13.初识Pytorch 复现VGG16及卷积神经网络图的可视化(Tensorboard)
-
搭建VGG16网络
用黄框画出vgg中vgg16的部分，将此网络分为8个部分，s1(stage 1)，s2(stage 2)，s3(stage 3)，s4，s5，s6，s7，s8，其中stage 出自RetinaNet
-
分析
公式 out_size = 1+(in_size+2*padding_size-kernel_size)/stride (1)
1. input_1 224*224*3 -> 112*112*64
其后有2层卷积层与1层最大池化层
224*224*3 -> 224*224*64
所以input_1_1 ->(conv, relu)
代入公式(1) 224 = 1+(224+2*padding_size-kernel_size)/stride
假设stride=1 -> 224 = 1+(224+2*padding_size-kernel_size)
-> 0 = 1+2*padding_size-kernel_size
假设padding_size = 1,kernel_size = 3
nn.Conv2d(in_channels=3,out_channels=64,padding=1,kernel_size=3,stride=1)
self.relu1_1=nn.ReLU(inplace=False)[inplace为是否覆盖以前的结果，False为不覆盖，且用一个新的变量来接，True为覆盖]
224*224*64 -> 224*224*64
input_1_2 -> (conv,relu)
nn.Conv2d(in_channels=64,out_channels=64,padding=1,kernel_size=3,stride=1)
self.relu1_2 = nn.ReLU(inplace=False)
224*224*64 -> 112*112*64
再计算maxpool层 input1_3
代入公式(1) 112 = 1+(224+2*padding_size-kernel_size)/stride
假设stride=2 -> 112 = 1+(112+padding_size-kernel_size/2)
-> 0 = 1+padding_size-kernel_size/2
假设padding_size=0,kernel_size/2=1
-> kernel_size=2
nn.MaxPool2d(padding=0,kernel_size=2,stride=2)
2.input_2 112*112*64 -> 56*56*128
其后有2层卷积层与1层最大池化层
112*112*64 -> 112*112*128
input2_1 -> (conv,relu)
nn.Conv2d(in_channels=64,out_channels=128,padding=1,kernel_size=3,stride=1)
self.relu2_1 = nn.ReLU(inplace=False)
112*112*128 -> 112*112*128
input2_2 -> (conv,relu)
nn.Conv2d(in_channels=128,out_channels=128,padding=1,kernel_size=3,stride=1)
self.relu2_2 = nn.ReLU(inplace=False)
112*112*128 -> 56*56*128
input2_3 -> maxpool
nn.MaxPool2d(padding=0,kernel_size=2,stride=2)
3.input_3 56*56*128 -> 28*28*256
其后有3层卷积层与1层最大池化层
56*56*128 -> 56*56*256
input3_1 -> (conv,relu)
nn.Conv2d(in_channels=128,out_channels=256,padding=1,kernel_size=3,stride=1)
self.relu3_1 = nn.ReLU(inplace=False)
56*56*256 -> 56*56*256
input3_2 -> (conv,relu)
nn.Conv2d(in_channels=256,out_channels=256,padding=1,kernel_size=3,stride=1)
self.relu3_2 = nn.ReLU(inplace=False)
56*56*256 -> 56*56*256
input3_3 -> (conv,relu)
nn.Conv2d(in_channels=256,out_channels=256,padding=1,kernel_size=3,stride=1)
self.relu3_3 = nn.ReLU(inplace=False)
56*56*256 -> 28*28*256
input_3_4 -> maxpool
nn.MaxPool2d(padding=0,kernel_size=2,stride=2)
4.input_4 28*28*256 -> 14*14*512
其后有3层卷积层与1层最大池化层
28*28*256 -> 28*28*512
input_4_1 -> (conv,relu)
nn.Conv2d(in_channels=256,out_channels=512,padding=1,kernel_size=3,stride=1)
self.relu4_1 = nn.ReLU(inplace=False)
28*28*512 -> 28*28*512
input_4_2 -> (conv,relu)
nn.Conv2d(in_channels=512,out_channels=512,padding=1,kernel_size=3,stride=1)
self.relu4_2 = nn.ReLU(inplace=False)
28*28*512 -> 28*28*512
input_4_3 -> (conv,relu)
