dreamer4 0.1.2__tar.gz → 0.1.5__tar.gz

{dreamer4-0.1.2 → dreamer4-0.1.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: dreamer4
-Version: 0.1.2
+Version: 0.1.5
 Summary: Dreamer 4
 Project-URL: Homepage, https://pypi.org/project/dreamer4/
 Project-URL: Repository, https://github.com/lucidrains/dreamer4
@@ -57,6 +57,12 @@ Description-Content-Type: text/markdown
 
 Implementation of Danijar's [latest iteration](https://arxiv.org/abs/2509.24527v1) for his [Dreamer](https://danijar.com/project/dreamer4/) line of work
 
+[Discord channel](https://discord.gg/ab4BEk3W) for collaborating with other researchers interested in this work
+
+## Appreciation
+
+- [@dirkmcpherson](https://github.com/dirkmcpherson) for fixes to typo errors and unpassed arguments!
+
 ## Install
 
 ```bash
@@ -79,9 +85,16 @@ tokenizer = VideoTokenizer(
     image_width = 256
 )
 
+video = torch.randn(2, 3, 10, 256, 256)
+
+# learn the tokenizer
+
+loss = tokenizer(video)
+loss.backward() # ler
+
 # dynamics world model
 
-dynamics = DynamicsWorldModel(
+world_model = DynamicsWorldModel(
     dim = 512,
     dim_latent = 32,
     video_tokenizer = tokenizer,
@@ -97,7 +110,7 @@ rewards = torch.randn(2, 10)
 
 # learn dynamics / behavior cloned model
 
-loss = dynamics(
+loss = world_model(
     video = video,
     rewards = rewards,
     discrete_actions = discrete_actions
@@ -109,7 +122,7 @@ loss.backward()
 
 # then generate dreams
 
-dreams = dynamics.generate(
+dreams = world_model.generate(
     10,
     batch_size = 2,
     return_decoded_video = True,
@@ -118,7 +131,19 @@ dreams = dynamics.generate(
 
 # learn from the dreams
 
-actor_loss, critic_loss = dynamics.learn_from_experience(dreams)
+actor_loss, critic_loss = world_model.learn_from_experience(dreams)
+
+(actor_loss + critic_loss).backward()
+
+# learn from environment
+
+from dreamer4.mocks import MockEnv
+
+mock_env = MockEnv((256, 256), vectorized = True, num_envs = 4)
+
+experience = world_model.interact_with_env(mock_env, max_timesteps = 8, env_is_vectorized = True)
+
+actor_loss, critic_loss = world_model.learn_from_experience(experience)
 
 (actor_loss + critic_loss).backward()
 ```

{dreamer4-0.1.2 → dreamer4-0.1.5}/README.md

@@ -4,6 +4,12 @@
 
 Implementation of Danijar's [latest iteration](https://arxiv.org/abs/2509.24527v1) for his [Dreamer](https://danijar.com/project/dreamer4/) line of work
 
+[Discord channel](https://discord.gg/ab4BEk3W) for collaborating with other researchers interested in this work
+
+## Appreciation
+
+- [@dirkmcpherson](https://github.com/dirkmcpherson) for fixes to typo errors and unpassed arguments!
+
 ## Install
 
 ```bash
@@ -26,9 +32,16 @@ tokenizer = VideoTokenizer(
     image_width = 256
 )
 
+video = torch.randn(2, 3, 10, 256, 256)
+
+# learn the tokenizer
+
+loss = tokenizer(video)
+loss.backward() # ler
+
 # dynamics world model
 
-dynamics = DynamicsWorldModel(
+world_model = DynamicsWorldModel(
     dim = 512,
     dim_latent = 32,
     video_tokenizer = tokenizer,
@@ -44,7 +57,7 @@ rewards = torch.randn(2, 10)
 
 # learn dynamics / behavior cloned model
 
-loss = dynamics(
+loss = world_model(
     video = video,
     rewards = rewards,
     discrete_actions = discrete_actions
@@ -56,7 +69,7 @@ loss.backward()
 
 # then generate dreams
 
-dreams = dynamics.generate(
+dreams = world_model.generate(
     10,
     batch_size = 2,
     return_decoded_video = True,
@@ -65,7 +78,19 @@ dreams = dynamics.generate(
 
 # learn from the dreams
 
-actor_loss, critic_loss = dynamics.learn_from_experience(dreams)
+actor_loss, critic_loss = world_model.learn_from_experience(dreams)
+
+(actor_loss + critic_loss).backward()
+
+# learn from environment
+
+from dreamer4.mocks import MockEnv
+
+mock_env = MockEnv((256, 256), vectorized = True, num_envs = 4)
+
+experience = world_model.interact_with_env(mock_env, max_timesteps = 8, env_is_vectorized = True)
+
+actor_loss, critic_loss = world_model.learn_from_experience(experience)
 
 (actor_loss + critic_loss).backward()
 ```

{dreamer4-0.1.2 → dreamer4-0.1.5}/dreamer4/dreamer4.py

@@ -14,7 +14,7 @@ from torch.nested import nested_tensor
 from torch.distributions import Normal, kl
 from torch.nn import Module, ModuleList, Embedding, Parameter, Sequential, Linear, RMSNorm, Identity
 from torch import nn, cat, stack, arange, tensor, Tensor, is_tensor, full, zeros, ones, randint, rand, randn, randn_like, empty, full, linspace, arange
-from torch.utils._pytree import tree_flatten, tree_unflatten
+from torch.utils._pytree import tree_map, tree_flatten, tree_unflatten
 
 import torchvision
 from torchvision.models import VGG16_Weights
@@ -91,6 +91,14 @@ class Experience:
     agent_index: int = 0
     is_from_world_model: bool = True
 
