Image Datamodules

Image Datamodules#

Image datamodules in Anomalib are designed to handle image-based anomaly detection datasets. They provide a standardized interface for loading and processing image data for both training and inference.

Available Datamodules#

BTech

Surface defect detection in steel manufacturing.

anomalib.data.datamodules.image.BTech
Datumaro

Dataset format compatible with Intel Geti™.

anomalib.data.datamodules.image.Datumaro
Folder

Custom folder-based dataset organization.

anomalib.data.datamodules.image.Folder
Kolektor

Surface defect detection in electrical commutators.

anomalib.data.datamodules.image.Kolektor
MVTec

Industrial anomaly detection benchmark.

anomalib.data.datamodules.image.MVTec
Visa

Visual inspection of surface anomalies.

anomalib.data.datamodules.image.Visa

API Reference#

Anomalib Datasets.

This module provides datasets and data modules for anomaly detection tasks.

The module contains:

  • Data classes for representing different types of data (images, videos, etc.)

  • Dataset classes for loading and processing data

  • Data modules for use with PyTorch Lightning

  • Helper functions for data loading and validation

Example

>>> from anomalib.data import get_datamodule
>>> from omegaconf import DictConfig
>>> config = DictConfig({
...     "data": {
...         "class_path": "MVTec",
...         "init_args": {
...             "root": "./datasets/MVTec",
...             "category": "bottle",
...             "image_size": (256, 256)
...         }
...     }
... })
>>> datamodule = get_datamodule(config)
class anomalib.data.BTech(root='./datasets/BTech', category='01', train_batch_size=32, eval_batch_size=32, num_workers=8, train_augmentations=None, val_augmentations=None, test_augmentations=None, augmentations=None, test_split_mode=TestSplitMode.FROM_DIR, test_split_ratio=0.2, val_split_mode=ValSplitMode.SAME_AS_TEST, val_split_ratio=0.5, seed=None)#

Bases: AnomalibDataModule

BTech Lightning Data Module.

Parameters:

  • root (Path | str) – Path to the root of the dataset. Defaults to "./datasets/BTech".

  • category (str) – Category of the BTech dataset (e.g. "01", "02", or "03"). Defaults to "01".

  • train_batch_size (int, optional) – Training batch size. Defaults to 32.

  • eval_batch_size (int, optional) – Test batch size. Defaults to 32.

  • num_workers (int, optional) – Number of workers. Defaults to 8.

  • train_augmentations (Transform | None) – Augmentations to apply dto the training images Defaults to None.

  • val_augmentations (Transform | None) – Augmentations to apply to the validation images. Defaults to None.

  • test_augmentations (Transform | None) – Augmentations to apply to the test images. Defaults to None.

  • augmentations (Transform | None) – General augmentations to apply if stage-specific augmentations are not provided.

  • test_split_mode (TestSplitMode) – Setting that determines how the testing subset is obtained. Defaults to TestSplitMode.FROM_DIR.

  • test_split_ratio (float) – Fraction of images from the train set that will be reserved for testing. Defaults to 0.2.

  • val_split_mode (ValSplitMode) – Setting that determines how the validation subset is obtained. Defaults to ValSplitMode.SAME_AS_TEST.

  • val_split_ratio (float) – Fraction of train or test images that will be reserved for validation. Defaults to 0.5.

  • seed (int | None, optional) – Seed which may be set to a fixed value for reproducibility. Defaults to None.

