hongshaorou/dreamer4
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

use the agent embeds off the stored experience if available

Browse Source
This commit is contained in:
lucidrains 2025-10-28 09:14:02 -07:00
parent d476fa7b14
commit 903c43b770
2 changed files with 25 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

41
dreamer4/dreamer4.py
View File

@ -77,7 +77,7 @@ class Experience:
latents: Tensor
video: Tensor | None = None
proprio: Tensor | None = None
agent_embed: Tensor | None = None,
agent_embed: Tensor | None = None
rewards: Tensor | None = None
actions: tuple[Tensor, Tensor] | None = None
log_probs: tuple[Tensor, Tensor] | None = None
@ -2295,6 +2295,7 @@ class DynamicsWorldModel(Module):
old_log_probs = experience.log_probs
old_values = experience.values
rewards = experience.rewards
agent_embeds = experience.agent_embed
step_size = experience.step_size
agent_index = experience.agent_index
@ -2374,32 +2375,38 @@ class DynamicsWorldModel(Module):
advantage = F.layer_norm(advantage, advantage.shape, eps = eps)
# replay for the action logits and values
# but only do so if fine tuning the entire world model for RL
discrete_actions, continuous_actions = actions
with world_model_forward_context():
_, (agent_embed, _) = self.forward(
latents = latents,
signal_levels = self.max_steps - 1,
step_sizes = step_size,
rewards = rewards,
discrete_actions = discrete_actions,
continuous_actions = continuous_actions,
latent_is_noised = True,
return_pred_only = True,
return_intermediates = True
)
if (
not only_learn_policy_value_heads or
not exists(agent_embeds)
):
agent_embed = agent_embed[..., agent_index, :]
with world_model_forward_context():
_, (agent_embeds, _) = self.forward(
latents = latents,
signal_levels = self.max_steps - 1,
step_sizes = step_size,
rewards = rewards,
discrete_actions = discrete_actions,
continuous_actions = continuous_actions,
latent_is_noised = True,
return_pred_only = True,
return_intermediates = True
)
agent_embeds = agent_embeds[..., agent_index, :]
# maybe detach agent embed
if only_learn_policy_value_heads:
agent_embed = agent_embed.detach()
agent_embeds = agent_embeds.detach()
# ppo
policy_embed = self.policy_head(agent_embed)
policy_embed = self.policy_head(agent_embeds)
log_probs, entropies = self.action_embedder.log_probs(policy_embed, pred_head_index = 0, discrete_targets = discrete_actions, continuous_targets = continuous_actions, return_entropies = True)
@ -2448,7 +2455,7 @@ class DynamicsWorldModel(Module):
# value loss
value_bins = self.value_head(agent_embed)
value_bins = self.value_head(agent_embeds)
values = self.reward_encoder.bins_to_scalar_value(value_bins)
clipped_values = old_values + (values - old_values).clamp(-self.value_clip, self.value_clip)

2
pyproject.toml
View File

@ -1,6 +1,6 @@
[project]
name = "dreamer4"
version = "0.0.85"
version = "0.0.87"
description = "Dreamer 4"
authors = [
{ name = "Phil Wang", email = "lucidrains@gmail.com" }