data
DreamerOutput
extends
ModelOutputDreamerOutput(observation: Tensor, reward: Tensor, value: Tensor, posterior_stoch: Tensor, posterior_mean: Tensor, posterior_std: Tensor, prior_mean: Tensor, prior_std: Tensor, deter: Tensor, loss: Tensor | None = None, recon_loss: Tensor | None = None, reward_loss: Tensor | None = None, kl_loss: Tensor | None = None, pcont_loss: Tensor | None = None, behavior: DreamerBehaviorOutput | None = None)What DreamerModel returns after filtering a trajectory.
Attributes
observationTensorReconstruction from the posterior states,
(B, T, C, 64, 64).rewardTensorPredicted reward at each posterior state,
(B, T).valueTensorThe critic's estimate at each posterior state,
(B, T).posterior_stoch, posterior_mean, posterior_stdTensorThe filtered latent and the Gaussian it came from,
(B, T, S).prior_mean, prior_stdTensorThe Gaussian the dynamics predicted before seeing the frame,
(B, T, S).deterTensorThe deterministic path,
(B, T, D).loss, recon_loss, reward_loss, kl_loss, pcont_lossTensor or NoneWorld-model terms, set only by
DreamerForWorldModeling. loss is the world-model
loss alone — the behaviour losses live on
DreamerBehaviorOutput, because they belong to different
optimisers. pcont_loss is None unless the config asks for
a discount head.behaviorDreamerBehaviorOutput or NoneActor and critic terms, set only by
DreamerForWorldModeling.Examples
>>> import lucid
>>> from lucid.models.generative.dreamer import DreamerConfig, DreamerModel
>>> cfg = DreamerConfig(action_dim=2, cnn_depth=2, stoch_size=4,
... deter_size=8, hidden_size=8, actor_hidden=8,
... value_hidden=8, reward_hidden=8)
>>> model = DreamerModel(cfg)
>>> out = model(lucid.randn((1, 3, 3, 64, 64)), lucid.randn((1, 3, 2)))
>>> out.observation.shape, out.value.shape
((1, 3, 3, 64, 64), (1, 3))Used by 1
Constructors
1dunder
__init__
→None__init__(observation: Tensor, reward: Tensor, value: Tensor, posterior_stoch: Tensor, posterior_mean: Tensor, posterior_std: Tensor, prior_mean: Tensor, prior_std: Tensor, deter: Tensor, loss: Tensor | None = None, recon_loss: Tensor | None = None, reward_loss: Tensor | None = None, kl_loss: Tensor | None = None, pcont_loss: Tensor | None = None, behavior: DreamerBehaviorOutput | None = None)Initialise the actor. See the class docstring for parameters.