Example

To create the BTech datamodule, instantiate the class and call setup:

>>> from anomalib.data import BTech
>>> datamodule = BTech(
...     root="./datasets/BTech",
...     category="01",
...     train_batch_size=32,
...     eval_batch_size=32,
...     num_workers=8,
... )
>>> datamodule.setup()

Get the train dataloader and first batch:

>>> i, data = next(enumerate(datamodule.train_dataloader()))
>>> data.keys()
dict_keys(['image'])
>>> data["image"].shape
torch.Size([32, 3, 256, 256])

Access the validation dataloader and first batch:

>>> i, data = next(enumerate(datamodule.val_dataloader()))
>>> data.keys()
dict_keys(['image_path', 'label', 'mask_path', 'image', 'mask'])
>>> data["image"].shape, data["mask"].shape
(torch.Size([32, 3, 256, 256]), torch.Size([32, 256, 256]))

Access the test dataloader and first batch:

>>> i, data = next(enumerate(datamodule.test_dataloader()))
>>> data.keys()
dict_keys(['image_path', 'label', 'mask_path', 'image', 'mask'])
>>> data["image"].shape, data["mask"].shape
(torch.Size([32, 3, 256, 256]), torch.Size([32, 256, 256]))
prepare_data()#

Download the dataset if not available.

This method checks if the specified dataset is available in the file system. If not, it downloads and extracts the dataset into the appropriate directory.

Return type:

None

Example

Assume the dataset is not available on the file system. Here’s how the directory structure looks before and after calling prepare_data:

# Before
$ tree datasets
datasets
├── dataset1
└── dataset2

# Calling prepare_data
>>> datamodule = BTech(root="./datasets/BTech", category="01")
>>> datamodule.prepare_data()

# After
$ tree datasets
datasets
├── dataset1
├── dataset2
└── BTech
    ├── 01
    ├── 02
    └── 03
class anomalib.data.Datumaro(root, train_batch_size=32, eval_batch_size=32, num_workers=8, train_augmentations=None, val_augmentations=None, test_augmentations=None, augmentations=None, test_split_mode=TestSplitMode.FROM_DIR, test_split_ratio=0.5, val_split_mode=ValSplitMode.FROM_TEST, val_split_ratio=0.5, seed=None)#

Bases: AnomalibDataModule

Datumaro datamodule.

Parameters:

  • root (Path | str) – Path to the dataset root directory.

  • train_batch_size (int, optional) – Training batch size. Defaults to 32.

  • eval_batch_size (int, optional) – Test batch size. Defaults to 32.

  • num_workers (int, optional) – Number of workers. Defaults to 8.

  • train_augmentations (Transform | None) – Augmentations to apply dto the training images Defaults to None.

  • val_augmentations (Transform | None) – Augmentations to apply to the validation images. Defaults to None.

  • test_augmentations (Transform | None) – Augmentations to apply to the test images. Defaults to None.

  • augmentations (Transform | None) – General augmentations to apply if stage-specific augmentations are not provided.

  • image_size (tuple[int, int], optional) – Size to which input images should be resized. Defaults to None.

  • test_split_mode (TestSplitMode) – Setting that determines how the testing subset is obtained. Defaults to TestSplitMode.FROM_DIR.

  • test_split_ratio (float) – Fraction of images from the train set that will be reserved for testing. Defaults to 0.2.

  • val_split_mode (ValSplitMode) – Setting that determines how the validation subset is obtained. Defaults to ValSplitMode.SAME_AS_TEST.

  • val_split_ratio (float) – Fraction of train or test images that will be reserved for validation. Defaults to 0.5.

  • seed (int | None, optional) – Seed which may be set to a fixed value for reproducibility. Defaults to None.

Example

>>> from anomalib.data import Datumaro
>>> datamodule = Datumaro(
...     root="./datasets/datumaro",
...     train_batch_size=32,
...     eval_batch_size=32,
...     num_workers=8,
... )
>>> datamodule.setup()
>>> i, data = next(enumerate(datamodule.train_dataloader()))
>>> data.keys()
dict_keys(['image_path', 'label', 'image'])
class anomalib.data.Folder(name, normal_dir, root=None, abnormal_dir=None, normal_test_dir=None, mask_dir=None, normal_split_ratio=0.2, extensions=None, train_batch_size=32, eval_batch_size=32, num_workers=8, train_augmentations=None, val_augmentations=None, test_augmentations=None, augmentations=None, test_split_mode=TestSplitMode.FROM_DIR, test_split_ratio=0.2, val_split_mode=ValSplitMode.FROM_TEST, val_split_ratio=0.5, seed=None)#

Bases: AnomalibDataModule

Folder DataModule.

