5_2_1 = BasicConv2d(160, 64, kernel_size=1)self.conv5_2_2 = BasicConv2d(64, 64, kernel_size=(7, 1), padding=(3, 0))self.conv5_2_3 = BasicConv2d(64, 64, kernel_size=(1, 7), padding=(0, 3))self.conv5_2_4 = BasicConv2d(64, 96, kernel_size=3)# conv3*3(192 valid) & maxpool3*3(stride2 valid)self.conv6 = BasicConv2d(192, 192, kernel_size=3, stride=2)self.maxpool6 = nn.MaxPool2d(kernel_size=3, stride=2)def forward(self, x):x1_1 = self.maxpool4(self.conv3(self.conv2(self.conv1(x))))x1_2 = self.conv4(self.conv3(self.conv2(self.conv1(x))))x1 = torch.cat([x1_1, x1_2], 1)x2_1 = self.conv5_1_2(self.conv5_1_1(x1))x2_2 = self.conv5_2_4(self.conv5_2_3(self.conv5_2_2(self.conv5_2_1(y1))))x2 = torch.cat([x2_1, x2_2], 1)x3_1 = self.conv6(x2)x3_2 = self.maxpool6(x2)x3 = torch.cat([x3_1, x3_2], 1)return x3

Inception-A模块: 卷积层组+池化层

# InceptionV4A:BasicConv2d+MaxPool2d
class InceptionV4A(nn.Module):def __init__(self, in_channels, ch1x1, ch3x3red, ch3x3, ch3x3redX2, ch3x3X2, pool_proj):super(InceptionV4A, self).__init__()# conv1*1(96)self.branch1 = BasicConv2d(in_channels, ch1x1, kernel_size=1)# conv1*1(64)+conv3*3(96)self.branch2 = nn.Sequential(BasicConv2d(in_channels, ch3x3red, kernel_size=1),BasicConv2d(ch3x3red, ch3x3, kernel_size=3, padding=1)   # 保证输出大小等于输入大小)# conv1*1(64)+conv3*3(96)+conv3*3(96)self.branch3 = nn.Sequential(BasicConv2d(in_channels, ch3x3redX2, kernel_size=1),BasicConv2d(ch3x3redX2, ch3x3X2, kernel_size=3, padding=1),BasicConv2d(ch3x3X2, ch3x3X2, kernel_size=3, padding=1)         # 保证输出大小等于输入大小)# avgpool + conv1*1(96)self.branch4 = nn.Sequential(nn.MaxPool2d(kernel_size=3, stride=1, padding=1),BasicConv2d(in_channels, pool_proj, kernel_size=1))def forward(self, x):branch1 = self.branch1(x)branch2 = self.branch2(x)branch3 = self.branch3(x)branch4 = self.branch4(x)# 拼接outputs = [branch1, branch2, branch3, branch4]return torch.cat(outputs, 1)

