dreamer_v2(pretrained: bool = False, overrides: object = {})Construct DreamerV2's world model, actor and critic — no objectives.
Dreamer's architecture with a categorical latent: 32 variables of 32 classes, sampled as one-hots and carried as a 1024-wide sparse binary vector. Four-layer 400-unit ELU heads for reward, value and policy, plus a frozen copy of the critic to compute returns against.
Model Size
Parameters
pretrainedbool= FalseNo weights are published for this family; passing
True raises.**overridesobject= {}Optional
DreamerV2Config field overrides. action_dim
is set by the environment, not by the paper.Returns
DreamerV2ModelThe trunk, configured with the released defaults.
Notes
Reference: Hafner, Lillicrap, Norouzi, and Ba, "Mastering Atari with Discrete World Models", ICLR, 2021 (arXiv:2010.02193).
Examples
>>> from lucid.models import dreamer_v2
>>> model = dreamer_v2(action_dim=6).eval()
>>> model.config.stoch_size, model.config.discrete
(32, 32)