titans-pytorch 0.0.61__py3-none-any.whl → 0.0.62__py3-none-any.whl

titans_pytorch/__init__.py

@@ -1,7 +1,8 @@
 from titans_pytorch.titans import (
     NeuralMemory,
     MemoryMLP,
-    MemoryAttention
+    MemoryAttention,
+    FactorizedMemoryMLP
 )
 
 from titans_pytorch.mac_transformer import (

titans_pytorch/titans.py

@@ -6,7 +6,7 @@ from functools import partial
 import torch
 from torch import nn, Tensor
 import torch.nn.functional as F
-from torch.nn import Linear, Module
+from torch.nn import Linear, Module, Parameter, ParameterList
 from torch.func import functional_call, vmap, grad
 
 from tensordict import TensorDict
@@ -88,7 +88,7 @@ class MultiheadRMSNorm(Module):
     def __init__(self, dim, heads):
         super().__init__()
         self.rmsnorm = nn.RMSNorm(dim, elementwise_affine = False)
-        self.gamma = nn.Parameter(torch.zeros(heads, 1, dim))
+        self.gamma = Parameter(torch.zeros(heads, 1, dim))
 
     def forward(self, x):
         return self.rmsnorm(x) * (self.gamma + 1.)
@@ -102,7 +102,10 @@ class MemoryMLP(Module):
         depth
     ):
         super().__init__()
-        self.weights = nn.ParameterList([nn.Parameter(torch.randn(dim, dim)) for _ in range(depth)])
+        self.weights = ParameterList([Parameter(torch.randn(dim, dim)) for _ in range(depth)])
+
+        for weight in self.weights:
+            nn.init.xavier_uniform_(weight)
 
     def forward(
         self,
@@ -118,13 +121,50 @@ class MemoryMLP(Module):
 
         return x
 
+# memory mlp with factorized weights
+# so can tradeoff capacity for smaller chunk sizes
+
+class FactorizedMemoryMLP(Module):
+    def __init__(
+        self,
+        dim,
+        depth,
+        k = 32
+    ):
+        super().__init__()
+        self.weights = ParameterList([
+            ParameterList([
+                Parameter(torch.randn(dim, k)),
+                Parameter(torch.randn(k, dim)),
+            ]) for _ in range(depth)
+        ])
+
+        for weight1, weight2 in self.weights:
+            nn.init.xavier_uniform_(weight1)
+            nn.init.xavier_uniform_(weight2)
+
+    def forward(
+        self,
+        x
+    ):
+        for ind, (weight1, weight2) in enumerate(self.weights):
+            is_first = ind == 0
+
+            if not is_first:
+                x = F.silu(x)
+
+            x = x @ weight1 @ weight2
+
+        return x
+
 # improvised attention as memory module
 
 class MemoryAttention(Module):
     def __init__(
         self,
         dim,
-        scale = 8.
+        scale = 8.,
+        expansion_factor = 2.
     ):
         super().__init__()
         self.scale = scale
@@ -133,10 +173,13 @@ class MemoryAttention(Module):
             nn.Parameter(torch.randn(dim, dim)), # queries
             nn.Parameter(torch.randn(dim, dim)), # keys
             nn.Parameter(torch.randn(dim, dim)), # values
-            nn.Parameter(torch.randn(dim, dim * 2)), # ff w1
-            nn.Parameter(torch.randn(dim * 2, dim)), # ff w2
+            nn.Parameter(torch.randn(dim, dim * expansion_factor)), # ff w1
+            nn.Parameter(torch.randn(dim * expansion_factor, dim)), # ff w2
         ])
 
+        for weight in self.weights:
+            nn.init.xavier_uniform_(weight)
+
     def forward(self, x):
         wq, wk, wv, ffw1, ffw2 = self.weights
 
@@ -536,6 +579,7 @@ class NeuralMemory(Module):
 
         past_weights, _ = past_state
 
+
         retrieved = self.retrieve_memories(seq, past_weights + updates)
 
         if not return_aux_kv_loss:

{titans_pytorch-0.0.61.dist-info → titans_pytorch-0.0.62.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: titans-pytorch
-Version: 0.0.61
+Version: 0.0.62
 Summary: Titans
 Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
 Project-URL: Repository, https://github.com/lucidrains/titans-pytorch

titans_pytorch-0.0.62.dist-info/RECORD

@@ -0,0 +1,8 @@
+titans_pytorch/__init__.py,sha256=I-4oF1JPEmUvhLszEocM5cDgHYNFcNp0Q9nbDYSPFqU,195
+titans_pytorch/associative_scan.py,sha256=Y-iYqmFuG-NoCKu6kgql1mhowXTeJfyawi3eUIXamp0,2650
+titans_pytorch/mac_transformer.py,sha256=R9Xu-BjqAe9ZY60IGk4aNXBx_L8THsjJ4QrkbTnRNHo,15346
+titans_pytorch/titans.py,sha256=95J6UL44lOrdZSXdm7p36xC9tDeSmRBwdjig9T82PzI,17452
+titans_pytorch-0.0.62.dist-info/METADATA,sha256=08Blaa9Ehyv09rSA9uWguxbhKpbrd7C53Ya13E1VbpU,4457
+titans_pytorch-0.0.62.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+titans_pytorch-0.0.62.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+titans_pytorch-0.0.62.dist-info/RECORD,,

titans_pytorch-0.0.61.dist-info/RECORD

@@ -1,8 +0,0 @@
-titans_pytorch/__init__.py,sha256=Q0MQA3RS8vqzs-KzSGZkvLR7afQ6ZW9uMOq1MeNuFoY,170
-titans_pytorch/associative_scan.py,sha256=Y-iYqmFuG-NoCKu6kgql1mhowXTeJfyawi3eUIXamp0,2650
-titans_pytorch/mac_transformer.py,sha256=R9Xu-BjqAe9ZY60IGk4aNXBx_L8THsjJ4QrkbTnRNHo,15346
-titans_pytorch/titans.py,sha256=5wuAoDULbgXTM8Nbq8bXrW3Fd2nsn22kpERRfJOwZiU,16367
-titans_pytorch-0.0.61.dist-info/METADATA,sha256=Cfhqnse_9nnFNqVGo9p_kxO_LVawwv4uuZOx4anqhf0,4457
-titans_pytorch-0.0.61.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-titans_pytorch-0.0.61.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
-titans_pytorch-0.0.61.dist-info/RECORD,,