titans-pytorch 0.1.9__tar.gz → 0.1.11__tar.gz

{titans_pytorch-0.1.9 → titans_pytorch-0.1.11}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: titans-pytorch
-Version: 0.1.9
+Version: 0.1.11
 Summary: Titans
 Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
 Project-URL: Repository, https://github.com/lucidrains/titans-pytorch
@@ -35,7 +35,7 @@ Classifier: Programming Language :: Python :: 3.9
 Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
 Requires-Python: >=3.9
 Requires-Dist: accelerated-scan>=0.2.0
-Requires-Dist: axial-positional-embedding>=0.3.5
+Requires-Dist: axial-positional-embedding>=0.3.7
 Requires-Dist: einops>=0.8.0
 Requires-Dist: einx>=0.3.0
 Requires-Dist: hyper-connections>=0.1.8
@@ -62,7 +62,7 @@ Unofficial implementation of [Titans](https://arxiv.org/abs/2501.00663) in Pytor
 
 ## Appreciation
 
-- [@sentialx](https://github.com/sentialx) for sharing his early experimental results with me
+- [Eryk](https://github.com/sentialx) for sharing his early experimental results with me, positive for 2 layer MLP
 
 ## Install
 

{titans_pytorch-0.1.9 → titans_pytorch-0.1.11}/README.md

@@ -8,7 +8,7 @@ Unofficial implementation of [Titans](https://arxiv.org/abs/2501.00663) in Pytor
 
 ## Appreciation
 
-- [@sentialx](https://github.com/sentialx) for sharing his early experimental results with me
+- [Eryk](https://github.com/sentialx) for sharing his early experimental results with me, positive for 2 layer MLP
 
 ## Install
 

{titans_pytorch-0.1.9 → titans_pytorch-0.1.11}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "titans-pytorch"
-version = "0.1.9"
+version = "0.1.11"
 description = "Titans"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }
@@ -26,7 +26,7 @@ classifiers=[
 
 dependencies = [
     "accelerated-scan>=0.2.0",
-    "axial_positional_embedding>=0.3.5",
+    "axial_positional_embedding>=0.3.7",
     "einops>=0.8.0",
     "einx>=0.3.0",
     "hyper-connections>=0.1.8",

{titans_pytorch-0.1.9 → titans_pytorch-0.1.11}/titans_pytorch/mac_transformer.py

@@ -593,11 +593,9 @@ class MemoryAsContextTransformer(Module):
         # apply axial positional embedding
         # so intra and inter segment can be more easily discerned by the network
 
-        neural_mem_windows = ceil(seq_len_with_mem / neural_mem_segment_len)
+        pos_emb = self.axial_pos_emb.forward_with_seq_len(seq_len_with_mem, (neural_mem_segment_len,))
 
-        pos_emb = self.axial_pos_emb((neural_mem_windows, neural_mem_segment_len), flatten = True)
-
-        x = x + pos_emb[:seq_len_with_mem]
+        x = x + pos_emb
 
         # prep flex attention