Minmaxmetric

minmaxmetric

Metric MinMaxMetric from torchmetrics (torchmetrics.MinMaxMetric)

When to invoke this skill

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

Reference signature

from torchmetrics import MinMaxMetric

# MinMaxMetric(base_metric: torchmetrics.metric.Metric, **kwargs: Any) -> None

Library docstring

Wrapper metric that tracks both the minimum and maximum of a scalar/tensor across an experiment.

The min/max value will be updated each time ``.compute`` is called.

Args:
    base_metric:
        The metric of which you want to keep track of its maximum and minimum values.
    kwargs: Additional keyword arguments, see :ref:`Metric kwargs` for more info.

Raises:
    ValueError
        If ``base_metric` argument is not a subclasses instance of ``torchmetrics.Metric``

Example::
    >>> import torch
    >>> from torchmetrics.wrappers import MinMaxMetric
    >>> from torchmetrics.classification import BinaryAccuracy
    >>> from pprint import pprint
    >>> base_metric = BinaryAccuracy()
    >>> minmax_metric = MinMaxMetric(base_metric)
    >>> preds_1 = torch.Tensor([[0.1, 0.9], [0.2, 0.8]])
    >>> preds_2 = torch.Tensor([[0.9, 0.1], [0.2, 0.8]])
    >>> labels = torch.Tensor([[0, 1], [0, 1]]).long()
    >>> pprint(minmax_metric(preds_1, labels))
    {'max': tensor(1.), 'min': tensor(1.), 'raw': tensor(1.)}
    >>> pprint(minmax_metric.compute())
    {'max': tensor(1.), 'min': tensor(1.), 'raw': tensor(1.)}
    >>> minmax_metric.update(preds_2, labels)
    >>> pprint(minmax_metric.compute())
    {'max': tensor(1.), 'min': tensor(0.7500), 'raw': tensor(0.7500)}

Quick recipe

import torchmetrics as _m
score = _m.MinMaxMetric(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).