CategoricalFBetaScore metric

@synalinks_export(
    [
        "synalinks.metrics.CategoricalFBetaScore",
        "synalinks.metrics.ListFBetaScore",
    ]
)
class CategoricalFBetaScore(FBetaScore):
    """Computes F-Beta score on categorical (list / label) structures.

    Formula:

    ```python
    b2 = beta ** 2
    f_beta_score = (1 + b2) * (precision * recall) / (precision * b2 + recall)
    ```

    This is the weighted harmonic mean of precision and recall.
    Its output range is `[0, 1]`. It operates at a label level
    and can be used for **classification** or **retrieval pipelines**.

    The difference between this metric and `F1Score` is that this one considers
    each element of the list (or the string value) as **one label**, comparing
    label sets rather than tokenized words.

    If `labels` is provided, accumulation is performed per-label (sklearn-style):
    for each label `L`, `tp[L] += 1` when `L` appears in both `y_true` and
    `y_pred`, `fp[L] += 1` when it appears only in `y_pred`, `fn[L] += 1` when
    it appears only in `y_true`. This enables stable `macro`/`weighted`
    averaging across batches even when some labels are absent from a given
    sample, and lets `result()` return a `{label: score}` dict when
    `average=None`.

    If `labels` is `None`, a single global set-based TP/FP/FN is computed
    over the pooled label values; in that mode `average=None` returns one
    scalar (use `labels=...` for a per-label breakdown).

    Example:

    ```python

    # for single label classification

    class ListClassification(synalinks.DataModel):
        label: Literal["label", "label_1", "label_2"]

    # for multi label classification

    class ListClassification(synalinks.DataModel):
        labels: List[Literal["label", "label_1", "label_2"]]

    # or use it with retrieval pipelines, in that case make sure to mask
    # the correct fields.

    class AnswerWithReferences(synalinks.DataModel):
        sources: List[str]
        answer: str

    ```


    Compilation example:

    ```python
    program.compile(
        metrics=[
            synalinks.metrics.CategoricalFBetaScore(),
        ],
    )
    ```

    Args:
        average (str): Type of averaging to be performed across per-field results
            in the multi-field case.
            Acceptable values are `None`, `"micro"`, `"macro"` and
            `"weighted"`. Defaults to `None`.
            If `None`, no averaging is performed and `result()` will return
            the score for each field.
            If `"micro"`, compute metrics globally by counting the total
            true positives, false negatives and false positives.
            If `"macro"`, compute metrics for each label,
            and return their unweighted mean.
            This does not take label imbalance into account.
            If `"weighted"`, compute metrics for each label,
            and return their average weighted by support
            (the number of true instances for each label).
            This alters `"macro"` to account for label imbalance.
            It can result in an score that is not between precision and recall.
        beta (float): Determines the weight of given to recall
            in the harmonic mean between precision and recall. Defaults to `1`.
        labels (list): (Optional) Explicit list of label names to track.
            When provided, accumulation is per-label across all batches and
            `result()` returns a `{label: score}` dict for `average=None`.
        name (str): (Optional) string name of the metric instance.
        in_mask (list): (Optional) list of keys to keep to compute the metric.
        out_mask (list): (Optional) list of keys to remove to compute the metric.
        in_mask_pattern (str): (Optional) Regex pattern; fields whose names match
            are kept (combined with ``in_mask`` via OR).
        out_mask_pattern (str): (Optional) Regex pattern; fields whose names match
            are dropped (combined with ``out_mask`` via OR).
    """

    def __init__(
        self,
        average=None,
        beta=1.0,
        labels=None,
        name="categorical_fbeta_score",
        in_mask=None,
        out_mask=None,
        in_mask_pattern=None,
        out_mask_pattern=None,
    ):
        super().__init__(
            average=average,
            beta=beta,
            name=name,
            in_mask=in_mask,
            out_mask=out_mask,
            in_mask_pattern=in_mask_pattern,
            out_mask_pattern=out_mask_pattern,
        )
        if labels is not None:
            labels = [str(label) for label in labels]
        self.labels = labels

    async def update_state(self, y_true, y_pred):
        y_pred = tree.map_structure(lambda x: ops.convert_to_json_data_model(x), y_pred)
        y_true = tree.map_structure(lambda x: ops.convert_to_json_data_model(x), y_true)

