Xiaoyang Wu

Hey, I think num_classes should never be 1. As num_classes = 1 means every point only has one option. In your case, I think num_classes should be 1: [background, foreground]...

Hey guys. I think the balance loss weight for CrossEntropyLoss is a short-term solution. For the long term, we need a pre-trained backbone learned with object datasets. We are working...

Hi! Do you mean the validation phase or the testing phase after training? I found a TODO [here](https://github.com/Pointcept/Pointcept/blob/main/pointcept/engines/test.py#L233), I think I can add some support to this function in the...

The mIoU of a dataset is not the mean of each mean IoU, it is the mean IoU of accumulated points belonging to each category. Code here might be helpful:...

> Me personally, I like to think that one dataset is treated as one single BIG scene. This is also correct.

``` Does SparseUNet use an attention mechanism? ``` SparseUNet is UNet-style Sparse Conv, you can find more information in Chrischoy's paper, 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. ``` Why...

Hey, there are some operators like `spconv`, `torch_scatter`, that are non-deterministic in CUDA. A slight difference caused by non-deterministic operators won't be a problem in actual usage.

Hi, try the below new InformationWriter. But not ensure the same as the PTv3 paper. In the PTv3 paper, I use a subprocess to track GPU memory usage. Recently, I...

Here: https://github.com/Pointcept/Pointcept/blob/main/pointcept/models/default.py#L211-L217

You can use FPS and KNN to group points into N clusters, and then use torch_scatter to aggregate.