nn.Conv2d(in_channels=512,out_channels=512,padding=1,kernel_size=3,stride=1)
self.relu4_3 = nn.ReLU(inplace=False)
28*28*512 -> 14*14*512
input_4_4 -> maxpool
nn.MaxPool2d(padding=0,kernel_size=2,stride=2)
5.input_5 14*14*512 -> 7*7*512
其后有3层卷积层与1层最大池化层
14*14*512 -> 14*14*512
input5_1 -> (conv,relu)
nn.Conv2d(in_channels=512,out_channels=512,padding=1,kernel_size=3,stride=1)
self.relu5_1 = nn.ReLU(inplace=False)
14*14*512 -> 14*14*512
input5_2 -> (conv,relu)
nn.Conv2d(in_channels=512,out_channels=512,padding=1,kernel_size=3,stride=1)
self.relu5_2 = nn.ReLU(inplace=False)
14*14*512 -> 14*14*512
input5_3 -> (conv,relu)
nn.Conv2d(in_channels=512,out_channels=512,padding=1,kernel_size=3,stride=1
self.relu5_3 = nn.ReLU(inplace=False)
14*14*512 -> 7*7*512
input_5_4 -> maxpool
nn.MaxPool2d(padding=0,kernel_size=2,stride=2)
6.input_6 7*7*512 -> 1*1*4096
展平 nn.flatten()
nn.linear(7*7*512,1*1*4096)
self.relu6 = nn.ReLU(inplace=False)
nn.Dropout(0.5)
7.input_7 4096 -> 4096
nn.linear(4096,4096)
self.relu7 = nn.ReLU(inplace=False)
nn.Dropout(0.5)
8.input_8 1*1*4096 -> 1*1*100
nn.linear(4096,1000)
self.relu8 = nn.ReLU(inplace=False)
- 上代码 1
import torch
from torch import nn
class vgg_16(nn.Module):
def __init__(self):
super(vgg_16, self).__init__()
# input_1
# input_1_1 224*224*3 -> 224*224*64
self.conv1_1 = nn.Conv2d(in_channels=3, out_channels=64, padding=1, kernel_size=3, stride=1)
self.relu1_1 = nn.ReLU(inplace=False)
# input_1_2 224*224*64 -> 224*224*64
self.conv1_2 = nn.Conv2d(in_channels=64, out_channels=64, padding=1, kernel_size=3, stride=1)
self.relu1_2 = nn.ReLU(inplace=False)
# input_1_3 224*224*64 -> 112*112*64
self.maxpool1_3 = nn.MaxPool2d(padding=0, stride=2, kernel_size=2)
# input_2
# input_2_1 112*112*64 -> 112*112*128
self.conv2_1 = nn.Conv2d(in_channels=64, out_channels=128, padding=1, kernel_size=3, stride=1)
self.relu2_1 = nn.ReLU(inplace=False)
# input_2_2 112*112*128 -> 112*112*128
self.conv2_2 = nn.Conv2d(in_channels=128, out_channels=128, padding=1, kernel_size=3, stride=1)
self.relu2_2 = nn.ReLU(inplace=False)
# input_2_3 112*112*128 -> 56*56*128
self.maxpool2_3 = nn.MaxPool2d(padding=0, stride=2, kernel_size=2)
# input_3
# input_3_1 56*56*128 -> 56*56*256
self.conv3_1 = nn.Conv2d(in_channels=128, out_channels=256, padding=1, kernel_size=3, stride=1)
self.relu3_1 = nn.ReLU(inplace=False)
# input_3_2 56*56*256 -> 56*56*256
self.conv3_2 = nn.Conv2d(in_channels=256, out_channels=256, padding=1, kernel_size=3, stride=1)
self.relu3_2 = nn.ReLU(inplace=False)
# input_3_3 56*56*256 -> 56*56*256
self.conv3_3 = nn.Conv2d(in_channels=256, out_channels=256, padding=1, kernel_size=3, stride=1)
self.relu3_3 = nn.ReLU(inplace=False)
