metacontroller-pytorch 0.0.2__py3-none-any.whl → 0.0.4__py3-none-any.whl

metacontroller/metacontroller.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 from functools import partial
+from collections import namedtuple
 
 import torch
 from torch import nn, cat, stack, tensor
@@ -62,12 +63,12 @@ class MetaController(Module):
         self.bidirectional_temporal_compressor = GRU(dim_latent, dim_latent, bidirectional = True) # revisit naming
 
         self.emitter = GRU(dim_latent * 2, dim_latent * 2)
-        self.emitter_to_action_mean_log_var = LinearNoBias(dim_latent * 2, dim_latent * 2)
+        self.emitter_to_action_mean_log_var = Readout(dim_latent * 2, num_continuous = dim_latent)
 
         # internal rl phase substitutes the acausal + emitter with a causal ssm
 
         self.action_proposer = GRU(dim_latent, dim_latent)
-        self.action_proposer_mean_log_var = LinearNoBias(dim_latent, dim_latent * 2)
+        self.action_proposer_mean_log_var = Readout(dim_latent, num_continuous = dim_latent)
 
         # switching unit
 
@@ -123,24 +124,25 @@ class MetaController(Module):
             temporal_compressed = reduce(temporal_compressed, '... (two d) -> ... d', 'mean', two = 2)
 
             proposed_action_hidden, _ = self.emitter(cat((temporal_compressed, residual_stream), dim = -1))
-            proposed_action = self.emitter_to_action_mean_log_var(proposed_action_hidden)
+            readout = self.emitter_to_action_mean_log_var
 
         else: # else internal rl phase
             proposed_action_hidden, _ = self.action_proposer(residual_stream)
-            proposed_action = self.action_proposer_mean_log_var(proposed_action_hidden)
+            readout = self.action_proposer_mean_log_var
 
         # sample from the gaussian as the action from the meta controller
 
-        mean, log_var = proposed_action.chunk(2, dim = -1)
+        action_dist = readout(proposed_action_hidden)
 
-        std = (0.5 * log_var).exp()
-        sampled_action_intents = mean + torch.randn_like(mean) * std
+        sampled_action = readout.sample(action_dist)
 
         # need to encourage normal distribution
 
         vae_kl_loss = self.zero
 
         if discovery_phase:
+            mean, log_var = action_dist.unbind(dim = -1)
+
             vae_kl_loss = (0.5 * (
                 log_var.exp()
                 + mean.square()
@@ -150,13 +152,13 @@ class MetaController(Module):
 
         # switching unit timer
 
-        batch, _, dim = sampled_action_intents.shape
+        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()
 
-        action_intent_for_gating = rearrange(sampled_action_intents, 'b n d -> (b d) n')
+        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)
 
         forget = 1. - switch_beta
@@ -177,7 +179,7 @@ class MetaController(Module):
 
         modified_residual_stream = residual_stream + control_signal
 
-        return modified_residual_stream, vae_kl_loss
+        return modified_residual_stream, action_dist, sampled_action, vae_kl_loss
 
 # main transformer, which is subsumed into the environment after behavioral cloning
 
@@ -215,6 +217,8 @@ class Transformer(Module):
 
         self.meta_controller = meta_controller
 
+        self.register_buffer('zero', tensor(0.), persistent = False)
+
     def evolve(
         self,
         environment,
@@ -238,7 +242,7 @@ 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)
 
@@ -246,7 +250,10 @@ 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, 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
 

{metacontroller_pytorch-0.0.2.dist-info → metacontroller_pytorch-0.0.4.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: metacontroller-pytorch
-Version: 0.0.2
+Version: 0.0.4
 Summary: Transformer Metacontroller
 Project-URL: Homepage, https://pypi.org/project/metacontroller/
 Project-URL: Repository, https://github.com/lucidrains/metacontroller

metacontroller_pytorch-0.0.4.dist-info/RECORD

@@ -0,0 +1,6 @@
+metacontroller/__init__.py,sha256=lj7IOGpN--qMxJWbB-4SGqoPXG7Hd4mgtToTRSyTZ58,57
+metacontroller/metacontroller.py,sha256=fGbNUdxTYGgHBdINJziUINDYmANkNLy0yiTIt4dycKM,8162
+metacontroller_pytorch-0.0.4.dist-info/METADATA,sha256=wfeiKctuqzj_NlWq2Xg5hbgjs6bzMmgL-VdTCzgceS8,3706
+metacontroller_pytorch-0.0.4.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
+metacontroller_pytorch-0.0.4.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+metacontroller_pytorch-0.0.4.dist-info/RECORD,,

metacontroller_pytorch-0.0.2.dist-info/RECORD

@@ -1,6 +0,0 @@
-metacontroller/__init__.py,sha256=lj7IOGpN--qMxJWbB-4SGqoPXG7Hd4mgtToTRSyTZ58,57
-metacontroller/metacontroller.py,sha256=VkTl7xHwAW1RSkWNdaMpcJ9kSqcIME-5vK9otEzGocc,7899
-metacontroller_pytorch-0.0.2.dist-info/METADATA,sha256=76qLimsKlE06NLg6XdyE_eKsmBzqxwjL-F0zW0XEoGk,3706
-metacontroller_pytorch-0.0.2.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
-metacontroller_pytorch-0.0.2.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
-metacontroller_pytorch-0.0.2.dist-info/RECORD,,