MAITE Datasets

MAITE Datasets are a collection of public datasets wrapped in a MAITE compliant format.

Installation

To install and use maite-datasets you can use pip:

pip install maite-datasets

Extras

Optional features are installed as extras, e.g. pip install maite-datasets[tqdm].

Available Downloadable Datasets

The paired RGB-infrared datasets (DroneVehicle, M3FD) stack the infrared channel onto the RGB image. MilitaryAircraft and MilitaryVehicles expose a hierarchy attribute describing their class ontology. Every dataset accepts lazy=True to defer image decoding until the image is actually read, which keeps metadata-only passes over large datasets cheap.

Usage

Here is an example of how to import MNIST for usage with your workflow.

>>> from maite_datasets.image_classification import MNIST

>>> mnist = MNIST(root="data", download=True)
>>> print(mnist)
MNIST Dataset
-------------
    Corruption: None
    Transforms: []
    Image_set: train
    Metadata: {'id': 'MNIST_train', 'index2label': {0: 'zero', 1: 'one', 2: 'two', 3: 'three', 4: 'four', 5: 'five', 6: 'six', 7: 'seven', 8: 'eight', 9: 'nine'}, 'split': 'train'}
    Path: /home/user/maite-datasets/data/mnist
    Size: 60000

>>> print("tuple("+", ".join([str(type(t)) for t in mnist[0]])+")")
tuple(<class 'numpy.ndarray'>, <class 'numpy.ndarray'>, <class 'dict'>)

datamaite Interoperability

With the datamaite extra installed, any dataset can be written to disk in a datamaite wire format (COCO or YOLO) and read back as a datamaite dataset.

>>> from maite_datasets.object_detection import MILCO

# Export an existing dataset, then load it back through datamaite
>>> milco = MILCO(root="data", download=True)
>>> dm_milco = milco.to_datamaite(dest="exports/milco")

# Or download, convert and load in one step
>>> dm_milco = MILCO(root="data", download=True, as_datamaite=True)

The format follows from the task and is not selectable, because each task only has one that works. Object detection writes COCO — the only one of the two that records a datum's own metadata, a dataset's telemetry as per-image columns and its per-object values as detection attributes. Image classification writes YOLO, because datamaite's COCO writer is task-closed and refuses an IC dataset outright.

Reading is more permissive than writing: an existing directory in either format is sniffed and loaded, so a YOLO dataset already on disk still works with as_datamaite=True.

as_datamaite=True writes to <root>/<dataset>_datamaite, kept separate from the <root>/<dataset> folder raw downloads use so the two can coexist under one root. An existing export there is reused; otherwise a raw dataset already under root is converted in place, and only a genuinely missing dataset is downloaded (to a temporary directory that is discarded after conversion).

Two conversion details are worth knowing. Splits are folded onto train/val/test, because those are the only ones the wire formats read back: an image_set such as operational is exported as train, with a warning, and its original name survives in each sample's metadata, which object detection keeps and image classification cannot. And datasets carrying transforms export re-encoded pixels rather than referencing the original files, so the written images always match the written annotations.

Dataset Wrappers

Wrappers provide a way to convert datasets to allow usage of tools within specific backend frameworks.

Torchvision

TorchvisionWrapper is a convenience class that wraps any of the datasets and provides the capability to apply torchvision transforms to the dataset.

NOTE: TorchvisionWrapper requires torch and torchvision to be installed.

>>> from maite_datasets.object_detection import MILCO

>>> milco = MILCO(root="data", download=True)
>>> print(milco)
MILCO Dataset
-------------
    Transforms: []
    Image Set: train
    Metadata: {'id': 'MILCO_train', 'index2label': {0: 'MILCO', 1: 'NOMBO'}, 'split': 'train'}
    Path: /home/user/maite-datasets/data/milco
    Size: 261

>>> print(f"type={milco[0][0].__class__.__name__}, shape={milco[0][0].shape}")
type=ndarray, shape=(3, 1024, 1024)

>>> print(milco[0][1].boxes[0])
[ 75. 217. 130. 247.]