Inception-B模块: 卷积层组+池化层

# InceptionV4B:BasicConv2d+MaxPool2d
class InceptionV4B(nn.Module):def __init__(self, in_channels, ch1x1, ch3x3red, ch3x3_1, ch3x3_2, ch3x3redX2, ch3x3X2_1, ch3x3X2_2, pool_proj):super(InceptionV4B, self).__init__()# conv1*1(384)self.branch1 = BasicConv2d(in_channels, ch1x1, kernel_size=1)# conv1*1(192)+conv1*7(224)+conv1*7(256)self.branch2 = nn.Sequential(BasicConv2d(in_channels, ch3x3red, kernel_size=1),BasicConv2d(ch3x3red, ch3x3_1, kernel_size=[1, 7], padding=[0, 3]),BasicConv2d(ch3x3_1, ch3x3_2, kernel_size=[7, 1], padding=[3, 0])   # 保证输出大小等于输入大小)# conv1*1(192)+conv1*7(192)+conv7*1(224)+conv1*7(224)+conv7*1(256)self.branch3 = nn.Sequential(BasicConv2d(in_channels, ch3x3redX2, kernel_size=1),BasicConv2d(ch3x3redX2, ch3x3redX2, kernel_size=[1, 7], padding=[0, 3]),BasicConv2d(ch3x3redX2, ch3x3X2_1, kernel_size=[7, 1], padding=[3, 0]),BasicConv2d(ch3x3X2_1, ch3x3X2_1, kernel_size=[1, 7], padding=[0, 3]),BasicConv2d(ch3x3X2_1, ch3x3X2_2, kernel_size=[7, 1], padding=[3, 0])  # 保证输出大小等于输入大小)# avgpool+conv1*1(128)self.branch4 = nn.Sequential(nn.MaxPool2d(kernel_size=3, stride=1, padding=1),BasicConv2d(in_channels, pool_proj, kernel_size=1))def forward(self, x):branch1 = self.branch1(x)branch2 = self.branch2(x)branch3 = self.branch3(x)branch4 = self.branch4(x)# 拼接outputs = [branch1, branch2, branch3, branch4]return torch.cat(outputs, 1)

Inception-C模块: 卷积层组+池化层

# InceptionV4C:BasicConv2d+MaxPool2d
class InceptionV4C(nn.Module):def __init__(self, in_channels, ch1x1, ch3x3red, ch3x3, ch3x3redX2, ch3x3X2_1, ch3x3X2_2, ch3x3X2_3,pool_proj):super(InceptionV4C, self).__init__()# conv1*1(256)self.branch1 = BasicConv2d(in_channels, ch1x1, kernel_size=1)# conv1*1(384)+conv1*3(256) & conv3*1(256)self.branch2_0 = BasicConv2d(in_channels, ch3x3red, kernel_size=1)self.branch2_1 = BasicConv2d(ch3x3red, ch3x3, kernel_size=[1, 3], padding=[0, 1])self.branch2_2 = BasicConv2d(ch3x3red, ch3x3, kernel_size=[3, 1], padding=[1, 0])# conv1*1(384)+conv1*3(448)+conv3*1(512)+conv3*1(256) & conv7*1(256)self.branch3_0 = nn.Sequential(BasicConv2d(in_channels, ch3x3redX2, kernel_size=1),BasicConv2d(ch3x3redX2, ch3x3X2_1, kernel_size=[1, 3], padding=[0, 1]),BasicConv2d(ch3x3X2_1, ch3x3X2_2, kernel_size=[3, 1], padding=[1, 0]),)self.branch3_1 = BasicConv2d(ch3x3X2_2, ch3x3X2_3, kernel_size=[1, 3], padding=[0, 1])self.branch3_2 = BasicConv2d(ch3x3X2_2, ch3x3X2_3, kernel_size=[3, 1], padding=[1, 0])# avgpool+conv1*1(256)self.branch4 = nn.Sequential(nn.MaxPool2d(kernel_size=3, stride=1, padding=1),BasicConv2d(in_channels, pool_proj, kernel_size=1))def forward(self, x):branch1 = self.branch1(x)branch2_0 = self.branch2_0(x)branch2 = torch.cat([self.branch2_1(branch2_0), self.branch2_2(branch2_0)], dim=1)branch3_0 = self.branch3_0(x)branch3 = torch.cat([self.branch3_1(branch3_0), self.branch3_2(branch3_0)], dim=1)branch4 = self.branch4(x)# 拼接outputs = [branch1, branch2, branch3, branch4]return torch.cat(outputs, 1)

