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
|
"""Text decoder."""
import torch
from torch import Tensor, nn
from text_recognizer.networks.transformer.decoder import Decoder
class TextDecoder(nn.Module):
"""Decodes images to token logits."""
def __init__(
self,
dim: int,
num_classes: int,
pad_index: Tensor,
decoder: Decoder,
) -> None:
super().__init__()
self.dim = dim
self.num_classes = num_classes
self.pad_index = pad_index
self.decoder = decoder
self.token_embedding = nn.Embedding(
num_embeddings=self.num_classes, embedding_dim=self.dim
)
self.to_logits = nn.Linear(in_features=self.dim, out_features=self.num_classes)
def forward(self, tokens: Tensor, img_features: Tensor) -> Tensor:
"""Decodes latent images embedding into word pieces.
Args:
tokens (Tensor): Token indecies.
img_features (Tensor): Latent images embedding.
Shapes:
- tokens: :math: `(B, Sy)`
- img_features: :math: `(B, Sx, D)`
- logits: :math: `(B, Sy, C)`
where Sy is the length of the output, C is the number of classes
and D is the hidden dimension.
Returns:
Tensor: Sequence of logits.
"""
tokens = tokens.long()
mask = tokens != self.pad_index
tokens = self.token_embedding(tokens)
tokens = self.decoder(x=tokens, context=img_features, mask=mask)
logits = (
tokens @ torch.transpose(self.token_embedding.weight.to(tokens.dtype), 0, 1)
).float()
logits = self.to_logits(tokens) # [B, Sy, C]
logits = logits.permute(0, 2, 1) # [B, C, Sy]
return logits
|
