Potential error in the latent overshooting objective
jan1854 opened this issue · comments
Jan Schneider commented
Hi, I think there might be a small error in the implementation of the latent overshooting objective. In the following line the prior (overshooting_vars[4]
) is passed as initial state to the transition model.
Line 187 in dacf418
If I understand the code correctly, this prior state corresponds to st-d from equation (7) of the PlaNet paper. However, in the paper st-d is sampled from the posterior distribution q(st-d | o≤ t-d), not the prior.
The original implementation seems to use the posterior as initial state as well (see https://github.com/google-research/planet/blob/c04226b6db136f5269625378cd6a0aa875a92842/planet/tools/overshooting.py#L126-L134).
So I think the posterior rather than the prior should be used as initial state here.
Kai Arulkumaran commented
I think you're right, thanks for spotting. I won't be able to work on this now, so would you be able to send in a PR that fixes this?
Jan Schneider commented
Yes, sure