How is the motion classifier trained?

[Cr7st]

The recognition accuracy of real motions in the paper is very high (over 99% on HumanAct12 and NTU-RGB-D). However, STGCN only achives accuracy over 80% on NTU-RGB-D. I'm wondering how the motion classifier is trained.

[NinomiyaAsuka]

I also want to know this. I'm wondering if author can provide the code for training the classifier based on my own datasets.