Audio Multilabel Classifier

speechline.modules.audio_multilabel_classifier.AudioMultiLabelClassifier (AudioModule)

Generic AudioClassifier Module. Performs padded audio classification.

Parameters:

Name	Type	Description	Default
model_checkpoint	str	HuggingFace Hub model checkpoint.	required

Source code in speechline/modules/audio_multilabel_classifier.py

class AudioMultiLabelClassifier(AudioModule):
    """
    Generic AudioClassifier Module. Performs padded audio classification.

    Args:
        model_checkpoint (str):
            HuggingFace Hub model checkpoint.
    """

    def __init__(self, model_checkpoint: str, **kwargs) -> None:
        classifier = pipeline(
            "audio-classification",
            model=model_checkpoint,
            feature_extractor=model_checkpoint,
            device=0 if torch.cuda.is_available() else -1,
            pipeline_class=AudioMultiLabelClassificationPipeline,
            **kwargs,
        )
        super().__init__(pipeline=classifier)

    def inference(
        self, dataset: Dataset, threshold: float = 0.5
    ) -> List[Dict[str, Union[str, float]]]:
        """
        Inference function for audio classification.

        Args:
            dataset (Dataset):
                Dataset to be inferred.
            threshold (float):
                Threshold probability for predicted labels.
                Anything above this threshold will be considered as a valid prediction.

        Returns:
            List[Dict[str, Union[str, float]]]:
                List of predictions in the format of dictionaries,
                consisting of the predicted label and probability.
        """

        def _get_audio_array(
            dataset: Dataset,
        ) -> np.ndarray:
            for item in dataset:
                yield item["audio"]["array"]

        results = []

        for out in tqdm(
            self.pipeline(_get_audio_array(dataset), top_k=1),
            total=len(dataset),
            desc="Classifying Audios",
        ):
            ids = np.where(out >= threshold)[0].tolist()
            if len(ids) > 0:
                prediction = [
                    {
                        "label": self.pipeline.model.config.id2label[id],
                        "score": out[id],
                    }
                    for id in ids
                ]
                results.append(prediction)

        return results

inference(self, dataset, threshold=0.5)

Inference function for audio classification.

Parameters:

Name	Type	Description	Default
dataset	Dataset	Dataset to be inferred.	required
threshold	float	Threshold probability for predicted labels. Anything above this threshold will be considered as a valid prediction.	0.5

Returns:

Type	Description
List[Dict[str, Union[str, float]]]	List of predictions in the format of dictionaries, consisting of the predicted label and probability.

Source code in speechline/modules/audio_multilabel_classifier.py

def inference(
    self, dataset: Dataset, threshold: float = 0.5
) -> List[Dict[str, Union[str, float]]]:
    """
    Inference function for audio classification.

    Args:
        dataset (Dataset):
            Dataset to be inferred.
        threshold (float):
            Threshold probability for predicted labels.
            Anything above this threshold will be considered as a valid prediction.

    Returns:
        List[Dict[str, Union[str, float]]]:
            List of predictions in the format of dictionaries,
            consisting of the predicted label and probability.
    """

    def _get_audio_array(
        dataset: Dataset,
    ) -> np.ndarray:
        for item in dataset:
            yield item["audio"]["array"]

    results = []

    for out in tqdm(
        self.pipeline(_get_audio_array(dataset), top_k=1),
        total=len(dataset),
        desc="Classifying Audios",
    ):
        ids = np.where(out >= threshold)[0].tolist()
        if len(ids) > 0:
            prediction = [
                {
                    "label": self.pipeline.model.config.id2label[id],
                    "score": out[id],
                }
                for id in ids
            ]
            results.append(prediction)

    return results