Parameters:

  • name (str) – Name of the dataset. Used for logging/saving.

  • normal_dir (str | Path | Sequence) – Directory containing normal images.

  • root (str | Path | None) – Root folder containing normal and abnormal directories. Defaults to None.

  • abnormal_dir (str | Path | None | Sequence) – Directory containing abnormal images. Defaults to None.

  • normal_test_dir (str | Path | Sequence | None) – Directory containing normal test images. Defaults to None.

  • mask_dir (str | Path | Sequence | None) – Directory containing mask annotations. Defaults to None.

  • normal_split_ratio (float) – Ratio to split normal training images for test set when no normal test images exist. Defaults to 0.2.

  • extensions (tuple[str, ...] | None) – Image extensions to include. Defaults to None.

  • train_batch_size (int) – Training batch size. Defaults to 32.

  • eval_batch_size (int) – Validation/test batch size. Defaults to 32.

  • num_workers (int) – Number of workers for data loading. Defaults to 8.

  • train_augmentations (Transform | None) – Augmentations to apply dto the training images Defaults to None.

  • val_augmentations (Transform | None) – Augmentations to apply to the validation images. Defaults to None.

  • test_augmentations (Transform | None) – Augmentations to apply to the test images. Defaults to None.

  • augmentations (Transform | None) – General augmentations to apply if stage-specific augmentations are not provided.

  • test_split_mode (TestSplitMode) – Method to obtain test subset. Defaults to TestSplitMode.FROM_DIR.

  • test_split_ratio (float) – Fraction of train images for testing. Defaults to 0.2.

  • val_split_mode (ValSplitMode) – Method to obtain validation subset. Defaults to ValSplitMode.FROM_TEST.

  • val_split_ratio (float) – Fraction of images for validation. Defaults to 0.5.

  • seed (int | None) – Random seed for splitting. Defaults to None.

Example

Create and setup a folder datamodule:

>>> from anomalib.data import Folder
>>> datamodule = Folder(
...     name="custom",
...     root="./datasets/custom",
...     normal_dir="good",
...     abnormal_dir="defect",
...     mask_dir="mask"
... )
>>> datamodule.setup()

Get a batch from train dataloader:

>>> batch = next(iter(datamodule.train_dataloader()))
>>> batch.keys()
dict_keys(['image', 'label', 'mask', 'image_path', 'mask_path'])

Get a batch from test dataloader:

>>> batch = next(iter(datamodule.test_dataloader()))
>>> batch.keys()
dict_keys(['image', 'label', 'mask', 'image_path', 'mask_path'])
property name: str#

Get name of the datamodule.

Returns:

Name of the datamodule.

class anomalib.data.Kolektor(root='./datasets/kolektor', train_batch_size=32, eval_batch_size=32, num_workers=8, train_augmentations=None, val_augmentations=None, test_augmentations=None, augmentations=None, test_split_mode=TestSplitMode.FROM_DIR, test_split_ratio=0.2, val_split_mode=ValSplitMode.SAME_AS_TEST, val_split_ratio=0.5, seed=None)#

Bases: AnomalibDataModule

Kolektor Surface-Defect DataModule.

Parameters:

  • root (Path | str) – Path to the root of the dataset. Defaults to "./datasets/kolektor".

  • train_batch_size (int, optional) – Training batch size. Defaults to 32.

  • eval_batch_size (int, optional) – Test batch size. Defaults to 32.

  • num_workers (int, optional) – Number of workers. Defaults to 8.

  • train_augmentations (Transform | None) – Augmentations to apply dto the training images Defaults to None.

  • val_augmentations (Transform | None) – Augmentations to apply to the validation images. Defaults to None.

