metacontroller-pytorch 0.0.1__tar.gz → 0.0.3__tar.gz

{metacontroller_pytorch-0.0.1 → metacontroller_pytorch-0.0.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: metacontroller-pytorch
-Version: 0.0.1
+Version: 0.0.3
 Summary: Transformer Metacontroller
 Project-URL: Homepage, https://pypi.org/project/metacontroller/
 Project-URL: Repository, https://github.com/lucidrains/metacontroller

{metacontroller_pytorch-0.0.1 → metacontroller_pytorch-0.0.3}/metacontroller/metacontroller.py

@@ -49,6 +49,7 @@ class MetaController(Module):
         self,
         dim_latent,
         *,
+        switch_per_latent_dim = True,
         decoder_expansion_factor = 2.,
         decoder_depth = 1,
         hypernetwork_low_rank = 16,
@@ -70,8 +71,10 @@ class MetaController(Module):
 
         # switching unit
 
+        self.switch_per_latent_dim = switch_per_latent_dim
+
         self.switching_unit = GRU(dim_latent, dim_latent)
-        self.to_switching_unit_beta = nn.Linear(dim_latent, 1, bias = False)
+        self.to_switching_unit_beta = nn.Linear(dim_latent, dim_latent if switch_per_latent_dim else 1, bias = False)
 
         self.switch_gating = AssocScan(**assoc_scan_kwargs)
 
@@ -154,7 +157,7 @@ class MetaController(Module):
         switch_beta = self.to_switching_unit_beta(switching_unit_gru_out).sigmoid()
 
         action_intent_for_gating = rearrange(sampled_action_intents, 'b n d -> (b d) n')
-        switch_beta = repeat(switch_beta, 'b n 1 -> (b d) n', d = dim)
+        switch_beta = repeat(switch_beta, 'b n d -> (b r d) n', r = dim if not self.switch_per_latent_dim else 1)
 
         forget = 1. - switch_beta
         gated_action_intent = self.switch_gating(action_intent_for_gating * forget, switch_beta)
@@ -212,6 +215,8 @@ class Transformer(Module):
 
         self.meta_controller = meta_controller
 
+        self.register_buffer('zero', tensor(0.), persistent = False)
+
     def evolve(
         self,
         environment,
@@ -235,7 +240,7 @@ class Transformer(Module):
         discovery_phase = False,
         return_latents = False
     ):
-        meta_controller = default(meta_controller, self.meta_controller, Identity())
+        meta_controller = default(meta_controller, self.meta_controller)
 
         embed = self.embed(ids)
 
@@ -243,7 +248,10 @@ class Transformer(Module):
 
         # meta controller acts on residual stream here
 
-        modified_residual_stream, vae_aux_loss = meta_controller(residual_stream, discovery_phase = discovery_phase)
+        if exists(meta_controller):
+            modified_residual_stream, vae_aux_loss = meta_controller(residual_stream, discovery_phase = discovery_phase)
+        else:
+            modified_residual_stream, vae_aux_loss = residual_stream, self.zero
 
         # modified residual stream sent back
 

{metacontroller_pytorch-0.0.1 → metacontroller_pytorch-0.0.3}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "metacontroller-pytorch"
-version = "0.0.1"
+version = "0.0.3"
 description = "Transformer Metacontroller"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{metacontroller_pytorch-0.0.1 → metacontroller_pytorch-0.0.3}/tests/test_metacontroller.py

@@ -5,8 +5,10 @@ import torch
 from metacontroller.metacontroller import Transformer, MetaController
 
 @param('discovery_phase', (False, True))
+@param('switch_per_latent_dim', (False, True))
 def test_metacontroller(
-    discovery_phase
+    discovery_phase,
+    switch_per_latent_dim
 ):
 
     ids = torch.randint(0, 256, (1, 1024))
@@ -19,7 +21,10 @@ def test_metacontroller(
         readout = dict(num_discrete = 256)
     )
 
-    meta_controller = MetaController(512)
+    meta_controller = MetaController(
+        512,
+        switch_per_latent_dim = switch_per_latent_dim
+    )
 
     logits = model(ids, meta_controller = meta_controller, discovery_phase = discovery_phase)