def detect_features_modalities()

in services/worker/src/worker/job_runners/dataset/modalities.py [0:0]

• 37 lines of code
• 23 McCabe index (conditional complexity)

def detect_features_modalities(features: Features) -> set[DatasetModality]:
    """
    Detect modalities of a dataset using the features (column types).

    Args:
        features (`datasets.Features`):
            The features of a config.

    Returns:
        `set[DatasetModality]`: A set of modalities.
    """
    modalities: set[DatasetModality] = set()

    def classify_modality(feature: FeatureType) -> None:
        nonlocal modalities
        if isinstance(feature, Audio):
            modalities.add("audio")
        elif isinstance(feature, Image):
            modalities.add("image")
        elif isinstance(feature, Value) and feature.dtype in ("string", "large_string"):
            modalities.add("text")
        elif isinstance(feature, (Translation, TranslationVariableLanguages)):
            modalities.add("text")
        elif isinstance(feature, Pdf):
            modalities.add("document")

    _visit(features, classify_modality)

    # detection of tabular data: if there are at least two top-level numerical columns, and no "media" columns
    if (
        not ("audio" in modalities or "image" in modalities)
        and len(
            [
                feature
                for feature in features.values()
                if isinstance(feature, Value) and ("int" in feature.dtype or "float" in feature.dtype)
            ]
        )
        >= 2
    ):
        modalities.add("tabular")

    # detection of time series
    if any(
        "emb" not in column_name  # ignore lists of floats that may be embeddings
        and (
            (isinstance(feature, (LargeList, Sequence)) and feature.feature == Value("float32"))
            or (isinstance(feature, list) and feature[0] == Value("float32"))
        )
        for column_name, feature in features.items()
    ):
        modalities.add("timeseries")
    # other idea: detect datasets with only numerical columns and one timestamp column
    # (and ideally be able to detect dates/timestamps even from a column with string type)

    return modalities