-
Notifications
You must be signed in to change notification settings - Fork 0
/
entity.py
216 lines (154 loc) · 6.16 KB
/
entity.py
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
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
import json
from transformers import BertTokenizer, BertModel
from transformers import AdamW
from transformers import get_linear_schedule_with_warmup
import torch
from torch.utils.data import DataLoader
import torch.nn as nn
from sklearn import preprocessing
import joblib
TOKENIZER = BertTokenizer.from_pretrained('bert-base-cased', do_lower_case=True)
MODEL = BertModel.from_pretrained('bert-base-uncased')
DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
MAX_LEN = 61 + 2
EPOCH = 15
BATCH_SIZE = 32
PATH = "models/net.bin"
class get_data:
def __init__(self, path_file):
self.path_file = path_file
def get_json_data(self):
with open(self.path_file) as file:
data = json.load(file)
classes = []
Label = []
ner = []
documents = []
doc = []
for entity in data['entity']:
for pattern in entity['patterns']:
documents.append((pattern, entity['tag'], entity['sentence']))
classes.append(entity['tag'])
for l in classes:
for word in l:
if word not in Label:
Label.append(word)
for sentence, labels, num_sents in documents:
inp = TOKENIZER.tokenize(sentence)
max_len = len(inp)
for i in range(max_len):
doc.append((inp[i], labels[i]))
return doc, ner
class entity:
def __init__(self, text, entity, tokenizer, max_len):
self.text = text
self.entity = entity
self.tokenizer = tokenizer
self.max_len = max_len
def __len__(self):
return len(self.text)
def __getitem__(self, item):
text = self.text[item]
entity = self.entity[item]
ids = []
target_entity = []
for i, sentence in enumerate(text):
token_ids = self.tokenizer.encode(sentence, add_special_tokens=False)
word_piece_entity = [entity[i]] * len(token_ids)
ids.extend(token_ids)
target_entity.extend(word_piece_entity)
ids = ids[:self.max_len - 2]
target_entity = target_entity[:self.max_len - 2]
ids = [101] + ids + [102]
target_entity = [0] + target_entity + [0]
mask,token_type_id = [1]*len(ids),[0]*len(ids)
padding_len = MAX_LEN - len(ids)
ids = ids + ([0] * padding_len)
target_entity = target_entity + ([0] * padding_len)
mask = mask + ([0] * padding_len)
token_type_id = token_type_id + ([0] * padding_len)
return {
'ids' : torch.tensor(ids,dtype=torch.long),
'target_entity' : torch.tensor(target_entity, dtype=torch.long),
'mask' : torch.tensor(mask, dtype=torch.long),
'token_type_id' : torch.tensor(token_type_id, dtype=torch.long)
}
class entity_model(nn.Module):
def __init__(self, model, num_entity):
super(entity_model, self).__init__()
self.model = model
self.num_entity = num_entity
#better label linear
self.drop = nn.Dropout(0.3)
self.out_entity = nn.Linear(768, self.num_entity)
def forward(self, ids, mask, token_type_id):
#add better model (BertForTokenClassification) or my own one
out = self.model(input_ids = ids, attention_mask = mask, token_type_ids = token_type_id)
lhs = out['last_hidden_state']
lhs_entity = self.drop(lhs)
entity_hs_r = self.out_entity(lhs_entity)
return entity_hs_r
def entity_loss(logits, targets, mask, num_classes):
criterion = nn.CrossEntropyLoss()
active_loss = mask.view(-1) == 1
active_targets = torch.where(
active_loss,
targets.view(-1),
torch.tensor(criterion.ignore_index).type_as(targets)
)
logits = logits.view(-1, num_classes)
loss = criterion(logits, active_targets)
return loss
def transform(ner):
enc_entity = preprocessing.LabelEncoder()
entity_transform = enc_entity.fit_transform(ner)
return entity_transform, enc_entity
def train(model, train_data_loader, optimizer, scheduler):
final_loss = 0
model.train()
for bi, batch in enumerate(train_data_loader):
for k, v in batch.items():
batch[k] = v.to(DEVICE)
optimizer.zero_grad()
out = model(batch['ids'], batch['mask'], batch['token_type_id'])
loss = entity_loss(out, batch['target_entity'], batch['mask'], model.num_entity)
loss.backward()
optimizer.step()
scheduler.step()
final_loss += loss.item()
final_loss_r = final_loss / len(train_data_loader)
return final_loss_r
if __name__ == '__main__':
test_text = []
test_entity = []
data = get_data("ner.json")
doc, ner = data.get_json_data()
for text, label in doc:
ner.append(label)
test_text.append(text)
test, enc_entity = transform(ner=ner)
meta_data = {
"enc_tag": enc_entity
}
joblib.dump(meta_data, "meta.bin")
num_entity = len(enc_entity.classes_)
test = list(test)
test_text = [test_text]
#change batch-size
entity_data = entity(text=test_text, entity=[test], tokenizer=TOKENIZER, max_len=MAX_LEN)
train_data_loader = DataLoader(entity_data, batch_size=BATCH_SIZE) # num_workers= 4
net = entity_model(MODEL, num_entity=num_entity)
net.to(DEVICE)
optimizer = AdamW(net.parameters(), lr=2e-5, correct_bias=False)
#change that
num_train_steps = EPOCH * BATCH_SIZE #int(60 / EPOCH * BATCH_SIZE)
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=0,
num_training_steps = num_train_steps
)
for epoch in range(EPOCH):
print(f"{epoch + 1} on {EPOCH}")
final_loss_r = train(model = net, train_data_loader = train_data_loader, optimizer = optimizer, scheduler = scheduler)
print(f"loss: {final_loss_r}")
torch.save(net.state_dict(), PATH)