迁移学习之Multi-Domain Adaptation多领域自适应常用数据集PACS介绍

Struggle

PACS数据集

Paper：Self-supervised Domain Adaptation for Computer Vision Tasks
GitHub：https://github.com/robertofranceschi/Domain-adaptation-on-PACS-dataset
数据集下载：https://github.com/MachineLearning2020/Homework3-PACS/tree/master/PACS

• PACS数据集总共9991张图片，每张图片3x227x227
  
• 7 classes：Dog, Elephant, Giraffe, Guitar, Horse, House, Person
  
• 4 domains: Art painting, Cartoon, Photo and Sketch.
  
• Photo (1,670 images), Art Painting (2,048 images), Cartoon (2,344 images) and Sketch (3,929 images)
  

用Pytorch加载PACS数据集

PACS原始数据集目录结果：

1. from torch.utils.data import Dataset, DataLoader
2. import matplotlib.pyplot as plt
3. import os
4. import random
5. from PIL import Image
6. import torch
7. import numpy as np
8. from torchvision.transforms import transforms
9. from sklearn.model_selection import train_test_split
10. import os
11. import torchvision.transforms as transforms
12. from PIL import Image
13. from torch.utils.data import Dataset, DataLoader
15. class PACS(Dataset):
16.     def __init__(self, root_path, domain, train=True, transform=None, target_transform=None):
17.         self.root = f"{root_path}/{domain}"
18.         self.train = train
19.         self.transform = transform
20.         self.target_transform = target_transform
22.         label_name_list = os.listdir(self.root)
23.         self.label = []
24.         self.data = []
26.         if not os.path.exists(f"{root_path}/precessed"):
27.             os.makedirs(f"{root_path}/precessed")
28.         if os.path.exists(f"{root_path}/precessed/{domain}_data.pt") and os.path.exists(
29.                 f"{root_path}/precessed/{domain}_label.pt"):
30.             print(f"Load {domain} data and label from cache.")
31.             self.data = torch.load(f"{root_path}/precessed/{domain}_data.pt")
32.             self.label = torch.load(f"{root_path}/precessed/{domain}_label.pt")
33.         else:
34.             print(f"Getting {domain} datasets")
35.             for index, label_name in enumerate(label_name_list):
36.                 label_name_2_index = {
37.                     'dog': 0,
38.                     'elephant': 1,
39.                     'giraffe': 2,
40.                     'guitar': 3,
41.                     'horse': 4,
42.                     'house': 5,
43.                     'person': 6,
44.                 }
45.                 images_list = os.listdir(f"{self.root}/{label_name}")
46.                 for img_name in images_list:
47.                     img = Image.open(f"{self.root}/{label_name}/{img_name}").convert('RGB')
48.                     img = np.array(img)
49.                     self.label.append(label_name_2_index[label_name])
50.                     if self.transform is not None:
51.                         img = self.transform(img)
52.                     self.data.append(img)
53.             self.data = torch.stack(self.data)
54.             self.label = torch.tensor(self.label, dtype=torch.long)
55.             torch.save(self.data, f"{root_path}/precessed/{domain}_data.pt")
56.             torch.save(self.label, f"{root_path}/precessed/{domain}_label.pt")
58.     def __getitem__(self, index):
59.         img, target = self.data[index], self.label[index]
61.         if self.transform is not None:
62.             img = self.transform(img)
63.         if self.target_transform is not None:
64.             target = self.target_transform(target)
66.         return img, target
68.     def __len__(self):
69.         return len(self.data)
72. def get_pacs_domain(root_path=f"{DATA_PATH}/PACS", domain='art_painting', verbose=False):
73.     transform = transforms.Compose([
74.         transforms.ToPILImage(),
75.         transforms.ToTensor(),
76.         # transforms.Resize((224, 224)),
77.         transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))
78.     ])
79.     all_data = PACS(root_path, domain, transform=transform)
80.     # train:test=8:2
81.     x_train, x_test, y_train, y_test = train_test_split(all_data.data.numpy(), all_data.label.numpy(),
82.                                                         test_size=0.20, random_state=42)
84.     return x_train, y_train, x_test, y_test

复制代码

其他写法：https://github.com/ValerioDiEugenio/DomainAdaptation-PACSDataset/blob/main/DomainAdaptation.ipynb

---

最后以dog类别为例，用Python代码展示四种不同风格的图片
Python可视化图片数据集的代码

修改dir_path为对应的文件夹，dir_path = f"{DATA_PATH}/PACS/art_painting/dog"

1. import os
2. import matplotlib.pyplot as plt
3. import random
4. from PIL import Image
7. def plotPics(data, h=3, w=3, filename="out.jpg"):
8.     fig, ax_array = plt.subplots(h, w, figsize=(15, 15))
10.     axes = ax_array.flatten()
12.     for i, ax in enumerate(axes):
13.         ri = random.randint(0, len(data) - 1)
14.         ax.imshow(data[ri], cmap=plt.cm.gray)
16.     plt.setp(axes, xticks=[], yticks=[], frame_on=False)
17.     fig.tight_layout()
18.     fig.savefig(filename)
19.     plt.show()
22. DATA_PATH = os.path.abspath(os.path.join(os.path.dirname(__file__), "../data/raw_data"))
24. dir_path = f"{DATA_PATH}/PACS/art_painting/dog"
26. data = []
27. for pic in os.listdir(dir_path):
28.     data.append(Image.open(f"{dir_path}/{pic}"))
30. plotPics(data, h=5, w=5)

复制代码

art_painting风格

sketch风格

cartoon风格

photo风格

来源：https://blog.csdn.net/qq_43827595/article/details/121345640
免责声明：如果侵犯了您的权益，请联系站长，我们会及时删除侵权内容，谢谢合作！