Function binary_cross_entropy

imperative/python/megengine/functional/loss.py:214–256

r"""Computes the binary cross entropy loss (using logits by default). Args: pred: `(N, *)`, where `*` means any number of additional dimensions. label: `(N, *)`, same shape as the input. with_logits: bool, whether to apply sigmoid first. Default: True reduction:

(
    pred: Tensor, label: Tensor, with_logits: bool = True, reduction: str = "mean",
)

@_reduce_output
def binary_cross_entropy(
    pred: Tensor, label: Tensor, with_logits: bool = True, reduction: str = "mean",
) -> Tensor:
    r"""Computes the binary cross entropy loss (using logits by default).

    Args:
        pred: `(N, *)`, where `*` means any number of additional dimensions.
        label: `(N, *)`, same shape as the input.
        with_logits: bool, whether to apply sigmoid first. Default: True
        reduction: the reduction to apply to the output: 'none' | 'mean' | 'sum'.

    Returns:
        loss value.

    Examples:

        By default(``with_logitis`` is True), ``pred`` is assumed to be logits,
        class probabilities are given by softmax.
        It has better numerical stability compared with sequential calls to 
        :func:`~.sigmoid` and :func:`~.binary_cross_entropy`.

        >>> pred = Tensor([0.9, 0.7, 0.3])
        >>> label = Tensor([1., 1., 1.])
        >>> F.nn.binary_cross_entropy(pred, label)
        Tensor(0.4328984, device=xpux:0)
        >>> F.nn.binary_cross_entropy(pred, label, reduction="none")
        Tensor([0.3412 0.4032 0.5544], device=xpux:0)

        If the ``pred`` value has been probabilities, set ``with_logits`` to False:

        >>> pred = Tensor([0.9, 0.7, 0.3])
        >>> label = Tensor([1., 1., 1.])
        >>> F.nn.binary_cross_entropy(pred, label, with_logits=False)
        Tensor(0.5553361, device=xpux:0)
        >>> F.nn.binary_cross_entropy(pred, label, with_logits=False, reduction="none")
        Tensor([0.1054 0.3567 1.204 ], device=xpux:0) 

    """
    if not with_logits:
        return -(label * log(pred) + (1 - label) * log(1 - pred))
    # logsigmoid(pred) and logsigmoid(-pred) has common sub-expression
    # hopefully the backend would optimize this
    return -(label * logsigmoid(pred) + (1 - label) * logsigmoid(-pred))


@_reduce_output