Source code for deepparse.embeddings_models.magnitude_embeddings_model

from pymagnitudelight import Magnitude

from .embeddings_model import EmbeddingsModel, ndarray




class MagnitudeEmbeddingsModel(EmbeddingsModel):
    """
    FastText embeddings network from `Enriching Word Vectors with Subword Information
    <https://arxiv.org/abs/1607.04606>`_ using the magnitude mapping
    (`Magnitude <https://github.com/plasticityai/magnitude>`_), which reduces the memory footprint.

    Args:
        embeddings_path (str): Path to the Magnitude embeddings vector (.magnitude).
        verbose (bool): Either or not to make the loading of the embeddings verbose.
    """

    def __init__(self, embeddings_path: str, verbose: bool = True) -> None:
        super().__init__(verbose=verbose)
        self.model = Magnitude(path=embeddings_path, lazy_loading=-1, blocking=True)



    def __call__(self, words: str) -> ndarray:
        """
        Callable method to get the word vector of a complete address.

        Args:
            words (str): Address to get vector for words.

        Return:
            The FastText embedding for a list of words.
        """
        # We leverage the multiple-word query which is faster than a single word query
        return self.model.query(words.split())