Use and modification of prior network model

First there is the use of the network model

Use sample code ：

import torchvision
from torch import nn

#  Load network 

#  This sentence （ When pretrained Set to False when ） It is equivalent to replacing the network architecture here , The parameters of the network model are initialized , Are the default parameters 
vgg16_false = torchvision.models.vgg16(pretrained=False)

#  This sentence （ When pretrained Set to True when ） The parameters of the network model are all in ImageNet Trained on the dataset , Is in the ImageNet A good result can be achieved on the data set 
vgg16_true = torchvision.models.vgg16(pretrained=True)

vgg16 There are two common parameters for the use of , Namely pretrained and process.

• pretrained - by True Words , It shows that the network has been trained and has a good effect on the training data set if False It means that this network is not trained
• process - by True The progress bar for downloading neural network parameters will be displayed. If it is False The progress bar for downloading neural network parameters will not be displayed
  To understand pretrained, It's equivalent to something ？ For example, when building the convolution layer of neural network , You gave one kernel_size But it didn't kernel_size Parameters in ,pretrained=True When you get a convolution kernel with parameters ,pretrained=False When you only know the size of this convolution kernel .

First there is the modification of the network model （ How to use the existing network to change its structure ）

1. Add a network layer

The sample code is as follows ：

import torchvision
from torch import nn

#  Load network 
vgg16_false = torchvision.models.vgg16(pretrained=False)
print("vgg16_false:\n",vgg16_false)

vgg16_true = torchvision.models.vgg16(pretrained=True)
print("vgg16_true:\n",vgg16_true)

#  How to use the existing network to change one of its structures 

# 1. Add a network layer 

#  load CIFAR10 Data sets 
train_data = torchvision.datasets.CIFAR10("./CIFAR10",train=True,transform=torchvision.transforms.ToTensor())

#  take vgg16_true Model applied to CIFAR10 On dataset , Why add a in_feature=1000,out_feature=10 The linear layer of ？ because vgg16_true Network training ImageNet The dataset has 1000 A classification , and CIFAR10 Only 10 classification , So will vgg16_true Internet applications are in CIFAR10 What I said , Need to add a in_feature=1000,out_feature=10 The linear layer of .

#  The way 1： Add... Directly across the network 
# vgg16_true.add_module("add_linear",nn.Linear(1000,10))

#  The way 2： Add 
vgg16_true.classifier.add_module("add_linear",nn.Linear(1000,10))

print("vgg16_true:\n",vgg16_true)

Running results ：

Explain ： take vgg16_true Model applied to CIFAR10 On dataset , Why add a in_feature=1000,out_feature=10 The linear layer of ？ because vgg16_true Network training ImageNet The dataset has 1000 A classification , and CIFAR10 Only 10 classification , So will vgg16_true Internet applications are in CIFAR10 What I said , Need to add a in_feature=1000,out_feature=10 The linear layer of .

2. Modify network directly

The sample code is as follows ：

import torchvision
from torch import nn

#  Load network model 
vgg16_false = torchvision.models.vgg16(pretrained=False)
print("vgg16_false:\n",vgg16_false)

vgg16_true = torchvision.models.vgg16(pretrained=True)
print("vgg16_true:\n",vgg16_true)

#  How to use the existing network to change one of its structures 
# 2. Modify network directly 

#  load CIFAR10 Data sets 
train_data = torchvision.datasets.CIFAR10("./CIFAR10",train=True,transform=torchvision.transforms.ToTensor())

#  take vgg16_true Model applied to CIFAR10 On dataset , Why modify the last linear layer out_feature=10 Well ？ because vgg16_true Network training ImageNet The dataset has 1000 A classification , and CIFAR10 Only 10 classification , So will vgg16_true Internet applications are in CIFAR10 What I said , Need to modify the last linear layer out_feature=10.

#  Index the network sequentially , Modify the last linear layer  
vgg16_false.classifier[6] = nn.Linear(4096,10)
print("vgg16_false",vgg16_false)

Running results ：

Explain ： take vgg16_true Model applied to CIFAR10 On dataset , Why modify the last linear layer out_feature=10 Well ？ because vgg16_true Network training ImageNet The dataset has 1000 A classification , and CIFAR10 Only 10 classification , So will vgg16_true Internet applications are in CIFAR10 What I said , Need to modify the last linear layer out_feature=10.