2 files changed, 50 insertions, 10 deletions
diff --git a/text_recognizer/network/convformer.py b/text_recognizer/network/convformer.py
index 0ee5487..e2b0204 100644
--- a/text_recognizer/network/convformer.py
+++ b/text_recognizer/network/convformer.py
@@ -1,12 +1,10 @@
-from typing import Optional
 
 from einops.layers.torch import Rearrange
 from torch import Tensor, nn
 
-from text_recognizer.network.convnext.convnext import ConvNext
-from .transformer.embedding.token import TokenEmbedding
-from .transformer.embedding.sincos import sincos_2d
 from .transformer.decoder import Decoder
+from .transformer.embedding.sincos import sincos_2d
+from .transformer.embedding.token import TokenEmbedding
 from .transformer.encoder import Encoder
 
 
@@ -24,12 +22,10 @@ class Convformer(nn.Module):
         token_embedding: TokenEmbedding,
         tie_embeddings: bool,
         pad_index: int,
-        stem: Optional[ConvNext] = None,
         channels: int = 1,
     ) -> None:
         super().__init__()
         patch_dim = patch_height * patch_width * channels
-        self.stem = stem if stem is not None else nn.Identity()
         self.to_patch_embedding = nn.Sequential(
             Rearrange(
                 "b c (h ph) (w pw) -> b (h w) (ph pw c)",
@@ -53,11 +49,11 @@ class Convformer(nn.Module):
         self.decoder = decoder
         self.pad_index = pad_index
 
-    def encode(self, img: Tensor) -> Tensor:
-        x = self.stem(img)
+    def encode(self, images: Tensor) -> Tensor:
+        x = self.encoder(images)
         x = self.to_patch_embedding(x)
-        x += self.patch_embedding.to(img.device, dtype=img.dtype)
-        return self.encoder(x)
+        x = x + self.patch_embedding.to(images.device, dtype=images.dtype)
+        return x
 
     def decode(self, text: Tensor, img_features: Tensor) -> Tensor:
         text = text.long()
diff --git a/text_recognizer/network/cvit.py b/text_recognizer/network/cvit.py
new file mode 100644
index 0000000..f9abb8c
--- /dev/null
+++ b/text_recognizer/network/cvit.py
@@ -0,0 +1,44 @@
+from einops.layers.torch import Rearrange
+from torch import Tensor, nn
+
+from text_recognizer.network.convnext.convnext import ConvNext
+
+from .transformer.embedding.sincos import sincos_2d
+from .transformer.encoder import Encoder
+
+
+class CVit(nn.Module):
+    def __init__(
+        self,
+        image_height: int,
+        image_width: int,
+        patch_height: int,
+        patch_width: int,
+        dim: int,
+        encoder: Encoder,
+        stem: ConvNext,
+        channels: int = 1,
+    ) -> None:
+        super().__init__()
+        patch_dim = patch_height * patch_width * channels
+        self.stem = stem
+        self.to_patch_embedding = nn.Sequential(
+            Rearrange(
+                "b c (h ph) (w pw) -> b (h w) (ph pw c)",
+                ph=patch_height,
+                pw=patch_width,
+            ),
+            nn.LayerNorm(patch_dim),
+            nn.Linear(patch_dim, dim),
+            nn.LayerNorm(dim),
+        )
+        self.patch_embedding = sincos_2d(
+            h=image_height // patch_height, w=image_width // patch_width, dim=dim
+        )
+        self.encoder = encoder
+
+    def forward(self, img: Tensor) -> Tensor:
+        x = self.stem(img)
+        x = self.to_patch_embedding(x)
+        x += self.patch_embedding.to(img.device, dtype=img.dtype)
+        return self.encoder(x)