broccoli-ml 10.1.1__py3-none-any.whl → 11.0.0__py3-none-any.whl

broccoli/transformer.py

@@ -22,12 +22,26 @@ except ImportError:
 
 
 class LayerScale(nn.Module):
-    def __init__(self, dim, init_values=1e-4):
+    def __init__(self, dim, decay=False, init_values=1e-4):
         super().__init__()
-        self.nondecay_scale = nn.Parameter(init_values * torch.ones(dim))
+        self.dim = dim
+        self.decay = decay
+        self.init_values = init_values
+        self.reset_parameters()
 
     def forward(self, x):
-        return x * self.nondecay_scale
+        if self.decay:
+            return x * self.scale
+        else:
+            return x * self.nondecay_scale
+
+    def reset_parameters(self):
+        if self.decay:
+            self.scale = nn.Parameter(self.init_values * torch.ones(self.dim))
+            self.nondecay_scale = None
+        else:
+            self.nondecay_scale = nn.Parameter(self.init_values * torch.ones(self.dim))
+            self.scale = None
 
 
 def drop_path(
@@ -531,10 +545,18 @@ 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 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)
+
+        self.layerscale = layerscale
         if layerscale:
             self.layerscale1 = LayerScale(d_model)
             self.layerscale2 = LayerScale(d_model)
@@ -604,20 +626,31 @@ 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.layerscale1(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.layerscale2(self.ff(process_x)))
+
+        if self.post_norm:
+            x = self.post_mlp_norm(x)
 
         return x
 
@@ -625,20 +658,34 @@ 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()
 
+        if self.layerscale:
+            self.layerscale1.reset_parameters()
+            self.layerscale2.reset_parameters()
+
 
 class TransformerEncoder(nn.Module):
     """
@@ -711,9 +758,10 @@ class TransformerEncoder(nn.Module):
         self.return_utility_tokens = return_utility_tokens
 
         if layerscale:
-            self.layerscale = LayerScale(d_model)
+            rope_and_ape = absolute_position_embedding and relative_position_embedding
+            self.position_layerscale = LayerScale(d_model, decay=rope_and_ape)
         else:
-            self.layerscale = None
+            self.position_layerscale = None
 
         # Initialise utility tokens with normal init, like usual Pytorch embeddings
         if self._utility_tokens:
@@ -807,8 +855,8 @@ class TransformerEncoder(nn.Module):
                     0
                 )  # to shape (1, seq_len) to broadcast over batch
             )
-            if self.layerscale is not None:
-                position_embedding = self.layerscale(position_embedding)
+            if self.position_layerscale is not None:
+                position_embedding = self.position_layerscale(position_embedding)
             x += position_embedding
 
         return x

{broccoli_ml-10.1.1.dist-info → broccoli_ml-11.0.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: broccoli-ml
-Version: 10.1.1
+Version: 11.0.0
 Summary: Some useful Pytorch models, circa 2025
 License: MIT
 Author: Nicholas Bailey

{broccoli_ml-10.1.1.dist-info → broccoli_ml-11.0.0.dist-info}/RECORD

@@ -4,10 +4,10 @@ broccoli/cnn.py,sha256=WjoPDSpe3ttwxCBNfCVRdaCHvbeZ7G-a5_i8fUsK_d8,4889
 broccoli/linear.py,sha256=W-3aNpBjd_0xRyzbCKkmg4H1qmslQOIQhB-WDDay2nM,13125
 broccoli/rope.py,sha256=GRqApBNmYCFaDak0WL1xE_BC5CTTYKQU_PBdeTcQcjc,12557
 broccoli/tensor.py,sha256=um8mrxkYbvNDo-QvHlmJm8Aw6qcngOlUZPoAk_PMReA,4480
-broccoli/transformer.py,sha256=a2l1uRWr6wD6Q84XdYmTsimNqBxDDmbK719b6fwKCik,27982
+broccoli/transformer.py,sha256=urW9KGxK47Mns7ZowPKTgdEyp4Yd21uw3PNCpKx04cI,29223
 broccoli/utils.py,sha256=oOWzn6dJ5nC_9r4zq0emmfmaYACJXJNFS48AOpW2jqc,358
 broccoli/vit.py,sha256=EGbQb-atuzG3JAx7kdTaJEbWvQR-4XgyYvwjKkN5C38,22612
-broccoli_ml-10.1.1.dist-info/LICENSE,sha256=0BAzJE5BqQ7Iixp_AFdB2W1uO-HCRX-Qfun8PHt6yVM,1073
-broccoli_ml-10.1.1.dist-info/METADATA,sha256=fi_mTSGNka5gnIfIVuvaunXqzbppH7NXMZR4ierpVbY,1369
-broccoli_ml-10.1.1.dist-info/WHEEL,sha256=b4K_helf-jlQoXBBETfwnf4B04YC67LOev0jo4fX5m8,88
-broccoli_ml-10.1.1.dist-info/RECORD,,
+broccoli_ml-11.0.0.dist-info/LICENSE,sha256=0BAzJE5BqQ7Iixp_AFdB2W1uO-HCRX-Qfun8PHt6yVM,1073
+broccoli_ml-11.0.0.dist-info/METADATA,sha256=7ixQROHXr3LyCMhT5hkKmvip8C8Bytmtz-IoQCbC_NQ,1369
+broccoli_ml-11.0.0.dist-info/WHEEL,sha256=b4K_helf-jlQoXBBETfwnf4B04YC67LOev0jo4fX5m8,88
+broccoli_ml-11.0.0.dist-info/RECORD,,