Model does not demonstrate the same capabilities as in the paper

[nicolaskossmann]

Dear Author(s)/Research Team/@yhyang-myron ,

First of all great work.
I used the given PTv3 object as a checkpoint and trained the MLP on a 3D model (glb 4k). Here, using a scaling factor of 0.75 (with the default settings, 5k epochs) as an example, you can see that the model appears to have difficulty accurately determining the segment boundaries. In Figure 4 of the paper, you can see exactly how it can segment complex objects with clear segmentation boundaries. Does anyone have any idea what might be causing this?

[Tommy881003]

Hi, sorry for jumping in — I’m trying to get the project running on my device, but I’ve been running into a lot of dependency conflicts and missing models. Would you be willing to share either:

• the output of your pip list (or conda list), so I can match the package versions, or
• even better, if you happen to have a working environment.yml or Dockerfile, that would be a lifesaver.

I’ve already tried setting up with:

• Python 3.10 / PyTorch 2.1.0 (CUDA 12.1)
• Installing flash-attention and other requirements

…but I still keep hitting errors like missing libcublas.so.11 or undefined symbol.
Totally understand if you don’t have time, but even a rough list of the versions you used would help a lot. Thanks!