🤵 Author ：Horizon John
✨ 编程技巧篇：各种操作小结
🎇 机器视觉篇：会变魔术 OpenCV
💥 深度学习篇：简单入门 PyTorch
🏆 神经网络篇：经典网络模型
💻 算法篇：再忙也别忘了 LeetCode
[ 图像分类 ] 经典网络模型2——GoogLeNet 详解与复现
🚀 GoogLeNet
GoogLeNet 是2014年 Christian Szegedy 提出的一种 全新 的深度学习结构 ；
在这之前的 AlexNet 、VGG 等结构都是通过 增大网络深度 来获得更好的训练效果 ；
但层数的增加会带来很多负作用，如 overfit 、梯度消失 、梯度爆炸 等 ；
作为2014年 ImageNet 挑战赛(ILSVRC) classification 任务中的 冠军，超过了同年的 VGG 网络模型 ；
模型参数量较12年冠军 AlexNet 只有其 1/12 ；
叫 GoogLeNet 而不是 GoogleNet 是为了致敬之前的 LeNet 网络 ；
🔗 论文地址：Going deeper with convolutions
🚀 GoogLeNet 详解
🎨 Inception module
保持 计算成本 的同时增加网络的 深度和宽度 ，从而提升训练结果 ；
文章中提到的简单的增加网络的深度和宽度会有两个缺点 ：
（1）尺寸越大，参数的数量越多，网络更容易发生过拟合 ；
（2）计算资源的使用显著增加 ；
🚩 核心思想
如何将卷积网络中的最优局部稀疏结构近似化，并由现成的密集组件覆盖 ；
特点：
（1）使用 1x1 的卷积来进行 升降维 ；
（2）在多个尺寸上 同时进行卷积再聚合 ；
结构（a）由于 5 x 5 卷积 的存在会导致参数量的增加 ；
基于此提出了结构（b），在 3 x 3 和 5 x 5 卷积之前使用 1 x 1 卷积 ；
1 x 1 卷积：
受到 Lin 等人提出的 Network-in-Network 的影响，为提高神经网络的表示能力，每次的 1 x 1 卷积都使用校正线性激活；
它们主要用作 降维 模块，以消除计算瓶颈，同时达到增加 网络深度和宽度 的目的，而不会造成显著的性能损失;
🎨 GoogLeNet 网络结构
在 Inception (4a) 和 (4d) 模块设置了 softmax 输出作为辅助网络用来 梯度传播 ；
将其损失 与 具有折扣权值的网络 总损失相加（辅助分类器的损失加权为 0.3）；
在测试时，两个辅助网络会被去掉；
输入大小为（224 x 224 x 3）
conv1： 卷积（7 x 7）、步长为 2 ， 输出（112 x 112 x 64）
maxpool1： 卷积（3 x 3）、步长为 2 ，输出（56 x 56 x 64）
conv2： 卷积（1 x 1）、步长为 1 ，输出（56 x 56 x 64）
conv3： 卷积（3 x 3）、步长为 1 ，输出（56 x 56 x 192）
maxpool2： 卷积（3 x 3）、步长为 2 ，输出（28 x 28 x 192）
inception3a： 输出（28 x 28 x 256）
inception3b： 输出（28 x 28 x 480）
maxpool3： 输出（14 x 14 x 480）
inception4a： 输出（14 x 14 x 512）
inception4b： 输出（14 x 14 x 512）
inception4c： 输出（14 x 14 x 512）
inception4d： 输出（14 x 14 x 528）
inception4e： 输出（14 x 14 x 832）
maxpool4： 输出（7 x 7 x 832）
inception5a： 输出（7 x 7 x 832）
inception5b： 输出（7 x 7 x 1024）
aovgpol： 输出（1 x 1 x 1024）
dropout： 40%
Linear： num_classes
🚀 GoogLeNet 复现
# Here is the code ：
import torch
import torch.nn as nn
import torch.nn.functional as F
from torchinfo import summary
class BasicConv2d(nn.Module):
def __init__(self, in_channels, out_channels, **kwargs): # **kwargs: 不定长度键值对
super(BasicConv2d, self).__init__()
self.conv = nn.Conv2d(in_channels, out_channels, bias=False, **kwargs)
self.bn = nn.BatchNorm2d(out_channels, eps=0.001)
def forward(self, x):
x = self.conv(x)
x = self.bn(x)
return F.relu(x, inplace=True)
class Inception(nn.Module):
def __init__(self, in_channels, ch1x1, ch3x3red, ch3x3, ch5x5red, ch5x5, pool_proj):
super(Inception, self).__init__()
self.branch1 = BasicConv2d(in_channels, ch1x1, kernel_size=1)
self.branch2 = nn.Sequential(
BasicConv2d(in_channels, ch3x3red, kernel_size=1),
BasicConv2d(ch3x3red, ch3x3, kernel_size=3, padding=1)
