araffin/sbx

[Question] `MultiInputPolicy` not supported (DroQ)

jbirnick opened this issue · 1 comments

Hey, I'm trying to use SBX, specifically DroQ, with MultiInputPolicy. I get the error:

ValueError: Policy MultiInputPolicy unknown

Why is that? I thought SBX is compatible with SB3.

Why is that? I thought SBX is compatible with SB3.

"Stable Baselines Jax (SBX) is a proof of concept version of Stable-Baselines3 in Jax.
It provides a minimal number of features compared to SB3 but can be much faster"
from https://stable-baselines3.readthedocs.io/en/master/guide/sbx.html

Not all SB3 features are supported by SBX, CNN and full dict support is missing.
Only basic dict support is there (I would welcome a PR that add this line for Droq, that was an oversight):

sbx/sbx/sac/sac.py

Lines 45 to 46 in db6120b

# Minimal dict support using flatten()
"MultiInputPolicy": SACPolicy,

EDIT: as mentioned in the readme, DroQ can be seen as a special config of SAC, https://github.com/araffin/sbx?tab=readme-ov-file#note-about-droq, which in that case would have the MultiInputPolicy for simple dict spaces