[已解决]基于VGGNet网络的多标签分类任务

黎明丿晓小

在基于原项目代码的基础上，根据以下要求做出修改使其满足多标签分类任务，并高亮对应文件的相关代码段注明修改原因：

1.修改softmax激活函数为sigmoid激活函数使其在输出层输出多个标签类别。

2.修改loss函数CrossEntropyLoss为BCELoss。

3.增添评价指标Hamming Loss，Accuracyexam，Precisionexam，Recallexam，Fβexam，并输出这些评价结果。

4.可视化训练时的loss损失，横坐标为epochs，epochs为训练轮数，纵坐标分别为train_loss和val_loss，train_loss为训练过程中的损失，val_loss为验证过程中的损失。

5.可视化预测结果，展示的图片中有输出的多标签类别。

6.当涉及到变量及函数定义时，请使用项目文件中的相关变量及函数定义，并与上下代码段保证顺畅。新定义的变量及函数除外。

将修改好的项目代码文件逐个展示自己的完整内容，并说明每个文件的具体用法。

主要项目代码如下：
model.py：定义了VGG模型的架构，并提供了辅助函数和配置信息。

1. import torch.nn as nn
   
2. import torch
   
3. 
   
4. # official pretrain weights
   
5. model_urls = {
   
6.     'vgg11': 'https://download.pytorch.org/models/vgg11-bbd30ac9.pth',
   
7.     'vgg13': 'https://download.pytorch.org/models/vgg13-c768596a.pth',
   
8.     'vgg16': 'https://download.pytorch.org/models/vgg16-397923af.pth',
   
9.     'vgg19': 'https://download.pytorch.org/models/vgg19-dcbb9e9d.pth'
   
10. }
    
11. 
    
12. 
    
13. class VGG(nn.Module):
    
14.     def __init__(self, features, num_classes=1000, init_weights=False):
    
15.         super(VGG, self).__init__()
    
16.         self.features = features
    
17.         self.classifier = nn.Sequential(
    
18.             nn.Linear(512*7*7, 4096),
    
19.             nn.ReLU(True),
    
20.             nn.Dropout(p=0.5),
    
21.             nn.Linear(4096, 4096),
    
22.             nn.ReLU(True),
    
23.             nn.Dropout(p=0.5),
    
24.             nn.Linear(4096, num_classes)
    
25.         )
    
26.         if init_weights:
    
27.             self._initialize_weights()
    
28. 
    
29.     def forward(self, x):
    
30.         # N x 3 x 224 x 224
    
31.         x = self.features(x)
    
32.         # N x 512 x 7 x 7
    
33.         x = torch.flatten(x, start_dim=1)
    
34.         # N x 512*7*7
    
35.         x = self.classifier(x)
    
36.         return x
    
37. 
    
38.     def _initialize_weights(self):
    
39.         for m in self.modules():
    
40.             if isinstance(m, nn.Conv2d):
    
41.                 # nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
    
42.                 nn.init.xavier_uniform_(m.weight)
    
43.                 if m.bias is not None:
    
44.                     nn.init.constant_(m.bias, 0)
    
45.             elif isinstance(m, nn.Linear):
    
46.                 nn.init.xavier_uniform_(m.weight)
    
47.                 # nn.init.normal_(m.weight, 0, 0.01)
    
48.                 nn.init.constant_(m.bias, 0)
    
49. 
    
50. 
    
51. def make_features(cfg: list):
    
52.     layers = []
    
53.     in_channels = 3
    
54.     for v in cfg:
    
55.         if v == "M":
    
56.             layers += [nn.MaxPool2d(kernel_size=2, stride=2)]
    
57.         else:
    
58.             conv2d = nn.Conv2d(in_channels, v, kernel_size=3, padding=1)
    
59.             layers += [conv2d, nn.ReLU(True)]
    
60.             in_channels = v
    
61.     return nn.Sequential(*layers)
    
62. 
    
63. 
    
64. cfgs = {
    
65.     'vgg11': [64, 'M', 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M'],
    
66.     'vgg13': [64, 64, 'M', 128, 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M'],
    
67.     'vgg16': [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 'M', 512, 512, 512, 'M', 512, 512, 512, 'M'],
    
68.     'vgg19': [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 256, 'M', 512, 512, 512, 512, 'M', 512, 512, 512, 512, 'M'],
    
69. }
    
70. 
    
71. 
    
72. def vgg(model_name="vgg16", **kwargs):
    
73.     assert model_name in cfgs, "Warning: model number {} not in cfgs dict!".format(model_name)
    
74.     cfg = cfgs[model_name]
    
75. 
    
76.     model = VGG(make_features(cfg), **kwargs)
    
77.     return model
    

复制代码

predict.py：基于训练好的VGG模型，对输入的图像进行分类预测，并输出预测结果。

1. import os
   
2. import json
   
3. 
   
4. import torch
   
5. from PIL import Image
   
6. from torchvision import transforms
   
7. import matplotlib.pyplot as plt
   
8. 
   
9. from model import vgg
   
10. 
    
11. 
    
12. def main():
    
13.     device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    
14. 
    
15.     data_transform = transforms.Compose(
    
16.         [transforms.Resize((224, 224)),
    
17.          transforms.ToTensor(),
    
18.          transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])
    
19. 
    
20.     # load image
    
21.     img_path = "../tulip.jpg"
    
22.     assert os.path.exists(img_path), "file: '{}' dose not exist.".format(img_path)
    
23.     img = Image.open(img_path)
    
24.     plt.imshow(img)
    
25.     # [N, C, H, W]
    
26.     img = data_transform(img)
    
27.     # expand batch dimension
    
28.     img = torch.unsqueeze(img, dim=0)
    
29. 
    
30.     # read class_indict
    
31.     json_path = './class_indices.json'
    
32.     assert os.path.exists(json_path), "file: '{}' dose not exist.".format(json_path)
    
33. 
    
34.     with open(json_path, "r") as f:
    
35.         class_indict = json.load(f)
    
36.    
    
37.     # create model
    
38.     model = vgg(model_name="vgg16", num_classes=5).to(device)
    
39.     # load model weights
    
40.     weights_path = "./vgg16Net.pth"
    
41.     assert os.path.exists(weights_path), "file: '{}' dose not exist.".format(weights_path)
    
42.     model.load_state_dict(torch.load(weights_path, map_location=device))
    
43. 
    
44.     model.eval()
    
45.     with torch.no_grad():
    
46.         # predict class
    
47.         output = torch.squeeze(model(img.to(device))).cpu()
    
48.         predict = torch.softmax(output, dim=0)
    
49.         predict_cla = torch.argmax(predict).numpy()
    
50. 
    
