broccoli-ml 10.2.0__tar.gz → 12.0.0__tar.gz

{broccoli_ml-10.2.0 → broccoli_ml-12.0.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: broccoli-ml
-Version: 10.2.0
+Version: 12.0.0
 Summary: Some useful Pytorch models, circa 2025
 License: MIT
 Author: Nicholas Bailey

{broccoli_ml-10.2.0 → broccoli_ml-12.0.0}/broccoli/transformer.py

@@ -21,24 +21,6 @@ except ImportError:
     FLASH_ATTN = False
 
 
-class LayerScale(nn.Module):
-    def __init__(self, dim, decay=False, init_values=1e-4):
-        super().__init__()
-        self.decay = decay
-        if decay:
-            self.scale = nn.Parameter(init_values * torch.ones(dim))
-            self.nondecay_scale = None
-        else:
-            self.nondecay_scale = nn.Parameter(init_values * torch.ones(dim))
-            self.scale = None
-
-    def forward(self, x):
-        if self.decay:
-            return x * self.scale
-        else:
-            return x * self.nondecay_scale
-
-
 def drop_path(
     x, drop_prob: float = 0.0, training: bool = False, scale_by_keep: bool = True
 ):
@@ -522,7 +504,6 @@ class TransformerBlock(nn.Module):
         post_norm=False,
         normformer=False,
         checkpoint_ff=True,
-        layerscale=True,
     ):
         """
         Args:
@@ -540,16 +521,16 @@ class TransformerBlock(nn.Module):
 
         self.drop_path = DropPath(drop_prob=identity_probability, scale_by_keep=True)
 
-        self.layer_norm_1 = nn.LayerNorm(d_model)
-        self.layer_norm_2 = nn.LayerNorm(d_model)
-        self.layer_norm_3 = nn.LayerNorm(d_model)
+        if self.pre_norm:
+            self.pre_attention_norm = nn.LayerNorm(d_model)
+            self.pre_mlp_norm = nn.LayerNorm(d_model)
 
-        if layerscale:
-            self.layerscale1 = LayerScale(d_model)
-            self.layerscale2 = LayerScale(d_model)
-        else:
-            self.layerscale1 = nn.Identity()
-            self.layerscale2 = nn.Identity()
+        if normformer:
+            self.normformer_norm = nn.LayerNorm(d_model)
+
+        if self.post_norm:
+            self.post_attention_norm = nn.LayerNorm(d_model)
+            self.post_mlp_norm = nn.LayerNorm(d_model)
 
         if relative_position_embedding:
             max_freq = int(max(source_size) / 2)  # Suggested by Gemini!
@@ -613,20 +594,29 @@ class TransformerBlock(nn.Module):
     def forward(self, x):
 
         if self.pre_norm:
-            x = self.layer_norm_1(x)
-            x = x + self.drop_path(self.layerscale1(self.attn(x, x, x)))
-            x = self.layer_norm_2(x)
-            x = x + self.drop_path(self.layerscale2(self.ff(x)))
-            if self.post_norm:  # i.e. in addition! Pre and post.
-                x = self.layer_norm_3(x)
-        elif self.post_norm:  # i.e. only, not prenorm, just post
-            x = x + self.drop_path(self.layerscale1(self.attn(x, x, x)))
-            x = self.layer_norm_1(x)
-            x = x + self.drop_path(self.layerscale2(self.ff(x)))
-            x = self.layer_norm_2(x)
-        else:  # Not pre or post norm. Stand well back.
-            x = x + self.drop_path(self.layerscale1(self.attn(x, x, x)))
-            x = x + self.drop_path(self.layerscale2(self.ff(x)))
+            process_x = self.pre_attention_norm(x)
+        else:
+            process_x = x
+
+        processed = self.drop_path(self.attn(process_x, process_x, process_x))
+
+        if self.normformer:
+            processed = self.normformer_norm(processed)
+
+        x = x + processed
+
+        if self.post_norm:
+            x = self.post_attention_norm(x)
+
+        if self.pre_norm:
+            process_x = self.pre_mlp_norm(x)
+        else:
+            process_x = x
+
+        x = x + self.drop_path(self.ff(process_x))
+
+        if self.post_norm:
+            x = self.post_mlp_norm(x)
 
