Add axial encoder

1 files changed, 40 insertions, 0 deletions

diff --git a/text_recognizer/networks/transformer/axial_attention/self_attention.py b/text_recognizer/networks/transformer/axial_attention/self_attention.py
new file mode 100644
index 0000000..b5e4142
--- /dev/null
+++ b/text_recognizer/networks/transformer/axial_attention/self_attention.py
@@ -0,0 +1,40 @@
+"""Axial self attention module."""
+
+import torch
+from torch import nn
+from torch import Tensor
+
+
+class SelfAttention(nn.Module):
+    """Axial self attention module."""
+
+    def __init__(
+        self,
+        dim: int,
+        dim_head: int,
+        heads: int,
+    ) -> None:
+        super().__init__()
+        self.dim_hidden = heads * dim_head
+        self.heads = heads
+        self.dim_head = dim_head
+        self.to_q = nn.Linear(dim, self.dim_hidden, bias=False)
+        self.to_kv = nn.Linear(dim, 2 * self.dim_hidden, bias=False)
+        self.to_out = nn.Linear(self.dim_hidden, dim)
+
+    def forward(self, x: Tensor) -> Tensor:
+        """Applies self attention."""
+        q, k, v = (self.to_q(x), *self.to_kv(x).chunk(2, dim=-1))
+        b, _, d, h, e = *q.shape, self.heads, self.dim_head
+
+        merge_heads = (
+            lambda x: x.reshape(b, -1, h, e).transpose(1, 2).reshape(b * h, -1, e)
+        )
+        q, k, v = map(merge_heads, (q, k, v))
+
+        energy = torch.einsum("bie,bje->bij", q, k) * (e ** -0.5)
+        energy = energy.softmax(dim=-1)
+        attn = torch.einsum("bij,bje->bie", energy, v)
+
+        out = attn.reshape(b, h, -1, e).transpose(1, 2).reshape(b, -1, d)
+        return self.to_out(out)