planet_world_model

→PlaNetForWorldModeling

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

source edit

Construct PlaNet with its variational objective.

Same trunk as planet, wrapped with the bound of Hafner et al., 2019 — reconstruction and reward likelihoods against a KL between the posterior and the dynamics' own prior, floored at free_nats.

Model Size

Parameters

pretrainedbool= False

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

**overridesobject= {}

Optional PlaNetConfig field overrides. free_nats and kl_weight shape the divergence term; action_dim is set by the environment.

Returns

PlaNetForWorldModeling

The trunk plus the objective.

Notes

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

Training objective:

\mathcal{L} = \mathbb{E}_q\big[\log p(o_t \mid h_t, s_t) + \log p(r_t \mid h_t, s_t)\big] - \beta\,\mathrm{KL}\big(q(s_t \mid h_t, o_t) \,\|\, p(s_t \mid h_t)\big).

Examples

>>> from lucid.models.generative.planet import planet_world_model
>>> model = planet_world_model(action_dim=6).eval()
>>> model.config.free_nats
3.0

Used by 1

lucid.models.generative.planet