From 617c63a9ac1b10b2f093d115eef5d63d87a03658 Mon Sep 17 00:00:00 2001
From: Gustaf Rydholm <gustaf.rydholm@gmail.com>
Date: Wed, 27 Oct 2021 22:15:15 +0200
Subject: Add local attn in transformer layer

---
 text_recognizer/networks/transformer/layers.py | 47 +++++++++++++++++++++-----
 1 file changed, 38 insertions(+), 9 deletions(-)

(limited to 'text_recognizer/networks')

diff --git a/text_recognizer/networks/transformer/layers.py b/text_recognizer/networks/transformer/layers.py
index 2b8427d..b132522 100644
--- a/text_recognizer/networks/transformer/layers.py
+++ b/text_recognizer/networks/transformer/layers.py
@@ -1,7 +1,8 @@
 """Transformer attention layer."""
-from typing import Any, Dict, Optional, Tuple
+from typing import Any, Optional, Tuple, Type
 
 import attr
+from omegaconf.dictconfig import DictConfig
 from torch import nn, Tensor
 
 from text_recognizer.networks.transformer.embeddings.rotary import RotaryEmbedding
@@ -14,20 +15,24 @@ class AttentionLayers(nn.Module):
     """Standard transfomer layer."""
 
     def __attrs_pre_init__(self) -> None:
+        """Pre init constructor."""
         super().__init__()
 
     dim: int = attr.ib()
     depth: int = attr.ib()
     num_heads: int = attr.ib()
     attn_fn: str = attr.ib()
-    attn_kwargs: Dict = attr.ib()
+    attn_kwargs: DictConfig = attr.ib()
     norm_fn: str = attr.ib()
     ff_fn: str = attr.ib()
-    ff_kwargs: Dict = attr.ib()
+    ff_kwargs: DictConfig = attr.ib()
     rotary_emb: Optional[RotaryEmbedding] = attr.ib()
+    local_attn_fn: Optional[str] = attr.ib(default=None)
+    local_attn_kwargs: Optional[DictConfig] = attr.ib(default=None)
     causal: bool = attr.ib(default=False)
     cross_attend: bool = attr.ib(default=False)
     pre_norm: bool = attr.ib(default=True)
+    local_depth: Optional[int] = attr.ib(default=None)
     layer_types: Tuple[str, ...] = attr.ib(init=False)
     layers: nn.ModuleList = attr.ib(init=False)
 
@@ -36,25 +41,48 @@ class AttentionLayers(nn.Module):
         self.layer_types = self._get_layer_types() * self.depth
         self.layers = self._build_network()
 
+        if self.local_attn_kwargs is not None and self.local_attn_fn is not None:
+            if "depth" not in self.local_attn_kwargs:
+                ValueError("Local depth has to be specified")
+            self.local_depth = self.local_attn_kwargs.pop("depth")
+
     def _get_layer_types(self) -> Tuple:
         """Get layer specification."""
         if self.cross_attend:
             return "a", "c", "f"
         return "a", "f"
 
+    def _configure_causal_attn(self, i: int) -> Type[nn.Module]:
+        if self.local_depth is not None and i <= self.local_depth:
+            return load_partial_fn(
+                self.local_attn_fn,
+                dim=self.dim,
+                num_heads=self.num_heads,
+                **self.local_attn_kwargs,
+            )()
+        return load_partial_fn(
+            self.attn_fn,
+            causal=self.causal,
+            dim=self.dim,
+            num_heads=self.num_heads,
+            **self.attn_kwargs,
+        )()
+
     def _build_network(self) -> nn.ModuleList:
         """Configures transformer network."""
-        attn = load_partial_fn(
-            self.attn_fn, dim=self.dim, num_heads=self.num_heads, **self.attn_kwargs
-        )
         norm = load_partial_fn(self.norm_fn, normalized_shape=self.dim)
         ff = load_partial_fn(self.ff_fn, dim=self.dim, **self.ff_kwargs)
         layers = nn.ModuleList([])
-        for layer_type in self.layer_types:
+        for i, layer_type in enumerate(self.layer_types):
             if layer_type == "a":
-                layer = attn(causal=self.causal)
+                layer = self._configure_causal_attn(i)
             elif layer_type == "c":
-                layer = attn()
+                layer = load_partial_fn(
+                    self.attn_fn,
+                    dim=self.dim,
+                    num_heads=self.num_heads,
+                    **self.attn_kwargs,
+                )()
             elif layer_type == "f":
                 layer = ff()
             residual_fn = Residual()
@@ -68,6 +96,7 @@ class AttentionLayers(nn.Module):
         mask: Optional[Tensor] = None,
         context_mask: Optional[Tensor] = None,
     ) -> Tensor:
+        """Forward pass."""
         rotary_pos_emb = self.rotary_emb(x) if self.rotary_emb is not None else None
         for i, (layer_type, (norm, block, residual_fn)) in enumerate(
             zip(self.layer_types, self.layers)
-- 
cgit v1.2.3-70-g09d2