metacontroller-pytorch 0.0.4__tar.gz → 0.0.6__tar.gz

{metacontroller_pytorch-0.0.4 → metacontroller_pytorch-0.0.6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: metacontroller-pytorch
-Version: 0.0.4
+Version: 0.0.6
 Summary: Transformer Metacontroller
 Project-URL: Homepage, https://pypi.org/project/metacontroller/
 Project-URL: Repository, https://github.com/lucidrains/metacontroller
@@ -34,7 +34,7 @@ Classifier: License :: OSI Approved :: MIT License
 Classifier: Programming Language :: Python :: 3.9
 Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
 Requires-Python: >=3.9
-Requires-Dist: assoc-scan
+Requires-Dist: assoc-scan>=0.0.3
 Requires-Dist: discrete-continuous-embed-readout>=0.1.11
 Requires-Dist: einops>=0.8.1
 Requires-Dist: einx>=0.3.0

{metacontroller_pytorch-0.0.4 → metacontroller_pytorch-0.0.6}/metacontroller/metacontroller.py

@@ -43,6 +43,11 @@ def default(*args):
             return arg
     return None
 
+# tensor helpers
+
+def straight_through(src, tgt):
+    return tgt + src - src.detach()
+
 # meta controller
 
 class MetaController(Module):
@@ -116,7 +121,8 @@ class MetaController(Module):
     def forward(
         self,
         residual_stream,
-        discovery_phase = False
+        discovery_phase = False,
+        hard_switch = False
     ):
 
         if discovery_phase:
@@ -136,6 +142,16 @@ class MetaController(Module):
 
         sampled_action = readout.sample(action_dist)
 
+        # switching unit timer
+
+        batch, _, dim = sampled_action.shape
+
+        switching_unit_gru_out, switching_unit_gru_hidden = self.switching_unit(residual_stream)
+
+        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
@@ -148,34 +164,30 @@ class MetaController(Module):
                 + mean.square()
                 - log_var
                 - 1.
-            )).sum(dim = -1).mean()
+            ))
 
-        # switching unit timer
+            vae_kl_loss = vae_kl_loss * switch_beta
+            vae_kl_loss = vae_kl_loss.sum(dim = -1).mean()
 
-        batch, _, dim = sampled_action.shape
+        # maybe hard switch, then use associative scan
 
-        switching_unit_gru_out, switching_unit_gru_hidden = self.switching_unit(residual_stream)
-
-        switch_beta = self.to_switching_unit_beta(switching_unit_gru_out).sigmoid()
-
-        action_intent_for_gating = rearrange(sampled_action, 'b n d -> (b d) n')
-        switch_beta = repeat(switch_beta, 'b n d -> (b r d) n', r = dim if not self.switch_per_latent_dim else 1)
+        if hard_switch:
+            hard_switch = (switch_beta > 0.5).float()
+            switch_beta = straight_through(switch_beta, hard_switch)
 
         forget = 1. - switch_beta
-        gated_action_intent = self.switch_gating(action_intent_for_gating * forget, switch_beta)
-
-        gated_action_intent = rearrange(gated_action_intent, '(b d) n -> b n d', b = batch)
+        gated_action = self.switch_gating(switch_beta, sampled_action * forget)
 
         # decoder
 
-        decoder_out = self.decoder(gated_action_intent)
+        decoder_out = self.decoder(gated_action)
 
         w1, w2 = self.to_hyper_network_weights(decoder_out)
         hypernetwork_weight = einsum(w1, w2, '... i r, ... j r -> ... i j')
 
         # generating the residual stream controlling signal
 
-        control_signal = einsum(gated_action_intent, hypernetwork_weight, '... d1, ... d1 d2 -> ... d1')
+        control_signal = einsum(gated_action, hypernetwork_weight, '... d1, ... d1 d2 -> ... d1')
 
         modified_residual_stream = residual_stream + control_signal
 

{metacontroller_pytorch-0.0.4 → metacontroller_pytorch-0.0.6}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "metacontroller-pytorch"
-version = "0.0.4"
+version = "0.0.6"
 description = "Transformer Metacontroller"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }
@@ -24,7 +24,7 @@ classifiers=[
 ]
 
 dependencies = [
-    "assoc-scan",
+    "assoc-scan>=0.0.3",
     "einx>=0.3.0",
     "einops>=0.8.1",
     "discrete-continuous-embed-readout>=0.1.11",