dreamer(pretrained: bool = False, overrides: object = {})Construct Dreamer's world model, actor and critic — no objectives.
PlaNet's architecture unchanged — a four-layer stride-2 convolutional
encoder, a recurrent state-space model with a 200-unit deterministic
path and a 30-unit stochastic state, a mirrored decoder — plus the two
heads that replace PlaNet's planner: a tanh-squashed Gaussian actor
and a scalar critic, each three 300-unit ELU layers.
Model Size
Parameters
pretrainedbool= FalseNo weights are published for this family; passing
True raises
rather than returning a randomly initialised model.**overridesobject= {}Optional
DreamerConfig field overrides. action_dim is
the one you almost always need — it is set by the environment, not
by the paper.Returns
DreamerModelThe trunk, configured with the paper defaults and any overrides.
Notes
Reference: Hafner, Lillicrap, Ba, and Norouzi, "Dream to Control: Learning Behaviors by Latent Imagination", ICLR, 2020 (arXiv:1912.01603).
Examples
>>> from lucid.models.generative.dreamer import dreamer
>>> model = dreamer(action_dim=6).eval()
>>> model.config.horizon, model.config.lambda_
(15, 0.95)