metacontroller-pytorch 0.0.5__tar.gz → 0.0.8__tar.gz

{metacontroller_pytorch-0.0.5 → metacontroller_pytorch-0.0.8}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: metacontroller-pytorch
-Version: 0.0.5
+Version: 0.0.8
 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-0.0.5 → metacontroller_pytorch-0.0.8}/metacontroller/metacontroller.py

@@ -113,9 +113,7 @@ 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(
@@ -150,9 +148,6 @@ class MetaController(Module):
 
         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)
-
         # need to encourage normal distribution
 
         vae_kl_loss = self.zero
@@ -165,10 +160,10 @@ class MetaController(Module):
                 + mean.square()
                 - log_var
                 - 1.
-            )).sum(dim = -1)
+            ))
 
             vae_kl_loss = vae_kl_loss * switch_beta
-            vae_kl_loss = vae_kl_loss.mean()
+            vae_kl_loss = vae_kl_loss.sum(dim = -1).mean()
 
         # maybe hard switch, then use associative scan
 
@@ -177,20 +172,18 @@ class MetaController(Module):
             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
 
@@ -236,13 +229,15 @@ 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

{metacontroller_pytorch-0.0.5 → metacontroller_pytorch-0.0.8}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "metacontroller-pytorch"
-version = "0.0.5"
+version = "0.0.8"
 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",

{metacontroller_pytorch-0.0.5 → metacontroller_pytorch-0.0.8}/tests/test_metacontroller.py

@@ -29,3 +29,6 @@ def test_metacontroller(
     logits = model(ids, meta_controller = meta_controller, discovery_phase = discovery_phase)
 
     assert logits.shape == (1, 1024, 256)
+
+    model.meta_controller = meta_controller
+    model.evolve(1, lambda _: 1., noise_population_size = 2)