MSDN_base.py 文件源码

def build_loss_bbox(self, bbox_pred, roi_data):
        bbox_targets, bbox_inside_weights, bbox_outside_weights = roi_data[2:]
        bbox_targets = torch.mul(bbox_targets, bbox_inside_weights)
        bbox_pred = torch.mul(bbox_pred, bbox_inside_weights)
        fg_cnt = torch.sum(bbox_inside_weights[:, 0].data.ne(0)) 
        loss_box = F.smooth_l1_loss(bbox_pred, bbox_targets, size_average=False) / (fg_cnt + 1e-5)
        return loss_box