Xiaoyang Wu

Hi, apology for waiting. I will release the config and model weight a few days later. I am off for the Chinese New Year holiday, and the config and weight...

Within today.

Hi, I just pushed some experiment records (containing joint training configs and weights) in the HuggingFace model repo (https://huggingface.co/Pointcept/PointTransformerV3) and I will update Pointcept tomorrow

I haven't encountered a similar issue before. What about first disabling the Hilbert-related setting for debugging? Also, most settings in Pointcept request 4 GPUs for best performance.

> I met the same problem. It seems that pytorch don't allow viewing a tensor as a new dtype with a different number of bytes. My pytorch version 1.10.0 I...

So sorry for the long wait. Please give me some time, I will check this after the CVPR deadline. There would be an update of Pointcept after that.

Hi, the default S3DIS + PTv3 config we provide disabled FlashAttention to enable RPE, as in our reproduction, we found that RPE is still needed for S3DIS (our reproduction result...

Okay, RPE is not efficient, you can refer our configs for ScanNet? Compare it with the S3DIS version, you will clearly know how to enable FlashAttention.

> just to make sure i understand this correctly, rpe is used to get the main results in table 5 (for higher performance), but then to show efficiency in table...

I think just some data scale issue, S3DIS is too small. Yet, for our PPT multi-dataset setting, the model is unified and stronger even with one weight. Anyway, the good...