dreamer4 0.0.74__tar.gz → 0.0.76__tar.gz

{dreamer4-0.0.74 → dreamer4-0.0.76}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: dreamer4
-Version: 0.0.74
+Version: 0.0.76
 Summary: Dreamer 4
 Project-URL: Homepage, https://pypi.org/project/dreamer4/
 Project-URL: Repository, https://github.com/lucidrains/dreamer4

{dreamer4-0.0.74 → dreamer4-0.0.76}/dreamer4/dreamer4.py

@@ -284,7 +284,7 @@ def create_multi_token_prediction_targets(
     batch, seq_len, device = *t.shape[:2], t.device
 
     batch_arange = arange(batch, device = device)
-    seq_arange = arange(seq_len, device = device)[1:]
+    seq_arange = arange(seq_len, device = device)
     steps_arange = arange(steps_future, device = device)
 
     indices = add('t, steps -> t steps', seq_arange, steps_arange)
@@ -2870,6 +2870,13 @@ class DynamicsWorldModel(Module):
 
             space_tokens = self.latents_to_spatial_tokens(noised_latents)
 
+            # maybe add view embedding
+
+            if self.video_has_multi_view:
+                space_tokens = add('b t v ... d, v d', space_tokens, self.view_emb)
+
+            # merge spatial tokens
+
             space_tokens, inverse_pack_space_per_latent = pack_one(space_tokens, 'b t * d')
 
             num_spatial_tokens = space_tokens.shape[-2]
@@ -3093,7 +3100,7 @@ class DynamicsWorldModel(Module):
 
             reward_pred = rearrange(reward_pred, 'mtp b t l -> b l t mtp')
 
-            reward_targets, reward_loss_mask = create_multi_token_prediction_targets(two_hot_encoding, self.multi_token_pred_len)
+            reward_targets, reward_loss_mask = create_multi_token_prediction_targets(two_hot_encoding[:, :-1], self.multi_token_pred_len)
 
             reward_targets = rearrange(reward_targets, 'b t mtp l -> b l t mtp')
 
@@ -3119,6 +3126,15 @@ class DynamicsWorldModel(Module):
         ):
             assert self.action_embedder.has_actions
 
+            # handle actions having time vs time - 1 length
+            # remove the first action if it is equal to time (as it would come from some agent token in the past)
+
+            if exists(discrete_actions) and discrete_actions.shape[1] == time:
+                discrete_actions = discrete_actions[:, 1:]
+
+            if exists(continuous_actions) and continuous_actions.shape[1] == time:
+                continuous_actions = continuous_actions[:, 1:]
+
             # only for 1 agent
 
             agent_tokens = rearrange(agent_tokens, 'b t 1 d -> b t d')

{dreamer4-0.0.74 → dreamer4-0.0.76}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "dreamer4"
-version = "0.0.74"
+version = "0.0.76"
 description = "Dreamer 4"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }