Xiaoyang Wu

Just as the GridSample. For grid sampling, we cluster point clouds into uneven clusters for each point cloud, but you can use Farthest Point Sampling to get N clusters for...

I will update one tuned from our latest Sonata (https://github.com/facebookresearch/sonata) soon

Yes, one known issue. Two suggestions: 1. disable AMP; 2. Enable clip on gradient. (I haven't try this)

This issue is caused by Flash Attention occasionally. I recently changed the dtype from fp16 to bp16, not sure whether it will help. Also, some advanced learning rate layer-wise scaler...

Alright, I have some time to reproduce these results now. I am pretty sure these number in the paper is exactly observed with my experiments. Unfortunately, I cannot access the...

In the old version, the training script could only provide an imprecise validation result on the point cloud after grid sampling. Also, the number is not the best, as we...

Hi, I'm the author you're waiting for :) Well, I usually use save_point_cloud in `pointcept.utils.visualization` to export the prediction to a PLY and then check it with Meshlab. You have...

> If you run the provided bash script, it will use the code from that specific folder. Thx

Hi Lukas, one most direct solution is to overload the following config: ```python hooks = [ dict(type="CheckpointLoader", keywords="module.seg_head.", replacement="module.seg_head_duplicate."), dict(type="IterationTimer", warmup_iter=2), dict(type="InformationWriter"), dict(type="SemSegEvaluator"), dict(type="CheckpointSaver", save_freq=None), dict(type="PreciseEvaluator", test_last=False), ] ``` Replacing...

Thanks to the efforts from @Yujia-Zhang0913. The following code should be helpful: https://github.com/Pointcept/Pointcept/commit/bebf24bef0f3730b0c570fae080479bbbc8b68f9 https://github.com/Pointcept/Concerto/blob/main/demo/0_pca_outdoor.py @vk7cp5ccvg-blip I prefer interacting with contributors using identifiable accounts, as it helps maintain constructive discussions. Please...