redutionA模块: 卷积层组+池化层

# redutionA:BasicConv2d+MaxPool2d
class redutionA(nn.Module):def __init__(self, in_channels, k, l, m, n):super(redutionA, self).__init__()# conv3*3(n stride2 valid)self.branch1 = nn.Sequential(BasicConv2d(in_channels, n, kernel_size=3, stride=2),)# conv1*1(k)+conv3*3(l)+conv3*3(m stride2 valid)self.branch2 = nn.Sequential(BasicConv2d(in_channels, k, kernel_size=1),BasicConv2d(k, l, kernel_size=3, padding=1),BasicConv2d(l, m, kernel_size=3, stride=2))# maxpool3*3(stride2 valid)self.branch3 = nn.Sequential(nn.MaxPool2d(kernel_size=3, stride=2))def forward(self, x):branch1 = self.branch1(x)branch2 = self.branch2(x)branch3 = self.branch3(x)# 拼接outputs = [branch1,branch2, branch3]return torch.cat(outputs, 1)

redutionB模块: 卷积层组+池化层

# redutionB:BasicConv2d+MaxPool2d
class redutionB(nn.Module):def __init__(self, in_channels, ch3x3red, ch3x3, ch3x3redX2, ch3x3X2):super(redutionB, self).__init__()# conv1*1(192)+conv3*3(192 stride2 valid)self.branch1 = nn.Sequential(BasicConv2d(in_channels, ch3x3red, kernel_size=1),BasicConv2d(ch3x3red, ch3x3, kernel_size=3, stride=2))# conv1*1(256)+conv1*7(256)+conv7*1(320)+conv3*3(320 stride2 valid)self.branch2 = nn.Sequential(BasicConv2d(in_channels, ch3x3redX2, kernel_size=1),BasicConv2d(ch3x3redX2, ch3x3redX2, kernel_size=(1, 7), padding=(0, 3)),# 保证输出大小等于输入大小BasicConv2d(ch3x3redX2, ch3x3X2, kernel_size=(7, 1), padding=(3, 0)),BasicConv2d(ch3x3X2, ch3x3X2, kernel_size=3, stride=2))#  maxpool3*3(stride2 valid)self.branch3 = nn.Sequential(nn.MaxPool2d(kernel_size=3, stride=2))def forward(self, x):branch1 = self.branch1(x)branch2 = self.branch2(x)branch3 = self.branch3(x)# 拼接outputs = [branch1,branch2, branch3]return torch.cat(outputs, 1)

