• 【pytorch】模型常用函数（conv2d、linear、loss、maxpooling等）

    1、二维卷积函数——cnv2d():

    1. '''
    2. in_channels (int): 输入通道数
    3. out_channels (int): 输出通道数
    4. kernel_size (int or tuple): 卷积核大小
    5. stride (int or tuple, optional): 步长 Default: 1
    6. padding (int, tuple or str, optional): 填充 Default: 0
    7. padding_mode (str, optional): 填充模式  Default: 'zeros'
    8. dilation (int or tuple, optional):  Default: 1
    9. groups (int, optional):  Default: 1
    10. bias (bool, optional): 偏置 Default: ``True``
    11. '''

    从数据集中加载数据（batch_size=64）

    1. dataset = torchvision.datasets.CIFAR10(root="./train_dataset",train=False,transform=torchvision.transforms.ToTensor(),download=True)
    2.  
    3. dataloader = DataLoader(dataset,batch_size=64)

     建立2维卷积网络模型

    1. class Diviner(nn.Module):
    2.  
    3.     def __init__(self):
    4.         super(Diviner,self).__init__()
    5.         self.conv1 = Conv2d(in_channels=3,out_channels=6, kernel_size=(3,3) , stride=(1,1),padding=0)
    6.  
    7.     def forward(self,x):
    8.         x = self.conv1(x)
    9.         return x

    实例化网络模型，并将卷积后得到的图片在tensorboard中展示 

    1. diviner = Diviner()
    2. writer = SummaryWriter("conv")
    3. step = 0
    4. for data in dataloader:
    5.     imgs,target = data
    6.     output = diviner(imgs)
    7.  
    8.     writer.add_images("input",imgs,step)
    9.     output = torch.reshape(output,(-1,3,30,30))
    10.     writer.add_images("output",output,step)
    11.     step = step + 1
    12.  
    13. writer.close()

    694f99e0d142406eb2533339dfc1c329.png 72b8586a3155425bba970645f04f4085.png

     2、线性层函数——Linear()

    1. '''
    2. in_features: size of each input sample
    3. out_features: size of each output sample
    4. bias: If set to ``False``, the layer will not learn an additive bias. Default: ``True``
    5. '''

    加载数据集（略）

     建立线性网络模型

    1. class Diviner(nn.Module):
    2.     def __init__(self):
    3.        super(Diviner, self).__init__()
    4.        self.linear1 = Linear(196608,10)
    5.  
    6.     def forward(self,input):
    7.         output = self.linear1(input)
    8.         return output

    3、最大池化函数——maxpooling()

     加载数据集（略）

     建立最大池化层网络模型

    1. class Diviner(nn.Module):
    2.     def __init__(self):
    3.         super(Diviner, self).__init__()
    4.         self.maxpool1 = MaxPool2d(kernel_size=3,ceil_mode=True)
    5.  
    6.     def forward(self,input):
    7.         output = self.maxpool1(input)
    8.         return output

     实例化网络模型，并将池化后得到的图片在tensorboard中展示 

    1. writer = SummaryWriter("maxpooling")
    2. step  = 0
    3.  
    4. for data in dataloader:
    5.     imgs,targets = data
    6.     output = diviner(imgs)
    7.     writer.add_images("input", imgs, step)
    8.     writer.add_images("output",output,step)
    9.     step = step + 1
    10.  
    11. writer.close()

     d5e6cacfc26c43549be0e347d456bbee.png8db187777305456e9df6fe8db1609d14.png

    4、激活函数——sigmoid()、relu()

      加载数据集（略）

     建立激活函数网络模型

    1. class Diviner(nn.Module):
    2.     def __init__(self):
    3.         super(Diviner, self).__init__()
    4.         self.relu1 = ReLU()
    5.         self.sigmoid1 = Sigmoid()
    6.  
    7.     def forward(self, input):
    8.         output = self.sigmoid1(input)
    9.         return output