)
self.branch3 = nn.Sequential(
BasicConv2d(in_channels, ch5x5red, kernel_size=1),
BasicConv2d(ch5x5red, ch5x5, kernel_size=3, padding=1)
)
self.branch4 = nn.Sequential(
nn.MaxPool2d(kernel_size=3, stride=1, padding=1, ceil_mode=True),
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)
outputs = torch.cat((branch1, branch2, branch3, branch4), 1)
return outputs
class InceptionAux(nn.Module):
def __init__(self, in_channels, num_classes):
super(InceptionAux, self).__init__()
self.avgPool = nn.AvgPool2d(kernel_size=5, stride=3)
self.conv = BasicConv2d(in_channels, 128, kernel_size=1)
self.fc1 = nn.Linear(2048, 1024)
self.fc2 = nn.Linear(1024, num_classes)
def forward(self, x):
x = self.avgPool(x)
x = self.conv(x)
x = torch.flatten(x, 1)
x = F.relu(self.fc1(x), inplace=True)
x = F.dropout(x, 0.7, training=self.training)
out = self.fc2(x)
return out
class GoogLeNet(nn.Module):
def __init__(self, num_classes=1000, aux_logits=True, init_weights=False):
super(GoogLeNet, self).__init__()
self.aux_logits = aux_logits
self.conv1 = BasicConv2d(3, 64, kernel_size=7, stride=2, padding=3)
self.maxpool1 = nn.MaxPool2d(3, stride=2, ceil_mode=True)
self.conv2 = BasicConv2d(64, 64, kernel_size=1)
self.conv3 = BasicConv2d(64, 192, kernel_size=3, padding=1)
self.maxpool2 = nn.MaxPool2d(3, stride=2, ceil_mode=True)
self.inception3a = Inception(192, 64, 96, 128, 16, 32, 32)
self.inception3b = Inception(256, 128, 128, 192, 32, 96, 64)
self.maxpool3 = nn.MaxPool2d(3, stride=2, ceil_mode=True)
self.inception4a = Inception(480, 192, 96, 208, 16, 48, 64)
self.inception4b = Inception(512, 160, 112, 224, 24, 64, 64)
self.inception4c = Inception(512, 128, 128, 256, 24, 64, 64)
self.inception4d = Inception(512, 112, 144, 288, 32, 64, 64)
self.inception4e = Inception(528, 256, 160, 320, 32, 128, 128)
self.maxpool4 = nn.MaxPool2d(2, stride=2, ceil_mode=True)
self.inception5a = Inception(832, 256, 160, 320, 32, 128, 128)
self.inception5b = Inception(832, 384, 192, 384, 48, 128, 128)
if aux_logits:
self.aux1 = InceptionAux(512, num_classes)
self.aux2 = InceptionAux(528, num_classes)
else:
self.aux1 = None
self.aux2 = None
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
self.dropout = nn.Dropout(0.4)
self.fc = nn.Linear(1024, num_classes)
if init_weights:
self._initialize_weights()
def initialize_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d) or isinstance(m, nn.Linear):
import scipy.stats as stats
X = stats.truncnorm(-2, 2, scale=0.01)
