take care of storing the old action logits and mean log var, and calculate kl div for pmpo based off that during learn from experience

dreamer4/dreamer4.py

@@ -81,6 +81,7 @@ class Experience:
     rewards: Tensor | None = None
     actions: tuple[Tensor, Tensor] | None = None
     log_probs: tuple[Tensor, Tensor] | None = None
+    old_action_unembeds: tuple[Tensor, Tensor] | None = None
     values: Tensor | None = None
     step_size: int | None = None
     lens: Tensor | None = None
@@ -1887,6 +1888,7 @@ class DynamicsWorldModel(Module):
         gae_lambda = 0.95,
         ppo_eps_clip = 0.2,
         pmpo_pos_to_neg_weight = 0.5, # pos and neg equal weight
+        pmpo_kl_div_loss_weight = 1.,
         value_clip = 0.4,
         policy_entropy_weight = .01,
         gae_use_accelerated = False
@@ -2093,6 +2095,7 @@ class DynamicsWorldModel(Module):
         # pmpo related
 
         self.pmpo_pos_to_neg_weight = pmpo_pos_to_neg_weight
+        self.pmpo_kl_div_loss_weight = pmpo_kl_div_loss_weight
 
         # rewards related
 
@@ -2222,7 +2225,8 @@ class DynamicsWorldModel(Module):
         max_timesteps = 16,
         env_is_vectorized = False,
         use_time_kv_cache = True,
-        store_agent_embed = False
+        store_agent_embed = False,
+        store_old_action_unembeds = False,
     ):
         assert exists(self.video_tokenizer)
 
@@ -2248,6 +2252,7 @@ class DynamicsWorldModel(Module):
         latents = None
 
         acc_agent_embed = None
+        acc_policy_embed = None
 
         # keep track of termination, for setting the `is_truncated` field on Experience and for early stopping interaction with env
 
@@ -2300,6 +2305,9 @@ class DynamicsWorldModel(Module):
 
             policy_embed = self.policy_head(one_agent_embed)
 
+            if store_old_action_unembeds:
+                acc_policy_embed = safe_cat((acc_policy_embed, policy_embed), dim = 1)
+
             # sample actions
 
             sampled_discrete_actions, sampled_continuous_actions = self.action_embedder.sample(policy_embed, pred_head_index = 0, squeeze = True)
@@ -2383,6 +2391,7 @@ class DynamicsWorldModel(Module):
             actions = (discrete_actions, continuous_actions),
             log_probs = (discrete_log_probs, continuous_log_probs),
             values = values,
+            old_action_unembeds = self.action_embedder.unembed(acc_policy_embed, pred_head_index = 0) if store_old_action_unembeds else None,
             agent_embed = acc_agent_embed if store_agent_embed else None,
             step_size = step_size,
             agent_index = agent_index,
@@ -2411,6 +2420,7 @@ class DynamicsWorldModel(Module):
         old_values = experience.values
         rewards = experience.rewards
         agent_embeds = experience.agent_embed
+        old_action_unembeds = experience.old_action_unembeds
 
         step_size = experience.step_size
         agent_index = experience.agent_index
@@ -2489,6 +2499,7 @@ class DynamicsWorldModel(Module):
         if use_pmpo:
             pos_advantage_mask = advantage >= 0.
             neg_advantage_mask = ~pos_advantage_mask
+
         else:
             advantage = F.layer_norm(advantage, advantage.shape, eps = eps)
 
@@ -2552,6 +2563,25 @@ class DynamicsWorldModel(Module):
 
             policy_loss = -(α * pos + (1. - α) * neg)
 
+            # take care of kl
+
+            if self.pmpo_kl_div_loss_weight > 0.:
+                new_unembedded_actions = self.action_embedder.unembed(policy_embed, pred_head_index = 0)
+
+                discrete_kl_div, continuous_kl_div = self.action_embedder.kl_div(new_unembedded_actions, old_action_unembeds)
+
+                # accumulate discrete and continuous kl div
+
+                kl_div_loss = 0.
+
+                if exists(discrete_kl_div):
+                    kl_div_loss = kl_div_loss + discrete_kl_div[mask].mean()
+
+                if exists(continuous_kl_div):
+                    kl_div_loss = kl_div_loss + continuous_kl_div[mask].mean()
+
+                policy_loss = policy_loss + kl_div_loss * self.pmpo_kl_div_loss_weight
+
         else:
             # ppo clipped surrogate loss
 
@@ -2694,6 +2724,10 @@ class DynamicsWorldModel(Module):
 
         acc_agent_embed = None
 
+        # maybe store old actions for kl
+
+        acc_policy_embed = None
+
         # maybe return rewards
 
         decoded_rewards = None
@@ -2818,6 +2852,13 @@ class DynamicsWorldModel(Module):
 
                 policy_embed = self.policy_head(one_agent_embed)
 
+                # maybe store old actions
+
+                if store_old_action_unembeds:
+                    acc_policy_embed = safe_cat((acc_policy_embed, policy_embed))
+
+                # sample actions
+
                 sampled_discrete_actions, sampled_continuous_actions = self.action_embedder.sample(policy_embed, pred_head_index = 0, squeeze = True)
 
                 decoded_discrete_actions = safe_cat((decoded_discrete_actions, sampled_discrete_actions), dim = 1)
@@ -2906,6 +2947,7 @@ class DynamicsWorldModel(Module):
             video = video,
             proprio = proprio if has_proprio else None,
             agent_embed = acc_agent_embed if store_agent_embed else None,
+            old_action_unembeds = self.action_embedder.unembed(acc_policy_embed, pred_head_index = 0) if store_old_action_unembeds else None,
             step_size = step_size,
             agent_index = agent_index,
             lens = experience_lens,

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "dreamer4"
-version = "0.0.92"
+version = "0.0.93"
 description = "Dreamer 4"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }