关于Focal loss 和 GHM

2019-12-30  本文已影响0人  _从前从前_

这里直接贴一个知乎链接,可以说讲得很好了https://zhuanlan.zhihu.com/p/80594704

在单阶段的目标检测任务中,由于前景和背景不平衡的问题导致效果不如两阶段的方法(这个观点是否正确还有待研究)。

Focal loss

通过两个参数,一个控制平衡前景和背景的Loss贡献,一个控制简单样本的loss贡献。在目标检测任务中,简单样本虽然产生的loss很小,但是大部分样本均是简单样本,所以梯度仍然是由简单样本主导的。Focal Loss引入了两个超参数尝试去解决了这个问题。

GHM

通过划分bins,将梯度标准化,去解决上述提到的问题。GHM的想法是,我们确实不应该过多关注易分样本,但是特别难分的样本(outliers,离群点)也不该关注啊!那怎么同时衰减易分样本和特别难分的样本呢?太简单了,谁的数量多衰减谁呗!那怎么衰减数量多的呢?简单啊,定义一个变量,让这个变量能衡量出一定梯度范围内的样本数量——这不就是物理上密度的概念吗?
其中密度的估计用到了EMA,考虑了全局的样本分布,这也是focal loss不具备的。只是网络的前期是否能直接使用GHM?这也是否会导致前期训练的不稳定?
在自己的数据集上。GHM的表现不如focal loss,还需要更多的探索。

GHMC mmdetection代码解析

# 注册loss函数
@LOSSES.register_module
class GHMC(nn.Module):
    """GHM Classification Loss.

    Details of the theorem can be viewed in the paper
    "Gradient Harmonized Single-stage Detector".
    https://arxiv.org/abs/1811.05181

    Args:
        bins (int): Number of the unit regions for distribution calculation.
        momentum (float): The parameter for moving average.
        use_sigmoid (bool): Can only be true for BCE based loss now.
        loss_weight (float): The weight of the total GHM-C loss.
    """

    def __init__(self, bins=10, momentum=0, use_sigmoid=True, loss_weight=1.0):
        super(GHMC, self).__init__()
        self.bins = bins
        self.momentum = momentum
        edges = torch.arange(bins + 1).float() / bins
        self.register_buffer('edges', edges)
        self.edges[-1] += 1e-6
        if momentum > 0:
            acc_sum = torch.zeros(bins)
            self.register_buffer('acc_sum', acc_sum)
        self.use_sigmoid = use_sigmoid
        if not self.use_sigmoid:
            raise NotImplementedError
        self.loss_weight = loss_weight

    def forward(self, pred, target, label_weight, *args, **kwargs):
        """Calculate the GHM-C loss.

        Args:
            pred (float tensor of size [batch_num, class_num]):
                The direct prediction of classification fc layer.
            target (float tensor of size [batch_num, class_num]):
                Binary class target for each sample.
            label_weight (float tensor of size [batch_num, class_num]):
                the value is 1 if the sample is valid and 0 if ignored.
        Returns:
            The gradient harmonized loss.
        """
        # the target should be binary class label
        if pred.dim() != target.dim():
            target, label_weight = _expand_binary_labels(
            target, label_weight, pred.size(-1))
        target, label_weight = target.float(), label_weight.float()
        edges = self.edges
        mmt = self.momentum
        weights = torch.zeros_like(pred)

        # gradient length
        # sigmoid梯度计算
        g = torch.abs(pred.sigmoid().detach() - target)
        # 有效的label的位置
        valid = label_weight > 0
        # 有效的label的数量
        tot = max(valid.float().sum().item(), 1.0)
        n = 0  # n valid bins
        for i in range(self.bins):
            # 将对应的梯度值划分到对应的bin中, 0-1
            inds = (g >= edges[i]) & (g < edges[i + 1]) & valid
            # 该bin中存在多少个样本
            num_in_bin = inds.sum().item()
            if num_in_bin > 0:
                if mmt > 0:
                    # moment计算num bin
                    self.acc_sum[i] = mmt * self.acc_sum[i] \
                        + (1 - mmt) * num_in_bin
                    # 权重等于总数/num bin
                    weights[inds] = tot / self.acc_sum[i]
                else:
                    weights[inds] = tot / num_in_bin
                n += 1
        if n > 0:
            # scale系数
            weights = weights / n

        loss = F.binary_cross_entropy_with_logits(
            pred, target, weights, reduction='sum') / tot
        return loss * self.loss_weight
上一篇下一篇

猜你喜欢

热点阅读