>>> from maite_datasets.wrappers import TorchvisionWrapper
>>> from torchvision.transforms.v2 import Resize

>>> milco_torch = TorchvisionWrapper(milco, transforms=Resize(224))
>>> print(milco_torch)
Torchvision Wrapped MILCO Dataset
---------------------------
    Transforms: Resize(size=[224], interpolation=InterpolationMode.BILINEAR, antialias=True)

MILCO Dataset
-------------
    Transforms: []
    Image Set: train
    Metadata: {'id': 'MILCO_train', 'index2label': {0: 'MILCO', 1: 'NOMBO'}, 'split': 'train'}
    Path: /home/user/maite-datasets/data/milco
    Size: 261

>>> print(f"type={milco_torch[0][0].__class__.__name__}, shape={milco_torch[0][0].shape}")
type=Image, shape=torch.Size([3, 224, 224])

>>> print(milco_torch[0][1].boxes[0])
tensor([16.4062, 47.4688, 28.4375, 54.0312], dtype=torch.float64)

Dataset Adapters

Adapters provide a way to read in datasets from other popular formats.

Huggingface

Hugging face datasets can be adapted into MAITE compliant format using the from_huggingface adapter.

>>> from datasets import load_dataset
>>> from maite_datasets.adapters import from_huggingface

>>> cppe5 = load_dataset("cppe-5")
>>> m_cppe5 = from_huggingface(cppe5["train"])
>>> print(m_cppe5)
HFObjectDetection Dataset
-------------------------
    Source: Dataset({
    features: ['image_id', 'image', 'width', 'height', 'objects'],
    num_rows: 1000
})
    Metadata: {'id': 'cppe-5', 'index2label': {0: 'Coverall', 1: 'Face_Shield', 2: 'Gloves', 3: 'Goggles', 4: 'Mask'}, 'description': '', 'citation': '', 'homepage': '', 'license': '', 'features': {'image_id': Value('int64'), 'image': Image(mode=None, decode=True), 'width': Value('int32'), 'height': Value('int32'), 'objects': {'id': List(Value('int64')), 'area': List(Value('int64')), 'bbox': List(List(Value('float32'), length=4)), 'category': List(ClassLabel(names=['Coverall', 'Face_Shield', 'Gloves', 'Goggles', 'Mask']))}}, 'post_processed': None, 'supervised_keys': None, 'builder_name': 'parquet', 'dataset_name': 'cppe-5', 'config_name': 'default', 'version': 0.0.0, 'splits': {'train': SplitInfo(name='train', num_bytes=240478590, num_examples=1000, shard_lengths=None, dataset_name='cppe-5'), 'test': SplitInfo(name='test', num_bytes=4172706, num_examples=29, shard_lengths=None, dataset_name='cppe-5')}, 'download_checksums': {'hf://datasets/cppe-5@66f6a5efd474e35bd7cb94bf15dea27d4c6ad3f8/data/train-00000-of-00001.parquet': {'num_bytes': 237015519, 'checksum': None}, 'hf://datasets/cppe-5@66f6a5efd474e35bd7cb94bf15dea27d4c6ad3f8/data/test-00000-of-00001.parquet': {'num_bytes': 4137134, 'checksum': None}}, 'download_size': 241152653, 'post_processing_size': None, 'dataset_size': 244651296, 'size_in_bytes': 485803949}

>>> image = m_cppe5[0][0]
>>> print(f"type={image.__class__.__name__}, shape={image.shape}")
type=ndarray, shape=(3, 663, 943)

>>> target = m_cppe5[0][1]
>>> print(f"box={target.boxes[0]}, label={target.labels[0]}")
box=[302.0, 109.0, 73.0, 52.0], label=4

>>> print(m_cppe5[0][2])
{'id': [114, 115, 116, 117], 'image_id': 15, 'width': 943, 'height': 663, 'area': [3796, 1596, 152768, 81002]}

Additional Information

For more information on the MAITE protocol, check out their documentation.

Acknowledgement

CDAO Funding Acknowledgement

This material is based upon work supported by the Chief Digital and Artificial Intelligence Office under Contract No. W519TC-23-9-2033. The views and conclusions contained herein are those of the author(s) and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the U.S. Government.