51.     print_res = "class: {}   prob: {:.3}".format(class_indict[str(predict_cla)],
    
52.                                                  predict[predict_cla].numpy())
    
53.     plt.title(print_res)
    
54.     for i in range(len(predict)):
    
55.         print("class: {:10}   prob: {:.3}".format(class_indict[str(i)],
    
56.                                                   predict[i].numpy()))
    
57.     plt.show()
    
58. 
    
59. 
    
60. if __name__ == '__main__':
    
61.     main()
    

复制代码

train.py:基于给定的数据集，训练VGG模型，最终得到可用于图像分类的模型文件。

1. import os
   
2. import sys
   
3. import json
   
4. 
   
5. import torch
   
6. import torch.nn as nn
   
7. from torchvision import transforms, datasets
   
8. import torch.optim as optim
   
9. from tqdm import tqdm
   
10. 
    
11. from model import vgg
    
12. 
    
13. 
    
14. def main():
    
15.     device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    
16.     print("using {} device.".format(device))
    
17. 
    
18.     data_transform = {
    
19.         "train": transforms.Compose([transforms.RandomResizedCrop(224),
    
20.                                      transforms.RandomHorizontalFlip(),
    
21.                                      transforms.ToTensor(),
    
22.                                      transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]),
    
23.         "val": transforms.Compose([transforms.Resize((224, 224)),
    
24.                                    transforms.ToTensor(),
    
25.                                    transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])}
    
26. 
    
27.     data_root = os.path.abspath(os.path.join(os.getcwd(), "../"))  # get data root path
    
28.     image_path = os.path.join(data_root, "data_set", "PlantVillage")  # flower data set path
    
29.     assert os.path.exists(image_path), "{} path does not exist.".format(image_path)
    
30.     train_dataset = datasets.ImageFolder(root=os.path.join(image_path, "train"),
    
31.                                          transform=data_transform["train"])
    
32.     train_num = len(train_dataset)
    
33. 
    
34.     # {'daisy':0, 'dandelion':1, 'roses':2, 'sunflower':3, 'tulips':4}
    
35.     flower_list = train_dataset.class_to_idx
    
36.     cla_dict = dict((val, key) for key, val in flower_list.items())
    
37.     # write dict into json file
    
38.     json_str = json.dumps(cla_dict, indent=4)
    
39.     with open('class_indices.json', 'w') as json_file:
    
40.         json_file.write(json_str)
    
41. 
    
42.     batch_size = 32
    
43.     nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0, 8])  # number of workers
    
44.     print('Using {} dataloader workers every process'.format(nw))
    
45. 
    
46.     train_loader = torch.utils.data.DataLoader(train_dataset,
    
47.                                                batch_size=batch_size, shuffle=True,
    
48.                                                num_workers=nw)
    
49. 
    
50.     validate_dataset = datasets.ImageFolder(root=os.path.join(image_path, "val"),
    
51.                                             transform=data_transform["val"])
    
52.     val_num = len(validate_dataset)
    
53.     validate_loader = torch.utils.data.DataLoader(validate_dataset,
    
54.                                                   batch_size=batch_size, shuffle=False,
    
55.                                                   num_workers=nw)
    
56.     print("using {} images for training, {} images for validation.".format(train_num,
    
57.                                                                            val_num))
    
58. 
    
59.     # test_data_iter = iter(validate_loader)
    
60.     # test_image, test_label = test_data_iter.next()
    
61. 
    
62.     model_name = "vgg16"
    
63.     net = vgg(model_name=model_name, num_classes=4, init_weights=True)
    
64.     net.to(device)
    
65.     loss_function = nn.CrossEntropyLoss()
    
66.     optimizer = optim.Adam(net.parameters(), lr=0.0001)
    
67. 
    
68.     epochs = 30
    
69.     best_acc = 0.0
    
70.     save_path = './{}Net.pth'.format(model_name)
    
71.     train_steps = len(train_loader)
    
72.     for epoch in range(epochs):
    
73.         # train
    
74.         net.train()
    
75.         running_loss = 0.0
    
76.         train_bar = tqdm(train_loader, file=sys.stdout)
    
77.         for step, data in enumerate(train_bar):
    
78.             images, labels = data
    
79.             optimizer.zero_grad()
    
80.             outputs = net(images.to(device))
    
81.             loss = loss_function(outputs, labels.to(device))
    
82.             loss.backward()
    
83.             optimizer.step()
    
84. 
    
85.             # print statistics
    
86.             running_loss += loss.item()
    
87. 
    
88.             train_bar.desc = "train epoch[{}/{}] loss:{:.3f}".format(epoch + 1,
    
89.                                                                      epochs,
    
90.                                                                      loss)
    
91. 
    
92.         # validate
    
93.         net.eval()
    
94.         acc = 0.0  # accumulate accurate number / epoch
    
95.         with torch.no_grad():
    
96.             val_bar = tqdm(validate_loader, file=sys.stdout)
    
97.             for val_data in val_bar:
    
98.                 val_images, val_labels = val_data
    
99.                 outputs = net(val_images.to(device))
    
100.                 predict_y = torch.max(outputs, dim=1)[1]
     
101.                 acc += torch.eq(predict_y, val_labels.to(device)).sum().item()
     
102. 
     
103.         val_accurate = acc / val_num
     
104.         print('[epoch %d] train_loss: %.3f  val_accuracy: %.3f' %
     
105.               (epoch + 1, running_loss / train_steps, val_accurate))
     
106. 
     
107.         if val_accurate > best_acc:
     
108.             best_acc = val_accurate
     
109.             torch.save(net.state_dict(), save_path)
     
110. 
     
111.     print('Finished Training')
     
112. 
     
113. 
     
114. if __name__ == '__main__':
     
115.     main()
     

复制代码

最佳答案

月排行榜 / 总排行榜

isdkz

2023-4-22 10:33:52

黎明丿晓小 发表于 2023-4-22 16:42
在运行以下代码时，报错“RuntimeError: Error(s) in loading state_dict for VGG:
        size mismatch for  ...

这个错误是因为您试图将预训练模型的权重（输出层有4个神经元）加载到具有不同输出层结构的模型中（输出层有5个神经元）。这会导致权重和模型的形状不匹配，因此无法加载权重。

要解决这个问题，您需要确保模型的类别数量与预训练模型权重的类别数量一致。在这个例子中，您应该将模型的类别数量改为4：

1. model = vgg(model_name="vgg16", num_classes=4).to(device)
   

复制代码

然后，您需要更新`class_indices.json`文件，确保它包含正确的类别数量和类别名称。您可以手动编辑文件或在训练模型时自动生成该文件。

请注意，如果您需要处理5个类别，您需要重新训练模型，以适应5个类别的任务。在这种情况下，您需要使用适当数量的类别重新训练模型，并确保`class_indices.json`文件包含正确的类别信息。

跳转到最佳答案楼层

黎明丿晓小

isdkz 发表于 2023-4-22 10:39
在基于原项目代码的基础上，根据要求做出修改以满足多标签分类任务。以下是修改后的代码文件及修改原因：
...

