Multilabelf1score

multilabelf1score

Metric MultilabelF1Score from torchmetrics (torchmetrics.classification.MultilabelF1Score)

When to invoke this skill

The user has predictions + ground truth and asks to evaluate with MultilabelF1Score, or mentions torchmetrics.classification.MultilabelF1Score directly, or wants the standard torchmetrics implementation.

Reference signature

from torchmetrics.classification import MultilabelF1Score

# MultilabelF1Score(num_labels: int, threshold: float = 0.5, average: Optional[Literal['micro', 'macro', 'weighted', 'none']] = 'macro', multidim_average: Literal['global', 'samplewise'] = 'global', ignore_index: Optional[int] = None, validate_args: bool = True, zero_division: float = 0, **kwargs: Any) -> None

Library docstring

Compute F-1 score for multilabel tasks.

.. math::
    F_{1} = 2\frac{\text{precision} * \text{recall}}{(\text{precision}) + \text{recall}}

The metric is only proper defined when :math:`\text{TP} + \text{FP} \neq 0 \wedge \text{TP} + \text{FN} \neq 0`
where :math:`\text{TP}`, :math:`\text{FP}` and :math:`\text{FN}` represent the number of true positives, false
positives and false negatives respectively. If this case is encountered for any label, the metric for that label
will be set to `zero_division` (0 or 1, default is 0) and the overall metric may therefore be affected in turn.

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

- ``preds`` (:class:`~torch.Tensor`): An int or float tensor of shape ``(N, C, ...)``.
  If preds is a floating point tensor with values outside [0,1] range we consider the input to be logits and
  will auto apply sigmoid per element. Additionally, we convert to int tensor with thresholding using the value
  in ``threshold``.
- ``target`` (:class:`~torch.Tensor`): An int tensor of shape ``(N, C, ...)``.

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

- ``mlf1s`` (:class:`~torch.Tensor`): A tensor whose returned shape depends on the ``average`` and
  ``multidim_average`` arguments:

    - If ``multidim_average`` is set to ``global``:

      - If ``average='micro'/'macro'/'weighted'``, the output will be a scalar tensor
      - If ``average=None/'none'``, the shape will be ``(C,)``

    - If ``multidim_average`` is set to ``samplewise``:

      - If ``average='micro'/'macro'/'weighted'``, the shape will be ``(N,)``
      - If ``average=None/'none'``, the shape will be ``(N, C)```

If ``multidim_average`` is set to ``samplewise`` we expect at least one additional dimension ``...`` to be present,
which the reduction will then be applied over instead of the sample dimension ``N``.

Args:
    num_labels: Integer specifying the number of labels
    threshold: Threshold for transforming probability to binary (0,1) predictions
    average:
        Defines the reduction that is applied over labels. Should be one of the following:

        - ``micro``: Sum statistics over all labe

Quick recipe

import torchmetrics.classification as _m
score = _m.MultilabelF1Score(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).