这是我们架构图
导入相关的包
import torch
from torch import nn
from torchinfo import summary
定义我们基础的卷积层
我们每一个基础卷积层都是经过了一个卷积层，然后进过bn层，然后再用ReLu函数进行激活
class BassicConv2d(nn.Module):
#**kwargs就是谁用全部的参数
def __init__(self,in_channels,out_channels,**kwargs):
super().__init__()
self.conv=nn.Sequential(
#除了bias也就是偏差我们不需要，其他的参数我们都传入默认值
nn.Conv2d(in_channels,out_channels,bias=False,**kwargs)
,nn.BatchNorm2d(out_channels)
,nn.ReLU(inplace=True)
)
def forward(self,x):
x=self.conv(x)
return x
#测试BassicConv2d
BassicConv2d(2,10,kernel_size=3)
BassicConv2d(
(conv): Sequential(
(0): Conv2d(2, 10, kernel_size=(3, 3), stride=(1, 1), bias=False)
(1): BatchNorm2d(10, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace=True)
)
)
定义inception层
这是我们的inception层中具体的定义，也就是说有四层并行的结构
class Inception(nn.Module):
def __init__(self
#限制输入的类型是int
,in_channels :int
,ch1x1 :int
,ch3x3red :int
,ch3x3 :int
,ch5x5red :int
,ch5x5 :int
,pool_proj :int
):
super().__init__()
#1x1
self.branch1=BassicConv2d(in_channels,ch1x1,kernel_size=1)
#1x1+3x3
self.branch2=nn.Sequential(
BassicConv2d(in_channels,ch3x3red,kernel_size=1)
,BassicConv2d(ch3x3red,ch3x3,kernel_size=3,padding=1)
)
#1x1+5x5
self.branch3=nn.Sequential(
BassicConv2d(in_channels,ch5x5red,kernel_size=1)
,BassicConv2d(ch5x5red,ch5x5,kernel_size=5,padding=2)
)
#pool+1x1
self.branch4=nn.Sequential(
nn.MaxPool2d(kernel_size=3,stride=1,padding=1,ceil_mode=True)
,BassicConv2d(in_channels,pool_proj,kernel_size=1)
)
def forward(self,x):
#并行的四层结构
branch1=self.branch1(x)
branch2=self.branch2(x)
branch3=self.branch3(x)
branch4=self.branch4(x)
outputs=[branch1,branch2,branch3,branch4]
return torch.cat(outputs,1)#纵向进行合并
#测试
in3a=Inception(192,64,96,128,16,32,32)
data=torch.ones(10,192,28,28)
in3a(data).size()
torch.Size([10, 256, 28, 28])
定义辅助分类器
| 层 | 核尺寸/步长 | 输入4a输出尺寸 | 输入4d输出尺寸 |
|---|---|---|---|
| 平均池化层 | 5x5/3 | 4x4x512 | 4x4x528 |
| 卷积层+ReLU | 1x1/1 | 4x4x128 | 4x4x128 |
| 全连接层+ReLU | 1024 | 1024 | |
| Dropout(70%) | 1024 | 1024 | |
| 全连接层+Softmax | 1000 | 1000 |
#auxiliay classifier
#辅助分类器
class AuxClf(nn.Module):
def __init__(self,in_channels :int ,num_classes:int ,**kwargs):
super().__init__()
self.feature_=nn.Sequential(
nn.AvgPool2d(kernel_size=5,stride=3)
,BassicConv2d(in_channels,128,kernel_size=1)
)
self.clf_=nn.Sequential(
nn.Linear(4*4*128,1024)
,nn.ReLU(inplace=True)
,nn.Dropout(0.7)
,nn.Linear(1024,num_classes)
)
def forward(self,x):
x=self.feature_(x)
x=x.view(-1,4*4*128)
x=self.clf_(x)
return x
#4a后的辅助分类器
AuxClf(512,1000)
AuxClf(
(feature_): Sequential(
(0): AvgPool2d(kernel_size=5, stride=3, padding=0)
(1): BassicConv2d(
(conv): Sequential(
(0): Conv2d(512, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace=True)
)
)
)
(clf_): Sequential(
(0): Linear(in_features=2048, out_features=1024, bias=True)
(1): ReLU(inplace=True)
(2): Dropout(p=0.7, inplace=False)
