一、Sequential的使用

image.png

Tips：实际Flatten和Fully connented ，Fully connented 和output层之间是全连接层

  1. 导包

    1. import torch
    2. from torch import nn
    3. from torch.nn import Conv2d, MaxPool2d,Flatten, Linear
    4. from tensorboardX import SummaryWriter

  2. 定义网络 ```python class Classify(nn.Module): def init(self):

    1.  super().__init__()
    2.  """
    3.  # 使用普通方式
    4.  self.conv1 = Conv2d(3, 32, 5, padding=2, stride=1)
    5.  self.maxpool1 = MaxPool2d(2)
    6.  self.conv2 = Conv2d(32,32,5, padding=2, stride=1)
    7.  self.maxpool2 = MaxPool2d(2)
    8.  self.conv3 = Conv2d(32, 64, 5, padding=2 ,stride=1)
    9.  self.maxpool3 = MaxPool2d(2)
    10.  self.flatten = Flatten()
    11.  self.linear1 = Linear(1024,64)
    12.  self.linear2 = Linear(64,10)
    13.  """
    14.  self.model1 = nn.Sequential(
    15.      Conv2d(3, 32, 5, padding=2, stride=1),
    16.      MaxPool2d(2),
    17.      Conv2d(32,32,5, padding=2, stride=1),
    18.      MaxPool2d(2),
    19.      Conv2d(32, 64, 5, padding=2 ,stride=1),
    20.      MaxPool2d(2),
    21.      Flatten(),
    22.      Linear(1024,64),
    23.      Linear(64,10)
    24.  )
  1. def forward(self,x):
  2.     """
  3.     x = self.conv1(x)
  4.     x = self.maxpool1(x)
  5.     x = self.conv2(x)
  6.     x = self.maxpool2(x)
  7.     x = self.conv3(x)
  8.     x = self.maxpool3(x)
  9.     x = self.flatten(x)
  10.     x = self.flatten(x)
  11.     x = self.linear1(x)
  12.     x = self.linear2(x)
  13.     """
  14.     x = self.model1(x)
  15.     return x
  2. 3. 运行
  3. ```python
  4. writer = SummaryWriter("logs")
  5. model = Classify()
  6. input = torch.ones((64,3,32,32))
  7. output = model(x)
  8. writer.add_graph(model, input)
  9. writer.close()

image.png

二、优化器和损失函数

  1. # 定义数据集和数据加载器
  2. dataset = torchvision.datasets.CIFAR10("./dataset", train = False, transform = torchvision.transforms.ToTensor(), download = False)
  3. dataloader = DataLoader(dataset,batch_size = 32)
  1. # 构造模型
  2. model = Classify()
  3. #----------------------------------#
  4. # 构造优化器（随机梯度下降）
  5. # model.parameters() 为网络的参数
  6. # lr为学习率
  7. #----------------------------------#
  8. optim = torch.optim.SGD(model.parameters(), lr=0.01)
  9. # 计算交叉熵损失
  10. loss = nn.CrossEntropyLoss()
  11. writer = SummaryWriter("logs")
  12. # 20个epoch
  13. for epoch in range(20):
  14.     # runing_loss为本次epoch中所有损失的和
  15.     runing_loss = 0.0
  16.     for idx,data in enumerate(dataloader):
  17.         imgs ,targets = data
  18.         outputs = model(imgs)
  19.         # 计算本次batch的损失
  20.         result_loss = loss(outputs, targets)
  21.         # 求和
  22.         runing_loss = runing_loss + result_loss
  23.         #-------------------#
  24.         # zero_grad()：清空所管理参数的梯度。
  25.         # 由于 PyTorch 的特性是张量的梯度不自动清零，因此每次反向传播之后都需要清空梯度
  26.         # optim为定义的随机梯度下降优化器
  27.         #-------------------#
  28.         optim.zero_grad()
  29.         # 反向传播（为batch，不是epoch）
  30.         result_loss.backward()
  31.         # 执行一步梯度更新
  32.         optim.step()
  33.     print(runing_loss)
  34. writer.close()