broccoli-ml 0.25.0__tar.gz → 0.27.0__tar.gz

{broccoli_ml-0.25.0 → broccoli_ml-0.27.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: broccoli-ml
-Version: 0.25.0
+Version: 0.27.0
 Summary: Some useful Pytorch models, circa 2025
 License: MIT
 Author: Nicholas Bailey

{broccoli_ml-0.25.0 → broccoli_ml-0.27.0}/broccoli/transformer.py

@@ -299,9 +299,12 @@ class TransformerBlock(nn.Module):
     ):
         super().__init__()
 
+        self.pre_norm = pre_norm
+
         self.identity_probability = identity_probability
 
-        self.layer_norm = nn.LayerNorm(d_model)
+        self.layer_norm_1 = nn.LayerNorm(d_model)
+        self.layer_norm_2 = nn.LayerNorm(d_model)
 
         if position_embedding_type == "relative":
             max_freq = int(max(source_size) / 2)  # Suggested by Gemini!
@@ -358,12 +361,21 @@ class TransformerBlock(nn.Module):
         identity_x = shuffled[:identity_count, :, :]
         process_x = shuffled[identity_count:, :, :]
 
-        norm_process_x = self.layer_norm(process_x)
-        process_x = process_x + self.attn(
-            norm_process_x, norm_process_x, norm_process_x
+        if self.pre_norm:
+            norm_process_x = self.layer_norm_1(process_x)
+            process_x = process_x + self.attn(
+                norm_process_x, norm_process_x, norm_process_x
+            )
+            process_x = process_x + self.ff(process_x)
+        else:  # post-norm
+            process_x = process_x + self.attn(process_x, process_x, process_x)
+            norm_process_x = self.layer_norm_1(process_x)
+            process_x = process_x + self.ff(process_x)
+
+        # Always post norm as eventually we reach the classification head!
+        x = self.layer_norm_2(
+            torch.cat([identity_x, process_x])[unshuffle_indices, :, :].contiguous()
         )
-        process_x = process_x + self.ff(process_x)
-        x = torch.cat([identity_x, process_x])[unshuffle_indices, :, :].contiguous()
 
         return x
 

{broccoli_ml-0.25.0 → broccoli_ml-0.27.0}/broccoli/vit.py

@@ -306,7 +306,8 @@ class ViTEncoder(nn.Module):
                 activation_kwargs=transformer_activation_kwargs,
                 dropout=transformer_mlp_dropout,
                 linear_module=linear_module,
-                raw_input=not cnn,
+                pre_norm=transformer_pre_norm,
+                normformer=transformer_normformer,
             )
         else:
             self.initial_ff = nn.Identity()

{broccoli_ml-0.25.0 → broccoli_ml-0.27.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "broccoli-ml"
-version = "0.25.0"
+version = "0.27.0"
 description = "Some useful Pytorch models, circa 2025"
 authors = [
     {name = "Nicholas Bailey"}