titans-pytorch 0.1.33__py3-none-any.whl → 0.1.34__py3-none-any.whl

titans_pytorch/mac_transformer.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 from typing import Callable
 
 from math import ceil
+from copy import deepcopy
 from functools import partial
 from collections import namedtuple
 
@@ -485,11 +486,13 @@ class MemoryAsContextTransformer(Module):
         heads = 8,
         ff_mult = 4,
         num_residual_streams = 4,
+        neural_memory_model: Module | None = None,
         neural_memory_kwargs: dict = dict(),
         neural_memory_layers: tuple[int, ...] | None = None,
         aux_kv_recon_loss_weight = 0.,
         use_flex_attn = False,
-        sliding_window_attn = False
+        sliding_window_attn = False,
+        weight_tie_memory_model = False
     ):
         super().__init__()
 
@@ -523,6 +526,15 @@ class MemoryAsContextTransformer(Module):
 
         neural_memory_layers = default(neural_memory_layers, layers)
 
+        # weight tying neural memory model
+
+        maybe_copy = deepcopy if not weight_tie_memory_model else identity
+
+        if weight_tie_memory_model:
+            assert exists(neural_memory_model), '`neural_memory_model` must be explicitly set'
+
+        self.weight_tie_memory_model = weight_tie_memory_model
+
         # mem, attn, and feedforward layers
 
         for layer in layers:
@@ -551,6 +563,7 @@ class MemoryAsContextTransformer(Module):
                 mem = NeuralMemory(
                     dim = dim,
                     chunk_size = self.neural_memory_segment_len,
+                    model = maybe_copy(neural_memory_model),
                     **neural_memory_kwargs
                 )
 
@@ -683,7 +696,7 @@ class MemoryAsContextTransformer(Module):
 
         # math
 
-        batch, seq_len, neural_mem_segment_len, segment_len, num_longterm_mem_tokens, attn_window_size = *x.shape, self.neural_memory_segment_len, self.segment_len, self.num_longterm_mem_tokens, self.attn_window_size
+        batch, seq_len, neural_mem_segment_len, segment_len, num_longterm_mem_tokens, attn_window_size, weight_tie_memory_model = *x.shape, self.neural_memory_segment_len, self.segment_len, self.num_longterm_mem_tokens, self.attn_window_size, self.weight_tie_memory_model
 
         seq_len_with_mem = self.seq_len_with_longterm_mem(seq_len)
 
@@ -736,6 +749,10 @@ class MemoryAsContextTransformer(Module):
         next_kv_caches = []
         next_neural_mem_caches = []
 
+        # weight tied neural memory
+
+        neural_memory_updates = None
+
         # value residual
 
         value_residual = None
@@ -769,7 +786,8 @@ class MemoryAsContextTransformer(Module):
                 if not is_inferencing:
                     (retrieved, next_neural_mem_cache), mem_kv_aux_loss = mem(
                         mem_input,
-                        return_aux_kv_loss = True
+                        return_aux_kv_loss = True,
+                        prev_layer_updates = neural_memory_updates
                     )
 
                     kv_recon_losses = kv_recon_losses + mem_kv_aux_loss
@@ -777,9 +795,13 @@ class MemoryAsContextTransformer(Module):
                 else:
                     retrieved, next_neural_mem_cache = mem.forward_inference(
                         mem_input,
-                        state = next(neural_mem_caches, None)
+                        state = next(neural_mem_caches, None),
+                        prev_layer_updates = neural_memory_updates
                     )
 
+                if weight_tie_memory_model:
+                    neural_memory_updates = next_neural_mem_cache.updates
+
                 if self.gate_attn_output:
                     attn_out_gates = retrieved.sigmoid()
                 else:

titans_pytorch/neural_memory.py

@@ -822,6 +822,7 @@ class NeuralMemory(Module):
         self,
         token: Tensor,
         state = None,
+        prev_layer_updates: dict[str, Tensor] | None = None
     ):
 
         # unpack previous state
@@ -863,12 +864,17 @@ class NeuralMemory(Module):
         else:
             updates = updates.apply(lambda t: t[:, -1:])
 
+        if exists(prev_layer_updates):
+            prev_layer_updates = TensorDict(prev_layer_updates)
+            prev_layer_updates = prev_layer_updates.apply(lambda t: t[:, -1:])
+
         if store_seq_cache_len == self.chunk_size:
 
             next_updates, next_states, _ = self.store_memories(
                 cache_store_seq,
                 weights,
-                past_state = past_states
+                past_state = past_states,
+                prev_layer_updates = prev_layer_updates,
             )
 
             updates = next_updates

{titans_pytorch-0.1.33.dist-info → titans_pytorch-0.1.34.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: titans-pytorch
-Version: 0.1.33
+Version: 0.1.34
 Summary: Titans
 Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
 Project-URL: Repository, https://github.com/lucidrains/titans-pytorch

titans_pytorch-0.1.34.dist-info/RECORD

@@ -0,0 +1,8 @@
+titans_pytorch/__init__.py,sha256=rMT99CPQFH4Gudp0FmVPWGKfhBf6xksGEaFEcOVdqjs,230
+titans_pytorch/associative_scan.py,sha256=Y-iYqmFuG-NoCKu6kgql1mhowXTeJfyawi3eUIXamp0,2650
+titans_pytorch/mac_transformer.py,sha256=JvA4mhQaW9LD4j6boRUfLfjyzDCtjqybIr4Ajeio8n8,25708
+titans_pytorch/neural_memory.py,sha256=nNAxhkubuHCGs3bty_eA_yBhWqepPZJgKKvkWXO6IK4,28653
+titans_pytorch-0.1.34.dist-info/METADATA,sha256=pVgjCX_YTT9_5WPcFfXpoaBvzrg1-esvwS0kPpeJAYU,6826
+titans_pytorch-0.1.34.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+titans_pytorch-0.1.34.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+titans_pytorch-0.1.34.dist-info/RECORD,,

titans_pytorch-0.1.33.dist-info/RECORD

@@ -1,8 +0,0 @@
-titans_pytorch/__init__.py,sha256=rMT99CPQFH4Gudp0FmVPWGKfhBf6xksGEaFEcOVdqjs,230
-titans_pytorch/associative_scan.py,sha256=Y-iYqmFuG-NoCKu6kgql1mhowXTeJfyawi3eUIXamp0,2650
-titans_pytorch/mac_transformer.py,sha256=Cui-hCl6X4UVGmuyoKCSKWbag9Yrc-a2MrfVkHM-z0A,24828
-titans_pytorch/neural_memory.py,sha256=9dXpSaQYomc-ur-nEwej1nG9M5NqS0c3LBBP9jUIMPU,28352
-titans_pytorch-0.1.33.dist-info/METADATA,sha256=A9BBoe0Sas2kxUcUi7w_RFl8-SIF1TLzPIRGuZlauFM,6826
-titans_pytorch-0.1.33.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-titans_pytorch-0.1.33.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
-titans_pytorch-0.1.33.dist-info/RECORD,,