我有以下问题需要你来回答：
1.np.vstack中的np是numpy吗？
2.val_loss是如何定义的？
3.请在基于predict.py代码的基础上，修改部分代码使其满足以下图片要求：
   
   其中，red和dress为多标签预测，后面对应的数字为置信度，也就是概率，可以设置一个阈值threshold=0.5，当预测一个图片的输出结果时，某一个标签的置信度大于阈值即可显示在输出结果图片上面。
   

黎明丿晓小

isdkz 发表于 2023-4-22 11:38
1. 是的，`np`通常是`numpy`库的缩写。在Python程序中，您经常会看到这样的导入语句：`import numpy as n ...

我的意思是val_loss在以下代码是如何定义的：

1. import os
   
2. import sys
   
3. import json
   
4. 
   
5. import torch
   
6. import torch.nn as nn
   
7. from torchvision import transforms, datasets
   
8. import torch.optim as optim
   
9. from tqdm import tqdm
   
10. from sklearn.metrics import hamming_loss,accuracy_score,precision_score,recall_score,fbeta_score
    
11. import numpy as np
    
12. import matplotlib.pyplot as plt
    
13. 
    
14. from model import vgg
    
15. 
    
16. # Define the evaluation function
    
17. def evaluate(y_true, y_pred):
    
18.     hamming = hamming_loss(y_true, y_pred)
    
19.     accuracy = accuracy_score(y_true, y_pred)
    
20.     precision = precision_score(y_true, y_pred, average='micro')
    
21.     recall = recall_score(y_true, y_pred, average='micro')
    
22.     fbeta = fbeta_score(y_true, y_pred, beta=1, average='micro')
    
23. 
    
24.     return hamming, accuracy, precision, recall, fbeta
    
25. 
    
26. def main():
    
27.     device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    
28.     print("using {} device.".format(device))
    
29. 
    
30.     data_transform = {
    
31.         "train": transforms.Compose([transforms.RandomResizedCrop(224),
    
32.                                      transforms.RandomHorizontalFlip(),
    
33.                                      transforms.ToTensor(),
    
34.                                      transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]),
    
35.         "val": transforms.Compose([transforms.Resize((224, 224)),
    
36.                                    transforms.ToTensor(),
    
37.                                    transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])}
    
38. 
    
39.     data_root = os.path.abspath(os.path.join(os.getcwd(), "../"))  # get data root path
    
40.     image_path = os.path.join(data_root, "data_set", "PlantVillage")  # flower data set path
    
41.     assert os.path.exists(image_path), "{} path does not exist.".format(image_path)
    
42.     train_dataset = datasets.ImageFolder(root=os.path.join(image_path, "train"),
    
43.                                          transform=data_transform["train"])
    
44.     train_num = len(train_dataset)
    
45. 
    
46.     # {'daisy':0, 'dandelion':1, 'roses':2, 'sunflower':3, 'tulips':4}
    
47.     flower_list = train_dataset.class_to_idx
    
48.     cla_dict = dict((val, key) for key, val in flower_list.items())
    
49.     # write dict into json file
    
50.     json_str = json.dumps(cla_dict, indent=4)
    
51.     with open('class_indices.json', 'w') as json_file:
    
52.         json_file.write(json_str)
    
53. 
    
54.     batch_size = 32
    
55.     nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0, 8])  # number of workers
    
56.     print('Using {} dataloader workers every process'.format(nw))
    
57. 
    
58.     train_loader = torch.utils.data.DataLoader(train_dataset,
    
59.                                                batch_size=batch_size, shuffle=True,
    
60.                                                num_workers=nw)
    
61. 
    
62.     validate_dataset = datasets.ImageFolder(root=os.path.join(image_path, "val"),
    
63.                                             transform=data_transform["val"])
    
64.     val_num = len(validate_dataset)
    
65.     validate_loader = torch.utils.data.DataLoader(validate_dataset,
    
66.                                                   batch_size=batch_size, shuffle=False,
    
67.                                                   num_workers=nw)
    
68.     print("using {} images for training, {} images for validation.".format(train_num,
    
69.                                                                            val_num))
    
70. 
    
71.     # test_data_iter = iter(validate_loader)
    
72.     # test_image, test_label = test_data_iter.next()
    
73. 
    
74.     model_name = "vgg16"
    
75.     net = vgg(model_name=model_name, num_classes=4, init_weights=True)
    
76.     net.to(device)
    
77.     # Change the loss function to BCELoss
    
78.     loss_function = nn.BCEWithLogitsLoss()
    
79.     optimizer = optim.Adam(net.parameters(), lr=0.0001)
    
80. 
    
81.     train_losses = []
    
82.     val_losses = []
    
83. 
    
84.     # Modify the training and validation loop to calculate evaluation metrics
    
85.     epochs = 30
    
86.     best_acc = 0.0
    
87.     save_path = './{}Net.pth'.format(model_name)
    
88.     train_steps = len(train_loader)
    
89.     for epoch in range(epochs):
    
90.         # train
    
91.         net.train()
    
92.         running_loss = 0.0
    
93.         train_bar = tqdm(train_loader, file=sys.stdout)
    
94.         for step, data in enumerate(train_bar):
    
95.             images, labels = data
    
96.             optimizer.zero_grad()
    
97.             outputs = net(images.to(device))
    
98.             loss = loss_function(outputs, labels.to(device))
    
99.             loss.backward()
    
100.             optimizer.step()
     
101. 
     
102.             # print statistics
     
103.             running_loss += loss.item()
     
104. 
     
105.             train_bar.desc = "train epoch[{}/{}] loss:{:.3f}".format(epoch + 1,
     
106.                                                                      epochs,
     
107.                                                                      loss)
     
108.             train_losses.append(running_loss / train_steps)
     
109. 
     