values = torch.as_tensor(X.rvs(m.weight.numel()), dtype=m.weight.dtype)
values = values.view(m.weight.size())
with torch.no_grad():
m.weight.copy_(values)
elif isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
def forward(self, x):
x = self.conv1(x)
x = self.maxpool1(x)
x = self.conv2(x)
x = self.conv3(x)
x = self.maxpool2(x)
x = self.inception3a(x)
x = self.inception3b(x)
x = self.maxpool3(x)
x = self.inception4a(x)
if self.training and self.aux_logits:
aux1 = self.aux1(x)
x = self.inception4b(x)
x = self.inception4c(x)
x = self.inception4d(x)
if self.training and self.aux_logits:
aux2 = self.aux2(x)
x = self.inception4e(x)
x = self.maxpool4(x)
x = self.inception5a(x)
x = self.inception5b(x)
x = self.avgpool(x)
x = torch.flatten(x, 1)
x = self.dropout(x)
out = self.fc(x)
if self.training and self.aux_logits:
return out, aux1, aux2
else:
return out
def test():
net = GoogLeNet()
out, aux1, aux2 = net(torch.randn(1, 3, 224, 224))
print(out.size())
summary(net, (1, 3, 224, 224))
if __name__ == '__main__':
test()
输出结果：
torch.Size([1, 1000])
==========================================================================================
Layer (type:depth-idx) Output Shape Param #
==========================================================================================
GoogLeNet -- --
├─BasicConv2d: 1-1 [1, 64, 112, 112] --
│ └─Conv2d: 2-1 [1, 64, 112, 112] 9,408
│ └─BatchNorm2d: 2-2 [1, 64, 112, 112] 128
├─MaxPool2d: 1-2 [1, 64, 56, 56] --
├─BasicConv2d: 1-3 [1, 64, 56, 56] --
│ └─Conv2d: 2-3 [1, 64, 56, 56] 4,096
│ └─BatchNorm2d: 2-4 [1, 64, 56, 56] 128
├─BasicConv2d: 1-4 [1, 192, 56, 56] --
│ └─Conv2d: 2-5 [1, 192, 56, 56] 110,592
│ └─BatchNorm2d: 2-6 [1, 192, 56, 56] 384
├─MaxPool2d: 1-5 [1, 192, 28, 28] --
├─Inception: 1-6 [1, 256, 28, 28] --
│ └─BasicConv2d: 2-7 [1, 64, 28, 28] --
│ │ └─Conv2d: 3-1 [1, 64, 28, 28] 12,288
│ │ └─BatchNorm2d: 3-2 [1, 64, 28, 28] 128
│ └─Sequential: 2-8 [1, 128, 28, 28] --
│ │ └─BasicConv2d: 3-3 [1, 96, 28, 28] 18,624
│ │ └─BasicConv2d: 3-4 [1, 128, 28, 28] 110,848
│ └─Sequential: 2-9 [1, 32, 28, 28] --
│ │ └─BasicConv2d: 3-5 [1, 16, 28, 28] 3,104
│ │ └─BasicConv2d: 3-6 [1, 32, 28, 28] 4,672
│ └─Sequential: 2-10 [1, 32, 28, 28] --
│ │ └─MaxPool2d: 3-7 [1, 192, 28, 28] --
│ │ └─BasicConv2d: 3-8 [1, 32, 28, 28] 6,208
├─Inception: 1-7 [1, 480, 28, 28] --
│ └─BasicConv2d: 2-11 [1, 128, 28, 28] --
│ │ └─Conv2d: 3-9 [1, 128, 28, 28] 32,768
│ │ └─BatchNorm2d: 3-10 [1, 128, 28, 28] 256
│ └─Sequential: 2-12 [1, 192, 28, 28] --
│ │ └─BasicConv2d: 3-11 [1, 128, 28, 28] 33,024
