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 | @registry.factory.register("box-text-embedding")
class BoxTextEmbedding(Module):
"""
A module that embeds the textual features of the blocks
"""
def __init__(
self,
pooler: Module,
size: Optional[int] = None,
):
"""
Parameters
----------
size: int
Size of the output box embedding
pooler: Dict
The module used to encode the textual features of the blocks
"""
super().__init__()
self.size = size
self.shape_voc = Vocabulary(["__unk__"], default=0)
self.prefix_voc = Vocabulary(["__unk__"], default=0)
self.suffix_voc = Vocabulary(["__unk__"], default=0)
self.norm_voc = Vocabulary(["__unk__"], default=0)
self.shape_embedding = None
self.prefix_embedding = None
self.suffix_embedding = None
self.norm_embedding = None
self.hpos_embedding = None
self.vpos_embedding = None
self.first_page_embedding = None
self.last_page_embedding = None
self.pooler = pooler
punct = "[:punct:]" + "\"'ˊ＂〃ײ᳓″״‶˶ʺ“”˝"
num_like = r"\d+(?:[.,]\d+)?"
default = rf"[^\d{punct}'\n[[:space:]]+(?:['ˊ](?=[[:alpha:]]|$))?"
self.word_regex = regex.compile(
rf"({num_like}|[{punct}]|[\n\r\t]|[^\S\r\n\t]+|{default})([^\S\r\n\t])?"
)
@property
def output_size(self):
return self.size
def initialize(self, gold_data, size: int = None, **kwargs):
super().initialize(gold_data, size=size, **kwargs)
self.pooler.initialize(gold_data, input_size=size)
shape_init = self.shape_voc.initialization()
prefix_init = self.prefix_voc.initialization()
suffix_init = self.suffix_voc.initialization()
norm_init = self.norm_voc.initialization()
with shape_init, prefix_init, suffix_init, norm_init: # noqa: E501
with self.no_cache():
for doc in gold_data:
self.preprocess(doc, supervision=True)
self.shape_embedding = torch.nn.Embedding(len(self.shape_voc), self.size)
self.prefix_embedding = torch.nn.Embedding(len(self.prefix_voc), self.size)
self.suffix_embedding = torch.nn.Embedding(len(self.suffix_voc), self.size)
self.norm_embedding = torch.nn.Embedding(len(self.norm_voc), self.size)
def preprocess(self, doc, supervision: bool = False):
text_boxes = doc.lines
tokens_shape = [[] for _ in text_boxes]
tokens_prefix = [[] for _ in text_boxes]
tokens_suffix = [[] for _ in text_boxes]
tokens_norm = [[] for _ in text_boxes]
for i, b in enumerate(text_boxes):
words = [m.group(0) for m in self.word_regex.finditer(b.text)]
for word in words:
# ascii_str = unidecode(word)
ascii_str = word
tokens_shape[i].append(self.shape_voc.encode(word_shape(ascii_str)))
tokens_prefix[i].append(self.prefix_voc.encode(ascii_str.lower()[:3]))
tokens_suffix[i].append(self.suffix_voc.encode(ascii_str.lower()[-3:]))
tokens_norm[i].append(self.norm_voc.encode(ascii_str.lower()))
return {
"tokens_shape": tokens_shape,
"tokens_prefix": tokens_prefix,
"tokens_suffix": tokens_suffix,
"tokens_norm": tokens_norm,
}
def collate(self, batch, device: torch.device):
shapes = pad_2d(flatten(batch["tokens_shape"]), pad=-1, device=device)
mask = shapes != -1
shapes[~mask] = 0
return {
"tokens_shape": shapes,
"tokens_prefix": pad_2d(
flatten(batch["tokens_prefix"]), pad=0, device=device
),
"tokens_suffix": pad_2d(
flatten(batch["tokens_suffix"]), pad=0, device=device
),
"tokens_norm": pad_2d(flatten(batch["tokens_norm"]), pad=0, device=device),
"tokens_mask": mask,
}
def forward(self, batch, supervision=False):
text_embeds = self.pooler(
embeds=(
self.shape_embedding(batch["tokens_shape"])
+ self.prefix_embedding(batch["tokens_prefix"])
+ self.suffix_embedding(batch["tokens_suffix"])
+ self.norm_embedding(batch["tokens_norm"])
),
mask=batch["tokens_mask"],
)
return text_embeds
|