google-research/torchsde

Latent SDE failed to generate longer samples

xdong99 opened this issue · 0 comments

Hi,

Thank you for this great work! I've been playing with and testing the latent_sde_lorenz example. The original model you provided, trained with 2 seconds of data sampling form the L63 system, fails to generate samples for a bit longer time. Below is the x trajectory of one generated sample comparing to one real sample.

smpx_traj1

Might this due to the insufficient training of the drift function "h"? Could you please provide some insights on how to design the latent SDE model so it can have higher generative capacity? Thank you!