Problems about part segmentation

[jerryfeng2003]

Thank you for your great work!
I have tried to reproduce the results of PointGPT-L's part segmentation in paper, but find out there are some problems.

	Cls.mIoU	Inst.mIoU
paper	84.8	86.6
reproduce	83.7	85.7

It seems that there is a consistent gap in the evaluation metrics. I conducted the experiments on a single NVIDIA RTX 3090 GPU, with batch size 8, learning 0.0001. Other settings are set to their default values in the code.
Could you help me with the problem? Any insights would be greatly appreciated.

[KisAaki]

Hi, Sorry to bother you. I have a question. When doing the few-shot learning on Modelnet40, the accuracy is produced by the training process or the test process? I have tried so many times to test the few-shot learning and the result is so bad.
@jerryfeng2003

[jerryfeng2003]

Hi, Sorry to bother you. I have a question. When doing the few-shot learning on Modelnet40, the accuracy is produced by the training process or the test process? I have tried so many times to test the few-shot learning and the result is so bad. @jerryfeng2003

I have replied to your questions in another issue.