Skip to content

FBetaScore metric

Bases: Metric

Computes F-Beta score.

Formula:

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 word level and can be used for QA systems.

If y_true and y_pred contains multiple fields The JSON object's fields are flattened and the score computed for each one independently.

Example:

program.compile(
    metrics=[
        synalinks.metrics.FBetaScore(),
    ],
)

Parameters:

Name Type Description Default
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 class. 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.

None
beta float

Determines the weight of given to recall in the harmonic mean between precision and recall (see pseudocode equation above). Defaults to 1.

1.0
name str

(Optional) string name of the metric instance.

'fbeta_score'
in_mask list

(Optional) list of keys to keep to compute the metric.

None
out_mask list

(Optional) list of keys to remove to compute the metric.

None
in_mask_pattern str

(Optional) Regex pattern; fields whose names match are kept (combined with in_mask via OR).

None
out_mask_pattern str

(Optional) Regex pattern; fields whose names match are dropped (combined with out_mask via OR).

None
Source code in synalinks/src/metrics/f_score_metrics.py
@synalinks_export("synalinks.metrics.FBetaScore")
class FBetaScore(Metric):
    """Computes F-Beta score.

    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 word level
    and can be used for **QA systems**.

    If `y_true` and `y_pred` contains multiple fields
    The JSON object's fields are flattened and the score
    computed for each one independently.


    Example:

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

    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 class.
            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 (see pseudocode
            equation above). Defaults to `1`.
        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).
    """

    direction = "up"

    def __init__(
        self,
        average=None,
        beta=1.0,
        name="fbeta_score",
        in_mask=None,
        out_mask=None,
        in_mask_pattern=None,
        out_mask_pattern=None,
    ):
        super().__init__(
            name=name,
            in_mask=in_mask,
            out_mask=out_mask,
            in_mask_pattern=in_mask_pattern,
            out_mask_pattern=out_mask_pattern,
        )
        if average not in (None, "micro", "macro", "weighted"):
            raise ValueError(
                "Invalid `average` argument value. Expected one of: "
                "[None, 'micro', 'macro', 'weighted']. "
                f"Received: average={average}"
            )

        if not isinstance(beta, float):
            raise ValueError(
                "Invalid `beta` argument value. "
                "It should be a Python float. "
                f"Received: beta={beta} of type '{type(beta)}'"
            )
        self.state = self.add_variable(
            data_model=FBetaState,
            name="state_" + self.name,
        )
        self.average = average
        self.beta = beta
        self.axis = None
        if self.average != "micro":
            self.axis = 0
        # Subclasses (Precision, Recall) override this to switch the result
        # formula while reusing TP/FP/FN state and update_state.
        self._formula = "fbeta"

    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: str(x), y_true.get_json()))
        y_pred = tree.flatten(tree.map_structure(lambda x: str(x), y_pred.get_json()))

        true_positives = []
        false_positives = []
        false_negatives = []
        intermediate_weights = []
        # For each field of y_true and y_pred. SQuAD-style multiset (Counter)
        # intersection, needed so identical strings with repeated tokens
        # score 1.0.
        # zip_longest, not zip: unmatched leaves (pred and gold structures can
        # disagree on variable-length arrays) must be scored; the "" fill has
        # no tokens, so they land entirely in false positives/negatives.
        for yt, yp in zip_longest(y_true, y_pred, fillvalue=""):
            y_true_tokens = nlp_utils.normalize_and_tokenize(str(yt))
            y_pred_tokens = nlp_utils.normalize_and_tokenize(str(yp))
            num_common = sum((Counter(y_true_tokens) & Counter(y_pred_tokens)).values())
            true_positives.append(num_common)
            false_positives.append(len(y_pred_tokens) - num_common)
            false_negatives.append(len(y_true_tokens) - num_common)
            intermediate_weights.append(len(y_true_tokens))

        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):
        if (
            self.state.get("true_positives") is None
            and self.state.get("false_positives") is None
            and self.state.get("false_negatives") is None
        ):
            return 0.0
        tp = np.convert_to_tensor(self.state.get("true_positives"))
        fp = np.convert_to_tensor(self.state.get("false_positives"))
        fn = np.convert_to_tensor(self.state.get("false_negatives"))

        # Keras/sklearn "micro": aggregate TP/FP/FN across all fields first,
        # *then* compute precision/recall. Without this collapse, "micro"
        # would degenerate to a mean over per-field scores (i.e. macro).
        if self.average == "micro":
            tp = np.sum(tp)
            fp = np.sum(fp)
            fn = np.sum(fn)

        precision = np.convert_to_tensor(
            np.divide(tp, np.add(tp, fp) + backend.epsilon())
        )
        recall = np.convert_to_tensor(np.divide(tp, np.add(tp, fn) + backend.epsilon()))

        formula = getattr(self, "_formula", "fbeta")
        if formula == "precision":
            score = precision
        elif formula == "recall":
            score = recall
        else:
            mul_value = precision * recall
            add_value = ((self.beta**2) * precision) + recall
            mean = np.divide(mul_value, add_value + backend.epsilon())
            score = mean * (1 + (self.beta**2))

        return self._aggregate(score)

    def _aggregate(self, score):
        """Apply `average` reduction over per-field scores."""
        score = np.convert_to_tensor(score)
        if self.average == "weighted":
            intermediate_weights = self.state.get("intermediate_weights")
            weights = np.divide(
                intermediate_weights,
                np.sum(intermediate_weights) + backend.epsilon(),
            )
            score = np.sum(score * weights)
        elif self.average is not None:  # [micro, macro]
            score = np.mean(score, self.axis)
        # numpy 1.25+ deprecates float() on a >0-D array even when size == 1,
        # so go through .item() / .tolist() to always hand back Python scalars.
        score_arr = np.convert_to_numpy(score)
        if score_arr.size == 1:
            return float(score_arr.item())
        return [float(v) for v in score_arr.tolist()]

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

        Returns:
            (dict): The config dict.
        """
        config = {
            "name": self.name,
            "average": self.average,
            "beta": self.beta,
        }
        base_config = super().get_config()
        return {**base_config, **config}

get_config()

Return the serializable config of the metric.

Returns:

Type Description
dict

The config dict.

Source code in synalinks/src/metrics/f_score_metrics.py
def get_config(self):
    """Return the serializable config of the metric.

    Returns:
        (dict): The config dict.
    """
    config = {
        "name": self.name,
        "average": self.average,
        "beta": self.beta,
    }
    base_config = super().get_config()
    return {**base_config, **config}