Multilabelaverageprecision

multilabelaverageprecision

Metric MultilabelAveragePrecision from torchmetrics (torchmetrics.classification.MultilabelAveragePrecision)

When to invoke this skill

The user has predictions + ground truth and asks to evaluate with MultilabelAveragePrecision, or mentions torchmetrics.classification.MultilabelAveragePrecision directly, or wants the standard torchmetrics implementation.

Reference signature

from torchmetrics.classification import MultilabelAveragePrecision

# MultilabelAveragePrecision(num_labels: int, average: Optional[Literal['micro', 'macro', 'weighted', 'none']] = 'macro', thresholds: Union[int, list[float], torch.Tensor, NoneType] = None, ignore_index: Optional[int] = None, validate_args: bool = True, **kwargs: Any) -> None

Library docstring

Compute the average precision (AP) score for multilabel tasks.

The AP score summarizes a precision-recall curve as an weighted mean of precisions at each threshold, with the
difference in recall from the previous threshold as weight:

.. math::
    AP = \sum_{n} (R_n - R_{n-1}) P_n

where :math:`P_n, R_n` is the respective precision and recall at threshold index :math:`n`. This value is
equivalent to the area under the precision-recall curve (AUPRC).

As input to ``forward`` and ``update`` the metric accepts the following input:

- ``preds`` (:class:`~torch.Tensor`): A float tensor of shape ``(N, C, ...)`` containing probabilities or logits
  for each observation. If preds has values outside [0,1] range we consider the input to be logits and will auto
  apply sigmoid per element.
- ``target`` (:class:`~torch.Tensor`): An int tensor of shape ``(N, C, ...)`` containing ground truth labels, and
  therefore only contain {0,1} values (except if `ignore_index` is specified).

As output to ``forward`` and ``compute`` the metric returns the following output:

- ``mlap`` (:class:`~torch.Tensor`): If `average=None|"none"` then a 1d tensor of shape (n_classes, ) will be
  returned with AP score per class. If `average="micro|macro"|"weighted"` then a single scalar is returned.

Additional dimension ``...`` will be flattened into the batch dimension.

The implementation both supports calculating the metric in a non-binned but accurate version and a binned
version that is less accurate but more memory efficient. Setting the `thresholds` argument to `None` will activate
the non-binned  version that uses memory of size :math:`\mathcal{O}(n_{samples})` whereas setting the
`thresholds` argument to either an integer, list or a 1d tensor will use a binned version that uses memory of
size :math:`\mathcal{O}(n_{thresholds} \times n_{labels})` (constant memory).

Args:
    num_labels: Integer specifying the number of labels
    average:
        Defines the reduction that is applied over labels. Should be one of the following:

        - ``micro``: Sum score over all labels
        - ``macro``: Calculate score for each label and average them
        - ``weighted``: calculates score fo

Quick recipe

import torchmetrics.classification as _m
score = _m.MultilabelAveragePrecision(y_true, y_pred)

Don'ts

• Don't reimplement when the library version handles edge cases (NaN, ties, empty inputs) better than a hand-rolled formula.
• Always check the library version's argument order — sklearn is (y_true, y_pred) while torchmetrics is (preds, target).