110.         # validate, Modify the validation part
     
111.         net.eval()
     
112.         acc = 0.0  # accumulate accurate number / epoch
     
113.         val_steps = len(validate_loader)
     
114.         val_true_labels = []
     
115.         val_pred_labels = []
     
116.         with torch.no_grad():
     
117.             val_bar = tqdm(validate_loader, file=sys.stdout)
     
118.             for val_data in val_bar:
     
119.                 val_images, val_labels = val_data
     
120.                 outputs = net(val_images.to(device))
     
121.                 val_true_labels.append(val_labels.cpu().numpy())
     
122.                 val_pred_labels.append((outputs > 0.5).float().cpu().numpy())
     
123.                 val_losses.append(val_loss / val_steps)
     
124.                 # predict_y = torch.max(outputs, dim=1)[1]
     
125.                 # acc += torch.eq(predict_y, val_labels.to(device)).sum().item()
     
126.         val_true_labels = np.vstack(val_true_labels)
     
127.         val_pred_labels = np.vstack(val_pred_labels)
     
128.         hamming, accuracy, precision, recall, fbeta = evaluate(val_true_labels, val_pred_labels)
     
129.         print('[epoch %d] train_loss: %.3f val_hamming: %.3f val_accuracy: %.3f val_precision: %.3f val_recall: %.3f')
     
130. 
     
131.         # val_accurate = acc / val_num
     
132.         # print('[epoch %d] train_loss: %.3f  val_accuracy: %.3f' %
     
133.         #       (epoch + 1, running_loss / train_steps, val_accurate))
     
134.         #
     
135.         # if val_accurate > best_acc:
     
136.         #     best_acc = val_accurate
     
137.         #     torch.save(net.state_dict(), save_path)
     
138.     plt.figure(figsize=(10, 5))
     
139.     plt.plot(train_losses, label='Training Loss')
     
140.     plt.plot(val_losses, label='Validation Loss')
     
141.     plt.xlabel('Epochs')
     
142.     plt.ylabel('Loss')
     
143.     plt.legend()
     
144.     plt.show()
     
145. 
     
146.     print('Finished Training')
     
147. 
     
148. 
     
149. if __name__ == '__main__':
     
150.     main()
     

复制代码

另外，请在基于以下代码的基础上，修改部分代码使其满足以下图片要求：
其中，red和dress为多标签预测，后面对应的数字为置信度，也就是概率，可以设置一个阈值threshold=0.5，当预测一个图片的输出结果时，某一个标签的置信度大于阈值即可显示在输出结果图片上面。
  

1. import os
   
2. import json
   
3. 
   
4. import torch
   
5. from PIL import Image
   
6. from torchvision import transforms
   
7. import matplotlib.pyplot as plt
   
8. 
   
9. from model import vgg
   
10. 
    
11. 
    
12. def main():
    
13.     device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    
14. 
    
15.     data_transform = transforms.Compose(
    
16.         [transforms.Resize((224, 224)),
    
17.          transforms.ToTensor(),
    
18.          transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])
    
19. 
    
20.     # load image
    
21.     img_path = "./test.jpg"
    
22.     assert os.path.exists(img_path), "file: '{}' dose not exist.".format(img_path)
    
23.     img = Image.open(img_path)
    
24.     plt.imshow(img)
    
25.     # [N, C, H, W]
    
26.     img = data_transform(img)
    
27.     # expand batch dimension
    
28.     img = torch.unsqueeze(img, dim=0)
    
29. 
    
30.     # read class_indict
    
31.     json_path = './class_indices.json'
    
32.     assert os.path.exists(json_path), "file: '{}' dose not exist.".format(json_path)
    
33. 
    
34.     with open(json_path, "r") as f:
    
35.         class_indict = json.load(f)
    
36.    
    
37.     # create model
    
38.     model = vgg(model_name="vgg16", num_classes=5).to(device)
    
39.     # load model weights
    
40.     weights_path = "./vgg16Net.pth"
    
41.     assert os.path.exists(weights_path), "file: '{}' dose not exist.".format(weights_path)
    
42.     model.load_state_dict(torch.load(weights_path, map_location=device))
    
43. 
    
44.     model.eval()
    
45.     with torch.no_grad():
    
46.         # predict class
    
47.         output = torch.squeeze(model(img.to(device))).cpu()
    
48.         predict = torch.softmax(output, dim=0)
    
49.         predict_cla = torch.argmax(predict).numpy()
    
50. 
    
51.     print_res = "class: {}   prob: {:.3}".format(class_indict[str(predict_cla)],
    
52.                                                  predict[predict_cla].numpy())
    
53.     plt.title(print_res)
    
54.     for i in range(len(predict)):
    
55.         print("class: {:10}   prob: {:.3}".format(class_indict[str(i)],
    
56.                                                   predict[i].numpy()))
    
57.     plt.show()
    
58. 
    
59. 
    
60. if __name__ == '__main__':
    
61.     main()
    

复制代码

黎明丿晓小

isdkz 发表于 2023-4-22 11:50
首先，针对您提供的第一个代码片段，关于`val_loss`的定义问题。实际上，在这段代码中，并没有显式地定义 ...

在训练以下代码时报错ValueError：Target size (torch.Size([32])) must be the same as input size (torch.Size([32,4]))

1. import os
   
2. import sys
   
3. import json
   
4. 
   
5. import torch
   
6. import torch.nn as nn
   
7. from torchvision import transforms, datasets
   
8. import torch.optim as optim
   
9. from tqdm import tqdm
   
10. from sklearn.metrics import hamming_loss,accuracy_score,precision_score,recall_score,fbeta_score
    
11. import numpy as np
    
12. import matplotlib.pyplot as plt
    
13. 
    
14. from model import vgg
    
15. 
    
16. # Define the evaluation function
    
17. def evaluate(y_true, y_pred):
    
18.     hamming = hamming_loss(y_true, y_pred)
    
19.     accuracy = accuracy_score(y_true, y_pred)
    
20.     precision = precision_score(y_true, y_pred, average='micro')
    
21.     recall = recall_score(y_true, y_pred, average='micro')
    
22.     fbeta = fbeta_score(y_true, y_pred, beta=1, average='micro')
    
23. 
    
24.     return hamming, accuracy, precision, recall, fbeta
    
25. 
    
26. def main():
    
27.     device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    
28.     print("using {} device.".format(device))
    
29. 
    
30.     data_transform = {
    
31.         "train": transforms.Compose([transforms.RandomResizedCrop(224),
    
32.                                      transforms.RandomHorizontalFlip(),
    
33.                                      transforms.ToTensor(),
    
34.                                      transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]),
    
35.         "val": transforms.Compose([transforms.Resize((224, 224)),
    
36.                                    transforms.ToTensor(),
    
37.                                    transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])}
    
38. 
    
39.     data_root = os.path.abspath(os.path.join(os.getcwd(), "../"))  # get data root path
    
40.     image_path = os.path.join(data_root, "data_set", "PlantVillage")  # flower data set path
    
41.     assert os.path.exists(image_path), "{} path does not exist.".format(image_path)
    
42.     train_dataset = datasets.ImageFolder(root=os.path.join(image_path, "train"),
    
43.                                          transform=data_transform["train"])
    
44.     train_num = len(train_dataset)
    
45. 
    
