broccoli-ml 0.39.0__tar.gz → 1.0.0__tar.gz

{broccoli_ml-0.39.0 → broccoli_ml-1.0.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: broccoli-ml
-Version: 0.39.0
+Version: 1.0.0
 Summary: Some useful Pytorch models, circa 2025
 License: MIT
 Author: Nicholas Bailey

{broccoli_ml-0.39.0 → broccoli_ml-1.0.0}/broccoli/vit.py

@@ -85,6 +85,14 @@ class SequencePoolClassificationHead(ClassificationHead):
     def __init__(self, d_model, linear_module, out_dim, batch_norm=True):
         super().__init__(d_model, linear_module, out_dim, batch_norm=batch_norm)
         self.summarize = SequencePool(d_model, linear_module)
+        # Rebuild the classification process with the correct summary module:
+        self.classification_process = nn.Sequential(
+            *[
+                self.summarize,
+                self.projection,
+                self.batch_norm,
+            ]
+        )
 
 
 class ViTEncoder(nn.Module):
@@ -120,6 +128,7 @@ 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 +316,12 @@ 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 +363,7 @@ 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 +390,7 @@ 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 +453,7 @@ 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-0.39.0 → broccoli_ml-1.0.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "broccoli-ml"
-version = "0.39.0"
+version = "1.0.0"
 description = "Some useful Pytorch models, circa 2025"
 authors = [
     {name = "Nicholas Bailey"}