Add axial transformer

4 files changed, 205 insertions, 0 deletions

diff --git a/text_recognizer/networks/transformer/axial_attention/__init__.py b/text_recognizer/networks/transformer/axial_attention/__init__.py
new file mode 100644
index 0000000..e69de29
--- /dev/null
+++ b/text_recognizer/networks/transformer/axial_attention/__init__.py
diff --git a/text_recognizer/networks/transformer/axial_attention/encoder.py b/text_recognizer/networks/transformer/axial_attention/encoder.py
new file mode 100644
index 0000000..3082bd6
--- /dev/null
+++ b/text_recognizer/networks/transformer/axial_attention/encoder.py
@@ -0,0 +1,80 @@
+"""Axial transformer encoder."""
+
+from typing import List
+
+import attr
+from torch import nn, Tensor
+
+from text_recognizer.networks.transformer.axial_attention.self_attention import (
+    SelfAttention,
+)
+from text_recognizer.networks.transformer.axial_attention.utils import (
+    calculate_permutations,
+    PermuteToForm,
+    Sequential,
+)
+from text_recognizer.networks.transformer.norm import PreNorm
+
+
+@attr.s(eq=False)
+class AxialEncoder(nn.Module):
+    """Axial transfomer encoder."""
+
+    def __attrs_pre_init__(self) -> None:
+        super().__init__()
+
+    shape: List[int] = attr.ib()
+    dim: int = attr.ib()
+    depth: int = attr.ib()
+    heads: int = attr.ib()
+    dim_head: int = attr.ib()
+    dim_index: int = attr.ib()
+    fn: nn.Sequential = attr.ib(init=False)
+
+    def __attrs_post_init__(self) -> None:
+        self._build()
+
+    def _build(self) -> None:
+        permutations = calculate_permutations(2, self.dim_index)
+        get_ff = lambda: nn.Sequential(
+            nn.LayerNorm([self.dim, *self.shape]),
+            nn.Conv2d(
+                in_channels=self.dim,
+                out_channels=4 * self.dim,
+                kernel_size=3,
+                padding=1,
+            ),
+            nn.Mish(inplace=True),
+            nn.Conv2d(
+                in_channels=4 * self.dim,
+                out_channels=self.dim,
+                kernel_size=3,
+                padding=1,
+            ),
+        )
+
+        layers = nn.ModuleList([])
+        for _ in range(self.depth):
+            attns = nn.ModuleList(
+                [
+                    PermuteToForm(
+                        permutation=permutation,
+                        fn=PreNorm(
+                            self.dim,
+                            SelfAttention(
+                                dim=self.dim, heads=self.heads, dim_head=self.dim_head
+                            ),
+                        ),
+                    )
+                    for permutation in permutations
+                ]
+            )
+            convs = nn.ModuleList([get_ff(), get_ff()])
+            layers.append(attns)
+            layers.append(convs)
+
+        self.fn = Sequential(layers)
+
+    def forward(self, x: Tensor) -> Tensor:
+        """Applies fn to input."""
+        return self.fn(x)
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..ba162be
--- /dev/null
+++ b/text_recognizer/networks/transformer/axial_attention/self_attention.py
@@ -0,0 +1,45 @@
+"""Axial self attention module."""
+
+import attr
+import torch
+from torch import nn
+from torch import Tensor
+
+
+@attr.s(eq=False)
+class SelfAttention(nn.Module):
+    """Axial self attention module."""
+
+    def __attrs_pre_init__(self) -> None:
+        super().__init__()
+
+    dim: int = attr.ib()
+    dim_head: int = attr.ib()
+    heads: int = attr.ib()
+    dim_hidden: int = attr.ib(init=False)
+    to_q: nn.Linear = attr.ib(init=False)
+    to_kv: nn.Linear = attr.ib(init=False)
+    to_out: nn.Linear = attr.ib(init=False)
+
+    def __attrs_post_init__(self) -> None:
+        self.dim_hidden = self.heads * self.dim_head
+        self.to_q = nn.Linear(self.dim, self.dim_hidden, bias=False)
+        self.to_kv = nn.Linear(self.dim, 2 * self.dim_hidden, bias=False)
+        self.to_out = nn.Linear(self.dim_hidden, self.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)
diff --git a/text_recognizer/networks/transformer/axial_attention/utils.py b/text_recognizer/networks/transformer/axial_attention/utils.py
new file mode 100644
index 0000000..534ef4e
--- /dev/null
+++ b/text_recognizer/networks/transformer/axial_attention/utils.py
@@ -0,0 +1,80 @@
+"""Helper functions for axial attention."""
+from operator import itemgetter
+from typing import Callable, List, Tuple
+
+import attr
+from torch import nn, Tensor
+
+
+def _map_el_ind(arr: Tensor, ind: int) -> List:
+    return list(map(itemgetter(ind), arr))
+
+
+def _sort_indices(arr: Tensor) -> Tuple[List[int], List[int]]:
+    indices = [i for i in range(len(arr))]
+    arr = zip(arr, indices)
+    arr = sorted(arr)
+    return _map_el_ind(arr, 0), _map_el_ind(arr, 1)
+
+
+def calculate_permutations(num_dims: int, emb_dim: int) -> List[List[int]]:
+    """Returns permutations of tensor."""
+    total_dims = num_dims + 2
+    axial_dims = [i for i in range(1, total_dims) if i != emb_dim]
+
+    permutations = []
+
+    for axial_dim in axial_dims:
+        last_two_dims = [axial_dim, emb_dim]
+        dims_rest = set(range(0, total_dims)) - set(last_two_dims)
+        permutation = [*dims_rest, *last_two_dims]
+        permutations.append(permutation)
+
+    return permutations
+
+
+@attr.s(eq=False)
+class PermuteToForm(nn.Module):
+    """Helper class for applying axial attention."""
+
+    def __attrs_pre_init__(self) -> None:
+        super().__init__()
+
+    fn: Callable = attr.ib()
+    permutation: List[List[int]] = attr.ib()
+    inv_permutation: List[List[int]] = attr.ib(init=False)
+
+    def __attrs_post_init__(self) -> None:
+        _, self.inv_permutation = _sort_indices(self.permutation)
+
+    def forward(self, x: Tensor) -> Tensor:
+        """Permutes tensor, applies axial attention, permutes tensor back."""
+        x = x.permute(*self.permutation).contiguous()
+        shape = x.shape
+        *_, t, d = shape
+
+        # Merge all but axial dimension
+        x = x.reshape(-1, t, d)
+
+        # Apply attention
+        x = self.fn(x)
+
+        # Restore original shape and permutation
+        x = x.reshape(*shape)
+        x = x.permute(*self.inv_permutation).contiguous()
+        return x
+
+
+class Sequential(nn.Module):
+    """Applies a list of paired functions to input."""
+
+    def __init__(self, fns: nn.ModuleList) -> None:
+        super().__init__()
+        self.fns = fns
+
+    def forward(self, x: Tensor) -> Tensor:
+        """Applies blocks to input."""
+        for f, g in self.fns:
+            x = x + f(x)
+            x = x + g(x)
+        return x