│ │ └─BasicConv2d: 3-12 [1, 192, 28, 28] 221,568
│ └─Sequential: 2-13 [1, 96, 28, 28] --
│ │ └─BasicConv2d: 3-13 [1, 32, 28, 28] 8,256
│ │ └─BasicConv2d: 3-14 [1, 96, 28, 28] 27,840
│ └─Sequential: 2-14 [1, 64, 28, 28] --
│ │ └─MaxPool2d: 3-15 [1, 256, 28, 28] --
│ │ └─BasicConv2d: 3-16 [1, 64, 28, 28] 16,512
├─MaxPool2d: 1-8 [1, 480, 14, 14] --
├─Inception: 1-9 [1, 512, 14, 14] --
│ └─BasicConv2d: 2-15 [1, 192, 14, 14] --
│ │ └─Conv2d: 3-17 [1, 192, 14, 14] 92,160
│ │ └─BatchNorm2d: 3-18 [1, 192, 14, 14] 384
│ └─Sequential: 2-16 [1, 208, 14, 14] --
│ │ └─BasicConv2d: 3-19 [1, 96, 14, 14] 46,272
│ │ └─BasicConv2d: 3-20 [1, 208, 14, 14] 180,128
│ └─Sequential: 2-17 [1, 48, 14, 14] --
│ │ └─BasicConv2d: 3-21 [1, 16, 14, 14] 7,712
│ │ └─BasicConv2d: 3-22 [1, 48, 14, 14] 7,008
│ └─Sequential: 2-18 [1, 64, 14, 14] --
│ │ └─MaxPool2d: 3-23 [1, 480, 14, 14] --
│ │ └─BasicConv2d: 3-24 [1, 64, 14, 14] 30,848
├─Inception: 1-10 [1, 512, 14, 14] --
│ └─BasicConv2d: 2-19 [1, 160, 14, 14] --
│ │ └─Conv2d: 3-25 [1, 160, 14, 14] 81,920
│ │ └─BatchNorm2d: 3-26 [1, 160, 14, 14] 320
│ └─Sequential: 2-20 [1, 224, 14, 14] --
│ │ └─BasicConv2d: 3-27 [1, 112, 14, 14] 57,568
│ │ └─BasicConv2d: 3-28 [1, 224, 14, 14] 226,240
│ └─Sequential: 2-21 [1, 64, 14, 14] --
│ │ └─BasicConv2d: 3-29 [1, 24, 14, 14] 12,336
│ │ └─BasicConv2d: 3-30 [1, 64, 14, 14] 13,952
│ └─Sequential: 2-22 [1, 64, 14, 14] --
│ │ └─MaxPool2d: 3-31 [1, 512, 14, 14] --
│ │ └─BasicConv2d: 3-32 [1, 64, 14, 14] 32,896
├─Inception: 1-11 [1, 512, 14, 14] --
│ └─BasicConv2d: 2-23 [1, 128, 14, 14] --
│ │ └─Conv2d: 3-33 [1, 128, 14, 14] 65,536
│ │ └─BatchNorm2d: 3-34 [1, 128, 14, 14] 256
│ └─Sequential: 2-24 [1, 256, 14, 14] --
│ │ └─BasicConv2d: 3-35 [1, 128, 14, 14] 65,792
│ │ └─BasicConv2d: 3-36 [1, 256, 14, 14] 295,424
│ └─Sequential: 2-25 [1, 64, 14, 14] --
│ │ └─BasicConv2d: 3-37 [1, 24, 14, 14] 12,336
│ │ └─BasicConv2d: 3-38 [1, 64, 14, 14] 13,952
│ └─Sequential: 2-26 [1, 64, 14, 14] --
│ │ └─MaxPool2d: 3-39 [1, 512, 14, 14] --
│ │ └─BasicConv2d: 3-40 [1, 64, 14, 14] 32,896
├─Inception: 1-12 [1, 528, 14, 14] --
│ └─BasicConv2d: 2-27 [1, 112, 14, 14] --
│ │ └─Conv2d: 3-41 [1, 112, 14, 14] 57,344
│ │ └─BatchNorm2d: 3-42 [1, 112, 14, 14] 224
│ └─Sequential: 2-28 [1, 288, 14, 14] --
│ │ └─BasicConv2d: 3-43 [1, 144, 14, 14] 74,016
│ │ └─BasicConv2d: 3-44 [1, 288, 14, 14] 373,824
│ └─Sequential: 2-29 [1, 64, 14, 14] --
│ │ └─BasicConv2d: 3-45 [1, 32, 14, 14] 16,448
│ │ └─BasicConv2d: 3-46 [1, 64, 14, 14] 18,560
