Reformatting of positional encodings and ViT working

1 files changed, 16 insertions, 0 deletions

diff --git a/text_recognizer/networks/transformer/positional_encodings/absolute_embedding.py b/text_recognizer/networks/transformer/positional_encodings/absolute_embedding.py
new file mode 100644
index 0000000..9466f6e
--- /dev/null
+++ b/text_recognizer/networks/transformer/positional_encodings/absolute_embedding.py
@@ -0,0 +1,16 @@
+"""Absolute positional embedding."""
+from torch import nn, Tensor
+
+
+class AbsolutePositionalEmbedding(nn.Module):
+    def __init__(self, dim: int, max_seq_len: int) -> None:
+        super().__init__()
+        self.emb = nn.Embedding(max_seq_len, dim)
+        self._weight_init()
+
+    def _weight_init(self) -> None:
+        nn.init.normal_(self.emb.weight, std=0.02)
+
+    def forward(self, x: Tensor) -> Tensor:
+        n = torch.arange(x.shape[1], device=x.device)
+        return self.emb(n)[None, :, :]