PyTorch常见预训练模型的下载链接及使用指南

六大分类模型下载方式和使用方法&＃xff1a;
Resnet
inception
Densenet
Alexnet
vggnet

Resnet:
model_urls &＃061; {
&＃039;resnet18&＃039;: &＃039;https://download.pytorch.org/models/resnet18-5c106cde.pth&＃039;,
&＃039;resnet34&＃039;: &＃039;https://download.pytorch.org/models/resnet34-333f7ec4.pth&＃039;,
&＃039;resnet50&＃039;: &＃039;https://download.pytorch.org/models/resnet50-19c8e357.pth&＃039;,
&＃039;resnet101&＃039;: &＃039;https://download.pytorch.org/models/resnet101-5d3b4d8f.pth&＃039;,
&＃039;resnet152&＃039;: &＃039;https://download.pytorch.org/models/resnet152-b121ed2d.pth&＃039;,
}
inception:
model_urls &＃061; {
# Inception v3 ported from TensorFlow
&＃039;inception_v3_google&＃039;: &＃039;https://download.pytorch.org/models/inception_v3_google-1a9a5a14.pth&＃039;,
}
Densenet:
model_urls &＃061; {
&＃039;densenet121&＃039;: &＃039;https://download.pytorch.org/models/densenet121-a639ec97.pth&＃039;,
&＃039;densenet169&＃039;: &＃039;https://download.pytorch.org/models/densenet169-b2777c0a.pth&＃039;,
&＃039;densenet201&＃039;: &＃039;https://download.pytorch.org/models/densenet201-c1103571.pth&＃039;,
&＃039;densenet161&＃039;: &＃039;https://download.pytorch.org/models/densenet161-8d451a50.pth&＃039;,
}
Alexnet:
model_urls &＃061; {
&＃039;alexnet&＃039;: &＃039;https://download.pytorch.org/models/alexnet-owt-4df8aa71.pth&＃039;,
}
vggnet:
model_urls &＃061; {
&＃039;vgg11&＃039;: &＃039;https://download.pytorch.org/models/vgg11-bbd30ac9.pth&＃039;,
&＃039;vgg13&＃039;: &＃039;https://download.pytorch.org/models/vgg13-c768596a.pth&＃039;,
&＃039;vgg16&＃039;: &＃039;https://download.pytorch.org/models/vgg16-397923af.pth&＃039;,
&＃039;vgg19&＃039;: &＃039;https://download.pytorch.org/models/vgg19-dcbb9e9d.pth&＃039;,
&＃039;vgg11_bn&＃039;: &＃039;https://download.pytorch.org/models/vgg11_bn-6002323d.pth&＃039;,
&＃039;vgg13_bn&＃039;: &＃039;https://download.pytorch.org/models/vgg13_bn-abd245e5.pth&＃039;,
&＃039;vgg16_bn&＃039;: &＃039;https://download.pytorch.org/models/vgg16_bn-6c64b313.pth&＃039;,
&＃039;vgg19_bn&＃039;: &＃039;https://download.pytorch.org/models/vgg19_bn-c79401a0.pth&＃039;,
}

学习内容&＃xff1a;测试实现预训练模型的使用&＃xff0c;并牢记该方式-拿为己用

关键步骤讲述&＃xff1a;

1. 默认已经安装好环境和pytorch框架&＃xff0c;以及torchvision等需要的库。

   
   



2. import torchvision.models as models 所有成熟网络模型几乎都在里面

   
   



3. # 初始化模型 model &＃061; models.resnet18()此处应用ResNet18来分类。

   
   



4. 修改尾巴&＃xff0c;毕竟你的输出不一定和原版&＃xff08;1000&＃xff09;一模一样。
   # 修改网络结构&＃xff0c;将fc层1000个输出改为9个输出。
# 获取最后一层的输入特征层信息。 fc_input_feature &＃061; model.fc.in_features
# 取代原来输出层为新的nn。 model.fc &＃061; nn.Linear(fc_input_feature, 9)到这里&＃xff0c;网络就构建好了。

   
   



5. 下载预训练参数&＃xff0c;为己所用。# load除最后一层的预训练权重 pretrained_weight &＃061; torch.hub.load_state_dict_from_url( url&＃061;&＃039;https://download.pytorch.org/models/resnet18-5c106cde.pth&＃039;, progress&＃061;True)到这里&＃xff0c;下载的是原版的1000分类的参数&＃xff0c;我们需要删除不需要的尾巴&＃xff0c;并训练自己的尾巴。del pretrained_weight[&＃039;fc.weight&＃039;]
del pretrained_weight[&＃039;fc.bias&＃039;]因为分类就是用的线性函数&＃xff0c;包括权重w和偏移b,只需删除尾巴。

