Xiaoyang Wu

We don't crop real-world point clouds during preprocessing. Batched point cloud in pointcept no need to be even. https://github.com/Pointcept/Pointcept?tab=readme-ov-file#offset

> i concate intensity to "color" in your codebase, which means the "color" information is my intensity information. Hi, I think you can try to assign the name "strength" to...

> If we have batch size = 12 in config, does that mean we have 12 real-world point clouds (without cropping) in 1 batch? Yes, 12 real-world point clouds in...

Hi, I will update the code for object classification. I will give you a response later.

Hi, I think Spatio-Temporal Segmentation by Chrischoy is the first introduced ElasticDistortion. And I think the order of ElasticDistortion and GridSampling doesn't matter.

Hi, syncing does need a bit longer time, as multiple GPUs are separated testing without syncing each step. If you want to accelerate the testing process, you can try our...

> Hi, syncing does need a bit longer time, as multiple GPUs are separated testing without syncing each step. If you want to accelerate the testing process, you can try...

Oh, actually, I also don't fully understand the meaning of these parameters. I just transfer them from Chris Choy's repo.

Hi, enable amp (fp32) and flash attention (fp16) may cause NaN. Classical multi-head self-attention (MSA) is not stable. In your case, I think disabling AMP initially is a good choice,...

> Hi @JinyuanShao have you solved the problem? The updated config should be much faster. Also, the aforementioned case is not a bug. Some datasets do need a lot of...