【无标题】

• 🍨 本文为🔗365天深度学习训练营中的学习记录博客
• 🍖 原作者：K同学啊

• 一、GPU设置

• import torch
  import torch.nn as nn
  import torchvision.transforms as transforms
  import torchvision
  from torchvision import transforms, datasets
  import os,PIL,pathlib,warnings

  warnings.filterwarnings("ignore")             #忽略警告信息

  device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
  device

• 二、数据导入与拆分数据集

• import os,PIL,random,pathlib

  data_dir = './PotatoPlants/'#根据文件名修改
  data_dir = pathlib.Path(data_dir)

  data_paths  = list(data_dir.glob('*'))
  classeNames = [str(path).split("\\")[1] for path in data_paths]
  classeNames

• train_transforms = transforms.Compose([
      transforms.Resize([224, 224]),  # 将输入图片resize成统一尺寸
      # transforms.RandomHorizontalFlip(), # 随机水平翻转
      transforms.ToTensor(),          # 将PIL Image或numpy.ndarray转换为tensor，并归一化到[0,1]之间
      transforms.Normalize(           # 标准化处理-->转换为标准正太分布（高斯分布），使模型更容易收敛
          mean=[0.485, 0.456, 0.406], 
          std=[0.229, 0.224, 0.225])  # 其中 mean=[0.485,0.456,0.406]与std=[0.229,0.224,0.225] 从数据集中随机抽样计算得到的。
  ])

  test_transform = transforms.Compose([
      transforms.Resize([224, 224]),  # 将输入图片resize成统一尺寸
      transforms.ToTensor(),          # 将PIL Image或numpy.ndarray转换为tensor，并归一化到[0,1]之间
      transforms.Normalize(           # 标准化处理-->转换为标准正太分布（高斯分布），使模型更容易收敛
          mean=[0.485, 0.456, 0.406], 
          std=[0.229, 0.224, 0.225])  # 其中 mean=[0.485,0.456,0.406]与std=[0.229,0.224,0.225] 从数据集中随机抽样计算得到的。
  ])

  total_data = datasets.ImageFolder("./PotatoPlants/",transform=train_transforms)
  total_data

• train_size = int(0.8 * len(total_data))
  test_size  = len(total_data) - train_size
  train_dataset, test_dataset = torch.utils.data.random_split(total_data, [train_size, test_size])
  train_dataset, test_dataset

• batch_size = 32

  train_dl = torch.utils.data.DataLoader(train_dataset,
                                             batch_size=batch_size,
                                             shuffle=True,
                                             num_workers=1)
  test_dl = torch.utils.data.DataLoader(test_dataset,
                                            batch_size=batch_size,
                                            shuffle=True,
                                            num_workers=1)

• for X, y in test_dl:
      print("Shape of X [N, C, H, W]: ", X.shape)
      print("Shape of y: ", y.shape, y.dtype)
      break

• import torch.nn.functional as F

• 三、搭建Vgg16，也可以直接调用

  class vgg16(nn.Module):
      def __init__(self):
          super(vgg16, self).__init__()
          # 卷积块1
          self.block1 = nn.Sequential(
              nn.Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
              nn.ReLU(),
              nn.Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
              nn.ReLU(),
              nn.MaxPool2d(kernel_size=(2, 2), stride=(2, 2))
          )
          # 卷积块2
          self.block2 = nn.Sequential(
              nn.Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
              nn.ReLU(),
              nn.Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
              nn.ReLU(),
              nn.MaxPool2d(kernel_size=(2, 2), stride=(2, 2))
          )
          # 卷积块3
          self.block3 = nn.Sequential(
              nn.Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
              nn.ReLU(),
              nn.Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
              nn.ReLU(),
              nn.Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
              nn.ReLU(),
              nn.MaxPool2d(kernel_size=(2, 2), stride=(2, 2))
          )
          # 卷积块4
          self.block4 = nn.Sequential(
              nn.Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
              nn.ReLU(),
              nn.Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
              nn.ReLU(),
              nn.Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
              nn.ReLU(),
              nn.MaxPool2d(kernel_size=(2, 2), stride=(2, 2))
          )
          # 卷积块5
          self.block5 = nn.Sequential(
              nn.Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
              nn.ReLU(),
              nn.Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
              nn.ReLU(),
              nn.Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
              nn.ReLU(),
              nn.MaxPool2d(kernel_size=(2, 2), stride=(2, 2))
          )
        

          # 全连接网络层，用于分类
          self.classifier = nn.Sequential(
              nn.Linear(in_features=512*7*7, out_features=4096),
              nn.ReLU(),
              nn.Linear(in_features=4096, out_features=4096),
              nn.ReLU(),
              nn.Linear(in_features=4096, out_features=3)
          )

      def forward(self, x):

          x = self.block1(x)
          x = self.block2(x)
          x = self.block3(x)
          x = self.block4(x)
          x = self.block5(x)
          x = torch.flatten(x, start_dim=1)
          x = self.classifier(x)

          return x

  device = "cuda" if torch.cuda.is_available() else "cpu"
  print("Using {} device".format(device))
    

  model = vgg16().to(device)
  model

• 四、查看模型

• # 统计模型参数量以及其他指标
  import torchsummary as summary
  summary.summary(model, (3, 224, 224))

