Xiaoyang Wu comments

Results 495 comments of


                                            Xiaoyang Wu

Model Zoo

In the current version we do indoor SSL and outdoor SSL separately (a bit ugly I know), but we will make them together soon. I will train one version of...

Model Zoo

I see, then I think the indoor version should work. The normal can be acured by view direction which also can be estimated with LiDAR. I always believe outdoor point...

Model Zoo

I think tuning the model is necessary, as indoor contents and outdoor contents still have some gap. Yet for outdoor, we usually don’t need such a high resolution of perception,...

Model Zoo

And if possible, do Sonata pretraining with your own data. Joint training with existing assets would be better.

Question about Masked Scene Contrast evaluation

> If I understand correctly, in tab. 2(a) first row, MSC represents the results of pretraining on ScanNet + ARKitScenes and finetuning on ScanNet for semantic segmentation. SC denotes train...

Cannot allocate memory (12). Is there any way to optimize this error

Hi, strange issue. I haven't encountered a similar issue. Maybe reinstalling the environment can solve this problem.

PointTransformerV1 RuntimeError: Function 'AddmmBackward0' returned nan values in its 2th output.

What about training your custom dataset with PointTransformerV2? And could you provide the detailed training config and a sample point cloud of your custom dataset? The transformer is sensitive to...

PointTransformerV1 RuntimeError: Function 'AddmmBackward0' returned nan values in its 2th output.

With the PTv1?

PointTransformerV1 RuntimeError: Function 'AddmmBackward0' returned nan values in its 2th output.

> Thank you for your reply, I have resolved the issue. We need to make sure that each data in the dataset contains at least 256 points in order to...

RuntimeError: CUDA error: device-side assert triggered CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1. Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.

不建议打开rpe（关闭FlashAttention）时候单卡训练，我们马上会有一个无监督预训练的PTv3，这个PTv3完全不需要rpe或者bn layer，可以在S3DIS上达到非常高的结果，等待我们一个月后的更新吧。