how to reproduce the classification results

[M-leng]

I trained the Hengshuang's model. Under the default setting, I only got 89.7177 (instance average Acc). How to reproduce the result "91.7" listed in the table.

[qq456cvb]

Hi, Leng, you may look at the issue #13, and use the hyperparameters there. See if you can reproduce the results.

[yuchenlichuck]

Hi, Leng, I found that I used the pre-trained model to achieve 91.7%

[MikailINTech]

Hello, I have to study the point transformer from Hengshuang for a project so I wanted to reproduce the classification results. I cloned the repo and executed the code but it looks like the model cannot go beyond 50% of accuracy (I tried to run it up to 120 epochs with the same config). Did anyone had the same problem ?

[agoodge]

Hi, Leng, I found that I used the pre-trained model to achieve 91.7%

Where is the pre-trained model available?

[cygao1215]

如何可视化分类结果

[qq456cvb]

The table numbers are the best instance accuracy over a full 200-epoch run, and run-to-run variance of 1 to 2% is normal from seed, augmentation, and LR. Make sure you train the full schedule rather than stopping early, and keep the default augmentation. As the README notes, the Hengshuang/Nico LRs probably aren't fully tuned, so a bit of tuning can help close the gap.