dreamer_v2_world_model

→DreamerV2ForWorldModeling

dreamer_v2_world_model(pretrained: bool = False, overrides: object = {})

source edit

Construct DreamerV2 with all three of its objectives.

Same trunk as dreamer_v2, wrapped with the world-model bound under a balanced divergence, an actor maximising imagined $V_\lambda$ with an entropy bonus, and a critic regressing onto a target copy of itself.

Model Size

Parameters

pretrainedbool= False

No weights are published for this family; passing True raises.

**overridesobject= {}

Optional DreamerV2Config field overrides. kl_balance and actor_grad are the two that change the algorithm rather than its size.

Returns

DreamerV2ForWorldModeling

The trunk plus the objectives.

Notes

Reference: Hafner, Lillicrap, Norouzi, and Ba, "Mastering Atari with Discrete World Models", ICLR, 2021 (arXiv:2010.02193).

The three losses take three separate optimisers; use lucid.models.DreamerV2ForWorldModeling.backward to fill their gradients, and call update_slow_target once per step.

Examples

>>> from lucid.models import dreamer_v2_world_model
>>> model = dreamer_v2_world_model(action_dim=6).eval()
>>> model.config.kl_balance, model.config.resolved_actor_grad
(0.8, 'dynamics')

Used by 1

lucid.models.generative.dreamer_v2