metacontroller-pytorch 0.0.3__py3-none-any.whl → 0.0.5__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
@@ -42,6 +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):
@@ -62,12 +68,12 @@ 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
 
@@ -115,7 +121,8 @@ class MetaController(Module):
     def forward(
         self,
         residual_stream,
-        discovery_phase = False
+        discovery_phase = False,
+        hard_switch = False
     ):
 
         if discovery_phase:
@@ -123,41 +130,51 @@ 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)
+
+        sampled_action = readout.sample(action_dist)
+
+        # switching unit timer
+
+        batch, _, dim = sampled_action.shape
 
-        std = (0.5 * log_var).exp()
-        sampled_action_intents = mean + torch.randn_like(mean) * std
+        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)
 
         # 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()
                 - log_var
                 - 1.
-            )).sum(dim = -1).mean()
+            )).sum(dim = -1)
 
-        # switching unit timer
+            vae_kl_loss = vae_kl_loss * switch_beta
+            vae_kl_loss = vae_kl_loss.mean()
 
-        batch, _, dim = sampled_action_intents.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_intents, '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)
@@ -177,7 +194,7 @@ 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
 
@@ -249,9 +266,9 @@ class Transformer(Module):
         # meta controller acts on residual stream here
 
         if exists(meta_controller):
-            modified_residual_stream, vae_aux_loss = meta_controller(residual_stream, discovery_phase = discovery_phase)
+            modified_residual_stream, action_dist, sampled_action, 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, action_dist, sampled_action, vae_aux_loss = residual_stream, None, None, self.zero
 
         # modified residual stream sent back
 

{metacontroller_pytorch-0.0.3.dist-info → metacontroller_pytorch-0.0.5.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: metacontroller-pytorch
-Version: 0.0.3
+Version: 0.0.5
 Summary: Transformer Metacontroller
 Project-URL: Homepage, https://pypi.org/project/metacontroller/
 Project-URL: Repository, https://github.com/lucidrains/metacontroller

metacontroller_pytorch-0.0.5.dist-info/RECORD

@@ -0,0 +1,6 @@
+metacontroller/__init__.py,sha256=lj7IOGpN--qMxJWbB-4SGqoPXG7Hd4mgtToTRSyTZ58,57
+metacontroller/metacontroller.py,sha256=-glgWcv6QQZ6wVAy6tK2Ye8QNXuWBiGmB1rs6DVWA1I,8573
+metacontroller_pytorch-0.0.5.dist-info/METADATA,sha256=CfXW_uO8B9gz31XkUO-2aVl4TN64iYicdZPtW7DzzHc,3706
+metacontroller_pytorch-0.0.5.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
+metacontroller_pytorch-0.0.5.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+metacontroller_pytorch-0.0.5.dist-info/RECORD,,

metacontroller_pytorch-0.0.3.dist-info/RECORD

@@ -1,6 +0,0 @@
-metacontroller/__init__.py,sha256=lj7IOGpN--qMxJWbB-4SGqoPXG7Hd4mgtToTRSyTZ58,57
-metacontroller/metacontroller.py,sha256=JZHAkyYbonP5aj_Ob-Xe41ziLEoJLL0KjJy3Dm_bXGY,8091
-metacontroller_pytorch-0.0.3.dist-info/METADATA,sha256=XO3zKqbfSpGmYe34P0uXetIygqkfD1QLuKSSClEeZhk,3706
-metacontroller_pytorch-0.0.3.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
-metacontroller_pytorch-0.0.3.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
-metacontroller_pytorch-0.0.3.dist-info/RECORD,,