build_semantic_dataset
Builds a semantic anomaly detection dataset for training or evaluation.
from laft.datasets import build_semantic_dataset
dataset = build_semantic_dataset(
name="color_mnist",
split="train",
root="./data",
transform=None,
config=None
)
Parameters
name
Literal['color_mnist', 'waterbirds', 'celeba']
required
Name of the semantic anomaly dataset to build.Supported datasets:
color_mnist: MNIST digits with color attributeswaterbirds: Birds in different backgroundsceleba: Celebrity faces with attributes
split
Literal['train', 'valid', 'test']
required
Dataset split to load.
Root directory where the dataset files are stored or will be downloaded.
transform
Callable | None
default:"None"
Optional transform to apply to images. Typically a torchvision transform pipeline.
config
dict | None
default:"None"
Dataset-specific configuration dictionary. When None, uses default configuration.For color_mnist, config structure:{
"number": {0: False, 1: False, ..., 9: True}, # False=normal, True=anomaly
"color": {"red": False, "green": True, "blue": True}
}
Additional keyword arguments. For color_mnist, accepts seed parameter for reproducible train/valid splits.
Returns
A dataset instance that inherits from SemanticAnomalyDataset.
Examples
Load ColorMNIST with default config
from laft.datasets import build_semantic_dataset
from torchvision import transforms
transform = transforms.Compose([
transforms.Resize(224),
transforms.ToTensor()
])
# Load training set (normal samples only)
train_dataset = build_semantic_dataset(
name="color_mnist",
split="train",
transform=transform
)
# Load test set (normal and anomaly samples)
test_dataset = build_semantic_dataset(
name="color_mnist",
split="test",
transform=transform
)
Load with custom config
# Custom configuration: digits 0-7 are normal, 8-9 are anomalies
custom_config = {
"number": {
0: False, 1: False, 2: False, 3: False,
4: False, 5: False, 6: False, 7: False,
8: True, 9: True
},
"color": {
"red": False,
"green": True,
"blue": True
}
}
dataset = build_semantic_dataset(
name="color_mnist",
split="train",
config=custom_config,
seed=123 # Custom seed for train/valid split
)
Load other semantic datasets
# Waterbirds dataset
waterbirds = build_semantic_dataset(
name="waterbirds",
split="test",
transform=transform
)
# CelebA dataset
celeba = build_semantic_dataset(
name="celeba",
split="test",
transform=transform
)
build_industrial_dataset
Builds an industrial anomaly detection dataset for a specific product category.
from laft.datasets import build_industrial_dataset
dataset = build_industrial_dataset(
name="mvtec",
category="bottle",
split="test",
root="./data"
)
Parameters
name
Literal['mvtec', 'visa']
required
Name of the industrial anomaly dataset.Supported datasets:
mvtec: MVTec Anomaly Detection datasetvisa: VisA (Visual Anomaly) dataset
Product category to load.For MVTec, supported categories:
bottle, cable, capsule, carpet, grid, hazelnut, leather, metal_nut, pill, screw, tile, toothbrush, transistor, wood, zipper
split
Literal['train', 'test']
required
Dataset split to load.
Root directory where the dataset files are stored.
transform
Callable | None
default:"None"
Optional transform to apply to images.
mask_transform
Callable | None
default:"None"
Optional transform to apply to anomaly masks.
Returns
A dataset instance that inherits from IndustrialAnomalyDataset.
Examples
Load MVTec dataset
from laft.datasets import build_industrial_dataset
from torchvision import transforms
image_transform = transforms.Compose([
transforms.Resize(224),
transforms.ToTensor()
])
mask_transform = transforms.Compose([
transforms.Resize(224)
])
# Load test set for bottle category
dataset = build_industrial_dataset(
name="mvtec",
category="bottle",
split="test",
root="./data",
transform=image_transform,
mask_transform=mask_transform
)
# Iterate through dataset
for image, mask, label in dataset:
print(f"Image shape: {image.shape}")
print(f"Mask shape: {mask.shape}")
print(f"Is anomaly: {label.item()}")
break
Load multiple categories
categories = ["bottle", "cable", "capsule"]
datasets = {}
for category in categories:
datasets[category] = build_industrial_dataset(
name="mvtec",
category=category,
split="test",
transform=image_transform
)
SemanticAnomalyDataset
Base class for semantic anomaly detection datasets.
Attributes
List of attribute names (e.g., ["number", "color"] for ColorMNIST).
Boolean tensor of shape [num_samples, num_attrs] where False indicates normal and True indicates anomaly for each attribute.
Root directory of the dataset.
Dataset split (train, valid, or test).
Transform applied to images.
Methods
get_normal_subset()
Returns a subset containing only normal samples (all attributes are False).
dataset = build_semantic_dataset("color_mnist", "train")
normal_subset = dataset.get_normal_subset()
print(f"Total samples: {len(dataset)}")
print(f"Normal samples: {len(normal_subset)}")
A PyTorch Subset containing only samples where all attributes are normal.
load_image(index: int)
Loads the image at the given index. Must be implemented by subclasses.
__getitem__(index: int)
Returns a tuple of (image, target) where target is the attribute tensor.
image, attrs = dataset[0]
print(f"Image shape: {image.shape}")
print(f"Attributes: {attrs}") # e.g., tensor([False, True])
Tuple of (transformed image, attribute tensor).
IndustrialAnomalyDataset
Base class for industrial anomaly detection datasets.
Attributes
Boolean tensor of shape [num_samples] where False indicates normal and True indicates anomaly.
Root directory of the dataset.
Dataset split (train or test).
Transform applied to images.
Transform applied to anomaly masks.
Methods
load_image(index: int)
Loads the image at the given index. Must be implemented by subclasses.
load_mask(index: int)
Loads the anomaly mask at the given index. Must be implemented by subclasses.
Boolean tensor indicating anomalous regions.
__getitem__(index: int)
Returns a tuple of (image, mask, label).
image, mask, label = dataset[0]
print(f"Image shape: {image.shape}")
print(f"Mask shape: {mask.shape}")
print(f"Is anomaly: {label.item()}")
item
tuple[Any, torch.Tensor, torch.Tensor]
Tuple of (transformed image, mask, label). For normal samples, mask is all zeros.
Example Usage
from laft.datasets import build_industrial_dataset
from torch.utils.data import DataLoader
dataset = build_industrial_dataset(
name="mvtec",
category="bottle",
split="test"
)
loader = DataLoader(dataset, batch_size=4, shuffle=False)
for images, masks, labels in loader:
anomaly_count = labels.sum().item()
print(f"Batch has {anomaly_count} anomalies")
break
