OHEM的pytorch代码实现细节

最新推荐文章于 2024-10-30 06:06:39 发布

一路狂奔的猪

最新推荐文章于 2024-10-30 06:06:39 发布

阅读量6.1k

点赞数 1

分类专栏：目标检测文章标签： ohem OHEM代码实现细节

本文链接：https://blog.csdn.net/qq_33547191/article/details/89476341

版权

目标检测专栏收录该内容

35 篇文章 0 订阅

订阅专栏

详细解读一下OHEM的实现代码：

def ohem_loss(
    batch_size, cls_pred, cls_target, loc_pred, loc_target, smooth_l1_sigma=1.0
):
    """
    Arguments:
        batch_size (int): number of sampled rois for bbox head training
        loc_pred (FloatTensor): [R, 4], location of positive rois
        loc_target (FloatTensor): [R, 4], location of positive rois
        pos_mask (FloatTensor): [R], binary mask for sampled positive rois
        cls_pred (FloatTensor): [R, C]
        cls_target (LongTensor): [R]

    Returns:
        cls_loss, loc_loss (FloatTensor)
    """
    ohem_cls_loss = F.cross_entropy(cls_pred, cls_target, reduction='none', ignore_index=-1)
    ohem_loc_loss = smooth_l1_loss(loc_pred, loc_target, sigma=smooth_l1_sigma, reduce=False)
    #这里先暂存下正常的分类loss和回归loss
    loss = ohem_cls_loss + ohem_loc_loss
    #然后对分类和回归loss求和

  
    sorted_ohem_loss, idx = torch.sort(loss, descending=True)
    #再对loss进行降序排列
    keep_num = min(sorted_ohem_loss.size()[0], batch_size)
    #得到需要保留的loss数量
    if keep_num < sorted_ohem_loss.size()[0]:
    #这句的作用是如果保留数目小于现有loss总数，则进行筛选保留，否则全部保留
        keep_idx_cuda = idx[:keep_num]
        #保留到需要keep的数目
        ohem_cls_loss = ohem_cls_loss[keep_idx_cuda]
        ohem_loc_loss = ohem_loc_loss[keep_idx_cuda]
        #分类和回归保留相同的数目
    cls_loss = ohem_cls_loss.sum() / keep_num
    loc_loss = ohem_loc_loss.sum() / keep_num
    #然后分别对分类和回归loss求均值
    return cls_loss, loc_loss