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
|
"""CNN encoder for the VQ-VAE."""
from typing import Sequence, Optional, Tuple, Type
import attr
from torch import nn
from torch import Tensor
from text_recognizer.networks.util import activation_function
from text_recognizer.networks.vqvae.residual import Residual
@attr.s(eq=False)
class Encoder(nn.Module):
"""A CNN encoder network."""
in_channels: int = attr.ib()
out_channels: int = attr.ib()
res_channels: int = attr.ib()
num_residual_layers: int = attr.ib()
embedding_dim: int = attr.ib()
activation: str = attr.ib()
encoder: nn.Sequential = attr.ib(init=False)
def __attrs_post_init__(self) -> None:
"""Post init configuration."""
super().__init__()
self.encoder = self._build_compression_block()
def _build_compression_block(self) -> nn.Sequential:
activation_fn = activation_function(self.activation)
block = [
nn.Conv2d(
in_channels=self.in_channels,
out_channels=self.out_channels // 2,
kernel_size=4,
stride=2,
padding=1,
),
activation_fn,
nn.Conv2d(
in_channels=self.out_channels // 2,
out_channels=self.out_channels,
kernel_size=4,
stride=2,
padding=1,
),
activation_fn,
nn.Conv2d(
in_channels=self.out_channels,
out_channels=self.out_channels,
kernel_size=3,
padding=1,
),
]
for _ in range(self.num_residual_layers):
block.append(
Residual(in_channels=self.out_channels, out_channels=self.res_channels)
)
block.append(
nn.Conv2d(
in_channels=self.out_channels,
out_channels=self.embedding_dim,
kernel_size=1,
)
)
return nn.Sequential(*block)
def forward(self, x: Tensor) -> Tuple[Tensor, Tensor]:
"""Encodes input into a discrete representation."""
return self.encoder(x)
|