  • test_augmentations (Transform | None) – Augmentations to apply to the test images. Defaults to None.

  • augmentations (Transform | None) – General augmentations to apply if stage-specific augmentations are not provided.

  • test_split_mode (TestSplitMode) – Setting that determines how the testing subset is obtained. Defaults to TestSplitMode.FROM_DIR.

  • test_split_ratio (float) – Fraction of images from the train set that will be reserved for testing. Defaults to 0.2.

  • val_split_mode (ValSplitMode) – Setting that determines how the validation subset is obtained. Defaults to ValSplitMode.SAME_AS_TEST.

  • val_split_ratio (float) – Fraction of train or test images that will be reserved for validation. Defaults to 0.5.

  • seed (int | None, optional) – Seed which may be set to a fixed value for reproducibility.

Example

>>> from anomalib.data import Kolektor
>>> datamodule = Kolektor(
...     root="./datasets/kolektor",
...     train_batch_size=32,
...     eval_batch_size=32,
...     num_workers=8,
... )
>>> datamodule.setup()
>>> i, data = next(enumerate(datamodule.train_dataloader()))
>>> data.keys()
dict_keys(['image', 'label', 'mask', 'image_path', 'mask_path'])
prepare_data()#

Download the dataset if not available.

This method checks if the specified dataset is available in the file system. If not, it downloads and extracts the dataset into the appropriate directory.

Return type:

None

Example

Assume the dataset is not available on the file system. Here’s how the directory structure looks before and after calling the prepare_data method:

Before:

$ tree datasets
datasets
├── dataset1
└── dataset2

Calling the method:

>>> datamodule = Kolektor(root="./datasets/kolektor")
>>> datamodule.prepare_data()

After:

$ tree datasets
datasets
├── dataset1
├── dataset2
└── kolektor
    ├── kolektorsdd
    ├── kos01
    ├── ...
    └── kos50
        ├── Part0.jpg
        ├── Part0_label.bmp
        └── ...
class anomalib.data.MVTec(root='./datasets/MVTec', category='bottle', train_batch_size=32, eval_batch_size=32, num_workers=8, train_augmentations=None, val_augmentations=None, test_augmentations=None, augmentations=None, test_split_mode=TestSplitMode.FROM_DIR, test_split_ratio=0.2, val_split_mode=ValSplitMode.SAME_AS_TEST, val_split_ratio=0.5, seed=None)#

Bases: AnomalibDataModule

MVTec Datamodule.

Parameters:

  • root (Path | str) – Path to the root of the dataset. Defaults to "./datasets/MVTec".

  • category (str) – Category of the MVTec dataset (e.g. "bottle" or "cable"). Defaults to "bottle".

  • train_batch_size (int, optional) – Training batch size. Defaults to 32.

  • eval_batch_size (int, optional) – Test batch size. Defaults to 32.

  • num_workers (int, optional) – Number of workers. Defaults to 8.

  • train_augmentations (Transform | None) – Augmentations to apply dto the training images Defaults to None.

  • val_augmentations (Transform | None) – Augmentations to apply to the validation images. Defaults to None.

  • test_augmentations (Transform | None) – Augmentations to apply to the test images. Defaults to None.

  • augmentations (Transform | None) – General augmentations to apply if stage-specific augmentations are not provided.

  • test_split_mode (TestSplitMode) – Method to create test set. Defaults to TestSplitMode.FROM_DIR.

  • test_split_ratio (float) – Fraction of data to use for testing. Defaults to 0.2.

  • val_split_mode (ValSplitMode) – Method to create validation set. Defaults to ValSplitMode.SAME_AS_TEST.

  • val_split_ratio (float) – Fraction of data to use for validation. Defaults to 0.5.

  • seed (int | None, optional) – Seed for reproducibility. Defaults to None.

