Haisong Liu
Haisong Liu
你说的好像很有道理。。要不你提一个Pull Request吧,谢谢啦
See https://mmcv.readthedocs.io/en/master/understand_mmcv/config.html#inherit-from-base-config-with-ignored-fields
有吧,比如mixing末尾就有: https://github.com/MCG-NJU/SparseBEV/blob/024778a5fb4eb98a1b72dd6e2e16b054fea87c94/models/sparsebev_transformer.py#L379 self attention也有,在multiHeadAttention的官方实现里面
没有的话 它会自己编译吧,如果实在编译不过的话,你就找他们预编译好的whl然后按他们的要求选cuda和torch版本
没遇到这情况,建议你单步调试看一下哪里的类型不对。盲猜是backbone那边的auto fp16
没问题啊,tau学成负数不就是等价的了
在3d空间中生成采样点明显更好呗 可以更贴近3d box形状
Ah yeah, I forgot to upload the pretrained model for lidar-only varients. I will fix this in a few days. Sorry.
Hi, I've released the pretrained weights for camliraft_l
你总得回到真实坐标吧。。