      实例化网络模型，并将激活后得到的图片在tensorboard中展示 

    1. diviner = Diviner()
    2. step = 0
    3. writer = SummaryWriter("relu")
    4. for data in dataloader:
    5.     imgs,targets = data
    6.     output = diviner(imgs)
    7.     writer.add_images("relu",output,step)
    8.     step = step + 1
    9.  
    10. writer.close()

    652f1b771f474cc8b13ae8603efff1c0.png

     5、损失函数——loss()

      加载数据集（略）

     建立一个网络模型

    1. class Diviner(nn.Module):
    2.     def __init__(self):
    3.         super(Diviner, self).__init__()
    4.         self.modle1 = Sequential(
    5.             Conv2d(3, 32, (5, 5), padding=2),
    6.             MaxPool2d(2),
    7.             Conv2d(32, 32, (5, 5), padding=2),
    8.             MaxPool2d(2),
    9.             Conv2d(32, 64, (5, 5), padding=2),
    10.             MaxPool2d(2),
    11.             Flatten(),
    12.             Linear(1024, 64),
    13.             Linear(64, 10)
    14.         )
    15.     def forward(self,x):
    16.         x = self.modle1(x)
    17.         return x

    实例化网络模型，定义损失函数和优化器（反向传播） 

    1. diviner = Diviner()
    2. loss = nn.CrossEntropyLoss()
    3. optim = torch.optim.SGD(diviner.parameters(),lr=0.05)

    我们进行迭代，并记录损失值 

    1.  
    2. for epoch in range(20):
    3.     running_loss = 0.0
    4.     for data in dataloader:
    5.         imgs,targets = data
    6.         outputs =diviner(imgs)
    7.         result_loss = loss(outputs,targets)
    8.         optim.zero_grad()
    9.         result_loss.backward()
    10.         optim.step()
    11.         running_loss += result_loss
    12.     print(running_loss)