(3): Linear(in_features=1024, out_features=1000, bias=True)
)
)
定义我们的整一张网络
class GoogLeNet(nn.Module):
def __init__(self,num_classes:int =1000,blocks=None):
super().__init__()
if blocks is None:
blocks=[BassicConv2d,Inception,AuxClf]
conv_block=blocks[0]
inception_block=blocks[1]
aux_clf_block=blocks[2]
#block1
self.conv1=conv_block(3,64,kernel_size=7,stride=2,padding=3)
#需要向上取整得到56
#(112-3)/2+1=55.5
self.maxpool1=nn.MaxPool2d(kernel_size=3,stride=2,ceil_mode=True)
#block2
self.conv2=conv_block(64,64,kernel_size=1)
self.conv3=conv_block(64,192,kernel_size=3,padding=1,)
self.maxpool2=nn.MaxPool2d(kernel_size=3,stride=2,ceil_mode=True)
#block3
self.inception3a=inception_block(192,64,96,128,16,32,32)
self.inception3b=inception_block(256,128,128,192,32,96,64)
self.maxpool3=nn.MaxPool2d(kernel_size=3,stride=2,ceil_mode=True)
#block4
self.inception4a=inception_block(480,192,96,208,16,48,64)
self.inception4b=inception_block(512,160,112,224,24,64,64)
self.inception4c=inception_block(512,128,128,256,24,64,64)
self.inception4d=inception_block(512,112,144,288,32,64,64)
self.inception4e=inception_block(528,256,160,320,32,128,128)
self.maxpool4=nn.MaxPool2d(kernel_size=3,stride=2,ceil_mode=True)
#block5
self.inception5a=inception_block(832,256,160,320,32,128,128)
self.inception5b=inception_block(832,384,192,384,48,128,128)
#clf
#我们这里需要输出的特征图尺寸是1*1
self.avgpool=nn.AdaptiveAvgPool2d((1,1))#我需要输出的特征图尺寸
self.dropout=nn.Dropout(0.4)
self.fc=nn.Linear(1024,num_classes)
#auxclf
#辅助分类器
self.aux1=aux_clf_block(512,num_classes)#4a
self.aux2=aux_clf_block(528,num_classes)#4d
def forward(self,x):
#block1
x=self.maxpool1(self.conv1(x))
#block2
x=self.maxpool2(self.conv3(self.conv2(x)))
#block3
x=self.inception3a(x)
x=self.inception3b(x)
x=self.maxpool3(x)
#block4
x=self.inception4a(x)
aux1=self.aux1(x)
x=self.inception4b(x)
x=self.inception4c(x)
x=self.inception4d(x)
aux2=self.aux2(x)
x=self.inception4e(x)
x=self.maxpool4(x)
#block5
x=self.inception5a(x)
x=self.inception5b(x)
#clf
x=self.avgpool(x)#在这个全局平均池化之后特征图尺寸就变成了1x1
x=torch.flatten(x,1)
x=self.dropout(x)
x=self.fc(x)
return x,aux2,aux1
#测试
data=torch.ones(10,3,224,224)
net=GoogLeNet(num_classes=1000)
fc2,fc1,fc0=net(data)
for i in [fc2,fc1,fc0]:
print(i.shape)
torch.Size([10, 1000])
torch.Size([10, 1000])
torch.Size([10, 1000])
summary(net,(10,3,224,224),device="cpu")
===============================================================================================
Layer (type:depth-idx) Output Shape Param #
===============================================================================================
GoogLeNet [10, 1000] --
├─BassicConv2d: 1-1 [10, 64, 112, 112] --
│ └─Sequential: 2-1 [10, 64, 112, 112] --
│ │ └─Conv2d: 3-1 [10, 64, 112, 112] 9,408
│ │ └─BatchNorm2d: 3-2 [10, 64, 112, 112] 128
│ │ └─ReLU: 3-3 [10, 64, 112, 112] --
├─MaxPool2d: 1-2 [10, 64, 56, 56] --
