throw out yet another memory model, gated residual mlp variant

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "titans-pytorch"
-version = "0.1.5"
+version = "0.1.6"
 description = "Titans"
 authors = [
     { name = "Phil Wang", email = "[email protected]" }

titans_pytorch/__init__.py

@@ -2,7 +2,8 @@
     NeuralMemory,
     MemoryMLP,
     MemoryAttention,
-    FactorizedMemoryMLP
+    FactorizedMemoryMLP,
+    GatedResidualMemoryMLP
 )
 
 from titans_pytorch.mac_transformer import (

titans_pytorch/titans.py

@@ -1,10 +1,11 @@
 from __future__ import annotations
 from typing import Callable
+
 import math
 from functools import partial
 
 import torch
-from torch import nn, Tensor
+from torch import nn, cat, Tensor
 import torch.nn.functional as F
 from torch.nn import Linear, Module, Parameter, ParameterList
 from torch.func import functional_call, vmap, grad
@@ -154,6 +155,46 @@ def forward(
 
         return x
 
+# memory mlp, but with gated residual + final projection
+
+class GatedResidualMemoryMLP(Module):
+    def __init__(
+        self,
+        dim,
+        depth
+    ):
+        super().__init__()
+        self.depth = depth
+
+        self.weights = ParameterList([
+            ParameterList([
+                Parameter(torch.randn(dim, dim)),
+                Parameter(torch.randn(dim * 2, dim)),
+            ]) for _ in range(depth)
+        ])
+
+        self.final_proj = Parameter(torch.randn(dim, dim))
+
+        for param in self.parameters():
+            nn.init.xavier_uniform_(param)
+
+    def forward(
+        self,
+        x
+    ):
+        for weight, to_gates in self.weights:
+            res = x
+
+            x = x @ weight
+            x = F.silu(x)
+
+            # gated residual
+
+            gates = cat((x, res), dim = -1) @ to_gates
+            x = res.lerp(x, gates.sigmoid())
+
+        return x @ self.final_proj
+
 # memory mlp with factorized weights
 # so can tradeoff capacity for smaller chunk sizes