1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
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)
|