Example

Create MVTec datamodule with default settings:

>>> datamodule = MVTec()
>>> datamodule.setup()
>>> i, data = next(enumerate(datamodule.train_dataloader()))
>>> data.keys()
dict_keys(['image_path', 'label', 'image', 'mask_path', 'mask'])

>>> data["image"].shape
torch.Size([32, 3, 256, 256])

Change the category:

>>> datamodule = MVTec(category="cable")

Create validation set from test data:

>>> datamodule = MVTec(
...     val_split_mode=ValSplitMode.FROM_TEST,
...     val_split_ratio=0.1
... )

Create synthetic validation set:

>>> datamodule = MVTec(
...     val_split_mode=ValSplitMode.SYNTHETIC,
...     val_split_ratio=0.2
... )
prepare_data()#

Download the dataset if not available.

This method checks if the specified dataset is available in the file system. If not, it downloads and extracts the dataset into the appropriate directory.

Return type:

None

Example

Assume the dataset is not available on the file system:

>>> datamodule = MVTec(
...     root="./datasets/MVTec",
...     category="bottle"
... )
>>> datamodule.prepare_data()

Directory structure after download:

datasets/
└── MVTec/
    ├── bottle/
    ├── cable/
    └── ...
class anomalib.data.Visa(root='./datasets/visa', category='capsules', train_batch_size=32, eval_batch_size=32, num_workers=8, train_augmentations=None, val_augmentations=None, test_augmentations=None, augmentations=None, test_split_mode=TestSplitMode.FROM_DIR, test_split_ratio=0.2, val_split_mode=ValSplitMode.SAME_AS_TEST, val_split_ratio=0.5, seed=None)#

Bases: AnomalibDataModule

VisA Datamodule.

Parameters:

  • root (Path | str) – Path to the root of the dataset. Defaults to "./datasets/visa".

  • category (str) – Category of the VisA dataset (e.g. "candle"). Defaults to "capsules".

  • train_batch_size (int, optional) – Training batch size. Defaults to 32.

  • eval_batch_size (int, optional) – Test batch size. Defaults to 32.

  • num_workers (int, optional) – Number of workers for data loading. Defaults to 8.

  • train_augmentations (Transform | None) – Augmentations to apply dto the training images Defaults to None.

  • val_augmentations (Transform | None) – Augmentations to apply to the validation images. Defaults to None.

  • test_augmentations (Transform | None) – Augmentations to apply to the test images. Defaults to None.

  • augmentations (Transform | None) – General augmentations to apply if stage-specific augmentations are not provided.

  • test_split_mode (TestSplitMode | str) – Method to create test set. Defaults to TestSplitMode.FROM_DIR.

  • test_split_ratio (float) – Fraction of data to use for testing. Defaults to 0.2.

  • val_split_mode (ValSplitMode | str) – Method to create validation set. Defaults to ValSplitMode.SAME_AS_TEST.

  • val_split_ratio (float) – Fraction of data to use for validation. Defaults to 0.5.

  • seed (int | None, optional) – Random seed for reproducibility. Defaults to None.

apply_cls1_split()#

Apply the 1-class subset splitting using the fixed split in the csv file.

Adapted from amazon-science/spot-diff.

Return type:

None

prepare_data()#

Download and prepare the dataset if not available.

This method checks if the dataset exists and is properly formatted. If not, it downloads and prepares the data in the following steps: :rtype: None

  1. If the processed dataset exists (visa_pytorch/{category}), do nothing

  2. If the raw dataset exists but isn’t processed, apply the train/test split

  3. If the dataset doesn’t exist, download, extract, and process it

The final directory structure will be:

datasets/
└── visa/
    ├── visa_pytorch/
    │   ├── candle/
    │   │   ├── train/
    │   │   │   └── good/
    │   │   ├── test/
    │   │   │   ├── good/
    │   │   │   └── bad/
    │   │   └── ground_truth/
    │   │       └── bad/
    │   └── ...
    └── VisA_20220922.tar