metacontroller-pytorch 0.0.16__tar.gz → 0.0.17__tar.gz

{metacontroller_pytorch-0.0.16 → metacontroller_pytorch-0.0.17}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: metacontroller-pytorch
-Version: 0.0.16
+Version: 0.0.17
 Summary: Transformer Metacontroller
 Project-URL: Homepage, https://pypi.org/project/metacontroller/
 Project-URL: Repository, https://github.com/lucidrains/metacontroller

{metacontroller_pytorch-0.0.16 → metacontroller_pytorch-0.0.17}/metacontroller/metacontroller.py

@@ -57,7 +57,8 @@ MetaControllerOutput = namedtuple('MetaControllerOutput', (
     'prev_hiddens',
     'action_dist',
     'actions',
-    'kl_loss'
+    'kl_loss',
+    'switch_loss'
 ))
 
 class MetaController(Module):
@@ -173,7 +174,7 @@ class MetaController(Module):
 
         # need to encourage normal distribution
 
-        kl_loss = self.zero
+        kl_loss = switch_loss = self.zero
 
         if discovery_phase:
             mean, log_var = action_dist.unbind(dim = -1)
@@ -188,6 +189,10 @@ class MetaController(Module):
             kl_loss = kl_loss * switch_beta
             kl_loss = kl_loss.sum(dim = -1).mean()
 
+            # encourage less switching
+
+            switch_loss = switch_beta.mean()
+
         # maybe hard switch, then use associative scan
 
         if hard_switch:
@@ -220,7 +225,7 @@ class MetaController(Module):
             next_switch_gated_action
         )
 
-        return modified_residual_stream, MetaControllerOutput(next_hiddens, action_dist, sampled_action, kl_loss)
+        return modified_residual_stream, MetaControllerOutput(next_hiddens, action_dist, sampled_action, kl_loss, switch_loss)
 
 # main transformer, which is subsumed into the environment after behavioral cloning
 
@@ -357,7 +362,7 @@ class Transformer(Module):
 
             action_recon_loss = self.action_readout.calculate_loss(dist_params, target_action_ids)
 
-            return action_recon_loss, next_meta_hiddens.kl_loss
+            return action_recon_loss, next_meta_hiddens.kl_loss, next_meta_hiddens.switch_loss
 
         # returning
 

{metacontroller_pytorch-0.0.16 → metacontroller_pytorch-0.0.17}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "metacontroller-pytorch"
-version = "0.0.16"
+version = "0.0.17"
 description = "Transformer Metacontroller"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{metacontroller_pytorch-0.0.16 → metacontroller_pytorch-0.0.17}/tests/test_metacontroller.py

@@ -44,8 +44,8 @@ def test_metacontroller(
 
     # discovery phase
 
-    (action_recon_loss, kl_loss) = model(state, actions, meta_controller = meta_controller, discovery_phase = True)
-    (action_recon_loss + kl_loss * 0.1).backward()
+    (action_recon_loss, kl_loss, switch_loss) = model(state, actions, meta_controller = meta_controller, discovery_phase = True)
+    (action_recon_loss + kl_loss * 0.1 + switch_loss * 0.2).backward()
 
     # internal rl