Xiaoyang Wu

Hi, I responded to your other issue. Your case might be one data has some issue during the preprocessing (not sure). How about processing these data one by one? Or...

Here (https://github.com/Pointcept/Pointcept/blob/main/configs/scannet/pretrain-msc-v1m1-0-spunet-base.py#L44) https://pytorch.org/docs/stable/generated/torch.optim.lr_scheduler.OneCycleLR.html Usually we set the `pct_start` number to 0.05

Hey, OctFormer provides sparse convolution by an oct-tree data structure, which supports group operations. It has been a long time since we did the ablation; therefore, I have already forgotten...

Thanks. That might be the reason we fail to observe additional S3DIS can promote ScanNet. Yet we do observe additional ScanNet & Structured3D can benefit S3DIS which is a bit...

Q1: Do you see nan in training? If no, please provide a detailed config Q1: Working on it now

> > Same question here > > As for Q2, you can use set `-r true` behind your instruction, just like `sh scripts/train.sh -g 4 -d s3dis -c semseg-minkunet34c-0-base -n...

@iTomxy Currently run the `train` script with `-r true`, it will automatically run the test

I also tried to solve the issue with this commit (https://github.com/Pointcept/Pointcept/commit/402256155511a21b38b2c16233c97ad7d18120d1) Could you guys check whether the new version of the code solves this issue?

I will check your config. Als,o you can try our latest self-supervised pre-trained PTv3 --- Sonata (https://github.com/facebookresearch/sonata). In this version, we change all BN to LN, and it will also...

Sorry, the preprocessing code for nuScenes was written 1.5 years ago. I already forgot the details. Maybe directly reading the preprocessing code would be helpful?