├─BassicConv2d: 1-3 [10, 64, 56, 56] --
│ └─Sequential: 2-2 [10, 64, 56, 56] --
│ │ └─Conv2d: 3-4 [10, 64, 56, 56] 4,096
│ │ └─BatchNorm2d: 3-5 [10, 64, 56, 56] 128
│ │ └─ReLU: 3-6 [10, 64, 56, 56] --
├─BassicConv2d: 1-4 [10, 192, 56, 56] --
│ └─Sequential: 2-3 [10, 192, 56, 56] --
│ │ └─Conv2d: 3-7 [10, 192, 56, 56] 110,592
│ │ └─BatchNorm2d: 3-8 [10, 192, 56, 56] 384
│ │ └─ReLU: 3-9 [10, 192, 56, 56] --
├─MaxPool2d: 1-5 [10, 192, 28, 28] --
├─Inception: 1-6 [10, 256, 28, 28] --
│ └─BassicConv2d: 2-4 [10, 64, 28, 28] --
│ │ └─Sequential: 3-10 [10, 64, 28, 28] 12,416
│ └─Sequential: 2-5 [10, 128, 28, 28] --
│ │ └─BassicConv2d: 3-11 [10, 96, 28, 28] 18,624
│ │ └─BassicConv2d: 3-12 [10, 128, 28, 28] 110,848
│ └─Sequential: 2-6 [10, 32, 28, 28] --
│ │ └─BassicConv2d: 3-13 [10, 16, 28, 28] 3,104
│ │ └─BassicConv2d: 3-14 [10, 32, 28, 28] 12,864
│ └─Sequential: 2-7 [10, 32, 28, 28] --
│ │ └─MaxPool2d: 3-15 [10, 192, 28, 28] --
│ │ └─BassicConv2d: 3-16 [10, 32, 28, 28] 6,208
├─Inception: 1-7 [10, 480, 28, 28] --
│ └─BassicConv2d: 2-8 [10, 128, 28, 28] --
│ │ └─Sequential: 3-17 [10, 128, 28, 28] 33,024
│ └─Sequential: 2-9 [10, 192, 28, 28] --
│ │ └─BassicConv2d: 3-18 [10, 128, 28, 28] 33,024
│ │ └─BassicConv2d: 3-19 [10, 192, 28, 28] 221,568
│ └─Sequential: 2-10 [10, 96, 28, 28] --
│ │ └─BassicConv2d: 3-20 [10, 32, 28, 28] 8,256
│ │ └─BassicConv2d: 3-21 [10, 96, 28, 28] 76,992
│ └─Sequential: 2-11 [10, 64, 28, 28] --
│ │ └─MaxPool2d: 3-22 [10, 256, 28, 28] --
│ │ └─BassicConv2d: 3-23 [10, 64, 28, 28] 16,512
├─MaxPool2d: 1-8 [10, 480, 14, 14] --
├─Inception: 1-9 [10, 512, 14, 14] --
│ └─BassicConv2d: 2-12 [10, 192, 14, 14] --
│ │ └─Sequential: 3-24 [10, 192, 14, 14] 92,544
│ └─Sequential: 2-13 [10, 208, 14, 14] --
│ │ └─BassicConv2d: 3-25 [10, 96, 14, 14] 46,272
│ │ └─BassicConv2d: 3-26 [10, 208, 14, 14] 180,128
│ └─Sequential: 2-14 [10, 48, 14, 14] --
│ │ └─BassicConv2d: 3-27 [10, 16, 14, 14] 7,712
│ │ └─BassicConv2d: 3-28 [10, 48, 14, 14] 19,296
│ └─Sequential: 2-15 [10, 64, 14, 14] --
│ │ └─MaxPool2d: 3-29 [10, 480, 14, 14] --
│ │ └─BassicConv2d: 3-30 [10, 64, 14, 14] 30,848
├─AuxClf: 1-10 [10, 1000] --
│ └─Sequential: 2-16 [10, 128, 4, 4] --
│ │ └─AvgPool2d: 3-31 [10, 512, 4, 4] --
│ │ └─BassicConv2d: 3-32 [10, 128, 4, 4] 65,792
│ └─Sequential: 2-17 [10, 1000] --
│ │ └─Linear: 3-33 [10, 1024] 2,098,176
│ │ └─ReLU: 3-34 [10, 1024] --
│ │ └─Dropout: 3-35 [10, 1024] --
│ │ └─Linear: 3-36 [10, 1000] 1,025,000
├─Inception: 1-11 [10, 512, 14, 14] --
│ └─BassicConv2d: 2-18 [10, 160, 14, 14] --
│ │ └─Sequential: 3-37 [10, 160, 14, 14] 82,240
│ └─Sequential: 2-19 [10, 224, 14, 14] --
│ │ └─BassicConv2d: 3-38 [10, 112, 14, 14] 57,568
│ │ └─BassicConv2d: 3-39 [10, 224, 14, 14] 226,240
│ └─Sequential: 2-20 [10, 64, 14, 14] --
│ │ └─BassicConv2d: 3-40 [10, 24, 14, 14] 12,336
