logit vs. log probability in powerset to multilabel conversion

[lukasstorck]

In the Powerset.to_multilabel(...) function with soft=True the docstring asks for "logit" values:

def to_multilabel(self, powerset: torch.Tensor, soft: bool = False) -> torch.Tensor:
    """Convert predictions from powerset to multi-label

    Parameter
    ---------
    powerset : (batch_size, num_frames, num_powerset_classes) torch.Tensor
        Soft predictions in "powerset" space.
    soft : bool, optional
        Return soft multi-label predictions. Defaults to False (i.e. hard predictions)
        Assumes that `powerset` are "logits" (not "probabilities").

    Returns
    -------
    multi_label : (batch_size, num_frames, num_classes) torch.Tensor
        Predictions in "multi-label" space.
    """

As I understand, logits (ranging from -inf to +inf) are the values in the model before applying the activation function. In case of the pyannote/segmentation-3.0 model, the output is in the powerset representation and the activation function of the last layer is "LogSoftmax", which results in log probabilities ranging from -inf to 0. These values are the values used by the powerset to multilabel conversion and not the values before the activation function.

I was confused by the use of the term "logit" in the docstring. Isn't "log probabilities" the correct term here? Thank you :D

[hbredin]

You are right. PR to fix the docstring very welcome!