46.     # {'daisy':0, 'dandelion':1, 'roses':2, 'sunflower':3, 'tulips':4}
    
47.     flower_list = train_dataset.class_to_idx
    
48.     cla_dict = dict((val, key) for key, val in flower_list.items())
    
49.     # write dict into json file
    
50.     json_str = json.dumps(cla_dict, indent=4)
    
51.     with open('class_indices.json', 'w') as json_file:
    
52.         json_file.write(json_str)
    
53. 
    
54.     batch_size = 32
    
55.     nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0, 8])  # number of workers
    
56.     print('Using {} dataloader workers every process'.format(nw))
    
57. 
    
58.     train_loader = torch.utils.data.DataLoader(train_dataset,
    
59.                                                batch_size=batch_size, shuffle=True,
    
60.                                                num_workers=nw)
    
61. 
    
62.     validate_dataset = datasets.ImageFolder(root=os.path.join(image_path, "val"),
    
63.                                             transform=data_transform["val"])
    
64.     val_num = len(validate_dataset)
    
65.     validate_loader = torch.utils.data.DataLoader(validate_dataset,
    
66.                                                   batch_size=batch_size, shuffle=False,
    
67.                                                   num_workers=nw)
    
68.     print("using {} images for training, {} images for validation.".format(train_num,
    
69.                                                                            val_num))
    
70. 
    
71.     # test_data_iter = iter(validate_loader)
    
72.     # test_image, test_label = test_data_iter.next()
    
73. 
    
74.     model_name = "vgg16"
    
75.     net = vgg(model_name=model_name, num_classes=4, init_weights=True)
    
76.     net.to(device)
    
77.     # Change the loss function to BCELoss
    
78.     loss_function = nn.BCEWithLogitsLoss()
    
79.     optimizer = optim.Adam(net.parameters(), lr=0.0001)
    
80. 
    
81.     train_losses = []
    
82.     val_losses = []
    
83. 
    
84.     # Modify the training and validation loop to calculate evaluation metrics
    
85.     epochs = 30
    
86.     best_acc = 0.0
    
87.     save_path = './{}Net.pth'.format(model_name)
    
88.     train_steps = len(train_loader)
    
89.     for epoch in range(epochs):
    
90.         # train
    
91.         net.train()
    
92.         running_loss = 0.0
    
93.         train_bar = tqdm(train_loader, file=sys.stdout)
    
94.         for step, data in enumerate(train_bar):
    
95.             images, labels = data
    
96.             optimizer.zero_grad()
    
97.             outputs = net(images.to(device))
    
98.             loss = loss_function(outputs, labels.to(device))
    
99.             loss.backward()
    
100.             optimizer.step()
     
101. 
     
102.             # print statistics
     
103.             running_loss += loss.item()
     
104. 
     
105.             train_bar.desc = "train epoch[{}/{}] loss:{:.3f}".format(epoch + 1,
     
106.                                                                      epochs,
     
107.                                                                      loss)
     
108.             train_losses.append(running_loss / train_steps)
     
109. 
     
110.         # validate, Modify the validation part
     
111.         net.eval()
     
112.         validate_loss = 0.0
     
113.         acc = 0.0  # accumulate accurate number / epoch
     
114.         val_steps = len(validate_loader)
     
115.         val_true_labels = []
     
116.         val_pred_labels = []
     
117.         with torch.no_grad():
     
118.             val_bar = tqdm(validate_loader, file=sys.stdout)
     
119.             for val_data in val_bar:
     
120.                 val_images, val_labels = val_data
     
121.                 outputs = net(val_images.to(device))
     
122.                 val_loss = loss_function(outputs,val_labels.to(device))
     
123.                 val_loss.backward()
     
124.                 optimizer.step()
     
125.                 validate_loss += val_loss.item()
     
126.                 val_bar.desc = "val epoch[{}/{}] loss:{:.3f}".format(epoch + 1,
     
127.                                                                      epochs,
     
128.                                                                      val_loss)
     
129.                 val_losses.append(validate_loss / val_steps)
     
130.                 val_true_labels.append(val_labels.cpu().numpy())
     
131.                 val_pred_labels.append((outputs > 0.5).float().cpu().numpy())
     
132. 
     
133.                 # predict_y = torch.max(outputs, dim=1)[1]
     
134.                 # acc += torch.eq(predict_y, val_labels.to(device)).sum().item()
     
135.         val_true_labels = np.vstack(val_true_labels)
     
136.         val_pred_labels = np.vstack(val_pred_labels)
     
137.         hamming, accuracy, precision, recall, fbeta = evaluate(val_true_labels, val_pred_labels)
     
138.         print('[epoch %d] train_loss: %.3f val_hamming: %.3f val_accuracy: %.3f val_precision: %.3f val_recall: %.3f')
     
139. 
     
140.         # val_accurate = acc / val_num
     
141.         # print('[epoch %d] train_loss: %.3f  val_accuracy: %.3f' %
     
142.         #       (epoch + 1, running_loss / train_steps, val_accurate))
     
143.         #
     
144.         # if val_accurate > best_acc:
     
145.         #     best_acc = val_accurate
     
146.         #     torch.save(net.state_dict(), save_path)
     
147.     plt.figure(figsize=(10, 5))
     
148.     plt.plot(train_losses, label='Training Loss')
     
149.     plt.plot(val_losses, label='Validation Loss')
     
150.     plt.xlabel('Epochs')
     
151.     plt.ylabel('Loss')
     
152.     plt.legend()
     
153.     plt.show()
     
154. 
     
155.     print('Finished Training')
     
156. 
     
157. 
     
158. if __name__ == '__main__':
     
159.     main()
     

复制代码

请指出错误原因并更改相应代码段

黎明丿晓小

isdkz 发表于 2023-4-22 12:25
根据您提供的代码，错误出现在处理多标签分类问题时的损失函数计算部分。`nn.BCEWithLogitsLoss`需要将标 ...

应该是“将标签的尺寸更改为与模型输入相同的尺寸”吧，另外，这样改会失去多标签分类任务的功能吗？

黎明丿晓小

isdkz 发表于 2023-4-22 12:43
抱歉，我在描述中的确有误，应该是“将标签的尺寸更改为与模型输出相同的尺寸”。

另外，不用担心，这 ...

运行以下代码时，在训练过程中报出警告“UndefinedMetricWarning: Precision is ill-defined and being set to 0.0 in labels with no predicted samples. Use zero_division parameter to control this behavior._warn_prf(average, modifier, msg_start, len(result))”，请解释报出警告的原因，并给出解决办法：

1. import os
   
2. import sys
   
3. import json
   
4. 
   
