Update Model/NER/VLSP2021/Predict_Ner.py
Browse files- Model/NER/VLSP2021/Predict_Ner.py +210 -210
Model/NER/VLSP2021/Predict_Ner.py
CHANGED
@@ -1,210 +1,210 @@
|
|
1 |
-
|
2 |
-
from vncorenlp import VnCoreNLP
|
3 |
-
|
4 |
-
from typing import Union
|
5 |
-
from transformers import AutoConfig, AutoTokenizer
|
6 |
-
from Model.NER.VLSP2021.Ner_CRF import PhoBertCrf,PhoBertSoftmax,PhoBertLstmCrf
|
7 |
-
import re
|
8 |
-
import os
|
9 |
-
import torch
|
10 |
-
import itertools
|
11 |
-
import numpy as np
|
12 |
-
|
13 |
-
MODEL_MAPPING = {
|
14 |
-
'vinai/phobert-base': {
|
15 |
-
'softmax': PhoBertSoftmax,
|
16 |
-
'crf': PhoBertCrf,
|
17 |
-
'lstm_crf': PhoBertLstmCrf
|
18 |
-
},
|
19 |
-
}
|
20 |
-
|
21 |
-
|
22 |
-
def normalize_text(txt: str) -> str:
|
23 |
-
# Remove special character
|
24 |
-
txt = re.sub("\xad|\u200b|\ufeff", "", txt)
|
25 |
-
# Normalize vietnamese accents
|
26 |
-
txt = re.sub(r"òa", "oà", txt)
|
27 |
-
txt = re.sub(r"óa", "oá", txt)
|
28 |
-
txt = re.sub(r"ỏa", "oả", txt)
|
29 |
-
txt = re.sub(r"õa", "oã", txt)
|
30 |
-
txt = re.sub(r"ọa", "oạ", txt)
|
31 |
-
txt = re.sub(r"òe", "oè", txt)
|
32 |
-
txt = re.sub(r"óe", "oé", txt)
|
33 |
-
txt = re.sub(r"ỏe", "oẻ", txt)
|
34 |
-
txt = re.sub(r"õe", "oẽ", txt)
|
35 |
-
txt = re.sub(r"ọe", "oẹ", txt)
|
36 |
-
txt = re.sub(r"ùy", "uỳ", txt)
|
37 |
-
txt = re.sub(r"úy", "uý", txt)
|
38 |
-
txt = re.sub(r"ủy", "uỷ", txt)
|
39 |
-
txt = re.sub(r"ũy", "uỹ", txt)
|
40 |
-
txt = re.sub(r"ụy", "uỵ", txt)
|
41 |
-
txt = re.sub(r"Ủy", "Uỷ", txt)
|
42 |
-
|
43 |
-
txt = re.sub(r'"', '”', txt)
|
44 |
-
|
45 |
-
# Remove multi-space
|
46 |
-
txt = re.sub(" +", " ", txt)
|
47 |
-
return txt.strip()
|
48 |
-
class ViTagger(object):
|
49 |
-
def __init__(self, model_path: Union[str or os.PathLike], no_cuda=False):
|
50 |
-
self.device = 'cuda' if not no_cuda and torch.cuda.is_available() else 'cpu'
|
51 |
-
print("[ViTagger] VnCoreNLP loading ...")
|
52 |
-
self.rdrsegmenter = VnCoreNLP("
|
53 |
-
print("[ViTagger] Model loading ...")
|
54 |
-
self.model, self.tokenizer, self.max_seq_len, self.label2id, self.use_crf = self.load_model(model_path, device=self.device)
|
55 |
-
self.id2label = {idx: label for idx, label in enumerate(self.label2id)}
|
56 |
-
print("[ViTagger] All ready!")
