broccoli-ml 0.2.0__tar.gz → 0.4.0__tar.gz

{broccoli_ml-0.2.0 → broccoli_ml-0.4.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: broccoli-ml
-Version: 0.2.0
+Version: 0.4.0
 Summary: Some useful Pytorch models, circa 2025
 License: MIT
 Author: Nicholas Bailey

{broccoli_ml-0.2.0 → broccoli_ml-0.4.0}/broccoli/vit.py

@@ -49,7 +49,7 @@ class SequencePool(nn.Module):
         return self.norm(projection)
 
 
-class CCTEncoder(nn.Module):
+class DCTEncoder(nn.Module):
     """
     Based on the Compact Convolutional Transformer (CCT) of [Hasani et al. (2021)
         *''Escaping the Big Data Paradigm with Compact Transformers''*](
@@ -63,25 +63,26 @@ class CCTEncoder(nn.Module):
         self,
         input_size=(32, 32),
         cnn_in_channels=3,
+        minimum_cnn_out_channels=16,
         cnn_kernel_size=3,
         cnn_kernel_stride=1,
         cnn_kernel_padding="same",
         cnn_kernel_dilation=1,
         cnn_kernel_groups=1,
-        cnn_activation: nn.Module = nn.ReLU,
+        cnn_activation: nn.Module = ReLU,
         cnn_activation_kwargs: Optional[dict] = None,
         cnn_dropout=0.0,
-        pooling_type="maxpool",
+        pooling_type="concat",  # maxpool or concat
         pooling_kernel_size=3,
         pooling_kernel_stride=2,
         pooling_kernel_padding=1,
-        transformer_position_embedding="absolute",  # absolute or relative
+        transformer_position_embedding="relative",  # absolute or relative
         transformer_embedding_size=256,
         transformer_layers=7,
         transformer_heads=4,
         transformer_mlp_ratio=2,
         transformer_bos_tokens=4,
-        transformer_activation: nn.Module = nn.GELU,
+        transformer_activation: nn.Module = SquaredReLU,
         transformer_activation_kwargs: Optional[dict] = None,
         mlp_dropout=0.0,
         msa_dropout=0.1,
@@ -155,8 +156,9 @@ class CCTEncoder(nn.Module):
         if pooling_type in ["maxpool", None]:
             cnn_out_channels = transformer_embedding_size
         elif pooling_type == "concat":
-            cnn_out_channels = math.floor(
-                transformer_embedding_size / pooling_kernel_voxels
+            cnn_out_channels = min(
+                math.floor(transformer_embedding_size / pooling_kernel_voxels),
+                minimum_cnn_out_channels,
             )
         else:
             raise NotImplementedError("Pooling type must be maxpool, concat or None")
@@ -289,8 +291,9 @@ class CCTEncoder(nn.Module):
         return self.encoder(x)
 
 
-class CCT(nn.Module):
+class DCT(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
@@ -301,25 +304,26 @@ class CCT(nn.Module):
         self,
         input_size=(32, 32),
         cnn_in_channels=3,
+        minimum_cnn_out_channels=16,
         cnn_kernel_size=3,
         cnn_kernel_stride=1,
         cnn_kernel_padding="same",
         cnn_kernel_dilation=1,
         cnn_kernel_groups=1,
-        cnn_activation: nn.Module = nn.ReLU,
+        cnn_activation: nn.Module = ReLU,
         cnn_activation_kwargs: Optional[dict] = None,
         cnn_dropout=0.0,
-        pooling_type="maxpool",
+        pooling_type="concat",  # maxpool or concat
         pooling_kernel_size=3,
         pooling_kernel_stride=2,
         pooling_kernel_padding=1,
-        transformer_position_embedding="absolute",  # absolute or relative
+        transformer_position_embedding="relative",  # absolute or relative
         transformer_embedding_size=256,
         transformer_layers=7,
         transformer_heads=4,
         transformer_mlp_ratio=2,
         transformer_bos_tokens=4,
-        transformer_activation: nn.Module = nn.GELU,
+        transformer_activation: nn.Module = SquaredReLU,
         transformer_activation_kwargs: Optional[dict] = None,
         mlp_dropout=0.0,
         msa_dropout=0.1,
@@ -347,9 +351,10 @@ class CCT(nn.Module):
                 "SwiGLU": SwiGLU,
             }[transformer_activation]
 
-        self.encoder = CCTEncoder(
+        self.encoder = DCTEncoder(
             input_size=input_size,
             cnn_in_channels=cnn_in_channels,
+            minimum_cnn_out_channels=minimum_cnn_out_channels,
             cnn_kernel_size=cnn_kernel_size,
             cnn_kernel_stride=cnn_kernel_stride,
             cnn_kernel_padding=cnn_kernel_padding,

{broccoli_ml-0.2.0 → broccoli_ml-0.4.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "broccoli-ml"
-version = "0.2.0"
+version = "0.4.0"
 description = "Some useful Pytorch models, circa 2025"
 authors = [
     {name = "Nicholas Bailey"}