From 2c5093dda783cf7618a8554ae649d32b92b84b4c Mon Sep 17 00:00:00 2001
From: Gustaf Rydholm <gustaf.rydholm@gmail.com>
Date: Mon, 15 Apr 2024 21:50:08 +0200
Subject: Update networks

---
 text_recognizer/network/cvit.py | 44 +++++++++++++++++++++++++++++++++++++++++
 1 file changed, 44 insertions(+)
 create mode 100644 text_recognizer/network/cvit.py

(limited to 'text_recognizer/network/cvit.py')

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)
-- 
cgit v1.2.3-70-g09d2