|
57 |
-
|
58 |
-
@staticmethod
|
59 |
-
def load_model(model_path: Union[str or os.PathLike], device='cpu'):
|
60 |
-
if device == 'cpu':
|
61 |
-
checkpoint_data = torch.load(model_path, map_location='cpu')
|
62 |
-
else:
|
63 |
-
checkpoint_data = torch.load(model_path)
|
64 |
-
args = checkpoint_data["args"]
|
65 |
-
max_seq_len = args.max_seq_length
|
66 |
-
use_crf = True if 'crf' in args.model_arch else False
|
67 |
-
tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, use_fast=False)
|
68 |
-
config = AutoConfig.from_pretrained(args.model_name_or_path, num_labels=len(args.label2id))
|
69 |
-
model_clss = MODEL_MAPPING[args.model_name_or_path][args.model_arch]
|
70 |
-
model = model_clss(config=config)
|
71 |
-
model.load_state_dict(checkpoint_data['model'],strict=False)
|
72 |
-
model.to(device)
|
73 |
-
model.eval()
|
74 |
-
|
75 |
-
return model, tokenizer, max_seq_len, args.label2id, use_crf
|
76 |
-
|
77 |
-
def preprocess(self, in_raw: str):
|
78 |
-
norm_text = normalize_text(in_raw)
|
79 |
-
sents = []
|
80 |
-
sentences = self.rdrsegmenter.tokenize(norm_text)
|
81 |
-
for sentence in sentences:
|
82 |
-
sents.append(sentence)
|
83 |
-
return sents
|
84 |
-
|
85 |
-
def convert_tensor(self, tokens):
|
86 |
-
seq_len = len(tokens)
|
87 |
-
encoding = self.tokenizer(tokens,
|
88 |
-
padding='max_length',
|
89 |
-
truncation=True,
|
90 |
-
is_split_into_words=True,
|
91 |
-
max_length=self.max_seq_len)
|
92 |
-
if 'vinai/phobert' in self.tokenizer.name_or_path:
|
93 |
-
print(' '.join(tokens))
|
94 |
-
subwords = self.tokenizer.tokenize(' '.join(tokens))
|
95 |
-
valid_ids = np.zeros(len(encoding.input_ids), dtype=int)
|
96 |
-
label_marks = np.zeros(len(encoding.input_ids), dtype=int)
|
97 |
-
i = 1
|
98 |
-
for idx, subword in enumerate(subwords[:self.max_seq_len - 2]):
|
99 |
-
if idx != 0 and subwords[idx - 1].endswith("@@"):
|
100 |
-
continue
|
101 |
-
if self.use_crf:
|
102 |
-
valid_ids[i - 1] = idx + 1
|
103 |
-
else:
|
104 |
-
valid_ids[idx + 1] = 1
|
105 |
-
i += 1
|
106 |
-
else:
|
107 |
-
valid_ids = np.zeros(len(encoding.input_ids), dtype=int)
|
108 |
-
label_marks = np.zeros(len(encoding.input_ids), dtype=int)
|
109 |
-
i = 1
|
110 |
-
word_ids = encoding.word_ids()
|
111 |
-
for idx in range(1, len(word_ids)):
|
112 |
-
if word_ids[idx] is not None and word_ids[idx] != word_ids[idx - 1]:
|
113 |
-
if self.use_crf:
|
114 |
-
valid_ids[i - 1] = idx
|
115 |
-
else:
|
116 |
-
valid_ids[idx] = 1
|
117 |
-
i += 1
|
118 |
-
if self.max_seq_len >= seq_len + 2:
|
119 |
-
label_marks[:seq_len] = [1] * seq_len
|
120 |
-
else:
|
121 |
-
label_marks[:-2] = [1] * (self.max_seq_len - 2)
|
122 |
-
if self.use_crf and label_marks[0] == 0:
|
123 |
-
raise f"{tokens} have mark == 0 at index 0!"