5. import torch
   
6. import torch.nn as nn
   
7. from torchvision import transforms, datasets
   
8. import torch.optim as optim
   
9. from tqdm import tqdm
   
10. from sklearn.metrics import hamming_loss,accuracy_score,precision_score,recall_score,fbeta_score
    
11. import numpy as np
    
12. import matplotlib.pyplot as plt
    
13. import warnings
    
14. 
    
15. from model import vgg
    
16. 
    
17. # Define the evaluation function
    
18. def evaluate(y_true, y_pred):
    
19.     hamming = hamming_loss(y_true, y_pred)
    
20.     accuracy = accuracy_score(y_true, y_pred)
    
21.     precision = precision_score(y_true, y_pred, average='micro')
    
22.     recall = recall_score(y_true, y_pred, average='micro')
    
23.     fbeta = fbeta_score(y_true, y_pred, beta=1, average='micro')
    
24. 
    
25.     return hamming, accuracy, precision, recall, fbeta
    
26. 
    
27. def main():
    
28.     device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    
29.     print("using {} device.".format(device))
    
30. 
    
31.     data_transform = {
    
32.         "train": transforms.Compose([transforms.RandomResizedCrop(224),
    
33.                                      transforms.RandomHorizontalFlip(),
    
34.                                      transforms.ToTensor(),
    
35.                                      transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]),
    
36.         "val": transforms.Compose([transforms.Resize((224, 224)),
    
37.                                    transforms.ToTensor(),
    
38.                                    transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])}
    
39. 
    
40.     data_root = os.path.abspath(os.path.join(os.getcwd(), "../.."))  # get data root path
    
41.     image_path = os.path.join(data_root, "data_set", "plantvillage_demo1")  # flower data set path
    
42.     assert os.path.exists(image_path), "{} path does not exist.".format(image_path)
    
43.     train_dataset = datasets.ImageFolder(root=os.path.join(image_path, "train"),
    
44.                                          transform=data_transform["train"])
    
45.     train_num = len(train_dataset)
    
46. 
    
47.     # {'daisy':0, 'dandelion':1, 'roses':2, 'sunflower':3, 'tulips':4}
    
48.     flower_list = train_dataset.class_to_idx
    
49.     cla_dict = dict((val, key) for key, val in flower_list.items())
    
50.     # write dict into json file
    
51.     json_str = json.dumps(cla_dict, indent=4)
    
52.     with open('class_indices.json', 'w') as json_file:
    
53.         json_file.write(json_str)
    
54. 
    
55.     batch_size = 32
    
56.     nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0, 8])  # number of workers
    
57.     print('Using {} dataloader workers every process'.format(nw))
    
58. 
    
59.     def label_to_onehot(labels, num_classes):
    
60.         batch_size = len(labels)
    
61.         one_hot = torch.zeros(batch_size, num_classes)
    
62.         one_hot[torch.arange(batch_size), labels] = 1
    
63.         return one_hot
    
64. 
    
65.     train_loader = torch.utils.data.DataLoader(train_dataset,
    
66.                                                batch_size=batch_size, shuffle=True,
    
67.                                                num_workers=nw)
    
68. 
    
69.     validate_dataset = datasets.ImageFolder(root=os.path.join(image_path, "val"),
    
70.                                             transform=data_transform["val"])
    
71.     val_num = len(validate_dataset)
    
72.     validate_loader = torch.utils.data.DataLoader(validate_dataset,
    
73.                                                   batch_size=batch_size, shuffle=False,
    
74.                                                   num_workers=nw)
    
75.     print("using {} images for training, {} images for validation.".format(train_num,
    
76.                                                                            val_num))
    
77. 
    
78.     # test_data_iter = iter(validate_loader)
    
79.     # test_image, test_label = test_data_iter.next()
    
80. 
    
81.     model_name = "vgg16"
    
82.     net = vgg(model_name=model_name, num_classes=4, init_weights=True)
    
83.     net.to(device)
    
84.     # Change the loss function to BCELoss
    
85.     loss_function = nn.BCEWithLogitsLoss()
    
86.     optimizer = optim.Adam(net.parameters(), lr=0.0001)
    
87. 
    
88.     train_losses = []
    
89.     #val_losses = []
    
90. 
    
91.     # Modify the training and validation loop to calculate evaluation metrics
    
92.     #warnings.filterwarnings("ignore")
    
93.     epochs = 30
    
94.     best_acc = 0.0
    
95.     save_path = './{}Net.pth'.format(model_name)
    
96.     train_steps = len(train_loader)
    
97.     for epoch in range(epochs):
    
98.         # train
    
99.         net.train()
    
100.         running_loss = 0.0
     
101.         train_bar = tqdm(train_loader, file=sys.stdout)
     
102.         for step, data in enumerate(train_bar):
     
103.             images, labels = data
     
104.             labels = label_to_onehot(labels, num_classes=4).to(device)
     
105.             optimizer.zero_grad()
     
106.             outputs = net(images.to(device))
     
107.             loss = loss_function(outputs, labels.to(device))
     
108.             loss.backward()
     
109.             optimizer.step()
     
110. 
     
111.             # print statistics
     
112.             running_loss += loss.item()
     
113. 
     
114.             train_bar.desc = "train epoch[{}/{}] loss:{:.3f}".format(epoch + 1,
     
115.                                                                      epochs,
     
116.                                                                      loss)
     
117.             train_losses.append(running_loss / train_steps)
     
118. 
     
119.         # validate, Modify the validation part
     
120.         net.eval()
     
121.         acc = 0.0  # accumulate accurate number / epoch
     
122.         val_true_labels = []
     
123.         val_pred_labels = []
     
124.         with torch.no_grad():
     
125.             val_bar = tqdm(validate_loader, file=sys.stdout)
     
126.             for val_data in val_bar:
     
127.                 val_images, val_labels = val_data
     
128.                 #val_labels = label_to_onehot(val_labels, num_classes=4).to(device)
     
129.                 outputs = net(val_images.to(device))
     
130.                 val_true_labels.append(label_to_onehot(val_labels,num_classes=4).cpu().numpy())
     
131.                 val_pred_labels.append((outputs > 0.5).float().cpu().numpy())
     
132. 
     
133.                 predict_y = torch.max(outputs, dim=1)[1]
     
134.                 acc += torch.eq(predict_y, val_labels.to(device)).sum().item()
     
135.         val_true_labels = np.vstack(val_true_labels)
     
136.         val_pred_labels = np.vstack(val_pred_labels)
     
137.         hamming, accuracy, precision, recall, fbeta = evaluate(val_true_labels, val_pred_labels)
     
138.         val_accurate = acc / val_num
     
139.         print('[epoch %d] train_loss: %.3f val_hamming: %.3f val_accuracy: %.3f val_precision: %.3f val_recall: %.3f fbeta: %.3f' %
     
140.               (epoch + 1, running_loss / train_steps, hamming, val_accurate, precision, recall, fbeta))
     
141. 
     