│ │ └─BassicConv2d: 3-41 [10, 64, 14, 14] 38,528
│ └─Sequential: 2-21 [10, 64, 14, 14] --
│ │ └─MaxPool2d: 3-42 [10, 512, 14, 14] --
│ │ └─BassicConv2d: 3-43 [10, 64, 14, 14] 32,896
├─Inception: 1-12 [10, 512, 14, 14] --
│ └─BassicConv2d: 2-22 [10, 128, 14, 14] --
│ │ └─Sequential: 3-44 [10, 128, 14, 14] 65,792
│ └─Sequential: 2-23 [10, 256, 14, 14] --
│ │ └─BassicConv2d: 3-45 [10, 128, 14, 14] 65,792
│ │ └─BassicConv2d: 3-46 [10, 256, 14, 14] 295,424
│ └─Sequential: 2-24 [10, 64, 14, 14] --
│ │ └─BassicConv2d: 3-47 [10, 24, 14, 14] 12,336
│ │ └─BassicConv2d: 3-48 [10, 64, 14, 14] 38,528
│ └─Sequential: 2-25 [10, 64, 14, 14] --
│ │ └─MaxPool2d: 3-49 [10, 512, 14, 14] --
│ │ └─BassicConv2d: 3-50 [10, 64, 14, 14] 32,896
├─Inception: 1-13 [10, 528, 14, 14] --
│ └─BassicConv2d: 2-26 [10, 112, 14, 14] --
│ │ └─Sequential: 3-51 [10, 112, 14, 14] 57,568
│ └─Sequential: 2-27 [10, 288, 14, 14] --
│ │ └─BassicConv2d: 3-52 [10, 144, 14, 14] 74,016
│ │ └─BassicConv2d: 3-53 [10, 288, 14, 14] 373,824
│ └─Sequential: 2-28 [10, 64, 14, 14] --
│ │ └─BassicConv2d: 3-54 [10, 32, 14, 14] 16,448
│ │ └─BassicConv2d: 3-55 [10, 64, 14, 14] 51,328
│ └─Sequential: 2-29 [10, 64, 14, 14] --
│ │ └─MaxPool2d: 3-56 [10, 512, 14, 14] --
│ │ └─BassicConv2d: 3-57 [10, 64, 14, 14] 32,896
├─AuxClf: 1-14 [10, 1000] --
│ └─Sequential: 2-30 [10, 128, 4, 4] --
│ │ └─AvgPool2d: 3-58 [10, 528, 4, 4] --
│ │ └─BassicConv2d: 3-59 [10, 128, 4, 4] 67,840
│ └─Sequential: 2-31 [10, 1000] --
│ │ └─Linear: 3-60 [10, 1024] 2,098,176
│ │ └─ReLU: 3-61 [10, 1024] --
│ │ └─Dropout: 3-62 [10, 1024] --
│ │ └─Linear: 3-63 [10, 1000] 1,025,000
├─Inception: 1-15 [10, 832, 14, 14] --
│ └─BassicConv2d: 2-32 [10, 256, 14, 14] --
│ │ └─Sequential: 3-64 [10, 256, 14, 14] 135,680
│ └─Sequential: 2-33 [10, 320, 14, 14] --
│ │ └─BassicConv2d: 3-65 [10, 160, 14, 14] 84,800
│ │ └─BassicConv2d: 3-66 [10, 320, 14, 14] 461,440
│ └─Sequential: 2-34 [10, 128, 14, 14] --
│ │ └─BassicConv2d: 3-67 [10, 32, 14, 14] 16,960
│ │ └─BassicConv2d: 3-68 [10, 128, 14, 14] 102,656
│ └─Sequential: 2-35 [10, 128, 14, 14] --
│ │ └─MaxPool2d: 3-69 [10, 528, 14, 14] --
│ │ └─BassicConv2d: 3-70 [10, 128, 14, 14] 67,840
├─MaxPool2d: 1-16 [10, 832, 7, 7] --
├─Inception: 1-17 [10, 832, 7, 7] --
│ └─BassicConv2d: 2-36 [10, 256, 7, 7] --
│ │ └─Sequential: 3-71 [10, 256, 7, 7] 213,504
│ └─Sequential: 2-37 [10, 320, 7, 7] --
│ │ └─BassicConv2d: 3-72 [10, 160, 7, 7] 133,440
│ │ └─BassicConv2d: 3-73 [10, 320, 7, 7] 461,440
│ └─Sequential: 2-38 [10, 128, 7, 7] --
│ │ └─BassicConv2d: 3-74 [10, 32, 7, 7] 26,688
│ │ └─BassicConv2d: 3-75 [10, 128, 7, 7] 102,656
│ └─Sequential: 2-39 [10, 128, 7, 7] --
│ │ └─MaxPool2d: 3-76 [10, 832, 7, 7] --
│ │ └─BassicConv2d: 3-77 [10, 128, 7, 7] 106,752
├─Inception: 1-18 [10, 1024, 7, 7] --
│ └─BassicConv2d: 2-40 [10, 384, 7, 7] --
│ │ └─Sequential: 3-78 [10, 384, 7, 7] 320,256