|
124 |
-
item = {key: torch.as_tensor([val]).to(self.device, dtype=torch.long) for key, val in encoding.items()}
|
125 |
-
item['valid_ids'] = torch.as_tensor([valid_ids]).to(self.device, dtype=torch.long)
|
126 |
-
item['label_masks'] = torch.as_tensor([valid_ids]).to(self.device, dtype=torch.long)
|
127 |
-
return item
|
128 |
-
|
129 |
-
def extract_entity_doc(self, in_raw: str):
|
130 |
-
sents = self.preprocess(in_raw)
|
131 |
-
print(sents)
|
132 |
-
entities_doc = []
|
133 |
-
for sent in sents:
|
134 |
-
item = self.convert_tensor(sent)
|
135 |
-
with torch.no_grad():
|
136 |
-
outputs = self.model(**item)
|
137 |
-
entity = None
|
138 |
-
if isinstance(outputs.tags[0], list):
|
139 |
-
tags = list(itertools.chain(*outputs.tags))
|
140 |
-
else:
|
141 |
-
tags = outputs.tags
|
142 |
-
for w, l in list(zip(sent, tags)):
|
143 |
-
w = w.replace("_", " ")
|
144 |
-
tag = self.id2label[l]
|
145 |
-
if not tag == 'O':
|
146 |
-
parts = tag.split('-', 1)
|
147 |
-
prefix = parts[0]
|
148 |
-
tag = parts[1] if len(parts) > 1 else ""
|
149 |
-
if entity is None:
|
150 |
-
entity = (w, tag)
|
151 |
-
else:
|
152 |
-
if entity[-1] == tag:
|
153 |
-
if prefix == 'I':
|
154 |
-
entity = (entity[0] + f' {w}', tag)
|
155 |
-
else:
|
156 |
-
entities_doc.append(entity)
|
157 |
-
entity = (w, tag)
|
158 |
-
else:
|
159 |
-
entities_doc.append(entity)
|
160 |
-
entity = (w, tag)
|
161 |
-
elif entity is not None:
|
162 |
-
entities_doc.append(entity)
|
163 |
-
if w != ' ':
|
164 |
-
entities_doc.append((w, 'O'))
|
165 |
-
entity = None
|
166 |
-
elif w != ' ':
|
167 |
-
entities_doc.append((w, 'O'))
|
168 |
-
entity = None
|
169 |
-
return entities_doc
|
170 |
-
|
171 |
-
|
172 |
-
def __call__(self, in_raw: str):
|
173 |
-
sents = self.preprocess(in_raw)
|
174 |
-
entites = []
|
175 |
-
for sent in sents:
|
176 |
-
item = self.convert_tensor(sent)
|
177 |
-
with torch.no_grad():
|
178 |
-
outputs = self.model(**item)
|
179 |
-
entity = None
|
180 |
-
if isinstance(outputs.tags[0], list):
|
181 |
-
tags = list(itertools.chain(*outputs.tags))
|
182 |
-
else:
|
183 |
-
tags = outputs.tags
|
184 |
-
for w, l in list(zip(sent, tags)):
|
185 |
-
w = w.replace("_", " ")
|
186 |
-
tag = self.id2label[l]
|
187 |
-
if not tag == 'O':
|
188 |
-
prefix, tag = tag.split('-')
|
189 |
-
if entity is None:
|
190 |
-
entity = (w, tag)
|
191 |
-
else:
|
192 |
-
if entity[-1] == tag:
|
193 |
-
if prefix == 'I':
|
194 |
-
entity = (entity[0] + f' {w}', tag)
|
195 |
-
else:
|
196 |
-
entites.append(entity)
|
197 |
-
entity = (w, tag)
|
198 |
-
else:
|
199 |
-
entites.append(entity)
|
200 |
-
entity = (w, tag)
|
201 |
-
elif entity is not None:
|
202 |
-
entites.append(entity)
|
203 |
-
entity = None
|
204 |
-
else:
|
205 |
-
entity = None
|
206 |
-
return entites
|
207 |
-
|
208 |
-
|
209 |
-
|
210 |
-
|
|
|
1 |
+
|
2 |
+
from vncorenlp import VnCoreNLP
|
3 |
+
|
4 |
+
from typing import Union
|
5 |
+
from transformers import AutoConfig, AutoTokenizer
|
6 |
+
from Model.NER.VLSP2021.Ner_CRF import PhoBertCrf,PhoBertSoftmax,PhoBertLstmCrf
|
7 |
+
import re
|
8 |
+
import os
|
9 |
+
import torch
|
10 |
+
import itertools
|
11 |
+
import numpy as np
|
12 |
+
|
13 |
+
MODEL_MAPPING = {
|
14 |
+
'vinai/phobert-base': {
|
15 |
+
'softmax': PhoBertSoftmax,
|
16 |
+
'crf': PhoBertCrf,
|
17 |
+
'lstm_crf': PhoBertLstmCrf
|
18 |
+
},
|
19 |
+
}
|
20 |
+
|
21 |
+
|
22 |
+
def normalize_text(txt: str) -> str:
|
23 |
+
# Remove special character
|
24 |
+
txt = re.sub("\xad|\u200b|\ufeff", "", txt)
|
25 |
+
# Normalize vietnamese accents
|
26 |
+
txt = re.sub(r"òa", "oà", txt)
|
27 |
+
txt = re.sub(r"óa", "oá", txt)
|
28 |
+
txt = re.sub(r"ỏa", "oả", txt)
|
29 |
+
txt = re.sub(r"õa", "oã", txt)
|
30 |
+
txt = re.sub(r"ọa", "oạ", txt)
|
31 |
+
txt = re.sub(r"òe", "oè", txt)
|
32 |
+
txt = re.sub(r"óe", "oé", txt)
|
33 |
+
txt = re.sub(r"ỏe", "oẻ", txt)
|
34 |
+
txt = re.sub(r"õe", "oẽ", txt)
|
35 |
+
txt = re.sub(r"ọe", "oẹ", txt)
|
36 |
+
txt = re.sub(r"ùy", "uỳ", txt)
|
37 |
+
txt = re.sub(r"úy", "uý", txt)
|
38 |
+
txt = re.sub(r"ủy", "uỷ", txt)
|
39 |
+
txt = re.sub(r"ũy", "uỹ", txt)
|
40 |
+
txt = re.sub(r"ụy", "uỵ", txt)
|
41 |
+
txt = re.sub(r"Ủy", "Uỷ", txt)
|
42 |
+
|
43 |
+
txt = re.sub(r'"', '”', txt)
|
44 |
+
|
45 |
+
# Remove multi-space
|
46 |
+
txt = re.sub(" +", " ", txt)
|
47 |
+
return txt.strip()
|
48 |
+
class ViTagger(object):
|
49 |
+
def __init__(self, model_path: Union[str or os.PathLike], no_cuda=False):
|
50 |
+
self.device = 'cuda' if not no_cuda and torch.cuda.is_available() else 'cpu'
|
51 |
+
print("[ViTagger] VnCoreNLP loading ...")
|
52 |
+
self.rdrsegmenter = VnCoreNLP("/VnCoreNLP/VnCoreNLP-1.1.1.jar", annotators="wseg", max_heap_size='-Xmx500m')
|
53 |
+
print("[ViTagger] Model loading ...")
|
54 |
+
self.model, self.tokenizer, self.max_seq_len, self.label2id, self.use_crf = self.load_model(model_path, device=self.device)
|
55 |
+
self.id2label = {idx: label for idx, label in enumerate(self.label2id)}
|
56 |
+
print("[ViTagger] All ready!")
