broccoli-ml 0.38.0__tar.gz → 0.40.0__tar.gz

{broccoli_ml-0.38.0 → broccoli_ml-0.40.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: broccoli-ml
-Version: 0.38.0
+Version: 0.40.0
 Summary: Some useful Pytorch models, circa 2025
 License: MIT
 Author: Nicholas Bailey

{broccoli_ml-0.38.0 → broccoli_ml-0.40.0}/broccoli/linear.py

@@ -34,7 +34,8 @@ class SpectralNormLinear(nn.Module):
 
     def reset_parameters(self) -> None:
         weights = torch.empty(self.out_features, self.in_features)
-        nn.init.kaiming_uniform_(weights, a=math.sqrt(5))
+        stdv = 1.0 / math.sqrt(self.in_features)
+        nn.init.uniform_(weights, a=-stdv, b=stdv)
         if self.use_bias:
             fan_in, _ = nn.init._calculate_fan_in_and_fan_out(weights)
             bound = 1 / math.sqrt(fan_in) if fan_in > 0 else 0
@@ -77,7 +78,8 @@ class AnchoredLinear(nn.Module):
 
     def reset_parameters(self) -> None:
         weights = torch.empty(self.out_features, self.in_features)
-        nn.init.kaiming_uniform_(weights, a=math.sqrt(5))
+        stdv = 1.0 / math.sqrt(self.in_features)
+        nn.init.uniform_(weights, a=-stdv, b=stdv)
         if self.use_bias:
             fan_in, _ = nn.init._calculate_fan_in_and_fan_out(weights)
             bound = 1 / math.sqrt(fan_in) if fan_in > 0 else 0
@@ -120,7 +122,8 @@ class WeightNormedLinear(nn.Module):
 
     def reset_parameters(self) -> None:
         weights = torch.empty(self.out_features, self.in_features)
-        nn.init.kaiming_uniform_(weights, a=math.sqrt(5))
+        stdv = 1.0 / math.sqrt(self.in_features)
+        nn.init.uniform_(weights, a=-stdv, b=stdv)
         if self.use_bias:
             fan_in, _ = nn.init._calculate_fan_in_and_fan_out(weights)
             bound = 1 / math.sqrt(fan_in) if fan_in > 0 else 0

{broccoli_ml-0.38.0 → broccoli_ml-0.40.0}/broccoli/vit.py

@@ -120,6 +120,7 @@ class ViTEncoder(nn.Module):
         transformer_initial_ff_residual_path=True,
         transformer_initial_ff_linear_module_up=None,
         transformer_initial_ff_linear_module_down=None,
+        transformer_initial_ff_mlp_dropout=None,
         transformer_pre_norm=True,
         transformer_normformer=False,
         transformer_post_norm=False,
@@ -307,7 +308,12 @@ class ViTEncoder(nn.Module):
                 transformer_embedding_size,
                 activation=transformer_activation,
                 activation_kwargs=transformer_activation_kwargs,
-                dropout=transformer_mlp_dropout,
+                dropout=(
+                    # First truthy assigned value
+                    transformer_initial_ff_mlp_dropout
+                    if transformer_initial_ff_mlp_dropout is not None
+                    else transformer_mlp_dropout
+                ),
                 linear_module_up=(
                     # First truthy assigned value
                     transformer_initial_ff_linear_module_up
@@ -349,11 +355,7 @@ class ViTEncoder(nn.Module):
 
 class ViT(nn.Module):
     """
-    Denoising convolutional transformer
-    Based on the Compact Convolutional Transformer (CCT) of [Hasani et al. (2021)
-        *''Escaping the Big Data Paradigm with Compact Transformers''*](
-        https://arxiv.org/abs/2104.05704). It's a convolutional neural network
-        leading into a transformer encoder, followed by a sequence pooling layer.
+    ...
     """
 
     def __init__(
@@ -380,6 +382,7 @@ class ViT(nn.Module):
         transformer_initial_ff_residual_path=True,
         transformer_initial_ff_linear_module_up=None,
         transformer_initial_ff_linear_module_down=None,
+        transformer_initial_ff_mlp_dropout=None,
         transformer_pre_norm=True,
         transformer_normformer=False,
         transformer_post_norm=False,
@@ -442,6 +445,7 @@ class ViT(nn.Module):
             transformer_initial_ff_residual_path=transformer_initial_ff_residual_path,
             transformer_initial_ff_linear_module_up=transformer_initial_ff_linear_module_up,
             transformer_initial_ff_linear_module_down=transformer_initial_ff_linear_module_down,
+            transformer_initial_ff_mlp_dropout=transformer_initial_ff_mlp_dropout,
             transformer_pre_norm=transformer_pre_norm,
             transformer_normformer=transformer_normformer,
             transformer_post_norm=transformer_post_norm,

{broccoli_ml-0.38.0 → broccoli_ml-0.40.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "broccoli-ml"
-version = "0.38.0"
+version = "0.40.0"
 description = "Some useful Pytorch models, circa 2025"
 authors = [
     {name = "Nicholas Bailey"}