mml.core.data_loading.modality_loaders

class AcceleratedPillowImageLoader

Bases: PillowImageLoader

Extended Pillow ImageLoader based on the article “Fast import of Pillow images to NumPy / OpenCV arrays” by Alex Karpinsky, see https://uploadcare.com/blog/fast-import-of-pillow-images-to-numpy-opencv-arrays/

static transform(image: Image) → ndarray

class ClassLoader

Bases: ModalityLoader

Loads a simple class entry.

__init__()

load(entry: int | List[int] | List[float] | str) → Any

setup(task_dataset: TaskDataset) → None

class CombinedModalityLoader

Bases: ModalityLoader

Combines multiple modality loaders to support diverse data setups.

__init__(loaders: List[ModalityLoader])

load(entry: int | List[int] | List[float] | str) → Any

matches(entry: int | List[int] | List[float] | str) → bool

setup(task_dataset: TaskDataset) → None

class ModalityLoader

Bases: ABC

A modality loader provides the implementation to load entries of the sample dicts for a specific modality.

__init__(modality: Modality, suffixes: List[str] | None, entry_type: type | None)

This init stores supported file suffixes and the modality. They are used for the default matches implementation.

Parameters:

• modality (Modality) – the modality this loader supports

• suffixes (Optional[List[str]]) – a list of supported file suffixes, used for loader matching. Provide None to support “all” suffixes.

• entry_type (Optional[type]) – if given, entries must be of this type to match

abstract load(entry: int | List[int] | List[float] | str) → Any

The load function is the main routine that will be called with the corresponding entry of a samples modality within.

matches(entry: int | List[int] | List[float] | str) → bool

This method may be used to find the correct loader for a modality. It is given the entry and returns True if those can be handled or False if the loader does not support the provided kinds.

Parameters:

entry (ModalityEntry) – the entry in the sample description corresponding to the modality

Returns:

whether the loader accepts or rejects this

abstract setup(task_dataset: TaskDataset) → None

class MultiLabelClassLoader

Bases: ModalityLoader

Loads multi-label classification entries and one-hot encodes them.

__init__()

load(entry: int | List[int] | List[float] | str) → Any

setup(task_dataset: TaskDataset) → None

class NonMappingOpenCVMaskLoader

Bases: ModalityLoader

Special loader of masks, that does not implement a class mapping. Used during preprocessing.

__init__()

load(entry: int | List[int] | List[float] | str) → Any

setup(task_dataset: TaskDataset) → None

class NumpyArrayImageLoader

Bases: ModalityLoader

__init__()

Loads images stored as numpy arrays.

load(entry: int | List[int] | List[float] | str) → Any

setup(task_dataset: TaskDataset) → None

class OpenCVImageLoader

Bases: ModalityLoader

__init__()

Default loader for images.

load(entry: int | List[int] | List[float] | str) → Any

setup(task_dataset: TaskDataset) → None

class OpenCVMaskLoader

Bases: ModalityLoader

Default loader for segmentation masks. Adds greyscale interpretation and class mapping on top of image loading.

__init__()

load(entry: int | List[int] | List[float] | str) → Any

setup(task_dataset: TaskDataset) → None

class PillowImageLoader

Bases: ModalityLoader

__init__()

Another loader for images.

load(entry: int | List[int] | List[float] | str) → Any

setup(task_dataset: TaskDataset) → None

static transform(image: Image) → ndarray

class PureTorchvisionImageLoader

Bases: ModalityLoader

__init__()

Another loader for images based purely on torchvision (no pillow).

load(entry: int | List[int] | List[float] | str) → Any

setup(task_dataset: TaskDataset) → None

class ScikitImageLoader

Bases: ModalityLoader

__init__()

Scikit-Image loader for images.

load(entry: int | List[int] | List[float] | str) → Any

setup(task_dataset: TaskDataset) → None

class SoftLabelClassLoader

Bases: ModalityLoader

Loads soft-label classification entries.

__init__()

load(entry: int | List[int] | List[float] | str) → Any

setup(task_dataset: TaskDataset) → None

class ValueLoader

Bases: ModalityLoader

Loads a simple value entry.

__init__()

load(entry: int | List[int] | List[float] | str) → Any

setup(task_dataset: TaskDataset) → None