File size: 5,042 Bytes
3dbf358
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134

import webdataset as wds
from pathlib import Path
import torch
from torchvision import transforms
from PIL import Image
import io

def identity(x):
    return x

def pil_decoder(key, data):
    """Decodes image data from bytes to a PIL Image."""
    if not key.endswith((".jpg", ".jpeg", ".png")):
        return None
    try:
        return Image.open(io.BytesIO(data)).convert("RGB")
    except Exception:
        return None

def cls_decoder(key, data):
    """Decodes class label from bytes."""
    if not key.endswith(".cls"):
        return None
    try:
        return int(data.decode('utf-8'))
    except (ValueError, UnicodeDecodeError):
        return None

class MiniImageNetCWebDataset(torch.utils.data.IterableDataset):
    """
    A PyTorch Dataset for the WebDataset version of MiniImageNet-C.

    Args:
        root (str): The root directory of the WebDataset shards.
        corruption (str): The corruption type to load (e.g., 'gaussian_noise').
        severity (int): The severity level (should be 5 for MiniImageNet-C).
        transform (callable, optional): A function/transform that takes in a PIL image
            and returns a transformed version. E.g, `transforms.ToTensor()`.
        target_transform (callable, optional): A function/transform that takes in the
            target and transforms it.
    """
    def __init__(self, root, corruption, severity=5, transform=None, target_transform=None):
        self.root = Path(root)
        self.corruption = corruption
        self.severity = severity
        self.transform = transform if transform is not None else identity
        self.target_transform = target_transform if target_transform is not None else identity

        self.shard_path = self.root / self.corruption / str(self.severity)
        if not self.shard_path.exists():
            raise FileNotFoundError(f"Shards not found at: {self.shard_path}")

        shard_urls = [str(p) for p in sorted(self.shard_path.glob("*.tar"))]
        if not shard_urls:
            raise FileNotFoundError(f"No .tar shards found in {self.shard_path}")

        self.dataset = (
            wds.WebDataset(shard_urls, shardshuffle=True)
            .decode(pil_decoder, cls_decoder)
            .to_tuple("jpg", "cls")
            .map(self.apply_transforms)
        )

    def apply_transforms(self, sample):
        image, target = sample
        return self.transform(image), self.target_transform(target)

    def __iter__(self):
        return iter(self.dataset)

    def __len__(self):
        # The length of a WebDataset is not trivially known beforehand.
        # You can estimate it or, if needed, iterate through it once to count.
        # For Mini-ImageNet-C, each class has 50 images, and there are 1000 classes.
        return 50 * 1000

# Example Usage
if __name__ == '__main__':
    print("Example of how to use MiniImageNetCWebDataset")

    # This assumes you have a 'mini-imagenet-c-webdataset' directory
    # created by the convert_to_webdataset.py script.
    dataset_root = "../data/mini-imagenet-c-webdataset" 
    if not Path(dataset_root).exists():
        print(f"\nERROR: Example dataset root '{dataset_root}' not found.")
        print("Please run 'python data/scripts/convert_to_webdataset.py' first.")
        exit()

    # Define transformations for the images
    image_transform = transforms.Compose([
        transforms.Resize((224, 224)),
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ])

    # 1. Create a dataset for a specific corruption
    try:
        corruption_type = 'gaussian_noise'
        print(f"\nLoading dataset for corruption: '{corruption_type}'")
        
        dataset = MiniImageNetCWebDataset(
            root=dataset_root,
            corruption=corruption_type,
            transform=image_transform
        )

        # 2. Create a DataLoader
        # WebDataset is designed for streaming, so shuffling is handled differently.
        # For shuffling, you typically shuffle the shard URLs and the samples within each shard.
        # The loader here provides a basic sequential stream.
        dataloader = torch.utils.data.DataLoader(dataset, batch_size=32, num_workers=4)

        # 3. Iterate through a few batches
        print("Iterating through a few batches...")
        for i, (images, labels) in enumerate(dataloader):
            if i >= 3:
                break
            print(f"  Batch {i+1}:")
            print(f"    Images shape: {images.shape}")
            print(f"    Labels shape: {labels.shape}")
            print(f"    Sample labels: {labels[:4].tolist()}")

        print("\nExample finished successfully!")

    except FileNotFoundError as e:
        print(f"\nERROR: Could not run example. {e}")
        print("Please ensure the WebDataset has been generated and the paths are correct.")

    except ImportError:
        print("\nERROR: 'webdataset' library not found.")
        print("Please install it by running: pip install webdataset")