16.初识Pytorch现有网络(existing network structures)的使用(use)与修改(modify)

在Pytorch官方文档中,给出了很多现成的网络架构供我们使用。
例如 分类的 AlexNet,VGG, ResNet等

In the official documentation of Pytorch, many ready-made network architectures are given for us to use. 

For example,
For classification: AlexNet, VGG, ResNet, etc. 

语义分割的 FCN,ResNet50等

For semantic segmentation: FCN, ResNet50, etc. 

目标检测的 SSD,RetinaNet等
实例分割的Mask R-CNN等

For object detection: SSD, RetinaNet, etc. 
For instance segmentation: Mask R-CNN, etc. 

这个网络架构，一般最常用的参数有两个：一个是Pretrained有无经过预训练,如果False,表示没有经过预训练，如果True,表示经过某个数据集的预训练；
另一个是progress有无进度条,如果True,表示显示进度条,如果False,表示不显示进度条；

There are two most commonly used parameters for those network architectures: one is whether Pretrained has been pre-trained. 
If it is False, it means that it has not been pre-trained. If it is True, it means that it has been pre-trained on a certain data set; 

The other is whether the progress has a progress bar. If it is True, it means that the progress bar is displayed. If it is False, it means that the progress bar is not displayed; 

现在，以VGG16为例，观察pretrained参数
（因为VGG16的预训练是使用ImageNet,下载ImageNet_2012_train需要100多G,所以就这样直接看参数）

Now, taking VGG16 as an example, we can observe the pretrained parameters 

(Because the pre-training of VGG16 uses ImageNet, 
it takes more than 100 G to download ImageNet_2012_train, 
so just look at the parameters directly) 

上代码(code):

import torchvision

vgg16_false = torchvision.models.vgg16(pretrained=False, progress=True)
print(vgg16_false)
vgg16_true = torchvision.models.vgg16(pretrained=True, progress=True)
print(vgg16_true)
print("ok")

断点(breakpoint)结果：

以vgg16为例,使用CIFAR数据集,但是此数据集只有10类,由于vgg16的预训练数据集是ImageNet有1000个类,所以需要修改原本的vgg16网络，这里有很多种的修改方式。

Taking vgg16 as an example, the CIFAR dataset is used, but this dataset has only 10 categories. 
Since the pre-training dataset of vgg16 is ImageNet with 1000 categories, it is necessary to modify the original vgg16 network. 
There are many ways to modify it. 

method1 在最后增加一层全连接层，输入为1000输出为10;
method2 修改最后一层全连接层，将1000个类改为是10各类[则输入是4096,输出10]

method1 adds a fully connected layer at the end, the input is 1000  and the output is 10;
method2 Modify the last fully connected layer and change the 1000 classes to 10 classes [the input is 4096, the output is 10].

code:

import torch
import torchvision
from torch import nn
from torchvision import transforms
from torch.utils.data import DataLoader

tran_tensor = transforms.ToTensor()
dataset = torchvision.datasets.CIFAR10(root="../dataset", train=True, transform=tran_tensor, download=True)
dataloader = DataLoader(dataset=dataset, batch_size=1, shuffle=True, num_workers=0, drop_last=False)

# method1 to modify vgg16
vgg16_false = torchvision.models.vgg16(pretrained=False)
vgg16_false.add_module("add_linear", nn.Linear(1000, 10))
print(vgg16_false)
print("***********************************************")
# method2 to modify vgg16
vgg16_true = torchvision.models.vgg16(pretrained=True)
vgg16_true.classifier[6] = nn.Linear(4096, 10)
print(vgg16_true)

result:

------------------------------------------------------------------------------------------------------------上一章 15.初识Pytorch反向传播(Backward)与优化器(optimizer)SGD
下一章 17.初识Pytorch保存模型(model save)与加载模型(model load)