        if self.in_mask or self.in_mask_pattern:
            y_pred = tree.map_structure(
                lambda x: (
                    x.in_mask(mask=self.in_mask, pattern=self.in_mask_pattern)
                    if x is not None
                    else x
                ),
                y_pred,
            )
            y_true = tree.map_structure(
                lambda x: (
                    x.in_mask(mask=self.in_mask, pattern=self.in_mask_pattern)
                    if x is not None
                    else x
                ),
                y_true,
            )
        if self.out_mask or self.out_mask_pattern:
            y_pred = tree.map_structure(
                lambda x: (
                    x.out_mask(mask=self.out_mask, pattern=self.out_mask_pattern)
                    if x is not None
                    else x
                ),
                y_pred,
            )
            y_true = tree.map_structure(
                lambda x: (
                    x.out_mask(mask=self.out_mask, pattern=self.out_mask_pattern)
                    if x is not None
                    else x
                ),
                y_true,
            )

        if y_true is None or y_pred is None:
            # A failed prediction yields `y_pred is None`; there is nothing to
            # compare, so skip the sample instead of calling `.get_json()` on None.
            return
        y_true = tree.flatten(tree.map_structure(lambda x: x, y_true.get_json()))
        y_pred = tree.flatten(tree.map_structure(lambda x: x, y_pred.get_json()))

        true_positives = []
        false_positives = []
        false_negatives = []
        intermediate_weights = []

        if self.labels is not None:
            y_true_set = {str(v) for v in y_true}
            y_pred_set = {str(v) for v in y_pred}
            for label in self.labels:
                t = label in y_true_set
                p = label in y_pred_set
                true_positives.append(1 if (t and p) else 0)
                false_positives.append(1 if (p and not t) else 0)
                false_negatives.append(1 if (t and not p) else 0)
                intermediate_weights.append(1 if t else 0)
        else:
            # Set-based TP/FP/FN over the full pool of labels:
            # position-independent, so that `["a","b"]` vs `["b","a"]` scores
            # 1.0. Produces a single entry per call; per-label tracking
            # requires `labels=...`.
            y_true_labels = [str(v) for v in y_true]
            y_pred_labels = [str(v) for v in y_pred]
            common_labels = set(y_true_labels) & set(y_pred_labels)
            true_positives.append(len(common_labels))
            false_positives.append(len(set(y_pred_labels)) - len(common_labels))
            false_negatives.append(len(set(y_true_labels)) - len(common_labels))
            intermediate_weights.append(len(y_true_labels))

        true_positives = np.convert_to_numpy(true_positives)
        false_positives = np.convert_to_numpy(false_positives)
        false_negatives = np.convert_to_numpy(false_negatives)
        intermediate_weights = np.convert_to_numpy(intermediate_weights)

        current_true_positives = self.state.get("true_positives")
        if current_true_positives:
            true_positives = ragged_add(current_true_positives, true_positives)

        current_false_positives = self.state.get("false_positives")
        if current_false_positives:
            false_positives = ragged_add(current_false_positives, false_positives)

        current_false_negatives = self.state.get("false_negatives")
        if current_false_negatives:
            false_negatives = ragged_add(current_false_negatives, false_negatives)

        current_intermediate_weights = self.state.get("intermediate_weights")
        if current_intermediate_weights:
            intermediate_weights = ragged_add(
                current_intermediate_weights, intermediate_weights
            )

        self.state.update(
            {
                "true_positives": true_positives.tolist(),
                "false_positives": false_positives.tolist(),
                "false_negatives": false_negatives.tolist(),
                "intermediate_weights": intermediate_weights.tolist(),
            }
        )

    def result(self):
        res = super().result()
        if self.labels is not None and self.average is None and isinstance(res, list):
            return {label: score for label, score in zip(self.labels, res)}
        return res

    def get_config(self):
        """Return the serializable config of the metric.

        Returns:
            (dict): The config dict.
        """
        config = {
            "beta": self.beta,
            "labels": list(self.labels) if self.labels is not None else None,
            "name": self.name,
        }
        base_config = super().get_config()
        return {**base_config, **config}