    6、使用GPU进行完整模型训练

    1. import torch
    2. import  torchvision
    3.  
    4. #准备数据集
    5. from torch import nn
    6. from torch.utils.data import DataLoader
    7. from torch.utils.tensorboard import SummaryWriter
    8.  
    9. train_data = torchvision.datasets.CIFAR10(root="./train_dataset",train=True,transform=torchvision.transforms.ToTensor(),download=True)
    10. test_data = torchvision.datasets.CIFAR10(root="./train_dataset",train=False,transform=torchvision.transforms.ToTensor(),download=True)
    11.  
    12. train_data_size = len(train_data)
    13. test_data_size = len(test_data)
    14.  
    15. print("训练数据库的长度{}".format(train_data_size))
    16. print("测试数据库的长度{}".format(test_data_size))
    17.  
    18. #利用dataloader来加载数据集
    19.  
    20. train_dataloader = DataLoader(train_data,batch_size=64)
    21. test_dataloader = DataLoader(test_data,batch_size=64)
    22.  
    23. #创建网络模型
    24. class Diviner(nn.Module):
    25.     def __init__(self):
    26.         super(Diviner, self).__init__()
    27.         self.model = nn.Sequential(
    28.             nn.Conv2d(3, 32, (5, 5), (1, 1), 2),
    29.             nn.MaxPool2d(2),
    30.             nn.Conv2d(32, 32, (5, 5), (1, 1), 2),
    31.             nn.MaxPool2d(2),
    32.             nn.Conv2d(32, 64, (5, 5), (1, 1), 2),
    33.             nn.MaxPool2d(2),
    34.             nn.Flatten(),
    35.             nn.Linear(64 * 4 * 4, 64),
    36.             nn.Linear(64, 10)
    37.         )
    38.  
    39.     def forward(self, input):
    40.         x = self.model(input)
    41.         return x
    42.  
    43. diviner = Diviner()
    44. diviner = diviner.cuda() #模型
    45.  
    46. #损失函数
    47. loss_fn  = nn.CrossEntropyLoss()
    48. loss_fn = loss_fn.cuda() #损失函数
    49. #优化器
    50. learn_rate = 0.01
    51. optimizer = torch.optim.SGD(diviner.parameters(),lr=learn_rate)
    52.  
    53. #设置训练网络的一些参数
    54. #记录训练的次数
    55. total_train_step = 0
    56. #记录测试的次数
    57. total_test_step = 0
    58. #训练的轮次
    59. epoch = 10
    60.  
    61. writer = SummaryWriter("train")
    62.  
    63. for i in range(epoch):
    64.     print("-------第{}轮训练开始了-------".format(i+1))
    65.  
    66.     #训练步骤开始：
    67.     diviner.train() #非必要，在特定层
    68.     for data in train_dataloader:
    69.         imgs,targets = data
    70.         
    71.         #数据
    72.         imgs = imgs.cuda()
    73.         targets = targets.cuda()
    74.         
    75.         outputs = diviner(imgs)
    76.         loss = loss_fn(outputs,targets)
    77.  
    78.         #优化器模型
    79.         optimizer.zero_grad()
    80.         loss.backward()
    81.         optimizer.step()
    82.  
    83.         total_train_step += 1
    84.         if(total_train_step%100 == 0):
    85.             print("训练次数:{},loss:{}".format(total_train_step,loss))
    86.             writer.add_scalar("train_loss",loss.item(),total_train_step)
    87.  
    88.     #测试步骤开始：
    89.     diviner.eval() #非必要 在特定层
    90.     total_test_loss = 0
    91.     total_accuracy = 0
    92.     with torch.no_grad():
    93.         for data in test_dataloader:
    94.             imgs,targets = data
    95.             imgs = imgs.cuda()
    96.             targets = targets.cuda()
    97.             outputs = diviner(imgs)
    98.             loss = loss_fn(outputs,targets)
    99.             total_test_loss += loss
    100.             accuracy = (outputs.argmax(1) == targets).sum()
    101.             total_accuracy += accuracy
    102.     print("整体测试集上的loss：{}".format(total_test_loss))
    103.     print("整体测试集上的正确率：{}".format(total_accuracy/test_data_size))
    104.     writer.add_scalar("test_loss",total_test_loss,total_test_step)
    105.     writer.add_scalar("test_accuracy", total_accuracy/test_data_size, total_test_step)
    106.     total_test_step += 1
    107.  
    108.     #保存模型
    109.     torch.save(diviner,"diviner_{}".format(i))
    110.  
    111. writer.close()

    4fdff2e5e5e44777bf174e5796f5772f.png f7470e3737214c089cb643bedfb4c560.png

    最后一轮数据： 

    1. -------第10轮训练开始了-------
    2. 训练次数:7100,loss:1.2293018102645874
    3. 训练次数:7200,loss:0.9501622319221497
    4. 训练次数:7300,loss:1.0970317125320435
    5. 训练次数:7400,loss:0.8500756025314331
    6. 训练次数:7500,loss:1.195753812789917
    7. 训练次数:7600,loss:1.2974092960357666
    8. 训练次数:7700,loss:0.8670048117637634
    9. 训练次数:7800,loss:1.2882726192474365
    10. 整体测试集上的loss：197.17103576660156
    11. 整体测试集上的正确率：0.5577999949455261

     

     

  • 相关阅读:
    零代码编程：用ChatGPT批量调整文件名称中的词汇顺序
    计算机网络【CN】TCP报文段格式【20B】
    梯度下降算法（Gradient Descent）
    Unlikely argument type for equals(): SysUser seems to be unrelated to String
    linux vim用法
    【BOOST C++ 16 语言扩展】（4） Boost.Conversion
    LeetCode讲解篇之面试题 01.08. 零矩阵
    Java 文件 & 文件操作
    系统架构设计：3 软件架构建模技术与应用
    13-网络篇-局域网与外网通信
  • 原文地址：https://blog.csdn.net/qq_47982709/article/details/132947628