完整代码

GoogLeNet(InceptionV4)的输入图像尺寸是299×299

import torch.nn as nn
import torch
from torchsummary import summaryclass GoogLeNetV4(nn.Module):def __init__(self, num_classes=1000, init_weights=False):super(GoogLeNetV4, self).__init__()# stem模块self.stem = Stem(3, 384)# InceptionA模块self.inceptionA = InceptionV4A(384, 96, 64, 96, 64, 96, 96)# RedutionA模块self.RedutionA = redutionA(384, 192, 224, 256, 384)# InceptionB模块self.InceptionB = InceptionV4B(1024, 384, 192, 224, 256, 192, 224,256,128)# RedutionB模块self.RedutionB = redutionB(1024,     192, 192, 256, 320)# InceptionC模块self.InceptionC = InceptionV4C(1536, 256, 384, 256, 384, 448, 512, 256,256)self.avgpool = nn.AdaptiveAvgPool2d((1, 1))self.dropout = nn.Dropout(0.8)self.fc = nn.Linear(1536, num_classes)if init_weights:self._initialize_weights()def forward(self, x):# Stem Module# N x 3 x 299 x 299x = self.stem(x)# InceptionA Module * 4# N x 384 x 26 x 26x = self.inceptionA(self.inceptionA(self.inceptionA(self.inceptionA(x))))# ReductionA Module# N x 384 x 26 x 26x = self.RedutionA(x)# InceptionB Module * 7# N x 1024 x 12 x 12x = self.InceptionB(self.InceptionB(self.InceptionB(self.InceptionB(self.InceptionB(self.InceptionB(self.InceptionB(x)))))))# ReductionB Module# N x 1024 x 12 x 12x = self.RedutionB(x)# InceptionC Module * 3# N x 1536 x 5 x 5x = self.InceptionC(self.InceptionC(self.InceptionC(x)))# Average Pooling# N x 1536 x 5 x 5x = self.avgpool(x)# N x 1536 x 1 x 1x = x.view(x.size(0), -1)# Dropout# N x 1536x = self.dropout(x)# Linear(Softmax)# N x 1536x = self.fc(x)# N x 1000return x# 对模型的权重进行初始化操作def _initialize_weights(self):for m in self.modules():if isinstance(m, nn.Conv2d):nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')if m.bias is not None:nn.init.constant_(m.bias, 0)elif isinstance(m, nn.Linear):nn.init.normal_(m.weight, 0, 0.01)nn.init.constant_(m.bias, 0)# InceptionV4A:BasicConv2d+MaxPool2d
class InceptionV4A(nn.Module):def __init__(self, in_channels, ch1x1, ch3x3red, ch3x3, ch3x3redX2, ch3x3X2, pool_proj):super(InceptionV4A, self).__init__()# conv1*1(96)self.branch1 = BasicConv2d(in_channels, ch1x1, kernel_size=1)# conv1*1(64)+conv3*3(96)self.branch2 = nn.Sequential(BasicConv2d(in_channels, ch3x3red, kernel_size=1),BasicConv2d(ch3x3red, ch3x3, kernel_size=3, padding=1)   # 保证输出大小等于输入大小)# conv1*1(64)+conv3*3(96)+conv3*3(96)self.branch3 = nn.Sequential(BasicConv2d(in_channels, ch3x3redX2, kernel_size=1),BasicConv2d(ch3x3redX2, ch3x3X2, kernel_size=3, padding=1),BasicConv2d(ch3x3X2, ch3x3X2, kernel_size=3, padding=1)         # 保证输出大小等于输入大小)# avgpool+conv1*1(96)self.branch4 = nn.Sequential(nn.MaxPool2d(kernel_size=3, stride=1, padding=1),BasicConv2d(in_channels, pool_proj, kernel_size=1))def forward(self, x):branch1 = self.branch1(x)branch2 = self.branch2(x)branch3 = self.branch3(x)branch4 = self.branch4(x)# 拼接outputs = [branch1, branch2, branch3, branch4]return torch.cat(outputs, 1)# InceptionV4B:BasicConv2d+MaxPool2d
