〖MMDetection〗解析文件：mmdet/models/losses/utils.py

《解析 MMDetection 中的损失函数辅助工具》

在 MMDetection 中，损失函数的计算通常需要一些辅助工具来进行灵活的处理。本文将深入解析 mmdet/models/losses/utils.py 文件中的函数。

一、reduce_loss 函数

def reduce_loss(loss: Tensor, reduction: str) -> Tensor:
    """Reduce loss as specified.

    Args:
        loss (Tensor): Elementwise loss tensor.
        reduction (str): Options are "none", "mean" and "sum".

    Return:
        Tensor: Reduced loss tensor.
    """
    reduction_enum = F._Reduction.get_enum(reduction)
    # none: 0, elementwise_mean:1, sum: 2
    if reduction_enum == 0:
        return loss
    elif reduction_enum == 1:
        return loss.mean()
    elif reduction_enum == 2:
        return loss.sum()

这个函数用于根据指定的reduction策略来减少损失张量。它接受一个损失张量和一个表示减少策略的字符串参数。通过使用 PyTorch 的F._Reduction.get_enum方法，将字符串转换为对应的枚举值。如果reduction是"none"，则直接返回原始损失张量；如果是"mean"，则计算损失张量的平均值；如果是"sum"，则计算损失张量的总和。

二、weight_reduce_loss 函数

def weight_reduce_loss(loss: Tensor,
                       weight: Optional[Tensor] = None,
                       reduction: str = 'mean',
                       avg_factor: Optional[float] = None) -> Tensor:
    """Apply element-wise weight and reduce loss.

    Args:
        loss (Tensor): Element-wise loss.
        weight (Optional[Tensor], optional): Element-wise weights.
            Defaults to None.
        reduction (str, optional): Same as built-in losses of PyTorch.
            Defaults to 'mean'.
        avg_factor (Optional[float], optional): Average factor when
            computing the mean of losses. Defaults to None.

    Returns:
        Tensor: Processed loss values.
    """
    # if weight is specified, apply element-wise weight
    if weight is not None:
        loss = loss * weight

    # if avg_factor is not specified, just reduce the loss
    if avg_factor is None:
        loss = reduce_loss(loss, reduction)
    else:
        # if reduction is mean, then average the loss by avg_factor
        if reduction == 'mean':
            # Avoid causing ZeroDivisionError when avg_factor is 0.0,
            # i.e., all labels of an image belong to ignore index.
            eps = torch.finfo(torch.float32).eps
            loss = loss.sum() / (avg_factor + eps)
        # if reduction is 'none', then do nothing, otherwise raise an error
        elif reduction!= 'none':
            raise ValueError('avg_factor can not be used with reduction="sum"')
    return loss

这个函数首先根据是否提供了权重张量，对损失进行逐元素加权。然后，根据是否提供了平均因子和指定的减少策略来进一步处理损失。如果没有提供平均因子，直接使用reduce_loss函数进行减少。如果减少策略是"mean"且提供了平均因子，则通过将损失总和除以平均因子加上一个极小值来避免除零错误。如果减少策略不是"none"且提供了平均因子，同时减少策略是"sum"，则会抛出一个值错误。

三、weighted_loss 装饰器

def weighted_loss(loss_func: Callable) -> Callable:
    """Create a weighted version of a given loss function.

    To use this decorator, the loss function must have the signature like
    `loss_func(pred, target, **kwargs)`. The function only needs to compute
    element-wise loss without any reduction. This decorator will add weight
    and reduction arguments to the function. The decorated function will have
    the signature like `loss_func(pred, target, weight=None, reduction='mean',
    avg_factor=None, **kwargs)`.

    :Example:

    >>> import torch
    >>> @weighted_loss
    >>> def l1_loss(pred, target):
    >>>     return (pred - target).abs()

    >>> pred = torch.Tensor([0, 2, 3])
    >>> target = torch.Tensor([1, 1, 1])
    >>> weight = torch.Tensor([1, 0, 1])

    >>> l1_loss(pred, target)
    tensor(1.3333)
    >>> l1_loss(pred, target, weight)
    tensor(1.)
    >>> l1_loss(pred, target, reduction='none')
    tensor([1., 1., 2.])
    >>> l1_loss(pred, target, weight, avg_factor=2)
    tensor(1.5000)
    """

    @functools.wraps(loss_func)
    def wrapper(pred: Tensor,
                target: Tensor,
                weight: Optional[Tensor] = None,
                reduction: str = 'mean',
                avg_factor: Optional[int] = None,
                **kwargs) -> Tensor:
        """
        Args:
            pred (Tensor): The prediction.
            target (Tensor): Target bboxes.
            weight (Optional[Tensor], optional): The weight of loss for each
                prediction. Defaults to None.
            reduction (str, optional): Options are "none", "mean" and "sum".
                Defaults to 'mean'.
            avg_factor (Optional[int], optional): Average factor that is used
                to average the loss. Defaults to None.

        Returns:
            Tensor: Loss tensor.
        """
        # get element-wise loss
        loss = loss_func(pred, target, **kwargs)
        loss = weight_reduce_loss(loss, weight, reduction, avg_factor)
        return loss

    return wrapper

这个装饰器用于创建一个带权重的损失函数版本。被装饰的损失函数应该只计算逐元素的损失，不进行任何减少操作。装饰器会为函数添加权重和减少策略的参数。在装饰后的函数中，首先调用原始的损失函数计算逐元素损失，然后使用weight_reduce_loss函数对损失进行加权和减少处理，并返回最终的损失张量。