   
   



6. 最后&＃xff0c;将剩下的模型参数load到我们的模型上即可。model.load_state_dict(pretrained_weight, strict&＃061;False)模型准备完毕&＃xff0c;剩下的操作和所有训练方法一样。参见详细训练代码。

   
   

import os
import torch
from torch.utils.data import DataLoader
from torch import nn
from torch import optim
import torchvision.models as models
import time
# use res18
# from resnet.resnetmini import ClassificModel as Model
from datasets.read_data_sleep import PlayPhoneData
def train(data_path&＃061;r"E:\Datasets\sleep_traindata"):
# 设置超参数
batch_size &＃061; 1 # 每次训练的数据量
LR &＃061; 0.01 # 学习率
STEP_SIZE &＃061; 5 # 控制学习率变化
MAX_EPOCH &＃061; 20 # 总的训练次数
num_print &＃061; 100 # 每n个batch打印一次
playPhoneData &＃061; PlayPhoneData(data_path)
# 利用dataloader加载数据集
train_loader &＃061; torch.utils.data.DataLoader(playPhoneData, batch_size&＃061;batch_size, shuffle&＃061;True, drop_last&＃061;True)
# 生成驱动器
use_gpu &＃061; torch.cuda.is_available()
if use_gpu:
print(&＃039;congratulation! You can use gpu to support acceleration&＃039;)
else:
print(&＃039;oppps, please use a small batch size&＃039;)
device &＃061; torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# 初始化模型
model &＃061; models.resnet18()
# 修改网络结构&＃xff0c;将fc层1000个输出改为9个输出
fc_input_feature &＃061; model.fc.in_features
model.fc &＃061; nn.Linear(fc_input_feature, 9)
# load除最后一层的预训练权重
pretrained_weight &＃061; torch.hub.load_state_dict_from_url(
url&＃061;&＃039;https://download.pytorch.org/models/resnet18-5c106cde.pth&＃039;, progress&＃061;True)
del pretrained_weight[&＃039;fc.weight&＃039;]
del pretrained_weight[&＃039;fc.bias&＃039;]
model.load_state_dict(pretrained_weight, strict&＃061;False)
model.to(device)
# net &＃061; Model(8).to(device) # class_num&＃061;8分八类&＃xff1a;睡岗(趴着睡&＃xff0c;躺着睡&＃xff0c;仰着睡&＃xff0c;低头睡)&＃xff0c;玩手机&＃xff08;俯视玩手机&＃xff0c;平视玩手机&＃xff0c;侧视玩手机&＃xff09;&＃xff0c;其他&＃061;[0&＃xff0c;1&＃xff0c;2&＃xff0c;3&＃xff0c;4&＃xff0c;5&＃xff0c;6&＃xff0c;7]
# net &＃061; Model(9).to(device) # class_num&＃061;9分九类&＃xff1a;睡岗(趴着睡&＃xff0c;躺着睡&＃xff0c;低头睡)&＃xff0c;站立&＃xff0c;半蹲&＃xff0c;坐&＃xff08;背坐&＃xff0c;正坐&＃xff0c;侧坐&＃xff09;&＃xff0c;其他&＃061;[0&＃xff0c;1&＃xff0c;2&＃xff0c;3&＃xff0c;4&＃xff0c;5&＃xff0c;6&＃xff0c;7&＃xff0c;8]
# 损失函数
get_loss &＃061; nn.CrossEntropyLoss() #交叉熵损失函数
# SGD优化器 第一个参数是输入需要优化的参数,第二个是学习率&＃xff0c;第三个是动量&＃xff0c;大致就是借助上一次导数结果&＃xff0c;加快收敛速度。
&＃039;&＃039;&＃039;
这一行代码里面实际上包含了多种优化:
一个是动量优化,增加了一个关于上一次迭代得到的系数的偏置&＃xff0c;借助上一次的指导&＃xff0c;减小梯度震荡&＃xff0c;加快收敛速度
一个是权重衰减&＃xff0c;通过对权重增加一个(正则项),该正则项会使得迭代公式中的权重按照比例缩减&＃xff0c;这么做的原因是&＃xff0c;过拟合的表现一般为参数浮动大&＃xff0c;使用小参数可以防止过拟合
&＃039;&＃039;&＃039;
optimizer &＃061; optim.SGD(model.parameters(), lr&＃061;LR, momentum&＃061;0.9, weight_decay&＃061;0.001)
# optimizer &＃061; optim.Adam(net.parameters(), lr&＃061;learn_rate)
# 动态调整学习率 StepLR 是等间隔调整学习率&＃xff0c;每step_size 令lr&＃061;lr*gamma
# 学习率衰减&＃xff0c;随着训练的加深&＃xff0c;目前的权重也越来越接近最优权重&＃xff0c;原本的学习率会使得&＃xff0c;loss上下震荡&＃xff0c;逐步减小学习率能加快收敛速度。
# scheduler &＃061; optim.lr_scheduler.StepLR(optimizer, step_size&＃061;STEP_SIZE, gamma&＃061;0.5, last_epoch&＃061;-1)
# Step:设置学习率下降策略
scheduler &＃061; torch.optim.lr_scheduler.StepLR(optimizer, step_size&＃061;10, gamma&＃061;0.1)
loss_list &＃061; []
start &＃061; time.time()
for epoch in range(MAX_EPOCH):
running_loss &＃061; 0.0
# enumerate()是python自带的函数&＃xff0c;用于迭代字典。参数1&＃xff0c;是需要迭代的对象&＃xff0c;第二参数是迭代的起始位置
for i, (inputs, labels) in enumerate(train_loader, 0):
inputs, labels &＃061; inputs.to(device), labels.to(device)
outputs &＃061; model(inputs) # 前向传播求出预测的值
optimizer.zero_grad() # 将梯度初始化为0
loss &＃061; get_loss(outputs, labels.long())
loss.backward() # 反向传播求梯度
optimizer.step() # 更新所有参数
running_loss &＃043;&＃061; loss.item() # loss是张量&＃xff0c;访问值时需要使用item()
loss_list.append(loss.item())
if i % num_print &＃061;&＃061; num_print - 1: # 每num_print打印平均loss
print(&＃039;[%d epoch, %d] loss: %.6f&＃039; % (epoch &＃043; 1, i &＃043; 1, running_loss / num_print))
running_loss &＃061; 0.0
lr &＃061; optimizer.param_groups[0][&＃039;lr&＃039;] # 查看目前的学习率
print(&＃039;learn_rate : %.5f&＃039; % lr)
scheduler.step() # 根据迭代epoch更新学习率
end &＃061; time.time()
print(&＃039;time:{}&＃039;.format(end - start))
torch.save(model, f&＃039;E:/model/playphone&＃043;sleepthepose/model_resnetmini_睡岗9分类{end}.pth&＃039;)
if __name__ &＃061;&＃061; "__main__":
train()