# input_3_4 56*56*256 -> 28*28*256
self.maxpool3_4 = nn.MaxPool2d(padding=0, stride=2, kernel_size=2)
# input_4
# input_4_1 28*28*256 -> 28*28*512
self.conv4_1 = nn.Conv2d(in_channels=256, out_channels=512, padding=1, kernel_size=3, stride=1)
self.relu4_1 = nn.ReLU(inplace=False)
# input_4_2 28*28*512 -> 28*28*512
self.conv4_2 = nn.Conv2d(in_channels=512, out_channels=512, padding=1, kernel_size=3, stride=1)
self.relu4_2 = nn.ReLU(inplace=False)
# input_4_3 28*28*512 -> 28*28*512
self.conv4_3 = nn.Conv2d(in_channels=512, out_channels=512, padding=1, kernel_size=3, stride=1)
self.relu4_3 = nn.ReLU(inplace=False)
# input_4_4 28*28*512 -> 14*14*512
self.maxpool4_4 = nn.MaxPool2d(padding=0, stride=2, kernel_size=2)
# input_5
# input_5_1 14*14*512 -> 14*14*512
self.conv5_1 = nn.Conv2d(in_channels=512, out_channels=512, padding=1, kernel_size=3, stride=1)
self.relu5_1 = nn.ReLU(inplace=False)
# input_5_2 14*14*512 -> 14*14*512
self.conv5_2 = nn.Conv2d(in_channels=512, out_channels=512, padding=1, kernel_size=3, stride=1)
self.relu5_2 = nn.ReLU(inplace=False)
# input_5_3 14*14*512 -> 14*14*512
self.conv5_3 = nn.Conv2d(in_channels=512, out_channels=512, padding=1, kernel_size=3, stride=1)
self.relu5_3 = nn.ReLU(inplace=False)
# input_5_4 14*14*512 -> 7*7*512
self.maxpool5_4 = nn.MaxPool2d(padding=0, kernel_size=2, stride=2)
self.flatten = nn.Flatten()
self.linear6 = nn.Linear(7 * 7 * 512, 4096)
self.relu6 = nn.ReLU(inplace=False)
self.drouout6 = nn.Dropout(0.5)
self.linear7 = nn.Linear(4096, 4096)
self.relu7 = nn.ReLU(inplace=False)
self.drouout7 = nn.Dropout(0.5)
self.linear8 = nn.Linear(4096, 1000)
self.relu8 = nn.ReLU(inplace=False)
def forward(self, x):
# stage 1
x = self.conv1_1(x)
x = self.relu1_1(x)
x = self.conv1_2(x)
x = self.conv1_2(x)
x = self.maxpool1_3(x)
# stage 2
x = self.conv2_1(x)
x = self.relu2_1(x)
x = self.conv2_2(x)
x = self.relu2_2(x)
x = self.maxpool2_3(x)
# stage 3
x = self.conv3_1(x)
x = self.relu3_1(x)
x = self.conv3_2(x)
x = self.relu3_2(x)
x = self.conv3_3(x)
x = self.relu3_3(x)
x = self.maxpool3_4(x)
# stage 4
x = self.conv4_1(x)
x = self.relu4_1(x)
x = self.conv4_2(x)
x = self.relu4_2(x)
x = self.conv4_3(x)
x = self.relu4_3(x)
x = self.maxpool4_4(x)
# stage 5
x = self.conv5_1(x)
x = self.relu5_1(x)
x = self.conv5_2(x)
x = self.relu5_2(x)
x = self.conv5_3(x)
x = self.relu5_3(x)
x = self.maxpool5_4(x)
# stage 6
x = self.flatten(x)
x = self.linear6(x)
x = self.relu6(x)
x = self.drouout6(x)
# stage 7
x = self.linear7(x)
x = self.relu7(x)
x = self.drouout7(x)
# stage 8
x = self.linear8(x)
x = self.relu8(x)
return x
if __name__ == "__main__":
vgg_16 = vgg_16()
print(vgg_16)
input = torch.ones(1, 3, 224, 224)
output = vgg_16(input)
print(output)
结果：
.
.
.
.
.
.
