summaryrefslogtreecommitdiff
path: root/text_recognizer/networks/conv_transformer.py
blob: b1a101e38cad6cc479de2833b21608688b3212f8 (plain)
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
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
"""Vision transformer for character recognition."""
import math
from typing import Tuple

from torch import nn, Tensor

from text_recognizer.networks.transformer.layers import Decoder
from text_recognizer.networks.transformer.positional_encodings import (
    PositionalEncoding,
    PositionalEncoding2D,
)


class ConvTransformer(nn.Module):
    """Convolutional encoder and transformer decoder network."""

    def __init__(
        self,
        input_dims: Tuple[int, int, int],
        encoder_dim: int,
        hidden_dim: int,
        dropout_rate: float,
        num_classes: int,
        pad_index: Tensor,
        encoder: nn.Module,
        decoder: Decoder,
    ) -> None:
        super().__init__()
        self.input_dims = input_dims
        self.encoder_dim = encoder_dim
        self.hidden_dim = hidden_dim
        self.dropout_rate = dropout_rate
        self.num_classes = num_classes
        self.pad_index = pad_index
        self.encoder = encoder
        self.decoder = decoder

        # Latent projector for down sampling number of filters and 2d
        # positional encoding.
        self.latent_encoder = nn.Sequential(
            nn.Conv2d(
                in_channels=self.encoder_dim,
                out_channels=self.hidden_dim,
                kernel_size=1,
            ),
            PositionalEncoding2D(
                hidden_dim=self.hidden_dim,
                max_h=self.input_dims[1],
                max_w=self.input_dims[2],
            ),
            nn.Flatten(start_dim=2),
        )

        # Token embedding.
        self.token_embedding = nn.Embedding(
            num_embeddings=self.num_classes, embedding_dim=self.hidden_dim
        )

        # Positional encoding for decoder tokens.
        self.token_pos_encoder = PositionalEncoding(
            hidden_dim=self.hidden_dim, dropout_rate=self.dropout_rate
        )
        # Head
        self.head = nn.Linear(
            in_features=self.hidden_dim, out_features=self.num_classes
        )

        # Initalize weights for encoder.
        self.init_weights()

    def init_weights(self) -> None:
        """Initalize weights for decoder network and head."""
        bound = 0.1
        self.token_embedding.weight.data.uniform_(-bound, bound)
        self.head.bias.data.zero_()
        self.head.weight.data.uniform_(-bound, bound)
        # TODO: Initalize encoder?

    def encode(self, x: Tensor) -> Tensor:
        """Encodes an image into a latent feature vector.

        Args:
            x (Tensor): Image tensor.

        Shape:
            - x: :math: `(B, C, H, W)`
            - z: :math: `(B, Sx, E)`

            where Sx is the length of the flattened feature maps projected from
            the encoder. E latent dimension for each pixel in the projected
            feature maps.

        Returns:
            Tensor: A Latent embedding of the image.
        """
        z = self.encoder(x)
        z = self.latent_encoder(z)

        # Permute tensor from [B, E, Ho * Wo] to [B, Sx, E]
        z = z.permute(0, 2, 1)
        return z

    def decode(self, z: Tensor, context: Tensor) -> Tensor:
        """Decodes latent images embedding into word pieces.

        Args:
            z (Tensor): Latent images embedding.
            context (Tensor): Word embeddings.

        Shapes:
            - z: :math: `(B, Sx, E)`
            - context: :math: `(B, Sy)`
            - out: :math: `(B, Sy, T)`

            where Sy is the length of the output and T is the number of tokens.

        Returns:
            Tensor: Sequence of word piece embeddings.
        """
        context = context.long()
        context_mask = context != self.pad_index
        context = self.token_embedding(context) * math.sqrt(self.hidden_dim)
        context = self.token_pos_encoder(context)
        out = self.decoder(x=context, context=z, mask=context_mask)
        logits = self.head(out)  # [B, Sy, T]
        logits = logits.permute(0, 2, 1)  # [B, T, Sy]
        return logits

    def forward(self, x: Tensor, context: Tensor) -> Tensor:
        """Encodes images into word piece logtis.

        Args:
            x (Tensor): Input image(s).
            context (Tensor): Target word embeddings.

        Shapes:
            - x: :math: `(B, C, H, W)`
            - context: :math: `(B, Sy, T)`

            where B is the batch size, C is the number of input channels, H is
            the image height and W is the image width.

        Returns:
            Tensor: Sequence of logits.
        """
        z = self.encode(x)
        logits = self.decode(z, context)
        return logits