         return x
 
@@ -634,16 +624,26 @@ class TransformerBlock(nn.Module):
         """
         Give back the attention scores used in this layer.
         """
+        # Fix: Use the correct attribute name 'pre_attention_norm'
         if self.pre_norm:
-            x = self.layer_norm_1(x)
+            # We must normalize the input before measuring attention logits
+            # to match what the model actually sees during forward()
+            x = self.pre_attention_norm(x)
             return self.attn.attention_logits(x, x, x)
         else:
             return self.attn.attention_logits(x, x, x)
 
     def reset_parameters(self):
-        self.layer_norm_1.reset_parameters()
-        self.layer_norm_2.reset_parameters()
-        self.layer_norm_3.reset_parameters()
+        if self.pre_norm:
+            self.pre_attention_norm.reset_parameters()
+            self.pre_mlp_norm.reset_parameters()
+
+        if self.post_norm:
+            self.post_attention_norm.reset_parameters()
+            self.post_mlp_norm.reset_parameters()
+
+        if self.normformer:
+            self.normformer_norm.reset_parameters()
 
         self.attn.reset_parameters()
         self.ff.reset_parameters()
@@ -684,7 +684,6 @@ class TransformerEncoder(nn.Module):
         normformer=False,
         msa_scaling="d",
         checkpoint_ff=True,
-        layerscale=True,
     ):
         """
         Args:
@@ -719,12 +718,6 @@ class TransformerEncoder(nn.Module):
         self._utility_tokens = utility_tokens
         self.return_utility_tokens = return_utility_tokens
 
-        if layerscale:
-            rope_and_ape = absolute_position_embedding and relative_position_embedding
-            self.position_layerscale = LayerScale(d_model, decay=rope_and_ape)
-        else:
-            self.position_layerscale = None
-
         # Initialise utility tokens with normal init, like usual Pytorch embeddings
         if self._utility_tokens:
             self._utility_token_embedding = nn.Parameter(
@@ -789,7 +782,6 @@ class TransformerEncoder(nn.Module):
                     post_norm=post_norm,
                     normformer=normformer,
                     checkpoint_ff=checkpoint_ff,
-                    layerscale=layerscale,
                 )
                 for i in range(n_layers)
             ]
@@ -817,8 +809,6 @@ class TransformerEncoder(nn.Module):
                     0
                 )  # to shape (1, seq_len) to broadcast over batch
             )
-            if self.position_layerscale is not None:
-                position_embedding = self.position_layerscale(position_embedding)
             x += position_embedding
 
         return x

{broccoli_ml-10.2.0 → broccoli_ml-12.0.0}/broccoli/vit.py

@@ -187,7 +187,6 @@ class ViTEncoder(nn.Module):
         transformer_msa_dropout=0.1,
         transformer_stochastic_depth=0.1,
         transformer_checkpoint_ff=True,
-        transformer_layerscale=True,
         linear_module=nn.Linear,
     ):
         super().__init__()
@@ -353,7 +352,6 @@ class ViTEncoder(nn.Module):
                 normformer=transformer_normformer,
                 post_norm=transformer_post_norm,
                 checkpoint_ff=transformer_checkpoint_ff,
-                layerscale=transformer_layerscale,
             )
         else:
             self.transformer = nn.Identity()
@@ -489,7 +487,6 @@ class ViT(nn.Module):
         transformer_msa_dropout=0.1,
         transformer_stochastic_depth=0.1,
         transformer_checkpoint_ff=True,
-        transformer_layerscale=True,
         head=SequencePoolClassificationHead,
         batch_norm_logits=True,
         logit_projection_layer=nn.Linear,
@@ -562,7 +559,6 @@ class ViT(nn.Module):
             transformer_msa_dropout=transformer_msa_dropout,
             transformer_stochastic_depth=transformer_stochastic_depth,
             transformer_checkpoint_ff=transformer_checkpoint_ff,
-            transformer_layerscale=transformer_layerscale,
             linear_module=linear_module,
         )
 

{broccoli_ml-10.2.0 → broccoli_ml-12.0.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "broccoli-ml"
-version = "10.2.0"
+version = "12.0.0"
 description = "Some useful Pytorch models, circa 2025"
 authors = [
     {name = "Nicholas Bailey"}