Logcosherror

logcosherror

Metric LogCoshError from torchmetrics (torchmetrics.LogCoshError)

When to invoke this skill

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

Reference signature

from torchmetrics import LogCoshError

# LogCoshError(num_outputs: int = 1, **kwargs: Any) -> None

Library docstring

Compute the `LogCosh Error`_.

.. math:: \text{LogCoshError} = \log\left(\frac{\exp(\hat{y} - y) + \exp(\hat{y - y})}{2}\right)

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`): Estimated labels with shape ``(batch_size,)``
  or ``(batch_size, num_outputs)``
- ``target`` (:class:`~torch.Tensor`): Ground truth labels with shape ``(batch_size,)``
  or ``(batch_size, num_outputs)``

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

- ``log_cosh_error`` (:class:`~torch.Tensor`): A tensor with the log cosh error

Args:
    num_outputs: Number of outputs in multioutput setting
    kwargs: Additional keyword arguments, see :ref:`Metric kwargs` for more info.

Example (single output regression)::
    >>> from torchmetrics.regression import LogCoshError
    >>> preds = torch.tensor([3.0, 5.0, 2.5, 7.0])
    >>> target = torch.tensor([2.5, 5.0, 4.0, 8.0])
    >>> log_cosh_error = LogCoshError()
    >>> log_cosh_error(preds, target)
    tensor(0.3523)

Example (multi output regression)::
    >>> from torchmetrics.regression import LogCoshError
    >>> preds = torch.tensor([[3.0, 5.0, 1.2], [-2.1, 2.5, 7.0]])
    >>> target = torch.tensor([[2.5, 5.0, 1.3], [0.3, 4.0, 8.0]])
    >>> log_cosh_error = LogCoshError(num_outputs=3)
    >>> log_cosh_error(preds, target)
    tensor([0.9176, 0.4277, 0.2194])

Quick recipe

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