dreamer_v2_atari(pretrained: bool = False, overrides: object = {})Construct DreamerV2 as the paper configures it for Atari.
A 600-unit deterministic state, a divergence scaled to 0.1, a discount of 0.999 weighted five times over, and a discrete policy — a one-hot over the game's buttons rather than a box.
The discount is the one number here the paper and the released
configuration disagree on: Table D.1 prints 0.995, the released
atari config uses 0.999, and the table's own note recommends
searching . The released value
is followed, because it is the one that produced the reported Atari
results.
Model Size
Parameters
pretrainedbool= FalseTrue raises.**overridesobject= {}DreamerV2Config field overrides. action_dim
is the number of actions the game exposes.Returns
DreamerV2ModelThe trunk, configured for Atari.
Notes
Reference: Hafner, Lillicrap, Norouzi, and Ba, "Mastering Atari with Discrete World Models", ICLR, 2021 (arXiv:2010.02193).
Action repeat of 4 is part of this configuration and is not a model
field — pass it to lucid.utils.rollout.rollout.
Examples
>>> from lucid.models import dreamer_v2_atari
>>> model = dreamer_v2_atari(action_dim=18).eval()
>>> model.config.action_space, model.config.resolved_actor_grad
('discrete', 'reinforce')