Add subsampler layer

3 files changed, 56 insertions, 1 deletions

diff --git a/text_recognizer/networks/conformer/__init__.py b/text_recognizer/networks/conformer/__init__.py
index 5f3c7b5..1886f85 100644
--- a/text_recognizer/networks/conformer/__init__.py
+++ b/text_recognizer/networks/conformer/__init__.py
@@ -3,3 +3,4 @@ from text_recognizer.networks.conformer.ff import Feedforward
 from text_recognizer.networks.conformer.glu import GLU
 from text_recognizer.networks.conformer.conformer import Conformer
 from text_recognizer.networks.conformer.conv import ConformerConv
+from text_recognizer.networks.conformer.subsampler import Subsampler
diff --git a/text_recognizer/networks/conformer/conformer.py b/text_recognizer/networks/conformer/conformer.py
index d56955e..8d0e98e 100644
--- a/text_recognizer/networks/conformer/conformer.py
+++ b/text_recognizer/networks/conformer/conformer.py
@@ -1,5 +1,6 @@
 """Conformer module."""
 from copy import deepcopy
+from typing import Type
 
 from torch import nn, Tensor
 
@@ -7,11 +8,18 @@ from text_recognizer.networks.conformer.block import ConformerBlock
 
 
 class Conformer(nn.Module):
-    def __init__(self, block: ConformerBlock, depth: int) -> None:
+    def __init__(
+        self,
+        subsampler: Type[nn.Module],
+        block: ConformerBlock,
+        depth: int,
+    ) -> None:
         super().__init__()
+        self.subsampler = subsampler
         self.blocks = nn.ModuleList([deepcopy(block) for _ in range(depth)])
 
     def forward(self, x: Tensor) -> Tensor:
+        x = self.subsampler(x)
         for fn in self.blocks:
             x = fn(x)
         return x
diff --git a/text_recognizer/networks/conformer/subsampler.py b/text_recognizer/networks/conformer/subsampler.py
new file mode 100644
index 0000000..2bc0445
--- /dev/null
+++ b/text_recognizer/networks/conformer/subsampler.py
@@ -0,0 +1,46 @@
+"""Simple convolutional network."""
+from typing import Tuple
+
+from torch import nn, Tensor
+
+from text_recognizer.networks.transformer import (
+    AxialPositionalEmbedding,
+)
+
+
+class Subsampler(nn.Module):
+    def __init__(
+        self,
+        channels: int,
+        depth: int,
+        pixel_pos_embedding: AxialPositionalEmbedding,
+        dropout: float = 0.1,
+    ) -> None:
+        super().__init__()
+        self.pixel_pos_embedding = pixel_pos_embedding
+        self.subsampler, self.projector = self._build(channels, depth, dropout)
+
+    def _build(
+        self, channels: int, depth: int, dropout: float
+    ) -> Tuple[nn.Sequential, nn.Sequential]:
+        subsampler = []
+        for i in range(depth):
+            subsampler.append(
+                nn.Conv2d(
+                    in_channels=1 if i == 0 else channels,
+                    out_channels=channels,
+                    kernel_size=3,
+                    stride=2,
+                )
+            )
+            subsampler.append(nn.Mish(inplace=True))
+        projector = nn.Sequential(
+            nn.Flatten(start_dim=2), nn.Linear(channels, channels), nn.Dropout(dropout)
+        )
+        return nn.Sequential(*subsampler), projector
+
+    def forward(self, x: Tensor) -> Tensor:
+        x = self.subsampler(x)
+        x = self.pixel_pos_embedding(x)
+        x = self.projector(x)
+        return x.permute(0, 2, 1)