class InceptionV4B(nn.Module):def __init__(self, in_channels, ch1x1, ch_red, ch_1, ch_2, ch_redX2, ch_X2_1, ch_X2_2, pool_proj):super(InceptionV4B, self).__init__()# conv1*1(384)self.branch1 = BasicConv2d(in_channels, ch1x1, kernel_size=1)# conv1*1(192)+conv1*7(224)+conv1*7(256)self.branch2 = nn.Sequential(BasicConv2d(in_channels, ch_red, kernel_size=1),BasicConv2d(ch_red, ch_1, kernel_size=[1, 7], padding=[0, 3]),BasicConv2d(ch_1, ch_2, kernel_size=[7, 1], padding=[3, 0])   # 保证输出大小等于输入大小)# conv1*1(192)+conv1*7(192)+conv7*1(224)+conv1*7(224)+conv7*1(256)self.branch3 = nn.Sequential(BasicConv2d(in_channels, ch_redX2, kernel_size=1),BasicConv2d(ch_redX2, ch_redX2, kernel_size=[1, 7], padding=[0, 3]),BasicConv2d(ch_redX2, ch_X2_1, kernel_size=[7, 1], padding=[3, 0]),BasicConv2d(ch_X2_1, ch_X2_1, kernel_size=[1, 7], padding=[0, 3]),BasicConv2d(ch_X2_1, ch_X2_2, kernel_size=[7, 1], padding=[3, 0])  # 保证输出大小等于输入大小)# avgpool+conv1*1(128)self.branch4 = nn.Sequential(nn.MaxPool2d(kernel_size=3, stride=1, padding=1),BasicConv2d(in_channels, pool_proj, kernel_size=1))def forward(self, x):branch1 = self.branch1(x)branch2 = self.branch2(x)branch3 = self.branch3(x)branch4 = self.branch4(x)# 拼接outputs = [branch1, branch2, branch3, branch4]return torch.cat(outputs, 1)# InceptionV4C:BasicConv2d+MaxPool2d
class InceptionV4C(nn.Module):def __init__(self, in_channels, ch1x1, ch3x3red, ch3x3, ch3x3redX2, ch3x3X2_1, ch3x3X2_2, ch3x3X2_3,pool_proj):super(InceptionV4C, self).__init__()# conv1*1(256)self.branch1 = BasicConv2d(in_channels, ch1x1, kernel_size=1)# conv1*1(384)+conv1*3(256) & conv3*1(256)self.branch2_0 = BasicConv2d(in_channels, ch3x3red, kernel_size=1)self.branch2_1 = BasicConv2d(ch3x3red, ch3x3, kernel_size=[1, 3], padding=[0, 1])self.branch2_2 = BasicConv2d(ch3x3red, ch3x3, kernel_size=[3, 1], padding=[1, 0])# conv1*1(384)+conv1*3(448)+conv3*1(512)+conv3*1(256) & conv7*1(256)self.branch3_0 = nn.Sequential(BasicConv2d(in_channels, ch3x3redX2, kernel_size=1),BasicConv2d(ch3x3redX2, ch3x3X2_1, kernel_size=[1, 3], padding=[0, 1]),BasicConv2d(ch3x3X2_1, ch3x3X2_2, kernel_size=[3, 1], padding=[1, 0]),)self.branch3_1 = BasicConv2d(ch3x3X2_2, ch3x3X2_3, kernel_size=[1, 3], padding=[0, 1])self.branch3_2 = BasicConv2d(ch3x3X2_2, ch3x3X2_3, kernel_size=[3, 1], padding=[1, 0])# avgpool+conv1*1(256)self.branch4 = nn.Sequential(nn.MaxPool2d(kernel_size=3, stride=1, padding=1),BasicConv2d(in_channels, pool_proj, kernel_size=1))def forward(self, x):branch1 = self.branch1(x)branch2_0 = self.branch2_0(x)branch2 = torch.cat([self.branch2_1(branch2_0), self.branch2_2(branch2_0)], dim=1)branch3_0 = self.branch3_0(x)branch3 = torch.cat([self.branch3_1(branch3_0), self.branch3_2(branch3_0)], dim=1)branch4 = self.branch4(x)# 拼接outputs = [branch1, branch2, branch3, branch4]return torch.cat(outputs, 1)# redutionA:BasicConv2d+MaxPool2d
class redutionA(nn.Module):def __init__(self, in_channels, k, l, m, n):super(redutionA, self).__init__()# conv3*3(n stride2 valid)self.branch1 = nn.Sequential(BasicConv2d(in_channels, n, kernel_size=3, stride=2),)# conv1*1(k)+conv3*3(l)+conv3*3(m stride2 valid)self.branch2 = nn.Sequential(BasicConv2d(in_channels, k, kernel_size=1),BasicConv2d(k, l, kernel_size=3, padding=1),BasicConv2d(l, m, kernel_size=3, stride=2))# maxpool3*3(stride2 valid)self.branch3 = nn.Sequential(nn.MaxPool2d(kernel_size=3, stride=2))def forward(self, x):branch1 = self.branch1(x)branch2 = self.branch2(x)branch3 = self.branch3(x)# 拼接outputs = [branch1,branch2, branch3]return torch.cat(outputs, 1)# redutionB:BasicConv2d+MaxPool2d