|
57 |
+
|
58 |
+
@staticmethod
|
59 |
+
def load_model(model_path: Union[str or os.PathLike], device='cpu'):
|
60 |
+
if device == 'cpu':
|
61 |
+
checkpoint_data = torch.load(model_path, map_location='cpu')
|
62 |
+
else:
|
63 |
+
checkpoint_data = torch.load(model_path)
|
64 |
+
args = checkpoint_data["args"]
|
65 |
+
max_seq_len = args.max_seq_length
|
66 |
+
use_crf = True if 'crf' in args.model_arch else False
|
67 |
+
tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, use_fast=False)
|
68 |
+
config = AutoConfig.from_pretrained(args.model_name_or_path, num_labels=len(args.label2id))
|
69 |
+
model_clss = MODEL_MAPPING[args.model_name_or_path][args.model_arch]
|
70 |
+
model = model_clss(config=config)
|
71 |
+
model.load_state_dict(checkpoint_data['model'],strict=False)
|
72 |
+
model.to(device)
|
73 |
+
model.eval()
|
74 |
+
|
75 |
+
return model, tokenizer, max_seq_len, args.label2id, use_crf
|
76 |
+
|
77 |
+
def preprocess(self, in_raw: str):
|
78 |
+
norm_text = normalize_text(in_raw)
|
79 |
+
sents = []
|
80 |
+
sentences = self.rdrsegmenter.tokenize(norm_text)
|
81 |
+
for sentence in sentences:
|
82 |
+
sents.append(sentence)
|
83 |
+
return sents
|
84 |
+
|
85 |
+
def convert_tensor(self, tokens):
|
86 |
+
seq_len = len(tokens)
|
87 |
+
encoding = self.tokenizer(tokens,
|
88 |
+
padding='max_length',
|
89 |
+
truncation=True,
|
90 |
+
is_split_into_words=True,
|
91 |
+
max_length=self.max_seq_len)
|
92 |
+
if 'vinai/phobert' in self.tokenizer.name_or_path:
|
93 |
+
print(' '.join(tokens))
|
94 |
+
subwords = self.tokenizer.tokenize(' '.join(tokens))
|
95 |
+
valid_ids = np.zeros(len(encoding.input_ids), dtype=int)
|
96 |
+
label_marks = np.zeros(len(encoding.input_ids), dtype=int)
|
97 |
+
i = 1
|
98 |
+
for idx, subword in enumerate(subwords[:self.max_seq_len - 2]):
|
99 |
+
if idx != 0 and subwords[idx - 1].endswith("@@"):
|
100 |
+
continue
|
101 |
+
if self.use_crf:
|
102 |
+
valid_ids[i - 1] = idx + 1
|
103 |
+
else:
|
104 |
+
valid_ids[idx + 1] = 1
|
105 |
+
i += 1
|
106 |
+
else:
|
107 |
+
valid_ids = np.zeros(len(encoding.input_ids), dtype=int)
|
108 |
+
label_marks = np.zeros(len(encoding.input_ids), dtype=int)
|
109 |
+
i = 1
|
110 |
+
word_ids = encoding.word_ids()
|
111 |
+
for idx in range(1, len(word_ids)):
|
112 |
+
if word_ids[idx] is not None and word_ids[idx] != word_ids[idx - 1]:
|
113 |
+
if self.use_crf:
|
114 |
+
valid_ids[i - 1] = idx
|
115 |
+
else:
|
116 |
+
valid_ids[idx] = 1
|
117 |
+
i += 1
|
118 |
+
if self.max_seq_len >= seq_len + 2:
|
119 |
+
label_marks[:seq_len] = [1] * seq_len
|
120 |
+
else:
|
121 |
+
label_marks[:-2] = [1] * (self.max_seq_len - 2)
|
122 |
+
if self.use_crf and label_marks[0] == 0:
|
123 |
+
raise f"{tokens} have mark == 0 at index 0!"
