Multioutputwrapper

multioutputwrapper

Metric MultioutputWrapper from torchmetrics (torchmetrics.MultioutputWrapper)

When to invoke this skill

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

Reference signature

from torchmetrics import MultioutputWrapper

# MultioutputWrapper(base_metric: torchmetrics.metric.Metric, num_outputs: int, output_dim: int = -1, remove_nans: bool = True, squeeze_outputs: bool = True) -> None

Library docstring

Wrap a base metric to enable it to support multiple outputs.

Several torchmetrics metrics, such as :class:`~torchmetrics.regression.spearman.SpearmanCorrCoef` lack support for
multioutput mode. This class wraps such metrics to support computing one metric per output.
Unlike specific torchmetric metrics, it doesn't support any aggregation across outputs.
This means if you set ``num_outputs`` to 2, ``.compute()`` will return a Tensor of dimension
``(2, ...)`` where ``...`` represents the dimensions the metric returns when not wrapped.

In addition to enabling multioutput support for metrics that lack it, this class also supports, albeit in a crude
fashion, dealing with missing labels (or other data). When ``remove_nans`` is passed, the class will remove the
intersection of NaN containing "rows" upon each update for each output. For example, suppose a user uses
`MultioutputWrapper` to wrap :class:`torchmetrics.regression.r2.R2Score` with 2 outputs, one of which occasionally
has missing labels for classes like ``R2Score`` is that this class supports removing ``NaN`` values
(parameter ``remove_nans``) on a per-output basis. When ``remove_nans`` is passed the wrapper will remove all rows

Args:
    base_metric: Metric being wrapped.
    num_outputs: Expected dimensionality of the output dimension.
        This parameter is used to determine the number of distinct metrics we need to track.
    output_dim:
        Dimension on which output is expected. Note that while this provides some flexibility, the output dimension
        must be the same for all inputs to update. This applies even for metrics such as `Accuracy` where the labels
        can have a different number of dimensions than the predictions. This can be worked around if the output
        dimension can be set to -1 for both, even if -1 corresponds to different dimensions in different inputs.
    remove_nans:
        Whether to remove the intersection of rows containing NaNs from the values passed through to each underlying
        metric. Proper operation requires all tensors passed to update to have dimension ``(N, ...)`` where N
        represents the length of the batch or dataset being passed in.
    

Quick recipe

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