metacontroller-pytorch 0.0.41__py3-none-any.whl → 0.0.43__py3-none-any.whl

metacontroller/metacontroller.py

@@ -66,6 +66,13 @@ MetaControllerOutput = namedtuple('MetaControllerOutput', (
     'switch_loss'
 ))
 
+GRPOOutput = namedtuple('GRPOOutput', (
+    'state',
+    'action',
+    'log_prob',
+    'switch_beta'
+))
+
 def z_score(t, eps = 1e-8):
     return (t - t.mean()) / (t.std() + eps)
 
@@ -107,6 +114,17 @@ def policy_loss(
 
     return masked_mean(losses, mask)
 
+def extract_grpo_data(meta_controller, transformer_output):
+    meta_output = transformer_output.prev_hiddens.meta_controller
+
+    state = meta_output.input_residual_stream
+    action = meta_output.actions
+    switch_beta = meta_output.switch_beta
+
+    log_prob = meta_controller.log_prob(meta_output.action_dist, action)
+
+    return GRPOOutput(state, action, log_prob, switch_beta)
+
 @save_load()
 class MetaController(Module):
     def __init__(
@@ -273,7 +291,7 @@ class MetaController(Module):
         else:
             # else during inference, use the previous sampled latent action
 
-            assert seq_len == 1, f'inference RL phase must be done one token at a time'
+            assert seq_len == 1, 'inference RL phase must be done one token at a time - if replaying for policy optimization, please use `get_action_dist_for_internal_rl`'
             z_prev = prev_sampled_latent_action
 
         # switch input is previous latent action and the embedding

metacontroller/metacontroller_with_binary_mapper.py

@@ -241,7 +241,7 @@ class MetaControllerWithBinaryMapper(Module):
         if discovery_phase:
             z_prev = cat((prev_sampled_code, sampled_codes[:, :-1]), dim = 1)
         else:
-            assert seq_len == 1, f'inference RL phase must be done one token at a time'
+            assert seq_len == 1, 'inference RL phase must be done one token at a time - if replaying for policy optimization, please use `get_action_dist_for_internal_rl`'
             z_prev = prev_sampled_code
 
         switch_input = torch.cat((meta_embed, z_prev), dim=-1)

{metacontroller_pytorch-0.0.41.dist-info → metacontroller_pytorch-0.0.43.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: metacontroller-pytorch
-Version: 0.0.41
+Version: 0.0.43
 Summary: Transformer Metacontroller
 Project-URL: Homepage, https://pypi.org/project/metacontroller/
 Project-URL: Repository, https://github.com/lucidrains/metacontroller
@@ -53,7 +53,7 @@ Description-Content-Type: text/markdown
 
 <img src="./fig1.png" width="400px"></img>
 
-## metacontroller (wip)
+## metacontroller
 
 Implementation of the MetaController proposed in [Emergent temporal abstractions in autoregressive models enable hierarchical reinforcement learning](https://arxiv.org/abs/2512.20605)
 

metacontroller_pytorch-0.0.43.dist-info/RECORD

@@ -0,0 +1,8 @@
+metacontroller/__init__.py,sha256=iSKbCDp3UrWhZg7SIJFYNjdVQU56u-vqZarE6qCSX74,70
+metacontroller/metacontroller.py,sha256=3L5vPXFm7WbYuuWuCd9RpkMVC--mNpAlsJxpLwMeQ8M,17959
+metacontroller/metacontroller_with_binary_mapper.py,sha256=QJG-o4ZuEC6BkYff6nlyhVcze8o1q8GtbtCP8MLPrvc,9927
+metacontroller/transformer_with_resnet.py,sha256=R49ycusbq3kEX97WHZ41WY2ONc2mYPOuRUCmaFcBOEo,5546
+metacontroller_pytorch-0.0.43.dist-info/METADATA,sha256=y1CIKBvy4jTaqsC_FBfiA1FK7pQrqTasiwA96M83luc,6816
+metacontroller_pytorch-0.0.43.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
+metacontroller_pytorch-0.0.43.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+metacontroller_pytorch-0.0.43.dist-info/RECORD,,

metacontroller_pytorch-0.0.41.dist-info/RECORD

@@ -1,8 +0,0 @@
-metacontroller/__init__.py,sha256=iSKbCDp3UrWhZg7SIJFYNjdVQU56u-vqZarE6qCSX74,70
-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-0.0.41.dist-info/METADATA,sha256=IvP-wC73xCnT8X1aul1IfcaC4fUwRq9Y2UB1h0JG5TI,6822
-metacontroller_pytorch-0.0.41.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
-metacontroller_pytorch-0.0.41.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
-metacontroller_pytorch-0.0.41.dist-info/RECORD,,