class redutionB(nn.Module):def __init__(self, in_channels, ch3x3red, ch3x3, ch_redX2, ch_X2):super(redutionB, self).__init__()# conv1*1(192)+conv3*3(192 stride2 valid)self.branch1 = nn.Sequential(BasicConv2d(in_channels, ch3x3red, kernel_size=1),BasicConv2d(ch3x3red, ch3x3, kernel_size=3, stride=2))# conv1*1(256)+conv1*7(256)+conv7*1(320)+conv3*3(320 stride2 valid)self.branch2 = nn.Sequential(BasicConv2d(in_channels, ch_redX2, kernel_size=1),BasicConv2d(ch_redX2, ch_redX2, kernel_size=(1, 7), padding=(0, 3)),# 保证输出大小等于输入大小BasicConv2d(ch_redX2, ch_X2, kernel_size=(7, 1), padding=(3, 0)),BasicConv2d(ch_X2, ch_X2, kernel_size=3, stride=2))#  maxpool3*3(stride2 valid)self.branch3 = nn.Sequential(nn.MaxPool2d(kernel_size=3, stride=2))def forward(self, x):branch1 = self.branch1(x)branch2 = self.branch2(x)branch3 = self.branch3(x)# 拼接outputs = [branch1,branch2, branch3]return torch.cat(outputs, 1)# Stem:BasicConv2d+MaxPool2d
class Stem(nn.Module):def __init__(self, in_channels, out_channels):super(Stem, self).__init__()# conv3*3(32 stride2 valid)self.conv1 = BasicConv2d(in_channels, 32, kernel_size=3, stride=2)# conv3*3(32 valid)self.conv2 = BasicConv2d(32, 32, kernel_size=3)# conv3*3(64)self.conv3 = BasicConv2d(32, 64, kernel_size=3, padding=1)# maxpool3*3(stride2 valid) & conv3*3(96 stride2 valid)self.maxpool4 = nn.MaxPool2d(kernel_size=3, stride=2)self.conv4 = BasicConv2d(64, 96, kernel_size=3, stride=2)# conv1*1(64)+conv3*3(96 valid)self.conv5_1_1 = BasicConv2d(160, 64, kernel_size=1)self.conv5_1_2 = BasicConv2d(64, 96, kernel_size=3)# conv1*1(64)+conv7*1(64)+conv1*7(64)+conv3*3(96 valid)self.conv5_2_1 = BasicConv2d(160, 64, kernel_size=1)self.conv5_2_2 = BasicConv2d(64, 64, kernel_size=(7, 1), padding=(3, 0))self.conv5_2_3 = BasicConv2d(64, 64, kernel_size=(1, 7), padding=(0, 3))self.conv5_2_4 = BasicConv2d(64, 96, kernel_size=3)# conv3*3(192 valid) & maxpool3*3(stride2 valid)self.conv6 = BasicConv2d(192, 192, kernel_size=3, stride=2)self.maxpool6 = nn.MaxPool2d(kernel_size=3, stride=2)def forward(self, x):x1_1 = self.maxpool4(self.conv3(self.conv2(self.conv1(x))))x1_2 = self.conv4(self.conv3(self.conv2(self.conv1(x))))x1 = torch.cat([x1_1, x1_2], 1)x2_1 = self.conv5_1_2(self.conv5_1_1(x1))x2_2 = self.conv5_2_4(self.conv5_2_3(self.conv5_2_2(self.conv5_2_1(x1))))x2 = torch.cat([x2_1, x2_2], 1)x3_1 = self.conv6(x2)x3_2 = self.maxpool6(x2)x3 = torch.cat([x3_1, x3_2], 1)return x3# 卷积组: Conv2d+BN+ReLU
class BasicConv2d(nn.Module):def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0):super(BasicConv2d, self).__init__()self.conv = nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding)self.bn = nn.BatchNorm2d(out_channels)self.relu = nn.ReLU(inplace=True)def forward(self, x):x = self.conv(x)x = self.bn(x)x = self.relu(x)return xif __name__ == '__main__':device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")model = GoogLeNetV4().to(device)summary(model, input_size=(3, 229, 229))