│ └─Sequential: 2-41 [10, 384, 7, 7] --
│ │ └─BassicConv2d: 3-79 [10, 192, 7, 7] 160,128
│ │ └─BassicConv2d: 3-80 [10, 384, 7, 7] 664,320
│ └─Sequential: 2-42 [10, 128, 7, 7] --
│ │ └─BassicConv2d: 3-81 [10, 48, 7, 7] 40,032
│ │ └─BassicConv2d: 3-82 [10, 128, 7, 7] 153,856
│ └─Sequential: 2-43 [10, 128, 7, 7] --
│ │ └─MaxPool2d: 3-83 [10, 832, 7, 7] --
│ │ └─BassicConv2d: 3-84 [10, 128, 7, 7] 106,752
├─AdaptiveAvgPool2d: 1-19 [10, 1024, 1, 1] --
├─Dropout: 1-20 [10, 1024] --
├─Linear: 1-21 [10, 1000] 1,025,000
===============================================================================================
Total params: 13,385,816
Trainable params: 13,385,816
Non-trainable params: 0
Total mult-adds (G): 15.91
===============================================================================================
Input size (MB): 6.02
Forward/backward pass size (MB): 517.24
Params size (MB): 53.54
Estimated Total Size (MB): 576.81
===============================================================================================
当然我们也可以仅仅是查看最外层的基础的GoogLeNet结构，不看内部的。
summary(net,(10,3,224,224),device="mps",depth=1)
===============================================================================================
Layer (type:depth-idx) Output Shape Param #
===============================================================================================
GoogLeNet [10, 1000] --
├─BassicConv2d: 1-1 [10, 64, 112, 112] 9,536
├─MaxPool2d: 1-2 [10, 64, 56, 56] --
├─BassicConv2d: 1-3 [10, 64, 56, 56] 4,224
├─BassicConv2d: 1-4 [10, 192, 56, 56] 110,976
├─MaxPool2d: 1-5 [10, 192, 28, 28] --
├─Inception: 1-6 [10, 256, 28, 28] 164,064
├─Inception: 1-7 [10, 480, 28, 28] 389,376
├─MaxPool2d: 1-8 [10, 480, 14, 14] --
├─Inception: 1-9 [10, 512, 14, 14] 376,800
├─AuxClf: 1-10 [10, 1000] 3,188,968
├─Inception: 1-11 [10, 512, 14, 14] 449,808
├─Inception: 1-12 [10, 512, 14, 14] 510,768
├─Inception: 1-13 [10, 528, 14, 14] 606,080
├─AuxClf: 1-14 [10, 1000] 3,191,016
├─Inception: 1-15 [10, 832, 14, 14] 869,376
├─MaxPool2d: 1-16 [10, 832, 7, 7] --
├─Inception: 1-17 [10, 832, 7, 7] 1,044,480
├─Inception: 1-18 [10, 1024, 7, 7] 1,445,344
├─AdaptiveAvgPool2d: 1-19 [10, 1024, 1, 1] --
├─Dropout: 1-20 [10, 1024] --
├─Linear: 1-21 [10, 1000] 1,025,000
===============================================================================================
Total params: 13,385,816
Trainable params: 13,385,816
Non-trainable params: 0
Total mult-adds (G): 15.91
===============================================================================================
Input size (MB): 6.02
Forward/backward pass size (MB): 517.24
Params size (MB): 53.54
Estimated Total Size (MB): 576.81
===============================================================================================
437
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
