From 49ca6ade1a19f7f9c702171537fe4be0dfcda66d Mon Sep 17 00:00:00 2001
From: Gustaf Rydholm <gustaf.rydholm@gmail.com>
Date: Fri, 25 Aug 2023 23:19:14 +0200
Subject: Rename and add flash atten

---
 text_recognizer/network/convnext/attention.py | 79 +++++++++++++++++++++++++++
 1 file changed, 79 insertions(+)
 create mode 100644 text_recognizer/network/convnext/attention.py

(limited to 'text_recognizer/network/convnext/attention.py')

diff --git a/text_recognizer/network/convnext/attention.py b/text_recognizer/network/convnext/attention.py
new file mode 100644
index 0000000..6bc9692
--- /dev/null
+++ b/text_recognizer/network/convnext/attention.py
@@ -0,0 +1,79 @@
+"""Convolution self attention block."""
+
+import torch.nn.functional as F
+from einops import rearrange
+from torch import Tensor, einsum, nn
+
+from text_recognizer.network.convnext.norm import LayerNorm
+from text_recognizer.network.convnext.residual import Residual
+
+
+def l2norm(t: Tensor) -> Tensor:
+    return F.normalize(t, dim=-1)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim: int, mult: int = 4) -> None:
+        super().__init__()
+        inner_dim = int(dim * mult)
+        self.fn = Residual(
+            nn.Sequential(
+                LayerNorm(dim),
+                nn.Conv2d(dim, inner_dim, 1, bias=False),
+                nn.GELU(),
+                LayerNorm(inner_dim),
+                nn.Conv2d(inner_dim, dim, 1, bias=False),
+            )
+        )
+
+    def forward(self, x: Tensor) -> Tensor:
+        return self.fn(x)
+
+
+class Attention(nn.Module):
+    def __init__(
+        self, dim: int, heads: int = 4, dim_head: int = 64, scale: int = 8
+    ) -> None:
+        super().__init__()
+        self.scale = scale
+        self.heads = heads
+        inner_dim = heads * dim_head
+        self.norm = LayerNorm(dim)
+
+        self.to_qkv = nn.Conv2d(dim, inner_dim * 3, 1, bias=False)
+        self.to_out = nn.Conv2d(inner_dim, dim, 1, bias=False)
+
+    def forward(self, x: Tensor) -> Tensor:
+        h, w = x.shape[-2:]
+
+        residual = x.clone()
+
+        x = self.norm(x)
+
+        q, k, v = self.to_qkv(x).chunk(3, dim=1)
+        q, k, v = map(
+            lambda t: rearrange(t, "b (h c) ... -> b h (...) c", h=self.heads),
+            (q, k, v),
+        )
+
+        q, k = map(l2norm, (q, k))
+
+        sim = einsum("b h i d, b h j d -> b h i j", q, k) * self.scale
+        attn = sim.softmax(dim=-1)
+
+        out = einsum("b h i j, b h j d -> b h i d", attn, v)
+
+        out = rearrange(out, "b h (x y) d -> b (h d) x y", x=h, y=w)
+        return self.to_out(out) + residual
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, attn: Attention, ff: FeedForward) -> None:
+        super().__init__()
+        self.attn = attn
+        self.ff = ff
+
+    def forward(self, x: Tensor) -> Tensor:
+        x = self.attn(x)
+        x = self.ff(x)
+        return x
-- 
cgit v1.2.3-70-g09d2