142. 
     
143.         # print('[epoch %d] train_loss: %.3f  val_accuracy: %.3f' %
     
144.         #       (epoch + 1, running_loss / train_steps, val_accurate))
     
145. 
     
146.         if val_accurate > best_acc:
     
147.             best_acc = val_accurate
     
148.             torch.save(net.state_dict(), save_path)
     
149.     plt.figure(figsize=(10, 5))
     
150.     plt.plot(train_losses, label='Training Loss')
     
151.     plt.plot(val_accurate, label='Validation Accuracy')
     
152.     plt.xlabel('Epochs')
     
153.     plt.ylabel('Loss/Accuracy')
     
154.     plt.legend()
     
155.     plt.show()
     
156. 
     
157.     print('Finished Training')
     
158. 
     
159. 
     
160. if __name__ == '__main__':
     
161.     main()
     

复制代码

黎明丿晓小

isdkz 发表于 2023-4-22 16:21
这个警告出现的原因是，在计算precision（精确度）时，有些标签在预测结果中没有出现，导致在计算这些标 ...

在运行以下代码时，报错“RuntimeError: Error(s) in loading state_dict for VGG:
        size mismatch for classifier.6.weight: copying a param with shape torch.Size([4, 4096]) from checkpoint, the shape in current model is torch.Size([5, 4096]).
        size mismatch for classifier.6.bias: copying a param with shape torch.Size([4]) from checkpoint, the shape in current model is torch.Size([5]).”
请给出错误原因以及修改意见

1. import os
   
2. import json
   
3. 
   
4. import torch
   
5. from PIL import Image
   
6. from torchvision import transforms
   
7. import matplotlib.pyplot as plt
   
8. 
   
9. from model import vgg
   
10. 
    
11. 
    
12. def main():
    
13.     device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    
14. 
    
15.     data_transform = transforms.Compose(
    
16.         [transforms.Resize((224, 224)),
    
17.          transforms.ToTensor(),
    
18.          transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])
    
19. 
    
20.     # load image
    
21.     img_path = "./test.jpg"
    
22.     assert os.path.exists(img_path), "file: '{}' dose not exist.".format(img_path)
    
23.     img = Image.open(img_path)
    
24.     plt.imshow(img)
    
25.     # [N, C, H, W]
    
26.     img = data_transform(img)
    
27.     # expand batch dimension
    
28.     img = torch.unsqueeze(img, dim=0)
    
29. 
    
30.     # read class_indict
    
31.     json_path = './class_indices.json'
    
32.     assert os.path.exists(json_path), "file: '{}' dose not exist.".format(json_path)
    
33. 
    
34.     with open(json_path, "r") as f:
    
35.         class_indict = json.load(f)
    
36.    
    
37.     # create model
    
38.     model = vgg(model_name="vgg16", num_classes=5).to(device)
    
39.     # load model weights
    
40.     weights_path = "./vgg16Net.pth"
    
41.     assert os.path.exists(weights_path), "file: '{}' dose not exist.".format(weights_path)
    
42.     model.load_state_dict(torch.load(weights_path, map_location=device))
    
43. 
    
44.     model.eval()
    
45.     threshold = 0.5
    
46.     with torch.no_grad():
    
47.         # predict class
    
48.         output = torch.squeeze(model(img.to(device))).cpu()
    
49.         probabilities = torch.sigmoid(output)
    
50.         predicted_labels = (probabilities > threshold).float().numpy()
    
51.         label_confidences = {class_indict[str(i)]: prob.item() for i, prob in enumerate(probabilities) if
    
52.                              predicted_labels[i] == 1}
    
53. 
    
54.     for label, confidence in label_confidences.items():
    
55.         print("Label: {}    Confidence: {:.3f}".format(label, confidence))
    
56. 
    
57.     plt.title("\n".join("{}: {:.3f}".format(label, confidence) for label, confidence in label_confidences.items()))
    
58.     plt.show()
    
59.     #     output = torch.squeeze(model(img.to(device))).cpu()
    
60.     #     predict = torch.softmax(output, dim=0)
    
61.     #     predict_cla = torch.argmax(predict).numpy()
    
62.     #
    
63.     # print_res = "class: {}   prob: {:.3}".format(class_indict[str(predict_cla)],
    
64.     #                                              predict[predict_cla].numpy())
    
65.     # plt.title(print_res)
    
66.     # for i in range(len(predict)):
    
67.     #     print("class: {:10}   prob: {:.3}".format(class_indict[str(i)],
    
68.     #                                               predict[i].numpy()))
    
69.     # plt.show()
    
70. 
    
71. 
    
72. if __name__ == '__main__':
    
73.     main()
    

复制代码

黎明丿晓小

黎明丿晓小 发表于 2023-4-22 16:16
运行以下代码时，在训练过程中报出警告“UndefinedMetricWarning: Precision is ill-defined and being ...

在运行此代码的过程中，hamming,  precision, recall, fbeta的输出分别一直为0.250,0.000,0.000,0.000。请说明原因并给出解决方法

黎明丿晓小

isdkz 发表于 2023-4-22 16:45
这个错误是因为您试图将预训练模型的权重（输出层有4个神经元）加载到具有不同输出层结构的模型中（输出 ...

请修改以下代码的数据读取方式以满足多标签分类任务的数据读取方式：

1. import os
   
2. import sys
   
3. import json
   
4. 
   
5. import torch
   
6. import torch.nn as nn
   
7. from torchvision import transforms, datasets
   
8. import torch.optim as optim
   
9. from tqdm import tqdm
   
10. from sklearn.metrics import hamming_loss,accuracy_score,precision_score,recall_score,fbeta_score
    
11. import numpy as np
    
12. import matplotlib.pyplot as plt
    
13. import warnings
    
14. 
    
15. from model import vgg
    
16. 
    
17. # Define the evaluation function
    
18. def evaluate(y_true, y_pred):
    
19.     hamming = hamming_loss(y_true, y_pred)
    
20.     accuracy = accuracy_score(y_true, y_pred)
    
21.     precision = precision_score(y_true, y_pred, average='micro')
    
22.     recall = recall_score(y_true, y_pred, average='micro')
    
23.     fbeta = fbeta_score(y_true, y_pred, beta=1, average='micro')
    
24. 
    
25.     return hamming, accuracy, precision, recall, fbeta
    
26. 
    
27. def main():
    
28.     device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    
29.     print("using {} device.".format(device))
    
30. 
    