summary可以打印网络结构和参数,方便查看搭建好的网络结构。


总结

尽可能简单、详细的介绍了InceptionV4的改进方案,讲解了GoogLeNet(InceptionV4)模型的结构和pytorch代码。

本文来自互联网用户投稿,该文观点仅代表作者本人,不代表本站立场。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。如若转载,请注明出处:http://www.mzph.cn/news/150184.shtml

如若内容造成侵权/违法违规/事实不符,请联系多彩编程网进行投诉反馈email:809451989@qq.com,一经查实,立即删除!

相关文章

准备篇(四)HTTP 基本原理

URI 和 URLURIURLURI vs URLHTTP 和 HTTPS超文本HTTPHTTP 请求与响应HTTPS你是否想过,在浏览器中敲入 URL 到 获取网页内容 之间发生了什么? 了解这些,有助于进一步了解爬虫的基本原理。 URI 和 URL URI(Uniform Resource Identifier),即统一资源标识符;URL(Universa…

Ubuntu 20.04 LTS设置系统虚拟内存大小

1、查看系统是否已经设置虚拟内存 swapon -s swapon --show 2、通过执行上述命令返回的结果来判断等下是否需要重启(看系统used没有) Filename Type Size Used Priority /swapfile fil…

宽瞬时带宽放大器SKY66051-11、SKY66052-11、SKY66041-11、SKY66317-11(RF)适用于通讯网络

一、2300至2700 MHz宽瞬时带宽高增益线性驱动放大器:SKY66051-11 SKY66051-11是一款具有高增益和高线性度的宽瞬时带宽、完全输入/输出匹配驱动放大器。通过使用外部元件,增益可在30dB至36dB范围内调整。紧凑型33 mm PA专为工作频率为2300至2700 MHz的4…

C++ 递增/递减运算符重载

作用&#xff1a; 通过重载递增运算符&#xff0c;实现自己的整型数据 总结&#xff1a; 前置递增返回引用&#xff0c;后置递增返回值 递增 #include<iostream> using namespace std;class MyInteger { private:int m_Num 0; public:friend ostream& operator<…

Skywalking流程分析_9(JDK类库中增强流程)

前言 之前的文章详细介绍了关于非JDK类库的静态方法、构造方法、实例方法的增强拦截流程&#xff0c;本文会详细分析JDK类库中的类是如何被增强拦截的 回到最开始的SkyWalkingAgent#premain try {/** 里面有个重点逻辑 把一些类注入到Boostrap类加载器中 为了解决Bootstrap类…

开源与闭源:大模型时代的技术交融与商业平衡

一、开源和闭源的优劣势比较 1.1 开源 优势&#xff1a; 1.技术共享与吸引人才&#xff1a; 开源促进了技术共享&#xff0c;吸引了全球范围内的人才参与大模型的发展&#xff0c;形成了庞大的开发者社区。 2.推动创新&#xff1a; 开源模式鼓励开发者共同参与&#xff0c;推动…

uni-app:如何配置uni.request请求的超时响应时间(全局+局部)

方法一&#xff1a;全局配置响应时间 一、进入项目的manifest.json的代码视图模块 二、写入代码 "networkTimeout": {"request": 5000 }, 表示现在request请求响应时间最多位5秒 方法二&#xff1a;局部设置响应时间 一、直接在uni.request中写入属性…

redission源码解读

可以参考这篇文章&#xff0c;讲的比较详细 https://www.cnblogs.com/throwable/p/14264804.html jedis和redission的区别&#xff1f; Jedis和Redisson都是Java语言中常用的Redis客户端库&#xff0c;它们之间的区别如下&#xff1a; 线程模型不同&#xff1a;Jedis是单线程…

Docker打包Python项目

1. 简介 Docker是一种开源的容器化平台&#xff0c;可以将应用程序及其依赖项打包到一个轻量级、可移植的容器中。通过使用Docker&#xff0c;可以简化Python项目的部署和运行&#xff0c;提高开发效率和应用程序的可移植性。 本文将介绍如何使用Docker来打包Python项目。我们…

微服务下整合knife4j接口文档

前言:本文旨在解决微服务下通过网关访问所用服务的knife4j文档&#xff0c;无需再通过其他服务单独访问 功能模块配置&#xff1a; 1.配置类&#xff1a; 在这个文件中注意下basePackage的扫描路径&#xff0c;修改为对应controller下的路径。 Configuration EnableSwagger…

