titans-pytorch 0.0.6__py3-none-any.whl → 0.0.8__py3-none-any.whl

titans_pytorch/titans.py

@@ -55,6 +55,10 @@ def pack_one_with_inverse(t, pattern):
 
     return packed, inverse
 
+def softclamp_max(t, max_value):
+    range_value = max_value / 2
+    return ((t / range_value).tanh() * range_value) + range_value
+
 # classes
 
 class MLP(Module):
@@ -91,10 +95,15 @@ class NeuralMemory(Module):
         dim,
         chunk_size = 1,
         model: Module | None = None,
-        store_memory_loss_fn: Callable = default_loss_fn
+        store_memory_loss_fn: Callable = default_loss_fn,
+        pre_rmsnorm = False,
+        max_adaptive_step_size = 1e-5
     ):
         super().__init__()
 
+        self.retrieve_norm = nn.RMSNorm(dim) if pre_rmsnorm else nn.Identity()
+        self.store_norm = nn.RMSNorm(dim) if pre_rmsnorm else nn.Identity()
+
         if not exists(model):
             model = MLP(dim, depth = 4)
 
@@ -140,6 +149,8 @@ class NeuralMemory(Module):
             Rearrange('... 1 -> ...')
         )
 
+        self.max_adaptive_step_size = max_adaptive_step_size
+
         # weight decay factor
 
         self.to_decay_factor = nn.Sequential(
@@ -161,6 +172,8 @@ class NeuralMemory(Module):
         past_state: tuple[dict[str, Tensor], dict[str, Tensor]]
     ):
 
+        seq = self.store_norm(seq)
+
         # curtail sequence by multiple of the chunk size
         # only a complete chunk of the sequence provides the memory for the next chunk
 
@@ -182,7 +195,7 @@ class NeuralMemory(Module):
 
         batch = seq.shape[0]
 
-        adaptive_lr = self.to_adaptive_step(seq).tanh() * 0.5 + 0.5
+        adaptive_lr = softclamp_max(self.to_adaptive_step(seq), self.max_adaptive_step_size)
 
         adaptive_momentum = self.to_momentum(seq).sigmoid()
         decay_factor = self.to_decay_factor(seq).sigmoid()
@@ -244,6 +257,8 @@ class NeuralMemory(Module):
         chunk_size = self.chunk_size
         batch, seq_len = seq.shape[:2]
 
+        seq = self.retrieve_norm(seq)
+
         assert seq_len >= chunk_size
 
         seq = seq[:, (chunk_size - 1):]

{titans_pytorch-0.0.6.dist-info → titans_pytorch-0.0.8.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: titans-pytorch
-Version: 0.0.6
+Version: 0.0.8
 Summary: Titans
 Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
 Project-URL: Repository, https://github.com/lucidrains/titans-pytorch
@@ -39,7 +39,8 @@ Requires-Dist: einx>=0.3.0
 Requires-Dist: tensordict>=0.6.2
 Requires-Dist: torch>=2.3
 Provides-Extra: examples
-Requires-Dist: local-attention>=1.9.15; extra == 'examples'
+Requires-Dist: local-attention>=1.10.0; extra == 'examples'
+Requires-Dist: taylor-series-linear-attention; extra == 'examples'
 Provides-Extra: test
 Requires-Dist: pytest; extra == 'test'
 Description-Content-Type: text/markdown
@@ -64,16 +65,16 @@ $ pip install titans-pytorch
 import torch
 from titans_pytorch import NeuralMemory
 
-x = torch.randn(2, 64, 32)
-
 mem = NeuralMemory(
-    dim = 32,
-    chunk_size = 2
-)
+    dim = 384,
+    chunk_size = 64,
+    pre_rmsnorm = True
+).cuda()
 
-out = mem(x)
+seq = torch.randn(2, 1024, 384).cuda()
+retrieved = mem(seq)
 
-assert x.shape == out.shape
+assert seq.shape == retrieved.shape
 ```
 
 ## Citations

titans_pytorch-0.0.8.dist-info/RECORD

@@ -0,0 +1,7 @@
+titans_pytorch/__init__.py,sha256=QKuJPCOJCdgtaPeKoHEkYkiQe65_LV9_8-cIMbBPU30,55
+titans_pytorch/associative_scan.py,sha256=Y-iYqmFuG-NoCKu6kgql1mhowXTeJfyawi3eUIXamp0,2650
+titans_pytorch/titans.py,sha256=I8o_wauKQre87wl645Gy1nchPn5ckfZq4Q6pf5IpToU,9988
+titans_pytorch-0.0.8.dist-info/METADATA,sha256=70ENsVn58zYyJM0UKanESujf0xSQWTcWXBh5HY1frKk,3219
+titans_pytorch-0.0.8.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+titans_pytorch-0.0.8.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+titans_pytorch-0.0.8.dist-info/RECORD,,

titans_pytorch-0.0.6.dist-info/RECORD

@@ -1,7 +0,0 @@
-titans_pytorch/__init__.py,sha256=QKuJPCOJCdgtaPeKoHEkYkiQe65_LV9_8-cIMbBPU30,55
-titans_pytorch/associative_scan.py,sha256=Y-iYqmFuG-NoCKu6kgql1mhowXTeJfyawi3eUIXamp0,2650
-titans_pytorch/titans.py,sha256=S8J8B9o7Rlnj2hU3FZgpn28GTmis3ZbenLqjB_uny54,9470
-titans_pytorch-0.0.6.dist-info/METADATA,sha256=t4HXD6sZT7_pgcwD8TBY6ojYHUHiZ05J6t19wRKtHNc,3092
-titans_pytorch-0.0.6.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-titans_pytorch-0.0.6.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
-titans_pytorch-0.0.6.dist-info/RECORD,,