Ruoxi

Meet the same problem today. What I did is to try to change these lines in the matmul example to make use of faster lower-precision reduction tensor core operations: ```python...

50k per phase.

256, and the size of Objaverse.

The generalization was worse in an experimental, unreleased version of the model. I am not 100% sure if this is the same case for the v1.2 architecture.

It is the default or paper value, perhaps 5.0

Please check the ME installation requirements at https://github.com/NVIDIA/MinkowskiEngine?tab=readme-ov-file#requirements . If you think you meet the requirements, please post more details so we can look into it (including OS, CUDA, PyTorch...

Did you meet any problems by just changing the model name?

you need to add some arguments to the command line; see the readme: replace main.py with example.py for what you need ``` python3 src/main.py --trial_name pointbert_all model.name=PointBERT model.scaling=4 model.use_dense=True training.lr=0.0005...

depth means point map which is the xyz part, rgb should be the feat part.