打造自己的3D模型AI 自动纹理工具

在线工具推荐&#xff1a; 三维数字孪生场景工具 - GLTF/GLB在线编辑器 - Three.js AI自动纹理化开发 - YOLO 虚幻合成数据生成器 - 3D模型在线转换 - 3D模型预览图生成服务 为 3D 模型创建纹理可能比您想象的要容易。虽然注意细节很重要&#xff0c;但有很多方法可以制…

基于SSM框架的WEB管理系统

一、考核题目 基于SSM框架的WEB管理系统的设计和编码: 1、图书管理&#xff0c;客户管理等,自选也可; 2、必须使用 SpringSpringMVCMybatis框架结构; 3、必须完成三个以上业务模块的增删改查功能; 4、项目文档&#xff1b; 二、考核工具、环境 1、使用Eclipse、MyEclips…

html常用的标签

基本结构标签 <!DOCTYPE>&#xff1a; 定义 HTML 文档类型。<html>&#xff1a; HTML 文档的根元素。<head>&#xff1a; 文档的头部&#xff0c;包含了元数据和引用的外部资源。<title>&#xff1a; 定义网页标题&#xff0c;显示在浏览器标签上。&l…

超详细~25考研规划~感恩现在努力的你!!!

25考研规划 俄语&#xff0c;翻译过来叫我爱你 考试时间 第一天 8.30-11.30政治——100分 2.00-5.00英语——100分 第二天 8.30-11.30数学——150分 2.00-5.00专业课——150分 1.什么是25考研 将在2024年12月参加考研&#xff0c;2025年本科毕业&#xff0c;9月读研究…

使用k8s部署一个简单MySQL8服务,但是不能挂载

创建mysql的yaml文件 cat << eof > mysql.yaml apiVersion: apps/v1 # 通过kubectl explain deployment命令查看版本 kind: Deployment # 资源类型 metadata:name: mysql-deployment # 资源…

java基础练习缺少项目?看这篇文章就够了(下)!

公众号&#xff1a;全干开发 。 专注分享简洁但高质量的动图技术文章&#xff01; 回顾 在上节内容中&#xff0c;我们实现了用户开户的功能createAccount public void start(){System.out.println("欢迎您进入到了ATM系统");System.out.println("1、用户登录&…

git基本用法和操作

文章目录 创建版本库方式&#xff1a;Git常用操作命令&#xff1a;远程仓库相关命令分支(branch)操作相关命令版本(tag)操作相关命令子模块(submodule)相关操作命令忽略一些文件、文件夹不提交其他常用命令 创建版本库方式&#xff1a; 创建文件夹 在目录下 右键 Git Bush H…

如何使用$APPEALS法,分析用户期待?

$APPEALS分析法是一种用于分析用户期待和需求的方法&#xff0c;它可以帮助企业全方位多角度地了解客户对产品的期望&#xff0c;有助于企业多维度有侧重地调整市场规划和产品改进策略&#xff0c;帮助企业打造优势产品&#xff0c;提高市场竞争力。 下面是使用$APPEALS分析法来…

键盘控制ROS车运动

键盘控制ROS车运动 上位机 使用pyseria库与stm32单片机进行通信控制 #!/usr/bin/env python # -*- coding: utf-8 -*import sys, select, termios, tty import serialmsg """ ---------------------------w a x ds w : x a : y s : -x …

广西柳州机械异形零部件三维扫描3D抄数全尺寸测绘建模-CASAIM中科广电

一、背景介绍 复杂机械异形零部件具有不规则的形状和复杂的结构&#xff0c;给生产制造带来了很大的检测难度。为了确保零部件的制造质量和精度&#xff0c;需要对零部件进行全面的尺寸检测和分析。 CASAIM三维扫描仪在机械异形零部件全尺寸检测应用可以实现对机械异形零部件…