varifocal loss（VFL）介绍

---

一、论文链接

https://arxiv.org/pdf/2008.13367.pdf

二、公式理解

简单说明下，这里的IACS是IoU-aware classification score的缩写。VFL原文里面这个target socre也就是q，是一个和IOU有关的软标签。对于挑选出的正样本（也就是和GT相交的样本）的分类标签分数不再是0、1两个极端，而是与IOU有关的标签q，这样的好处在于：比如IOU很大，标签就是0.9，IOU很小，标签就是0.1，直觉上也更有道理。

当q>0时，VFL对于正样本没有超参，也就是没有任何衰减；而当q=0时，VFL对于负样本是有超参项的，gama会减少负样本贡献，alpha是为了防止过度抑制，总体来说是会减少负样本的贡献的。这就是和focal loss的不同之处，focal loss两项都有超参，正样本的质量也是会被降低的。

代码

下面代码抄自PaddleDetection套件：

def varifocal_loss(pred,
                   target,
                   alpha=0.75,
                   gamma=2.0,
                   iou_weighted=True,
                   use_sigmoid=True):
    """`Varifocal Loss <https://arxiv.org/abs/2008.13367>`_

    Args:
        pred (Tensor): The prediction with shape (N, C), C is the
            number of classes
        target (Tensor): The learning target of the iou-aware
            classification score with shape (N, C), C is the number of classes.
        alpha (float, optional): A balance factor for the negative part of
            Varifocal Loss, which is different from the alpha of Focal Loss.
            Defaults to 0.75.
        gamma (float, optional): The gamma for calculating the modulating
            factor. Defaults to 2.0.
        iou_weighted (bool, optional): Whether to weight the loss of the
            positive example with the iou target. Defaults to True.
    """
    # pred and target should be of the same size
    assert pred.shape == target.shape
    if use_sigmoid:
        pred_new = F.sigmoid(pred)
    else:
        pred_new = pred
    target = target.cast(pred.dtype)
    if iou_weighted:
        focal_weight = target * (target > 0.0).cast('float32') + \
            alpha * (pred_new - target).abs().pow(gamma) * \
            (target <= 0.0).cast('float32')
    else:
        focal_weight = (target > 0.0).cast('float32') + \
            alpha * (pred_new - target).abs().pow(gamma) * \
            (target <= 0.0).cast('float32')

    if use_sigmoid:
        loss = F.binary_cross_entropy_with_logits(
            pred, target, reduction='none') * focal_weight
    else:
        loss = F.binary_cross_entropy(
            pred, target, reduction='none') * focal_weight
        loss = loss.sum(axis=1)
    return loss

和公式有一点点不同的地方是q=0时，VFL=alpha * (pred_new - target).abs().pow(gamma)，猜测target实际可能并不一定等于0，当然也兼容=0的情况。