idontlikelongname

Loss abnormal with waymo TransFusion_L config

1

Thanks for your great works. Im training TransFusion_L (lidar only) with waymo datasets. but the loss is abnormal, in particular the box size loss. Also the box size in test...

章老师你好，我使用mmrotated.models.losses.GDLoss作为尺寸回归的补充，训练几个iters后所有loss都会变成nan。我对比了mmrotated里面gwd_loss和你的实现，除了增加一个height维度的尺寸，没有发现太多不同，什么原因导致nan？ 具体配置参数如下。 mmrotated commit b78bab24dc35657fe4e349d 最新版 mmrotated GDLoss配置参数loss_gauss_iou=dict(type='GDLoss', loss_type='gwd', loss_weight=5.0) ```python @HEADS.register_module() class Anchor3DHeadGaussianLoss(Anchor3DHead): """Anchor head for SECOND/PointPillars/MVXNet/PartA2. Args: loss_corners (dict): Config of box corners loss. """ def __init__(self, loss_gauss_iou=dict(...