minami

> +1 the equivalent of jax.lax.cond would be amazing. I'm currently using something equivalent to torch.where, but ultimately my torch.vmap solution is about twice as slow as the original vectorized...

Their representation is based on HumanML3D, which I think is very odd and strange. I do not know why one has to bother to run FK given gt joint rotations...

you can refer to these code: https://github.com/OpenMotionLab/MotionGPT/blob/main/fit.py, https://github.com/setarehc/diffusion-motion-inbetweening/blob/main/visualize/vis_utils.py very very strange