Problems for reproducing traversal disentanglement experiments in Figure 3 (b)

[jameszhou-gl]

Hi Dr Shen. Thanks very much for your impressive paper and open-sourced codes. I met a problem when I tried to reproduce the traversal disentanglement experiments (Figure 3(b) in the paper).

Specifically, There is a 'model.module.traverse' function in train.py

When I run the traverse function with a trained model, I find that if I change the causal factor, like the pendulum_angle, the shadow will also change. I can't achieve disentanglement. So, could you please help how to reproduce disentanglement experiments?