Meanabsolutepercentageerror

meanabsolutepercentageerror

Metric MeanAbsolutePercentageError from torchmetrics (torchmetrics.MeanAbsolutePercentageError)

When to invoke this skill

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

Reference signature

from torchmetrics import MeanAbsolutePercentageError

# MeanAbsolutePercentageError(**kwargs: Any) -> None

Library docstring

Compute `Mean Absolute Percentage Error`_ (MAPE).

.. math:: \text{MAPE} = \frac{1}{n}\sum_{i=1}^n\frac{|   y_i - \hat{y_i} |}{\max(\epsilon, | y_i |)}

Where :math:`y` is a tensor of target values, and :math:`\hat{y}` is a tensor of predictions.

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

- ``preds`` (:class:`~torch.Tensor`): Predictions from model
- ``target`` (:class:`~torch.Tensor`): Ground truth values

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

- ``mean_abs_percentage_error`` (:class:`~torch.Tensor`): A tensor with the mean absolute percentage error over
  state

Args:
    kwargs: Additional keyword arguments, see :ref:`Metric kwargs` for more info.

Note:
    MAPE output is a non-negative floating point. Best result is ``0.0`` . But it is important to note that,
    bad predictions, can lead to arbitrarily large values. Especially when some ``target`` values are close to 0.
    This `MAPE implementation returns`_ a very large number instead of ``inf``.

Example:
    >>> from torch import tensor
    >>> from torchmetrics.regression import MeanAbsolutePercentageError
    >>> target = tensor([1, 10, 1e6])
    >>> preds = tensor([0.9, 15, 1.2e6])
    >>> mean_abs_percentage_error = MeanAbsolutePercentageError()
    >>> mean_abs_percentage_error(preds, target)
    tensor(0.2667)

Quick recipe

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