Mutualinfoscore

mutualinfoscore

Metric MutualInfoScore from torchmetrics (torchmetrics.clustering.MutualInfoScore)

When to invoke this skill

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

Reference signature

from torchmetrics.clustering import MutualInfoScore

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

Library docstring

Compute `Mutual Information Score`_.

.. math::
    MI(U,V) = \sum_{i=1}^{|U|} \sum_{j=1}^{|V|} \frac{|U_i\cap V_j|}{N}
    \log\frac{N|U_i\cap V_j|}{|U_i||V_j|}

Where :math:`U` is a tensor of target values, :math:`V` is a tensor of predictions,
:math:`|U_i|` is the number of samples in cluster :math:`U_i`, and :math:`|V_i|` is the number of samples in
cluster :math:`V_i`. The metric is symmetric, therefore swapping :math:`U` and :math:`V` yields the same mutual
information score.

This clustering metric is an extrinsic measure, because it requires ground truth clustering labels, which may not
be available in practice since clustering in generally is used for unsupervised learning.

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

- ``preds`` (:class:`~torch.Tensor`): single integer tensor with shape ``(N,)`` with predicted cluster labels
- ``target`` (:class:`~torch.Tensor`): single integer tensor with shape ``(N,)`` with ground truth cluster labels

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

- ``mi_score`` (:class:`~torch.Tensor`): A tensor with the Mutual Information Score

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

Example::
    >>> import torch
    >>> from torchmetrics.clustering import MutualInfoScore
    >>> preds = torch.tensor([2, 1, 0, 1, 0])
    >>> target = torch.tensor([0, 2, 1, 1, 0])
    >>> mi_score = MutualInfoScore()
    >>> mi_score(preds, target)
    tensor(0.5004)

Quick recipe

import torchmetrics.clustering as _m
score = _m.MutualInfoScore(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).