How is point-wise F1 score being calculated?

Dear author, great work on creating such a comprehensive benchmark for time series anomaly detection! I was wondering if the mechanism of calculating the point-wise F1 score has been full provided as I am trying to replicate the results presented in the online leaderboard. From the Github Website, I see that the PointF1 class is defined as following:

class PointF1(EvalInterface):
    """
    Using Traditional F1 score to evaluate the models.
    """
    def __init__(self) -> None:
        super().__init__()
        self.name = "point-wise f1"

    def calc(self, scores, labels, margins) -> type[MetricInterface]:
        '''
        Returns:
         A F1class (Evaluations.Metrics.F1class), including:\n
            best_f1: the value of best f1 score;\n
            precision: corresponding precision value;\n
            recall: corresponding recall value;
        '''
        print("scores: ", scores)
        print("labels: ", labels)
        prec = precision_score(labels, scores)
        rec = recall_score(labels, scores)
        f1 = f1_score(labels, scores)

        return F1class(
            name=self.name,
            p=prec,
            r=rec,
            f1=f1
        )

However, a threshold has not been defined for finding the anomaly points based on the scores, which makes it impossible to calculate the precision and recall of the results outputted by models. Is it because I haven't found the right function for calculating point-wise F1 score or is it still under development? Thanks!

dawnvince / EasyTSAD

How is point-wise F1 score being calculated? #6