AutoVI Datamodule

AutoVI Datamodule#

AutoVI Data Module.

This module provides a PyTorch Lightning DataModule for the AutoVI (Automotive Visual Inspection) dataset.

The dataset is hosted on Zenodo as six independent zip files — one per class. When AutoVI.prepare_data() is called, only the zip file that corresponds to the requested category is downloaded and extracted, so you don’t have to fetch the full 4 GB archive if you only need one class.

Usage:

from anomalib.data.datamodules.image import AutoVI

datamodule = AutoVI(
    root="./datasets/AutoVI",
    category="pipe_clip",
    train_batch_size=32,
    eval_batch_size=32,
    num_workers=8,
)
datamodule.setup()

for batch in datamodule.train_dataloader():
    print(batch["image"].shape)   # (32, 3, H, W)

Data License:: Copyright © 2023-2024 Renault Group. Released under Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0). https://creativecommons.org/licenses/by-nc-sa/4.0/
Reference:: Carvalho, P., Lafou, M., Durupt, A., Leblanc, A., & Grandvalet, Y. (2024). The Automotive Visual Inspection Dataset (AutoVI): A Genuine Industrial Production Dataset for Unsupervised Anomaly Detection [Dataset]. https://doi.org/10.5281/zenodo.10459003

class anomalib.data.datamodules.image.autovi.AutoVI(root=None, category='engine_wiring', 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

AutoVI Lightning DataModule.

Downloads (on demand) and wraps the AutoVI dataset for use with the anomalib training engine. Because the dataset ships as six separate zip files, only the zip that belongs to the requested category is fetched.

The resulting on-disk layout mirrors MVTec AD, except that ground-truth masks are stored one level deeper — each mask lives in a subdirectory named after the corresponding test image stem:

datasets/
└── AutoVI/
    ├── engine_wiring/
    │   ├── train/good/*.png
    │   ├── test/{good,<defect_type>}/*.png
    │   └── ground_truth/<defect_type>/<image_stem>/0000.png
    ├── pipe_clip/
    ├── pipe_staple/
    ├── tank_screw/
    ├── underbody_pipes/
    └── underbody_screw/

For example, the mask for test/unclipped/0089.png is stored at ground_truth/unclipped/0089/0000.png.

Parameters:

root (Path | str | None) – Root directory that will contain (or already contains) the per-category sub-directories. Defaults to None → <anomalib_datasets_dir>/AutoVI.
category (str) – One of the six AutoVI classes. Defaults to "engine_wiring".
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 DataLoader workers. Defaults to 8.
train_augmentations (Transform | None) – Augmentations for the training split. Defaults to None.
val_augmentations (Transform | None) – Augmentations for the validation split. Defaults to None.
test_augmentations (Transform | None) – Augmentations for the test split. Defaults to None.
augmentations (Transform | None) – Fallback augmentations used when no split-specific augmentation is supplied. Defaults to None.
test_split_mode (TestSplitMode | str) – How to obtain the 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) – How to obtain the 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) – Random seed for reproducibility. Defaults to None.

Example

>>> datamodule = AutoVI(
...     root="./datasets/AutoVI",
...     category="pipe_clip",
... )
>>> datamodule.setup()
>>> i, batch = next(enumerate(datamodule.train_dataloader()))
>>> batch["image"].shape
torch.Size([32, 3, 256, 256])

prepare_data()#

Download and extract the zip for the requested category, if needed.

Only the single zip that belongs to self.category is fetched from Zenodo — the remaining five classes are left untouched.

The check is: if <root>/<category>/ already exists as a directory we assume the data are present and skip the download.

Return type:: None

Example

>>> datamodule = AutoVI(
...     root="./datasets/AutoVI",
...     category="tank_screw",
... )
>>> datamodule.prepare_data()   # downloads tank_screw.zip only

Directory structure after a successful download:

datasets/
    └── AutoVI/
        ├── engine_wiring/
        │   ├── train/good/*.png
        │   ├── test/{good,<defect_type>}/*.png
        │   └── ground_truth/<defect_type>/<image_stem>/0000.png

AutoVI Datamodule

Contents

AutoVI Datamodule#