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
|
"""A Vision Transformer.
Inspired by:
https://openreview.net/pdf?id=YicbFdNTTy
"""
from typing import Optional, Tuple
from einops import rearrange, repeat
import torch
from torch import nn
from torch import Tensor
from text_recognizer.networks.transformer import Transformer
class ViT(nn.Module):
"""Transfomer for image to sequence prediction."""
def __init__(
self,
num_encoder_layers: int,
num_decoder_layers: int,
hidden_dim: int,
vocab_size: int,
num_heads: int,
expansion_dim: int,
patch_dim: Tuple[int, int],
image_size: Tuple[int, int],
dropout_rate: float,
trg_pad_index: int,
max_len: int,
activation: str = "gelu",
) -> None:
super().__init__()
self.trg_pad_index = trg_pad_index
self.patch_dim = patch_dim
self.num_patches = image_size[-1] // self.patch_dim[1]
# Encoder
self.patch_to_embedding = nn.Linear(
self.patch_dim[0] * self.patch_dim[1], hidden_dim
)
self.cls_token = nn.Parameter(torch.randn(1, 1, hidden_dim))
self.character_embedding = nn.Embedding(vocab_size, hidden_dim)
self.pos_embedding = nn.Parameter(torch.randn(1, max_len, hidden_dim))
self.dropout = nn.Dropout(dropout_rate)
self._init()
self.transformer = Transformer(
num_encoder_layers,
num_decoder_layers,
hidden_dim,
num_heads,
expansion_dim,
dropout_rate,
activation,
)
self.head = nn.Sequential(nn.Linear(hidden_dim, vocab_size),)
def _init(self) -> None:
nn.init.normal_(self.character_embedding.weight, std=0.02)
# nn.init.normal_(self.pos_embedding.weight, std=0.02)
def _create_trg_mask(self, trg: Tensor) -> Tensor:
# Move this outside the transformer.
trg_pad_mask = (trg != self.trg_pad_index)[:, None, None]
trg_len = trg.shape[1]
trg_sub_mask = torch.tril(
torch.ones((trg_len, trg_len), device=trg.device)
).bool()
trg_mask = trg_pad_mask & trg_sub_mask
return trg_mask
def encoder(self, src: Tensor) -> Tensor:
"""Forward pass with the encoder of the transformer."""
return self.transformer.encoder(src)
def decoder(self, trg: Tensor, memory: Tensor, trg_mask: Tensor) -> Tensor:
"""Forward pass with the decoder of the transformer + classification head."""
return self.head(
self.transformer.decoder(trg=trg, memory=memory, trg_mask=trg_mask)
)
def extract_image_features(self, src: Tensor) -> Tensor:
"""Extracts image features with a backbone neural network.
It seem like the winning idea was to swap channels and width dimension and collapse
the height dimension. The transformer is learning like a baby with this implementation!!! :D
Ohhhh, the joy I am experiencing right now!! Bring in the beers! :D :D :D
Args:
src (Tensor): Input tensor.
Returns:
Tensor: A input src to the transformer.
"""
# If batch dimension is missing, it needs to be added.
if len(src.shape) < 4:
src = src[(None,) * (4 - len(src.shape))]
patches = rearrange(
src,
"b c (h p1) (w p2) -> b (h w) (p1 p2 c)",
p1=self.patch_dim[0],
p2=self.patch_dim[1],
)
# From patches to encoded sequence.
x = self.patch_to_embedding(patches)
b, n, _ = x.shape
cls_tokens = repeat(self.cls_token, "() n d -> b n d", b=b)
x = torch.cat((cls_tokens, x), dim=1)
x += self.pos_embedding[:, : (n + 1)]
x = self.dropout(x)
return x
def target_embedding(self, trg: Tensor) -> Tuple[Tensor, Tensor]:
"""Encodes target tensor with embedding and postion.
Args:
trg (Tensor): Target tensor.
Returns:
Tuple[Tensor, Tensor]: Encoded target tensor and target mask.
"""
_, n = trg.shape
trg = self.character_embedding(trg.long())
trg += self.pos_embedding[:, :n]
return trg
def decode_image_features(self, h: Tensor, trg: Optional[Tensor] = None) -> Tensor:
"""Takes images features from the backbone and decodes them with the transformer."""
trg_mask = self._create_trg_mask(trg)
trg = self.target_embedding(trg)
out = self.transformer(h, trg, trg_mask=trg_mask)
logits = self.head(out)
return logits
def forward(self, x: Tensor, trg: Optional[Tensor] = None) -> Tensor:
"""Forward pass with CNN transfomer."""
h = self.extract_image_features(x)
logits = self.decode_image_features(h, trg)
return logits
|
