Wierd results with custom data

Hi! Thank you for your effort!

I encountered some weird results when running pretrained Ditto_s2m.ckpt on my custom data. Some of the revolute joints did well, but Ditto wrongly predicted almost all the prismatic joints as revolute joints. 

Just so you know, I directly modified the code through the provided demo.ipynb, so is there any clue what's going on?