31.     data_transform = {
    
32.         "train": transforms.Compose([transforms.RandomResizedCrop(224),
    
33.                                      transforms.RandomHorizontalFlip(),
    
34.                                      transforms.ToTensor(),
    
35.                                      transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]),
    
36.         "val": transforms.Compose([transforms.Resize((224, 224)),
    
37.                                    transforms.ToTensor(),
    
38.                                    transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])}
    
39. 
    
40.     data_root = os.path.abspath(os.path.join(os.getcwd(), "../.."))  # get data root path
    
41.     image_path = os.path.join(data_root, "data_set", "plantvillage_demo1")  # flower data set path
    
42.     assert os.path.exists(image_path), "{} path does not exist.".format(image_path)
    
43.     train_dataset = datasets.ImageFolder(root=os.path.join(image_path, "train"),
    
44.                                          transform=data_transform["train"])
    
45.     train_num = len(train_dataset)
    
46. 
    
47.     # {'daisy':0, 'dandelion':1, 'roses':2, 'sunflower':3, 'tulips':4}
    
48.     flower_list = train_dataset.class_to_idx
    
49.     cla_dict = dict((val, key) for key, val in flower_list.items())
    
50.     # write dict into json file
    
51.     json_str = json.dumps(cla_dict, indent=4)
    
52.     with open('class_indices.json', 'w') as json_file:
    
53.         json_file.write(json_str)
    
54. 
    
55.     batch_size = 32
    
56.     nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0, 8])  # number of workers
    
57.     print('Using {} dataloader workers every process'.format(nw))
    
58. 
    
59.     def label_to_onehot(labels, num_classes):
    
60.         batch_size = len(labels)
    
61.         one_hot = torch.zeros(batch_size, num_classes)
    
62.         one_hot[torch.arange(batch_size), labels] = 1
    
63.         return one_hot
    
64. 
    
65.     train_loader = torch.utils.data.DataLoader(train_dataset,
    
66.                                                batch_size=batch_size, shuffle=True,
    
67.                                                num_workers=nw)
    
68. 
    
69.     validate_dataset = datasets.ImageFolder(root=os.path.join(image_path, "val"),
    
70.                                             transform=data_transform["val"])
    
71.     val_num = len(validate_dataset)
    
72.     validate_loader = torch.utils.data.DataLoader(validate_dataset,
    
73.                                                   batch_size=batch_size, shuffle=False,
    
74.                                                   num_workers=nw)
    
75.     print("using {} images for training, {} images for validation.".format(train_num,
    
76.                                                                            val_num))
    
77. 
    
78.     # test_data_iter = iter(validate_loader)
    
79.     # test_image, test_label = test_data_iter.next()
    
80. 
    
81.     model_name = "vgg16"
    
82.     net = vgg(model_name=model_name, num_classes=4, init_weights=True)
    
83.     net.to(device)
    
84.     # Change the loss function to BCELoss
    
85.     loss_function = nn.BCEWithLogitsLoss()
    
86.     optimizer = optim.Adam(net.parameters(), lr=0.0001)
    
87. 
    
88.     train_losses = []
    
89.     #val_losses = []
    
90. 
    
91.     # Modify the training and validation loop to calculate evaluation metrics
    
92.     warnings.filterwarnings("ignore")
    
93.     epochs = 30
    
94.     best_acc = 0.0
    
95.     save_path = './{}Net.pth'.format(model_name)
    
96.     train_steps = len(train_loader)
    
97.     for epoch in range(epochs):
    
98.         # train
    
99.         net.train()
    
100.         running_loss = 0.0
     
101.         train_bar = tqdm(train_loader, file=sys.stdout)
     
102.         for step, data in enumerate(train_bar):
     
103.             images, labels = data
     
104.             labels = label_to_onehot(labels, num_classes=4).to(device)
     
105.             optimizer.zero_grad()
     
106.             outputs = net(images.to(device))
     
107.             loss = loss_function(outputs, labels.to(device))
     
108.             loss.backward()
     
109.             optimizer.step()
     
110. 
     
111.             # print statistics
     
112.             running_loss += loss.item()
     
113. 
     
114.             train_bar.desc = "train epoch[{}/{}] loss:{:.3f}".format(epoch + 1,
     
115.                                                                      epochs,
     
116.                                                                      loss)
     
117.             train_losses.append(running_loss / train_steps)
     
118. 
     
119.         # validate, Modify the validation part
     
120.         net.eval()
     
121.         acc = 0.0  # accumulate accurate number / epoch
     
122.         val_true_labels = []
     
123.         val_pred_labels = []
     
124.         with torch.no_grad():
     
125.             val_bar = tqdm(validate_loader, file=sys.stdout)
     
126.             for val_data in val_bar:
     
127.                 val_images, val_labels = val_data
     
128.                 #val_labels = label_to_onehot(val_labels, num_classes=4).to(device)
     
129.                 outputs = net(val_images.to(device))
     
130.                 val_true_labels.append(label_to_onehot(val_labels,num_classes=4).cpu().numpy())
     
131.                 val_pred_labels.append((outputs > 0.5).float().cpu().numpy())
     
132. 
     
133.                 predict_y = torch.max(outputs, dim=1)[1]
     
134.                 acc += torch.eq(predict_y, val_labels.to(device)).sum().item()
     
135.         val_true_labels = np.vstack(val_true_labels)
     
136.         val_pred_labels = np.vstack(val_pred_labels)
     
137.         hamming, accuracy, precision, recall, fbeta = evaluate(val_true_labels, val_pred_labels)
     
138.         val_accurate = acc / val_num
     
139.         print('[epoch %d] train_loss: %.3f val_hamming: %.3f val_accuracy: %.3f val_precision: %.3f val_recall: %.3f fbeta: %.3f' %
     
140.               (epoch + 1, running_loss / train_steps, hamming, val_accurate, precision, recall, fbeta))
     
141. 
     
142. 
     
143.         # print('[epoch %d] train_loss: %.3f  val_accuracy: %.3f' %
     
144.         #       (epoch + 1, running_loss / train_steps, val_accurate))
     
145. 
     
146.         if val_accurate > best_acc:
     
147.             best_acc = val_accurate
     
148.             torch.save(net.state_dict(), save_path)
     
149.     plt.figure(figsize=(10, 5))
     
150.     plt.plot(train_losses, label='Training Loss')
     
151.     plt.plot(val_accurate, label='Validation Accuracy')
     
152.     plt.xlabel('Epochs')
     
153.     plt.ylabel('Loss/Accuracy')
     
154.     plt.legend()
     
155.     plt.show()
     
156. 
     
157.     print('Finished Training')
     
158. 
     
159. 
     
160. if __name__ == '__main__':
     
161.     main()
     

复制代码