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
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
|
"""Vision transformer for character recognition."""
import math
from typing import Tuple, Type
import attr
import torch
from torch import nn, Tensor
from text_recognizer.data.mappings import AbstractMapping
from text_recognizer.networks.base import BaseNetwork
from text_recognizer.networks.encoders.efficientnet import EfficientNet
from text_recognizer.networks.transformer.layers import Decoder
from text_recognizer.networks.transformer.positional_encodings import (
PositionalEncoding,
PositionalEncoding2D,
)
@attr.s(auto_attribs=True)
class ConvTransformer(BaseNetwork):
# Parameters and placeholders,
input_dims: Tuple[int, int, int] = attr.ib()
hidden_dim: int = attr.ib()
dropout_rate: float = attr.ib()
max_output_len: int = attr.ib()
num_classes: int = attr.ib()
start_token: str = attr.ib()
start_index: Tensor = attr.ib(init=False)
end_token: str = attr.ib()
end_index: Tensor = attr.ib(init=False)
pad_token: str = attr.ib()
pad_index: Tensor = attr.ib(init=False)
# Modules.
encoder: EfficientNet = attr.ib()
decoder: Decoder = attr.ib()
mapping: Type[AbstractMapping] = attr.ib()
latent_encoder: nn.Sequential = attr.ib(init=False)
token_embedding: nn.Embedding = attr.ib(init=False)
token_pos_encoder: PositionalEncoding = attr.ib(init=False)
head: nn.Linear = attr.ib(init=False)
def __attrs_post_init__(self) -> None:
"""Post init configuration."""
self.start_index = self.mapping.get_index(self.start_token)
self.end_index = self.mapping.get_index(self.end_token)
self.pad_index = self.mapping.get_index(self.pad_token)
# Latent projector for down sampling number of filters and 2d
# positional encoding.
self.latent_encoder = nn.Sequential(
nn.Conv2d(
in_channels=self.encoder.out_channels,
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_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)
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
def predict(self, x: Tensor) -> Tensor:
"""Predicts text in image.
Args:
x (Tensor): Image(s) to extract text from.
Shapes:
- x: :math: `(B, H, W)`
- output: :math: `(B, S)`
Returns:
Tensor: A tensor of token indices of the predictions from the model.
"""
bsz = x.shape[0]
# Encode image(s) to latent vectors.
z = self.encode(x)
# Create a placeholder matrix for storing outputs from the network
output = torch.ones((bsz, self.max_output_len), dtype=torch.long).to(x.device)
output[:, 0] = self.start_index
for i in range(1, self.max_output_len):
context = output[:, :i] # (bsz, i)
logits = self.decode(z, context) # (i, bsz, c)
tokens = torch.argmax(logits, dim=-1) # (i, bsz)
output[:, i : i + 1] = tokens[-1:]
# Early stopping of prediction loop if token is end or padding token.
if (
output[:, i - 1] == self.end_index | output[: i - 1] == self.pad_index
).all():
break
# Set all tokens after end token to pad token.
for i in range(1, self.max_output_len):
idx = (
output[:, i - 1] == self.end_index | output[:, i - 1] == self.pad_index
)
output[idx, i] = self.pad_index
return output
|
