metacontroller-pytorch 0.0.3__tar.gz → 0.0.5__tar.gz

{metacontroller_pytorch-0.0.3 → metacontroller_pytorch-0.0.5}/PKG-INFO

@@ -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-0.0.3 → metacontroller_pytorch-0.0.5}/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 → metacontroller_pytorch-0.0.5}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "metacontroller-pytorch"
-version = "0.0.3"
+version = "0.0.5"
 description = "Transformer Metacontroller"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }