broccoli-ml 15.5.0__tar.gz → 15.6.0__tar.gz

{broccoli_ml-15.5.0 → broccoli_ml-15.6.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: broccoli-ml
-Version: 15.5.0
+Version: 15.6.0
 Summary: Some useful Pytorch models, circa 2025
 License: MIT
 Author: Nicholas Bailey

{broccoli_ml-15.5.0 → broccoli_ml-15.6.0}/broccoli/transformer.py

@@ -351,9 +351,9 @@ class MHAttention(nn.Module):
         self.q_proj.reset_parameters()
         self.k_proj.reset_parameters()
         self.v_proj.reset_parameters()
-        scale_parameters(self.v_proj, self.beta)
+        scale_parameters(self.v_proj, math.sqrt(6) * self.beta)
         self.out_proj.reset_parameters()
-        scale_parameters(self.out_proj, self.beta)
+        scale_parameters(self.out_proj, math.sqrt(6) * self.beta)
 
         if self.talking_heads:
             # Initialize close to identity
@@ -470,8 +470,8 @@ class FeedforwardBlock(nn.Module):
             if hasattr(module, "reset_parameters"):
                 module.reset_parameters()
 
-        scale_parameters(self.linear_in, self.beta)
-        scale_parameters(self.linear_out, self.beta)
+        scale_parameters(self.linear_in, math.sqrt(6) * self.beta)
+        scale_parameters(self.linear_out, math.sqrt(6) * self.beta)
 
 
 class EncoderBlock(nn.Module):

{broccoli_ml-15.5.0 → broccoli_ml-15.6.0}/broccoli/vit.py

@@ -521,7 +521,7 @@ class ViT(nn.Module):
             }[transformer_activation]
 
         self.alpha = (2 * transformer_layers) ** 0.25
-        self.beta = (8 * transformer_layers) ** 0.25
+        self.beta = (8 * transformer_layers) ** -0.25
 
         self.encoder = ViTEncoder(
             input_size=input_size,

{broccoli_ml-15.5.0 → broccoli_ml-15.6.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "broccoli-ml"
-version = "15.5.0"
+version = "15.6.0"
 description = "Some useful Pytorch models, circa 2025"
 authors = [
     {name = "Nicholas Bailey"}