- 上代码 2
import torch
from torch import nn
class vgg_16(nn.Module):
def __init__(self):
super(vgg_16, self).__init__()
self.module = nn.Sequential(
# input_1
# input_1_1 224*224*3 -> 224*224*64
nn.Conv2d(in_channels=3, out_channels=64, padding=1, kernel_size=3, stride=1),
nn.ReLU(inplace=True),
# input_1_2 224*224*64 -> 224*224*64
nn.Conv2d(in_channels=64, out_channels=64, padding=1, kernel_size=3, stride=1),
nn.ReLU(inplace=True),
# input_1_3 224*224*64 -> 112*112*64
nn.MaxPool2d(padding=0, stride=2, kernel_size=2),
# input_2
# input_2_1 112*112*64 -> 112*112*128
nn.Conv2d(in_channels=64, out_channels=128, padding=1, kernel_size=3, stride=1),
nn.ReLU(inplace=True),
# input_2_2 112*112*128 -> 112*112*128
nn.Conv2d(in_channels=128, out_channels=128, padding=1, kernel_size=3, stride=1),
nn.ReLU(inplace=True),
# input_2_3 112*112*128 -> 56*56*128
nn.MaxPool2d(padding=0, stride=2, kernel_size=2),
# input_3
# input_3_1 56*56*128 -> 56*56*256
nn.Conv2d(in_channels=128, out_channels=256, padding=1, kernel_size=3, stride=1),
nn.ReLU(inplace=True),
# input_3_2 56*56*256 -> 56*56*256
nn.Conv2d(in_channels=256, out_channels=256, padding=1, kernel_size=3, stride=1),
nn.ReLU(inplace=True),
# input_3_3 56*56*256 -> 56*56*256
nn.Conv2d(in_channels=256, out_channels=256, padding=1, kernel_size=3, stride=1),
nn.ReLU(inplace=True),
# input_3_4 56*56*256 -> 28*28*256
nn.MaxPool2d(padding=0, stride=2, kernel_size=2),
# input_4
# input_4_1 28*28*256 -> 28*28*512
nn.Conv2d(in_channels=256, out_channels=512, padding=1, kernel_size=3, stride=1),
nn.ReLU(inplace=True),
# input_4_2 28*28*512 -> 28*28*512
nn.Conv2d(in_channels=512, out_channels=512, padding=1, kernel_size=3, stride=1),
nn.ReLU(inplace=True),
# input_4_3 28*28*512 -> 28*28*512
nn.Conv2d(in_channels=512, out_channels=512, padding=1, kernel_size=3, stride=1),
nn.ReLU(inplace=True),
# input_4_4 28*28*512 -> 14*14*512
nn.MaxPool2d(padding=0, stride=2, kernel_size=2),
# input_5
# input_5_1 14*14*512 -> 14*14*512
nn.Conv2d(in_channels=512, out_channels=512, padding=1, kernel_size=3, stride=1),
nn.ReLU(inplace=True),
# input_5_2 14*14*512 -> 14*14*512
nn.Conv2d(in_channels=512, out_channels=512, padding=1, kernel_size=3, stride=1),
nn.ReLU(inplace=True),
# input_5_3 14*14*512 -> 14*14*512
nn.Conv2d(in_channels=512, out_channels=512, padding=1, kernel_size=3, stride=1),
nn.ReLU(inplace=True),
# input_5_4 14*14*512 -> 7*7*512
nn.MaxPool2d(padding=0, kernel_size=2, stride=2),
nn.Flatten(),
nn.Linear(7 * 7 * 512, 4096),
nn.ReLU(inplace=True),
nn.Dropout(0.5,inplace=True),
nn.Linear(4096, 4096),
nn.ReLU(inplace=True),
nn.Dropout(0.5, inplace=True),
nn.Linear(4096, 1000),
nn.ReLU(inplace=True),
)
def forward(self, x):
x = self.module(x)
return x
if __name__ == "__main__":
vgg_16 = vgg_16()
print(vgg_16)
input = torch.ones(1, 3, 224, 224)
output = vgg_16(input)
print(output)
结果：
.
.
.
