metacontroller-pytorch 0.0.6__py3-none-any.whl → 0.0.9__py3-none-any.whl

metacontroller/metacontroller.py

@@ -1,4 +1,6 @@
 from __future__ import annotations
+from contextlib import nullcontext
+
 from functools import partial
 from collections import namedtuple
 
@@ -113,9 +115,7 @@ class MetaController(Module):
     def internal_rl_parameters(self):
         return [
             *self.action_proposer.parameters(),
-            *self.action_proposer_mean_log_var.parameters(),
-            *self.decoder.parameters(),
-            *self.switch_gating
+            *self.action_proposer_mean_log_var.parameters()
         ]
 
     def forward(
@@ -150,8 +150,6 @@ class MetaController(Module):
 
         switch_beta = self.to_switching_unit_beta(switching_unit_gru_out).sigmoid()
 
-        # switch_beta = switch_beta.expand_as(sampled_action)
-
         # need to encourage normal distribution
 
         vae_kl_loss = self.zero
@@ -233,13 +231,15 @@ class Transformer(Module):
 
     def evolve(
         self,
+        num_generations,
         environment,
         **kwargs
     ):
-        assert exists(self.meta_controller), '`meta_controller` must be defined on init for evolutionary strategies to be straightforwardly applied'
+        assert exists(self.meta_controller), '`meta_controller` must be passed in or defined on init for evolutionary strategies to be straightforwardly applied'
 
         evo_strat = EvoStrategy(
             self,
+            num_generations = num_generations,
             environment = environment,
             params_to_optimize = self.meta_controller.internal_rl_parameters(),
             **kwargs
@@ -252,26 +252,50 @@ class Transformer(Module):
         ids,
         meta_controller: Module | None = None,
         discovery_phase = False,
-        return_latents = False
+        return_latents = False,
+        no_grad_transformer = None,
+        no_grad_meta_controller = None
     ):
         meta_controller = default(meta_controller, self.meta_controller)
 
-        embed = self.embed(ids)
+        meta_controlling = exists(meta_controller)
+
+        # by default, if meta controller is passed in, transformer is no grad
+
+        no_grad_transformer = default(no_grad_transformer, meta_controlling)
+        no_grad_meta_controller = default(no_grad_meta_controller, no_grad_transformer) # by default, if transformer is eval no grad then meta controller is being learnt
+
+        transformer_context = torch.no_grad if no_grad_transformer else nullcontext
+        meta_controller_context = torch.no_grad if no_grad_meta_controller else nullcontext
 
-        residual_stream = self.lower_body(embed)
+        # transformer lower body
+
+        with transformer_context():
+
+            embed = self.embed(ids)
+
+            residual_stream = self.lower_body(embed)
 
         # meta controller acts on residual stream here
 
-        if exists(meta_controller):
-            modified_residual_stream, action_dist, sampled_action, vae_aux_loss = meta_controller(residual_stream, discovery_phase = discovery_phase)
-        else:
-            modified_residual_stream, action_dist, sampled_action, vae_aux_loss = residual_stream, None, None, self.zero
+        with meta_controller_context():
+
+            if exists(meta_controller):
+                modified_residual_stream, action_dist, sampled_action, vae_aux_loss = meta_controller(residual_stream, discovery_phase = discovery_phase)
+            else:
+                modified_residual_stream, action_dist, sampled_action, vae_aux_loss = residual_stream, None, None, self.zero
+
+        # modified residual stream sent back to transformer upper body
+
+        with transformer_context():
+
+            attended = self.upper_body(modified_residual_stream)
 
-        # modified residual stream sent back
+            # head readout
 
-        attended = self.upper_body(modified_residual_stream)
+            dist_params = self.readout(attended)
 
-        dist_params = self.readout(attended)
+        # returning
 
         if not return_latents:
             return dist_params

{metacontroller_pytorch-0.0.6.dist-info → metacontroller_pytorch-0.0.9.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: metacontroller-pytorch
-Version: 0.0.6
+Version: 0.0.9
 Summary: Transformer Metacontroller
 Project-URL: Homepage, https://pypi.org/project/metacontroller/
 Project-URL: Repository, https://github.com/lucidrains/metacontroller

metacontroller_pytorch-0.0.9.dist-info/RECORD

@@ -0,0 +1,6 @@
+metacontroller/__init__.py,sha256=lj7IOGpN--qMxJWbB-4SGqoPXG7Hd4mgtToTRSyTZ58,57
+metacontroller/metacontroller.py,sha256=V2Nb7ByGj310CalTzho-grwNsoHMp55oN5spkedJihc,9189
+metacontroller_pytorch-0.0.9.dist-info/METADATA,sha256=BA4AHlFW8DsD_NPXNv8N8rmRPISZNTkcjvGautB7xJA,3713
+metacontroller_pytorch-0.0.9.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
+metacontroller_pytorch-0.0.9.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+metacontroller_pytorch-0.0.9.dist-info/RECORD,,

metacontroller_pytorch-0.0.6.dist-info/RECORD

@@ -1,6 +0,0 @@
-metacontroller/__init__.py,sha256=lj7IOGpN--qMxJWbB-4SGqoPXG7Hd4mgtToTRSyTZ58,57
-metacontroller/metacontroller.py,sha256=8AWkPlZWh2A1sRT6nMV0CGHuYhQ5pHEpd5bgyRmZelg,8316
-metacontroller_pytorch-0.0.6.dist-info/METADATA,sha256=vXT_-n3bHgpddnS5axyyc-cADGNk4l2enJv4g4cTJ7A,3713
-metacontroller_pytorch-0.0.6.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
-metacontroller_pytorch-0.0.6.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
-metacontroller_pytorch-0.0.6.dist-info/RECORD,,