Xiaoyang Wu

你好，可以参考这个issue (https://github.com/Pointcept/Pointcept/issues/108) 里的讨论。

你好，看起来似乎是你将已经处理后的ScanNet文件夹作为预处理的Raw（dataset_root）了。可能你下载的是我提供的已处理好的scannet数据集，而不是官方的原始数据集。

As discussed here: https://github.com/Pointcept/Pointcept?tab=readme-ov-file#scannet-v2 ``` # PROCESSED_SCANNET_DIR: the directory of the processed ScanNet dataset. mkdir data ln -s ${PROCESSED_SCANNET_DIR} ${CODEBASE_DIR}/data/scannet ```

第一次的相对路径是正确的，可以去检测第一次路径下，报错文件目录是否可以正确读取。

Hi, thanks for noting, I will fix it in the next update.

I checked the link in the v1.5.2 codebase; I think it still works. (I move it to huggingface) https://huggingface.co/datasets/Pointcept/nuscenes-compressed

Public now. I am not sure if it is allowed.

Hi, the memory efficiency is sourced from flash attention as all operation is carried out in the L2 cache of the GPU. The memory cost is huge if conducting large...

Hi, I am going to optimize data augmentation in the next version and a careful ablation is needed before releasing them. Currently, you can try to disable some augmentations, you...

``` However, I find that during training, the CPU utilization rate is high and GPU utilization fluctuates greatly ``` If data time is not huge, the issue might not be...