broccoli-ml 0.13.0__tar.gz → 0.13.1__tar.gz

{broccoli_ml-0.13.0 → broccoli_ml-0.13.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: broccoli-ml
-Version: 0.13.0
+Version: 0.13.1
 Summary: Some useful Pytorch models, circa 2025
 License: MIT
 Author: Nicholas Bailey

{broccoli_ml-0.13.0 → broccoli_ml-0.13.1}/broccoli/linear.py

@@ -10,7 +10,8 @@ from .tensor import SigmaReparamTensor
 
 class SpectralNormLinear(nn.Module):
     """
-    ...
+    Inspired by Apple's Spectral Normed Linear Layers
+        (https://github.com/apple/ml-sigma-reparam)
     """
 
     def __init__(self, in_features: int, out_features: int, bias: bool = True):

{broccoli_ml-0.13.0 → broccoli_ml-0.13.1}/broccoli/transformer.py

@@ -222,7 +222,7 @@ class MHAttention(nn.Module):
         return self.out_proj(output_without_heads)
 
 
-class FeedforwardLayer(nn.Module):
+class FeedforwardBlock(nn.Module):
     """
     ...
     """
@@ -324,7 +324,7 @@ class TransformerBlock(nn.Module):
         )
 
         # Submodules for the feedforward process
-        self.ff = FeedforwardLayer(
+        self.ff = FeedforwardBlock(
             d_model,
             mlp_ratio,
             d_model,

{broccoli_ml-0.13.0 → broccoli_ml-0.13.1}/broccoli/vit.py

@@ -1,9 +1,10 @@
 import math
 from typing import Optional
 
-from .transformer import TransformerEncoder, FeedforwardLayer
+from .transformer import TransformerEncoder, FeedforwardBlock
 from .cnn import SpaceToDepth, calculate_output_spatial_size, spatial_tuple
 from .activation import ReLU, SquaredReLU, GELU, SwiGLU
+from .linear import SpectralNormLinear
 from einops import einsum
 from einops.layers.torch import Rearrange
 import torch.nn as nn
@@ -103,6 +104,7 @@ class ViTEncoder(nn.Module):
     def __init__(
         self,
         input_size=(32, 32),
+        initial_batch_norm=True,
         cnn=True,
         cnn_in_channels=3,
         cnn_out_channels=16,
@@ -132,7 +134,6 @@ class ViTEncoder(nn.Module):
         transformer_msa_dropout=0.1,
         transformer_stochastic_depth=0.1,
         linear_module=nn.Linear,
-        initial_batch_norm=True,
     ):
         super().__init__()
 
@@ -294,7 +295,7 @@ class ViTEncoder(nn.Module):
 
         if intermediate_feedforward_layer:
             self.pooling_channels_padding = nn.Identity()
-            self.intermediate_feedforward_layer = FeedforwardLayer(
+            self.intermediate_feedforward_layer = FeedforwardBlock(
                 pooling_out_channels,
                 transformer_mlp_ratio,
                 transformer_embedding_size,
@@ -349,6 +350,8 @@ class ViT(nn.Module):
     def __init__(
         self,
         input_size=(32, 32),
+        image_classes=100,
+        initial_batch_norm=True,
         cnn=True,
         cnn_in_channels=3,
         cnn_out_channels=16,
@@ -378,9 +381,7 @@ class ViT(nn.Module):
         transformer_msa_dropout=0.1,
         transformer_stochastic_depth=0.1,
         batch_norm_outputs=True,
-        initial_batch_norm=True,
-        linear_module=nn.Linear,
-        image_classes=100,
+        linear_module=SpectralNormLinear,
         head=SequencePoolClassificationHead,
     ):
 
@@ -404,6 +405,7 @@ class ViT(nn.Module):
 
         self.encoder = ViTEncoder(
             input_size=input_size,
+            initial_batch_norm=initial_batch_norm,
             cnn=cnn,
             cnn_in_channels=cnn_in_channels,
             cnn_out_channels=cnn_out_channels,
@@ -433,7 +435,6 @@ class ViT(nn.Module):
             transformer_msa_dropout=transformer_msa_dropout,
             transformer_stochastic_depth=transformer_stochastic_depth,
             linear_module=linear_module,
-            initial_batch_norm=initial_batch_norm,
         )
 
         self.pool = head(

{broccoli_ml-0.13.0 → broccoli_ml-0.13.1}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "broccoli-ml"
-version = "0.13.0"
+version = "0.13.1"
 description = "Some useful Pytorch models, circa 2025"
 authors = [
     {name = "Nicholas Bailey"}