Working on attention layer configuration

2 files changed, 25 insertions, 2 deletions

diff --git a/text_recognizer/networks/transformer/attention_layers.py b/text_recognizer/networks/transformer/attention_layers.py
new file mode 100644
index 0000000..721fa27
--- /dev/null
+++ b/text_recognizer/networks/transformer/attention_layers.py
@@ -0,0 +1,19 @@
+"""Generates the attention layer architecture."""
+from typing import Type
+
+import torch
+from torch import nn, Tensor
+
+
+class AttentionLayers(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        depth: int,
+        num_heads: int,
+        norm_layer: Type[nn.Module],
+        causal: bool = False,
+        cross_attend: bool = False,
+        only_cross: bool = False,
+    ) -> None:
+        pass
diff --git a/text_recognizer/networks/transformer/nystromer/nystromer.py b/text_recognizer/networks/transformer/nystromer/nystromer.py
index 7cc889e..799a811 100644
--- a/text_recognizer/networks/transformer/nystromer/nystromer.py
+++ b/text_recognizer/networks/transformer/nystromer/nystromer.py
@@ -26,7 +26,8 @@ class Nystromer(nn.Module):
         residual: bool = True,
         residual_conv_kernel: int = 33,
         dropout_rate: float = 0.0,
-    ):
+        glu: bool = True,
+    ) -> None:
         super().__init__()
         self.layers = nn.ModuleList(
             [
@@ -45,7 +46,10 @@ class Nystromer(nn.Module):
                                 dropout_rate=dropout_rate,
                             ),
                         ),
-                        PreNorm(dim, FeedForward(dim=dim, dropout_rate=dropout_rate)),
+                        PreNorm(
+                            dim,
+                            FeedForward(dim=dim, glu=glu, dropout_rate=dropout_rate),
+                        ),
                     ]
                 )
                 for _ in range(depth)