Sourceaggregatedsignaldistortionratio

sourceaggregatedsignaldistortionratio

Metric SourceAggregatedSignalDistortionRatio from torchmetrics (torchmetrics.audio.SourceAggregatedSignalDistortionRatio)

When to invoke this skill

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

Reference signature

from torchmetrics.audio import SourceAggregatedSignalDistortionRatio

# SourceAggregatedSignalDistortionRatio(scale_invariant: bool = True, zero_mean: bool = False, **kwargs: Any) -> None

Library docstring

`Source-aggregated signal-to-distortion ratio`_ (SA-SDR).

The SA-SDR is proposed to provide a stable gradient for meeting style source separation, where
one-speaker and multiple-speaker scenes coexist.

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

- ``preds`` (:class:`~torch.Tensor`): float tensor with shape ``(..., spk, time)``
- ``target`` (:class:`~torch.Tensor`): float tensor with shape ``(..., spk, time)``

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

- ``sa_sdr`` (:class:`~torch.Tensor`): float scalar tensor with average SA-SDR value over samples

Args:
    preds: float tensor with shape ``(..., spk, time)``
    target: float tensor with shape ``(..., spk, time)``
    scale_invariant: if True, scale the targets of different speakers with the same alpha
    zero_mean: If to zero mean target and preds or not
    kwargs: Additional keyword arguments, see :ref:`Metric kwargs` for more info.

Example:
    >>> from torch import randn
    >>> from torchmetrics.audio import SourceAggregatedSignalDistortionRatio
    >>> preds = randn(2, 8000) # [..., spk, time]
    >>> target = randn(2, 8000)
    >>> sasdr = SourceAggregatedSignalDistortionRatio()
    >>> sasdr(preds, target)
    tensor(-50.8171)
    >>> # use with pit
    >>> from torchmetrics.audio import PermutationInvariantTraining
    >>> from torchmetrics.functional.audio import source_aggregated_signal_distortion_ratio
    >>> preds = randn(4, 2, 8000)  # [batch, spk, time]
    >>> target = randn(4, 2, 8000)
    >>> pit = PermutationInvariantTraining(source_aggregated_signal_distortion_ratio,
    ...     mode="permutation-wise", eval_func="max")
    >>> pit(preds, target)
    tensor(-43.9780)

Quick recipe

import torchmetrics.audio as _m
score = _m.SourceAggregatedSignalDistortionRatio(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).