Xiaoyang Wu
Xiaoyang Wu
Hey, which dataset are you working on?
Sorry for the late response. Here is my suggestion: 1. Use our SSL pre-trained model (Sonata/Concerto) and freeze the encoder. 2. Check whether LovaszLoss is enabled 3. Manually reduce the...
Hey, the very old version of Pointcept use pth. Newer version use multiple npy file.
Hey all. If GPU memory is not sufficient, adjust grid size to 0.04m is a good choice for S3DIS. We generally use 0.02 as our GPU has large GPU memorys...
> The paper mentions 32 GPUs; what exact GPUs are used? 32 A100 > I have access to two H100 with 96GB - is that enough to train Sonata? I...
> > Hello, I have access to two H100 with 96GB - is that enough to train Sonata? I'd probably also have less data since I'm in a different domain....
> What was the GPU RAM of each NVIDIA A100 used to train Sonata model? I am facing OOM error to replicate Sonata results only for S3DIS data. I have...
Thanks for noting. I updated that comment. First comment "spconv" in env and then manually install with pip
Make sure you clean the environment before installing it. Also, docker file is available here: https://github.com/Pointcept/Pointcept/blob/main/scripts/build_image.sh We also have some images in dockerhub
Yes, > For example, the points of a car and a pedestrian may overlap in the same space. If the model is able to distinguish objects even when they overlap...