Haisong Liu comments

Results 34 comments of


                                            Haisong Liu

关于预训练权重下载问题

您再仔细看看readme呢

关于预训练权重下载问题

https://github.com/open-mmlab/mmdetection3d/tree/main/configs/nuimages @WZWAsuna

KeyError: 'occ_path'

Oh no, you should not generate these `pkl` by yourself. You should download them from the link we provide.

why there is a re-implementation of gradient checkpointing

This implementation is exactly the same as the official code of torch 2.0. So, if you are using torch2.0, you can directly use torch.utils.checkpoint. However, for torch1.10 users, the official...

why there is a re-implementation of gradient checkpointing

Yes you are correct. Our checkpoint.py is directly copied from https://pytorch.org/docs/2.0/_modules/torch/utils/checkpoint.html#checkpoint with no modifications

About max points in msmv_sampling.cpp

there is no problem if you increase MAX_POINT.

Question about merging SparseBEV into the MMDection3D Framework.

Thanks for your effort! Looks like there is a lot of detail here. You can reach me by email: [email protected]

Cuda out of memory

Sorry for the late reply. I optimized the code and now you should be able to evaluate vov99 model on your GPU. By the way, you can further boost the...

Is there ablation study about the train_interval?

No. You can have a try.

Deployment questions

Hi, we haven't tried TensorRT yet, as we are not experts on this field. SparseBEV can be implemented by native pytorch (the CUDA operator is optional and only 15% faster),...