|
124 |
+
item = {key: torch.as_tensor([val]).to(self.device, dtype=torch.long) for key, val in encoding.items()}
|
125 |
+
item['valid_ids'] = torch.as_tensor([valid_ids]).to(self.device, dtype=torch.long)
|
126 |
+
item['label_masks'] = torch.as_tensor([valid_ids]).to(self.device, dtype=torch.long)
|
127 |
+
return item
|
128 |
+
|
129 |
+
def extract_entity_doc(self, in_raw: str):
|
130 |
+
sents = self.preprocess(in_raw)
|
131 |
+
print(sents)
|
132 |
+
entities_doc = []
|
133 |
+
for sent in sents:
|
134 |
+
item = self.convert_tensor(sent)
|
135 |
+
with torch.no_grad():
|
136 |
+
outputs = self.model(**item)
|
137 |
+
entity = None
|
138 |
+
if isinstance(outputs.tags[0], list):
|
139 |
+
tags = list(itertools.chain(*outputs.tags))
|
140 |
+
else:
|
141 |
+
tags = outputs.tags
|
142 |
+
for w, l in list(zip(sent, tags)):
|
143 |
+
w = w.replace("_", " ")
|
144 |
+
tag = self.id2label[l]
|
145 |
+
if not tag == 'O':
|
146 |
+
parts = tag.split('-', 1)
|
147 |
+
prefix = parts[0]
|
148 |
+
tag = parts[1] if len(parts) > 1 else ""
|
149 |
+
if entity is None:
|
150 |
+
entity = (w, tag)
|
151 |
+
else:
|
152 |
+
if entity[-1] == tag:
|
153 |
+
if prefix == 'I':
|
154 |
+
entity = (entity[0] + f' {w}', tag)
|
155 |
+
else:
|
156 |
+
entities_doc.append(entity)
|
157 |
+
entity = (w, tag)
|
158 |
+
else:
|
159 |
+
entities_doc.append(entity)
|
160 |
+
entity = (w, tag)
|
161 |
+
elif entity is not None:
|
162 |
+
entities_doc.append(entity)
|
163 |
+
if w != ' ':
|
164 |
+
entities_doc.append((w, 'O'))
|
165 |
+
entity = None
|
166 |
+
elif w != ' ':
|
167 |
+
entities_doc.append((w, 'O'))
|
168 |
+
entity = None
|
169 |
+
return entities_doc
|
170 |
+
|
171 |
+
|
172 |
+
def __call__(self, in_raw: str):
|
173 |
+
sents = self.preprocess(in_raw)
|
174 |
+
entites = []
|
175 |
+
for sent in sents:
|
176 |
+
item = self.convert_tensor(sent)
|
177 |
+
with torch.no_grad():
|
178 |
+
outputs = self.model(**item)
|
179 |
+
entity = None
|
180 |
+
if isinstance(outputs.tags[0], list):
|
181 |
+
tags = list(itertools.chain(*outputs.tags))
|
182 |
+
else:
|
183 |
+
tags = outputs.tags
|
184 |
+
for w, l in list(zip(sent, tags)):
|
185 |
+
w = w.replace("_", " ")
|
186 |
+
tag = self.id2label[l]
|
187 |
+
if not tag == 'O':
|
188 |
+
prefix, tag = tag.split('-')
|
189 |
+
if entity is None:
|
190 |
+
entity = (w, tag)
|
191 |
+
else:
|
192 |
+
if entity[-1] == tag:
|
193 |
+
if prefix == 'I':
|
194 |
+
entity = (entity[0] + f' {w}', tag)
|
195 |
+
else:
|
196 |
+
entites.append(entity)
|
197 |
+
entity = (w, tag)
|
198 |
+
else:
|
199 |
+
entites.append(entity)
|
200 |
+
entity = (w, tag)
|
201 |
+
elif entity is not None:
|
202 |
+
entites.append(entity)
|
203 |
+
entity = None
|
204 |
+
else:
|
205 |
+
entity = None
|
206 |
+
return entites
|
207 |
+
|
208 |
+
|
209 |
+
|
210 |
+
|