+    def cpu(self):
+        return self.to(torch.device('cpu'))
+
+    def to(self, device):
+        experience_dict = asdict(self)
+        experience_dict = tree_map(lambda t: t.to(device) if is_tensor(t) else t, experience_dict)
+        return Experience(**experience_dict)
+
 def combine_experiences(
     exps: list[Experiences]
 ) -> Experience:
@@ -1179,10 +1187,11 @@ def special_token_mask(q, k, seq_len, num_tokens, special_attend_only_itself = F
 
 def block_mask_special_tokens_right(
     seq_len,
-    num_tokens
+    num_tokens,
+    special_attend_only_itself = False
 ):
     def inner(b, h, q, k):
-        return special_token_mask(q, k, seq_len, num_tokens)
+        return special_token_mask(q, k, seq_len, num_tokens, special_attend_only_itself)
     return inner
 
 def compose_mask(mask1, mask2):
@@ -1493,7 +1502,8 @@ class AxialSpaceTimeTransformer(Module):
 
         # attend functions for space and time
 
-        use_flex = exists(flex_attention) and tokens.is_cuda
+        has_kv_cache = exists(kv_cache) 
+        use_flex = exists(flex_attention) and tokens.is_cuda and not has_kv_cache # KV cache shape breaks flex attention TODO: Fix
 
         attend_kwargs = dict(use_flex = use_flex, softclamp_value = self.attn_softclamp_value, special_attend_only_itself = self.special_attend_only_itself, device = device)
 
@@ -1505,7 +1515,6 @@ class AxialSpaceTimeTransformer(Module):
 
         time_attn_kv_caches = []
 
-        has_kv_cache = exists(kv_cache)
 
         if has_kv_cache:
             past_tokens, tokens = tokens[:, :-1], tokens[:, -1:]
@@ -1847,7 +1856,7 @@ class VideoTokenizer(Module):
 
         losses = (recon_loss, lpips_loss)
 
-        return total_loss, TokenizerLosses(losses)
+        return total_loss, TokenizerLosses(*losses)
 
 # dynamics model, axial space-time transformer
 
@@ -2104,7 +2113,7 @@ class DynamicsWorldModel(Module):
 
         self.ppo_eps_clip = ppo_eps_clip
         self.value_clip = value_clip
-        self.policy_entropy_weight = value_clip
+        self.policy_entropy_weight = policy_entropy_weight
 
         # pmpo related
 
@@ -2127,7 +2136,7 @@ class DynamicsWorldModel(Module):
         self.flow_loss_normalizer = LossNormalizer(1)
         self.reward_loss_normalizer = LossNormalizer(multi_token_pred_len)
         self.discrete_actions_loss_normalizer = LossNormalizer(multi_token_pred_len) if num_discrete_actions > 0 else None
-        self.continuous_actions_loss_normalizer = LossNormalizer(multi_token_pred_len) if num_discrete_actions > 0 else None
+        self.continuous_actions_loss_normalizer = LossNormalizer(multi_token_pred_len) if num_continuous_actions > 0 else None
 
         self.latent_flow_loss_weight = latent_flow_loss_weight
 
@@ -2358,6 +2367,9 @@ class DynamicsWorldModel(Module):
             elif len(env_step_out) == 4:
                 next_frame, reward, terminated, truncated = env_step_out
 
+            elif len(env_step_out) == 5:
+                next_frame, reward, terminated, truncated, info = env_step_out
+
             # update episode lens
 
             episode_lens = torch.where(is_terminated, episode_lens, episode_lens + 1)
@@ -2431,6 +2443,8 @@ class DynamicsWorldModel(Module):
     ):
         assert isinstance(experience, Experience)
 
+        experience = experience.to(self.device)
+
         latents = experience.latents
         actions = experience.actions
         old_log_probs = experience.log_probs
@@ -3085,8 +3099,8 @@ class DynamicsWorldModel(Module):
         if latents.ndim == 4:
             latents = rearrange(latents, 'b t v d -> b t v 1 d') # 1 latent edge case
 
-        assert latents.shape[-2:] == self.latent_shape
-        assert latents.shape[2] == self.num_video_views
+        assert latents.shape[-2:] == self.latent_shape, f'latents must have shape {self.latent_shape}, got {latents.shape[-2:]}'
+        assert latents.shape[2] == self.num_video_views, f'latents must have {self.num_video_views} views, got {latents.shape[2]}'
 
         # variables
 

{dreamer4-0.1.2 → dreamer4-0.1.5}/dreamer4/trainers.py

@@ -528,7 +528,7 @@ class SimTrainer(Module):
 
                 total_experience += num_experience
 
-                experiences.append(experience)
+                experiences.append(experience.cpu())
 
             combined_experiences = combine_experiences(experiences)
 

{dreamer4-0.1.2 → dreamer4-0.1.5}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "dreamer4"
-version = "0.1.2"
+version = "0.1.5"
 description = "Dreamer 4"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{dreamer4-0.1.2 → dreamer4-0.1.5}/tests/test_dreamer.py

@@ -680,6 +680,12 @@ def test_online_rl(
 
     combined_experience = combine_experiences([one_experience, another_experience])
 
+    # quick test moving the experience to different devices
+
+    if torch.cuda.is_available():
+        combined_experience = combined_experience.to(torch.device('cuda'))
+        combined_experience = combined_experience.to(world_model_and_policy.device)
+
     if store_agent_embed:
         assert exists(combined_experience.agent_embed)