训练情况&＃xff1a;

......
[3 epoch, 500] loss: 2.186424
[3 epoch, 600] loss: 2.192622
[3 epoch, 700] loss: 2.165229
[3 epoch, 800] loss: 2.125184
[3 epoch, 900] loss: 2.185377
learn_rate : 0.01000
[4 epoch, 100] loss: 2.138786
[4 epoch, 200] loss: 2.177925
[4 epoch, 300] loss: 2.103718
......


备注&＃xff1a;代码只是讲解工具&＃xff0c;并非可以运行的实例&＃xff0c;因为里面的数据集需要有并自己写数据集的代码。

直接拿来用固然不错&＃xff0c;但自己分装一遍再用&＃xff0c;显得更加标准&＃xff0c;有水平。
比如封装如下:


class ResNet18forClassify(nn.Module):
def __init__(self, phase&＃061;"train"):
super(ResNet18forClassify, self).__init__()
self.phase &＃061; phase
self.net &＃061; models.resnet18()
fc_input_feature &＃061; self.net.fc.in_features
self.net.fc &＃061; nn.Linear(fc_input_feature, 9)
pretrained_weight &＃061; torch.hub.load_state_dict_from_url(
url&＃061;&＃039;https://download.pytorch.org/models/resnet18-5c106cde.pth&＃039;, progress&＃061;True)
del pretrained_weight[&＃039;fc.weight&＃039;]
del pretrained_weight[&＃039;fc.bias&＃039;]
self.net.load_state_dict(pretrained_weight, strict&＃061;False)
self.softmax &＃061; nn.Softmax(dim&＃061;1)
def forward(self, input_img):
out &＃061; self.net(input_img)
if self.phase &＃061;&＃061; "test":
return self.softmax(out)
return out


备注&＃xff1a;封装成自己的网络模型&＃xff0c;更加方便。
其中,if self.phase &＃061;&＃061; "test": return self.softmax(out)&＃xff0c;分类时训练输出的是类别标签与实际标签做损失计算&＃xff1b;测试时&＃xff0c;预测结果由激活函数转换为–类型和该类型可能性概率。输出可能是该类别的概率值。

1.https://github.com/pytorch/vision/tree/master/torchvision/models
2.环境搭建&＃xff1a;NVIDIA&＃043;CUDA&＃043;cudaNN的配置与Anaconda虚拟环境的搭建–深度学习第一步
3.Parallax:常用预训练模型下载地址