- 可视化 上代码
import torch
from torch import nn
from torch.utils.tensorboard import SummaryWriter
class vgg_16(nn.Module):
def __init__(self):
super(vgg_16, self).__init__()
self.module = nn.Sequential(
# input_1
# input_1_1 224*224*3 -> 224*224*64
nn.Conv2d(in_channels=3, out_channels=64, padding=1, kernel_size=3, stride=1),
nn.ReLU(inplace=True),
# input_1_2 224*224*64 -> 224*224*64
nn.Conv2d(in_channels=64, out_channels=64, padding=1, kernel_size=3, stride=1),
nn.ReLU(inplace=True),
# input_1_3 224*224*64 -> 112*112*64
nn.MaxPool2d(padding=0, stride=2, kernel_size=2),
# input_2
# input_2_1 112*112*64 -> 112*112*128
nn.Conv2d(in_channels=64, out_channels=128, padding=1, kernel_size=3, stride=1),
nn.ReLU(inplace=True),
# input_2_2 112*112*128 -> 112*112*128
nn.Conv2d(in_channels=128, out_channels=128, padding=1, kernel_size=3, stride=1),
nn.ReLU(inplace=True),
# input_2_3 112*112*128 -> 56*56*128
nn.MaxPool2d(padding=0, stride=2, kernel_size=2),
# input_3
# input_3_1 56*56*128 -> 56*56*256
nn.Conv2d(in_channels=128, out_channels=256, padding=1, kernel_size=3, stride=1),
nn.ReLU(inplace=True),
# input_3_2 56*56*256 -> 56*56*256
nn.Conv2d(in_channels=256, out_channels=256, padding=1, kernel_size=3, stride=1),
nn.ReLU(inplace=True),
# input_3_3 56*56*256 -> 56*56*256
nn.Conv2d(in_channels=256, out_channels=256, padding=1, kernel_size=3, stride=1),
nn.ReLU(inplace=True),
# input_3_4 56*56*256 -> 28*28*256
nn.MaxPool2d(padding=0, stride=2, kernel_size=2),
# input_4
# input_4_1 28*28*256 -> 28*28*512
nn.Conv2d(in_channels=256, out_channels=512, padding=1, kernel_size=3, stride=1),
nn.ReLU(inplace=True),
# input_4_2 28*28*512 -> 28*28*512
nn.Conv2d(in_channels=512, out_channels=512, padding=1, kernel_size=3, stride=1),
nn.ReLU(inplace=True),
# input_4_3 28*28*512 -> 28*28*512
nn.Conv2d(in_channels=512, out_channels=512, padding=1, kernel_size=3, stride=1),
nn.ReLU(inplace=True),
# input_4_4 28*28*512 -> 14*14*512
nn.MaxPool2d(padding=0, stride=2, kernel_size=2),
# input_5
# input_5_1 14*14*512 -> 14*14*512
nn.Conv2d(in_channels=512, out_channels=512, padding=1, kernel_size=3, stride=1),
nn.ReLU(inplace=True),
# input_5_2 14*14*512 -> 14*14*512
nn.Conv2d(in_channels=512, out_channels=512, padding=1, kernel_size=3, stride=1),
nn.ReLU(inplace=True),
# input_5_3 14*14*512 -> 14*14*512
nn.Conv2d(in_channels=512, out_channels=512, padding=1, kernel_size=3, stride=1),
nn.ReLU(inplace=True),
# input_5_4 14*14*512 -> 7*7*512
nn.MaxPool2d(padding=0, kernel_size=2, stride=2),
nn.Flatten(),
nn.Linear(7 * 7 * 512, 4096),
nn.ReLU(inplace=True),
nn.Dropout(0.5, inplace=True),
nn.Linear(4096, 4096),
nn.ReLU(inplace=True),
nn.Dropout(0.5, inplace=True),
nn.Linear(4096, 1000),
nn.ReLU(inplace=True),
)
def forward(self, x):
x = self.module(x)
return x
if __name__ == "__main__":
vgg_16 = vgg_16()
print(vgg_16)
input = torch.ones(1, 3, 224, 224)
output = vgg_16(input)
print(output)
writer = SummaryWriter("logs")
writer.add_graph(vgg_16, input)
writer.close()
结果：
上一章 12.初识Pytorch搭建网络 LeNet-5复现(含nn.Sequential用法)
下一章 14.初识Pytorch损失函数