titans-pytorch 0.1.6__tar.gz → 0.1.7__tar.gz

{titans_pytorch-0.1.6 → titans_pytorch-0.1.7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: titans-pytorch
-Version: 0.1.6
+Version: 0.1.7
 Summary: Titans
 Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
 Project-URL: Repository, https://github.com/lucidrains/titans-pytorch

{titans_pytorch-0.1.6 → titans_pytorch-0.1.7}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "titans-pytorch"
-version = "0.1.6"
+version = "0.1.7"
 description = "Titans"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{titans_pytorch-0.1.6 → titans_pytorch-0.1.7}/titans_pytorch/titans.py

@@ -161,14 +161,16 @@ class GatedResidualMemoryMLP(Module):
     def __init__(
         self,
         dim,
-        depth
+        depth,
+        expansion_factor = 2.
     ):
         super().__init__()
-        self.depth = depth
+        dim_hidden = int(dim * expansion_factor)
 
         self.weights = ParameterList([
             ParameterList([
-                Parameter(torch.randn(dim, dim)),
+                Parameter(torch.randn(dim, dim_hidden)),
+                Parameter(torch.randn(dim_hidden, dim)),
                 Parameter(torch.randn(dim * 2, dim)),
             ]) for _ in range(depth)
         ])
@@ -182,16 +184,17 @@ class GatedResidualMemoryMLP(Module):
         self,
         x
     ):
-        for weight, to_gates in self.weights:
+        for weight1, weight2, to_gates in self.weights:
             res = x
 
-            x = x @ weight
-            x = F.silu(x)
+            hidden = x @ weight1
+            hidden = F.silu(hidden)
+            branch_out = hidden @ weight2
 
             # gated residual
 
-            gates = cat((x, res), dim = -1) @ to_gates
-            x = res.lerp(x, gates.sigmoid())
+            gates = cat((branch_out, res), dim = -1) @ to_gates
+            x = res.lerp(branch_out, gates.sigmoid())
 
         return x @ self.final_proj