metacontroller-pytorch 0.0.36__py3-none-any.whl → 0.0.38__py3-none-any.whl

metacontroller/metacontroller.py

@@ -126,6 +126,7 @@ class MetaController(Module):
         )
     ):
         super().__init__()
+        self.dim_model = dim_model
         dim_meta = default(dim_meta_controller, dim_model)
 
         # the linear that brings from model dimension 
@@ -171,6 +172,15 @@ class MetaController(Module):
 
         self.register_buffer('zero', tensor(0.), persistent = False)
 
+    @property
+    def replay_buffer_field_dict(self):
+        return dict(
+            states = ('float', self.dim_model),
+            log_probs = ('float', self.dim_latent),
+            switch_betas = ('float', self.dim_latent if self.switch_per_latent_dim else 1),
+            latent_actions = ('float', self.dim_latent)
+        )
+
     def discovery_parameters(self):
         return [
             *self.model_to_meta.parameters(),
@@ -408,6 +418,7 @@ class Transformer(Module):
         meta_controller: Module | None = None,
         cache: TransformerOutput | None = None,
         discovery_phase = False,
+        force_behavior_cloning = False,
         meta_controller_temperature = 1.,
         return_raw_action_dist = False,
         return_latents = False,
@@ -420,11 +431,15 @@ class Transformer(Module):
 
         meta_controller = default(meta_controller, self.meta_controller)
 
+        if force_behavior_cloning:
+            assert not discovery_phase, 'discovery phase cannot be set to True if force behavioral cloning is set to True'
+            meta_controller = None
+
         has_meta_controller = exists(meta_controller)
 
         assert not (discovery_phase and not has_meta_controller), 'meta controller must be made available during discovery phase'
 
-        behavioral_cloning = not has_meta_controller and not return_raw_action_dist
+        behavioral_cloning = force_behavior_cloning or (not has_meta_controller and not return_raw_action_dist)
 
         # by default, if meta controller is passed in, transformer is no grad
 
@@ -472,7 +487,7 @@ class Transformer(Module):
 
         with meta_controller_context():
 
-            if exists(meta_controller):
+            if exists(meta_controller) and not behavioral_cloning:
                 control_signal, next_meta_hiddens = meta_controller(residual_stream, cache = meta_hiddens, discovery_phase = discovery_phase, temperature = meta_controller_temperature, episode_lens = episode_lens)
             else:
                 control_signal, next_meta_hiddens = self.zero, None

metacontroller/metacontroller_with_binary_mapper.py

@@ -74,7 +74,7 @@ class MetaControllerWithBinaryMapper(Module):
         kl_loss_threshold = 0.
     ):
         super().__init__()
-
+        self.dim_model = dim_model
         assert not switch_per_code, 'switch_per_code is not supported for binary mapper'
 
         dim_meta = default(dim_meta_controller, dim_model)
@@ -126,6 +126,15 @@ class MetaControllerWithBinaryMapper(Module):
 
         self.register_buffer('zero', tensor(0.), persistent = False)
 
+    @property
+    def replay_buffer_field_dict(self):
+        return dict(
+            states = ('float', self.dim_model),
+            log_probs = ('float', self.dim_code_bits),
+            switch_betas = ('float', self.num_codes if self.switch_per_code else 1),
+            latent_actions = ('float', self.num_codes)
+        )
+
     def discovery_parameters(self):
         return [
             *self.model_to_meta.parameters(),

{metacontroller_pytorch-0.0.36.dist-info → metacontroller_pytorch-0.0.38.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: metacontroller-pytorch
-Version: 0.0.36
+Version: 0.0.38
 Summary: Transformer Metacontroller
 Project-URL: Homepage, https://pypi.org/project/metacontroller/
 Project-URL: Repository, https://github.com/lucidrains/metacontroller
@@ -94,6 +94,18 @@ $ pip install metacontroller-pytorch
 }
 ```
 
+```bibtex
+@misc{hwang2025dynamicchunkingendtoendhierarchical,
+    title   = {Dynamic Chunking for End-to-End Hierarchical Sequence Modeling},
+    author  = {Sukjun Hwang and Brandon Wang and Albert Gu},
+    year    = {2025},
+    eprint  = {2507.07955},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.LG},
+    url     = {https://arxiv.org/abs/2507.07955},
+}
+```
+
 ```bibtex
 @misc{fleuret2025freetransformer,
     title     = {The Free Transformer}, 

metacontroller_pytorch-0.0.38.dist-info/RECORD

@@ -0,0 +1,8 @@
+metacontroller/__init__.py,sha256=lj7IOGpN--qMxJWbB-4SGqoPXG7Hd4mgtToTRSyTZ58,57
+metacontroller/metacontroller.py,sha256=bhgCqqM-dfysGrMtZYe2w87lRVkf8fETjxUCdjrnI8Q,17386
+metacontroller/metacontroller_with_binary_mapper.py,sha256=Ce5-O95_pLuWNA3aZTlKrTGbc5cemb61tBtJBdSiLx4,9843
+metacontroller/transformer_with_resnet.py,sha256=R49ycusbq3kEX97WHZ41WY2ONc2mYPOuRUCmaFcBOEo,5546
+metacontroller_pytorch-0.0.38.dist-info/METADATA,sha256=xFax5wjB3faXWgHNeHoRG1uOxl2S3-xeQuqJvrY3YTo,5114
+metacontroller_pytorch-0.0.38.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
+metacontroller_pytorch-0.0.38.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+metacontroller_pytorch-0.0.38.dist-info/RECORD,,

metacontroller_pytorch-0.0.36.dist-info/RECORD

@@ -1,8 +0,0 @@
-metacontroller/__init__.py,sha256=lj7IOGpN--qMxJWbB-4SGqoPXG7Hd4mgtToTRSyTZ58,57
-metacontroller/metacontroller.py,sha256=rVYzBJ8jQx9tfkZ3B9NdxKTI7dyBxXtTl4kwfizYuis,16728
-metacontroller/metacontroller_with_binary_mapper.py,sha256=odZs49ZWY7_FfEweYkD0moX7Vn0jGd91FjFTxzjLyr8,9480
-metacontroller/transformer_with_resnet.py,sha256=R49ycusbq3kEX97WHZ41WY2ONc2mYPOuRUCmaFcBOEo,5546
-metacontroller_pytorch-0.0.36.dist-info/METADATA,sha256=eLKG8B0gSZyIMkaLvjYE8SvWVN387BuuNFoOC_6lmT4,4747
-metacontroller_pytorch-0.0.36.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
-metacontroller_pytorch-0.0.36.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
-metacontroller_pytorch-0.0.36.dist-info/RECORD,,