data

PlaNetConfig

extendsWorldModelConfig

PlaNetConfig(sample_size: int | tuple[int, int] = WORLD_MODEL_IMAGE_SIZE, in_channels: int = 3, out_channels: int = 3, act_fn: GenerativeActivation = 'relu', action_dim: int = 1, stoch_size: int = 30, deter_size: int = 200, hidden_size: int = 200, cnn_depth: int = 32, min_std: float = 0.1, mean_only: bool = False, free_nats: float = 3.0, kl_weight: float = 1.0, overshoot_distance: int | None = None, overshoot_weight: float = 1.0, overshoot_reward_weight: float = 1.0, reward_hidden: int = 200, reward_layers: int = 2, reward_loss_scale: float = 1.0)

source edit

Frozen configuration for the PlaNet family.

Defaults reproduce Hafner et al., 2019 on the DeepMind Control Suite.

Parameters

act_fn(silu, swish, relu, gelu, elu)= "silu"

Activation throughout. The paper uses ReLU, which is why this family overrides the generative-domain default of "silu". Dreamer uses ELU, which is why neither value sits on WorldModelConfig.

overshoot_distanceint or None= None

How many steps ahead latent overshooting trains the dynamics. None overshoots as far as each batch allows (the paper's full setting); 1 disables it, leaving the ordinary one-step bound. Costs one extra recurrence sweep per distance — see Notes.

overshoot_weightfloat= 1.0

Multiplier on the averaged multi-step divergence.

overshoot_reward_weightfloat= 1.0

Multiplier on reward prediction at the overshot states. Without it the reward head only ever sees posterior states while the planner only ever evaluates prior ones — see Notes.

reward_hiddenint= 200

Width of each reward-head hidden layer. Dreamer's is 300, so this stays here rather than on the shared base.

reward_layersint= 2

Number of reward-head hidden layers.

reward_loss_scalefloat= 1.0

Multiplier on the reward likelihood. The paper states none, so the default is 1; the released implementation uses 10 — see Notes.

Notes

The state geometry, the frame size, the action width and the two KL knobs are inherited from WorldModelConfig, which is where Dreamer reads the same values from.

Reference: Hafner, Lillicrap, Fischer, Villegas, Ha, Lee, and Davidson, "Learning Latent Dynamics for Planning from Pixels", ICML, 2019 (arXiv:1811.04551).

sample_size is pinned to 64 rather than made general. The encoder is four stride-2 convolutions with 4x4 kernels and the decoder four transposed convolutions with kernels 5, 5, 6, 6 — an irregular schedule that lands on 64 exactly (1 -> 5 -> 13 -> 30 -> 64) and on nothing else. Generalising it would mean inventing a schedule the paper does not give, so an unsupported size raises here instead of silently reconstructing to the wrong shape. Scale the model with cnn_depth / deter_size / stoch_size.

Examples

>>> from lucid.models.generative.planet import PlaNetConfig
>>> cfg = PlaNetConfig(action_dim=6)
>>> cfg.stoch_size, cfg.deter_size, cfg.free_nats
(30, 200, 3.0)
>>> cfg.latent_size
230

Used by 3

Constructors

dunder

init

→None

__init__(sample_size: int | tuple[int, int] = WORLD_MODEL_IMAGE_SIZE, in_channels: int = 3, out_channels: int = 3, act_fn: GenerativeActivation = 'relu', action_dim: int = 1, stoch_size: int = 30, deter_size: int = 200, hidden_size: int = 200, cnn_depth: int = 32, min_std: float = 0.1, mean_only: bool = False, free_nats: float = 3.0, kl_weight: float = 1.0, overshoot_distance: int | None = None, overshoot_weight: float = 1.0, overshoot_reward_weight: float = 1.0, reward_hidden: int = 200, reward_layers: int = 2, reward_loss_scale: float = 1.0)

source edit

PlaNetConfig(sample_size: int | tuple[int, int] = WORLD_MODEL_IMAGE_SIZE, in_channels: int = 3, out_channels: int = 3, act_fn: GenerativeActivation = 'relu', action_dim: int = 1, stoch_size: int = 30, deter_size: int = 200, hidden_size: int = 200, cnn_depth: int = 32, min_std: float = 0.1, mean_only: bool = False, free_nats: float = 3.0, kl_weight: float = 1.0, overshoot_distance: int | None = None, overshoot_weight: float = 1.0, overshoot_reward_weight: float = 1.0, reward_hidden: int = 200, reward_layers: int = 2, reward_loss_scale: float = 1.0)

__init__(sample_size: int | tuple[int, int] = WORLD_MODEL_IMAGE_SIZE, in_channels: int = 3, out_channels: int = 3, act_fn: GenerativeActivation = 'relu', action_dim: int = 1, stoch_size: int = 30, deter_size: int = 200, hidden_size: int = 200, cnn_depth: int = 32, min_std: float = 0.1, mean_only: bool = False, free_nats: float = 3.0, kl_weight: float = 1.0, overshoot_distance: int | None = None, overshoot_weight: float = 1.0, overshoot_reward_weight: float = 1.0, reward_hidden: int = 200, reward_layers: int = 2, reward_loss_scale: float = 1.0)