protomotions.agents.common.latent module
Shared latent-output keys and latent-specific loss helpers.
-
protomotions.agents.common.latent.compute_discrete_latent_ppo_loss(
- *,
- logits,
- selected,
- old_neglogp,
- advantages,
- e_clip,
- entropy_coef=0.0,
- temperature=1.0,
- top_p=1.0,
- prior_logits=None,
- log_prefix='actor',
)[source]
Compute PPO loss under the same distribution used during token rollout.