• 五、训练和测试

• # 训练循环
  def train(dataloader, model, loss_fn, optimizer):
      size = len(dataloader.dataset)  # 训练集的大小
      num_batches = len(dataloader)   # 批次数目, (size/batch_size，向上取整)

      train_loss, train_acc = 0, 0  # 初始化训练损失和正确率
    

      for X, y in dataloader:  # 获取图片及其标签
          X, y = X.to(device), y.to(device)
        

          # 计算预测误差
          pred = model(X)          # 网络输出
          loss = loss_fn(pred, y)  # 计算网络输出和真实值之间的差距，targets为真实值，计算二者差值即为损失
        

          # 反向传播
          optimizer.zero_grad()  # grad属性归零
          loss.backward()        # 反向传播
          optimizer.step()       # 每一步自动更新
        

          # 记录acc与loss
          train_acc  += (pred.argmax(1) == y).type(torch.float).sum().item()
          train_loss += loss.item()
            

      train_acc  /= size
      train_loss /= num_batches

      return train_acc, train_loss

• def test (dataloader, model, loss_fn):
      size        = len(dataloader.dataset)  # 测试集的大小
      num_batches = len(dataloader)          # 批次数目, (size/batch_size，向上取整)
      test_loss, test_acc = 0, 0
    

      # 当不进行训练时，停止梯度更新，节省计算内存消耗
      with torch.no_grad():
          for imgs, target in dataloader:
              imgs, target = imgs.to(device), target.to(device)
            

              # 计算loss
              target_pred = model(imgs)
              loss        = loss_fn(target_pred, target)
            

              test_loss += loss.item()
              test_acc  += (target_pred.argmax(1) == target).type(torch.float).sum().item()

      test_acc  /= size
      test_loss /= num_batches

      return test_acc, test_loss

• 六、正式训练

• import copy

  optimizer  = torch.optim.Adam(model.parameters(), lr= 1e-4)
  loss_fn    = nn.CrossEntropyLoss() # 创建损失函数

  epochs     = 40

  train_loss = []
  train_acc  = []
  test_loss  = []
  test_acc   = []

  best_acc = 0    # 设置一个最佳准确率，作为最佳模型的判别指标

  for epoch in range(epochs):
    

      model.train()
      epoch_train_acc, epoch_train_loss = train(train_dl, model, loss_fn, optimizer)
    

      model.eval()
      epoch_test_acc, epoch_test_loss = test(test_dl, model, loss_fn)
    

      # 保存最佳模型到 best_model
      if epoch_test_acc > best_acc:
          best_acc   = epoch_test_acc
          best_model = copy.deepcopy(model)
    

      train_acc.append(epoch_train_acc)
      train_loss.append(epoch_train_loss)
      test_acc.append(epoch_test_acc)
      test_loss.append(epoch_test_loss)
    

      # 获取当前的学习率
      lr = optimizer.state_dict()['param_groups'][0]['lr']
    

      template = ('Epoch:{:2d}, Train_acc:{:.1f}%, Train_loss:{:.3f}, Test_acc:{:.1f}%, Test_loss:{:.3f}, Lr:{:.2E}')
      print(template.format(epoch+1, epoch_train_acc*100, epoch_train_loss, 
                            epoch_test_acc*100, epoch_test_loss, lr))
    

  # 保存最佳模型到文件中
  PATH = './best_model.pth'  # 保存的参数文件名
  torch.save(model.state_dict(), PATH)

  print('Done')

• import matplotlib.pyplot as plt
  #隐藏警告
  import warnings
  warnings.filterwarnings("ignore")               #忽略警告信息
  plt.rcParams['font.sans-serif']    = ['SimHei'] # 用来正常显示中文标签
  plt.rcParams['axes.unicode_minus'] = False      # 用来正常显示负号
  plt.rcParams['figure.dpi']         = 100        #分辨率

  epochs_range = range(epochs)

  plt.figure(figsize=(12, 3))
  plt.subplot(1, 2, 1)

  plt.plot(epochs_range, train_acc, label='Training Accuracy')
  plt.plot(epochs_range, test_acc, label='Test Accuracy')
  plt.legend(loc='lower right')
  plt.title('Training and Validation Accuracy')

  plt.subplot(1, 2, 2)
  plt.plot(epochs_range, train_loss, label='Training Loss')
  plt.plot(epochs_range, test_loss, label='Test Loss')
  plt.legend(loc='upper right')
  plt.title('Training and Validation Loss')
  plt.show()

• 七、指定图片预测

• from PIL import Image 

  classes = list(total_data.class_to_idx)

  def predict_one_image(image_path, model, transform, classes):
    

      test_img = Image.open(image_path).convert('RGB')
      plt.imshow(test_img)  # 展示预测的图片

      test_img = transform(test_img)
      img = test_img.to(device).unsqueeze(0)
    

      model.eval()
      output = model(img)

      _,pred = torch.max(output,1)
      pred_class = classes[pred]
      print(f'预测结果是：{pred_class}')# 预测训练集中的某张照片
  predict_one_image(image_path='./PotatoPlants/Early_blight/1.JPG', 
                    model=model, 
                    transform=train_transforms, 
                    classes=classes)

• 八、学习感想

• 对于Vgg16网络有了更深的理解，关于将改代码优化器换成随机梯度算法下周想尝试一下，然后识别准确率还需进一步提高