│ └─Sequential: 2-30 [1, 64, 14, 14] --
│ │ └─MaxPool2d: 3-47 [1, 512, 14, 14] --
│ │ └─BasicConv2d: 3-48 [1, 64, 14, 14] 32,896
├─Inception: 1-13 [1, 832, 14, 14] --
│ └─BasicConv2d: 2-31 [1, 256, 14, 14] --
│ │ └─Conv2d: 3-49 [1, 256, 14, 14] 135,168
│ │ └─BatchNorm2d: 3-50 [1, 256, 14, 14] 512
│ └─Sequential: 2-32 [1, 320, 14, 14] --
│ │ └─BasicConv2d: 3-51 [1, 160, 14, 14] 84,800
│ │ └─BasicConv2d: 3-52 [1, 320, 14, 14] 461,440
│ └─Sequential: 2-33 [1, 128, 14, 14] --
│ │ └─BasicConv2d: 3-53 [1, 32, 14, 14] 16,960
│ │ └─BasicConv2d: 3-54 [1, 128, 14, 14] 37,120
│ └─Sequential: 2-34 [1, 128, 14, 14] --
│ │ └─MaxPool2d: 3-55 [1, 528, 14, 14] --
│ │ └─BasicConv2d: 3-56 [1, 128, 14, 14] 67,840
├─MaxPool2d: 1-14 [1, 832, 7, 7] --
├─Inception: 1-15 [1, 832, 7, 7] --
│ └─BasicConv2d: 2-35 [1, 256, 7, 7] --
│ │ └─Conv2d: 3-57 [1, 256, 7, 7] 212,992
│ │ └─BatchNorm2d: 3-58 [1, 256, 7, 7] 512
│ └─Sequential: 2-36 [1, 320, 7, 7] --
│ │ └─BasicConv2d: 3-59 [1, 160, 7, 7] 133,440
│ │ └─BasicConv2d: 3-60 [1, 320, 7, 7] 461,440
│ └─Sequential: 2-37 [1, 128, 7, 7] --
│ │ └─BasicConv2d: 3-61 [1, 32, 7, 7] 26,688
│ │ └─BasicConv2d: 3-62 [1, 128, 7, 7] 37,120
│ └─Sequential: 2-38 [1, 128, 7, 7] --
│ │ └─MaxPool2d: 3-63 [1, 832, 7, 7] --
│ │ └─BasicConv2d: 3-64 [1, 128, 7, 7] 106,752
├─Inception: 1-16 [1, 1024, 7, 7] --
│ └─BasicConv2d: 2-39 [1, 384, 7, 7] --
│ │ └─Conv2d: 3-65 [1, 384, 7, 7] 319,488
│ │ └─BatchNorm2d: 3-66 [1, 384, 7, 7] 768
│ └─Sequential: 2-40 [1, 384, 7, 7] --
│ │ └─BasicConv2d: 3-67 [1, 192, 7, 7] 160,128
│ │ └─BasicConv2d: 3-68 [1, 384, 7, 7] 664,320
│ └─Sequential: 2-41 [1, 128, 7, 7] --
│ │ └─BasicConv2d: 3-69 [1, 48, 7, 7] 40,032
│ │ └─BasicConv2d: 3-70 [1, 128, 7, 7] 55,552
│ └─Sequential: 2-42 [1, 128, 7, 7] --
│ │ └─MaxPool2d: 3-71 [1, 832, 7, 7] --
│ │ └─BasicConv2d: 3-72 [1, 128, 7, 7] 106,752
├─AdaptiveAvgPool2d: 1-17 [1, 1024, 1, 1] --
├─Dropout: 1-18 [1, 1024] --
├─Linear: 1-19 [1, 1000] 1,025,000
==========================================================================================
Total params: 6,624,904
Trainable params: 6,624,904
Non-trainable params: 0
Total mult-adds (G): 1.50
==========================================================================================
Input size (MB): 0.60
Forward/backward pass size (MB): 51.63
Params size (MB): 26.50
Estimated Total Size (MB): 78.73
==========================================================================================
5532
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
