-
Notifications
You must be signed in to change notification settings - Fork 2
/
inference.py
276 lines (244 loc) · 9.7 KB
/
inference.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
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
import argparse
import csv
import os
import tarfile
from typing import List
import torch
from dataset import NerDataset
from transformers import (AutoModelForTokenClassification, AutoTokenizer,
PreTrainedTokenizer)
from utils import read_data
KLUE_NER_OUTPUT = "output.csv" # the name of the output file should be output.csv
def load_model(model_dir, model_tar_file):
tarpath = os.path.join(model_dir, model_tar_file)
tar = tarfile.open(tarpath, "r:gz")
tar.extractall(path=model_dir)
model = AutoModelForTokenClassification.from_pretrained(model_dir)
return model
class OutputConvertor(object):
def __init__(
self,
tokenizer: PreTrainedTokenizer,
label_list: List[str],
strip_char: str,
max_length: int,
):
self.in_unk_token = "[+UNK]"
self.tokenizer = tokenizer
self.label_list = label_list
self.strip_char = strip_char
self.max_length = max_length
def tokenizer_out_aligner(self, t_in, t_out, strip_char="##"):
t_out_new = []
i, j = 0, 0
UNK_flag = False
while True:
if i == len(t_in) and j == len(t_out) - 1:
break
step_t_out = (
len(t_out[j].replace(strip_char, ""))
if t_out[j] != self.tokenizer.unk_token
else 1
)
if UNK_flag:
t_out_new.append(self.in_unk_token)
else:
t_out_new.append(t_out[j])
if (
j < len(t_out) - 1
and t_out[j] == self.tokenizer.unk_token
and t_out[j + 1] != self.tokenizer.unk_token
):
i += step_t_out
UNK_flag = True
if t_in[i] == t_out[j + 1][0]:
j += 1
UNK_flag = False
else:
i += step_t_out
j += 1
UNK_flag = False
if j == len(t_out):
UNK_flag = True
j -= 1
return t_out_new
def convert_into_character_pred(self, data, subword_preds):
text = data["text_a"]
original_sentence = text # 안녕 하세요 ^^
subword_preds = [int(x) for x in subword_preds]
character_preds = [subword_preds[0]] # [CLS]
character_preds_idx = 1
for word in original_sentence.split(" "): # 안녕 하세요
if character_preds_idx >= self.max_length - 1:
break
subwords = self.tokenizer.tokenize(word)
if self.tokenizer.unk_token in subwords: # 뻥튀기가 필요한 case!
# case1: ..찝찝..찝찝해 --> [".", ".", "[UNK]", ".", ".", "[UNK]"]
# case2: 미나藤井美菜27가 --> ['미나', '[UNK]', '[UNK]', '美', '[UNK]', '27', '##가']
unk_aligned_subwords = self.tokenizer_out_aligner(
word, subwords, self.strip_char
) # 복원 함수
# case1: [".", ".", "[UNK]", "[+UNK]", ".", ".", "[UNK]", "[+UNK]", "[+UNK]"]
# case2: ['미나', '[UNK]', '[UNK]', '美', '[UNK]', '27', '##가']
unk_flag = False
for subword in unk_aligned_subwords:
if character_preds_idx >= self.max_length - 1:
break
subword_pred = subword_preds[character_preds_idx]
subword_pred_label = self.label_list[subword_pred]
if subword == self.tokenizer.unk_token:
unk_flag = True
character_preds.append(subword_pred)
continue
elif subword == self.in_unk_token:
if subword_pred_label == "O":
character_preds.append(subword_pred)
else:
_, entity_category = subword_pred_label.split("-")
character_pred_label = "I-" + entity_category
character_pred = self.label_list.index(character_pred_label)
character_preds.append(character_pred)
continue
else:
if unk_flag:
character_preds_idx += 1
subword_pred = subword_preds[character_preds_idx]
character_preds.append(subword_pred)
unk_flag = False
else:
character_preds.append(subword_pred)
character_preds_idx += (
1 # TODO +unk가 끝나는 시점에서도 += 1 을 해줘야 다음 label로 넘어감
)
else:
for subword in subwords: # 안녕 -> 안, ##녕하, ##세요
if character_preds_idx >= self.max_length - 1:
break
subword = subword.replace(
self.strip_char, ""
) # xlm roberta: "▁" / others "##"
subword_pred = subword_preds[character_preds_idx]
subword_pred_label = self.label_list[subword_pred]
for i in range(0, len(subword)): # 안, 녕
if i == 0:
character_preds.append(subword_pred)
else:
if subword_pred_label == "O":
character_preds.append(subword_pred)
else:
_, entity_category = subword_pred_label.split("-")
character_pred_label = "I-" + entity_category
character_pred = self.label_list.index(
character_pred_label
)
character_preds.append(character_pred)
character_preds_idx += 1
character_preds.append(subword_preds[-1]) # [SEP] label
return character_preds
def return_char_pred_output(self, data_list, preds):
list_of_character_preds = []
for data, pred in zip(data_list, preds):
character_preds = self.convert_into_character_pred(data, pred)
list_of_character_preds.append(character_preds)
return list_of_character_preds
@torch.no_grad()
def inference(args):
# Set GPU
num_gpus = torch.cuda.device_count()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# Load model
model = load_model(args.model_dir, args.model_tar_file).to(device)
if num_gpus > 1:
model = torch.nn.DataParallel(model)
model.eval()
# Load tokenzier
kwargs = (
{"num_workers": num_gpus, "pin_memory": True}
if torch.cuda.is_available()
else {}
)
tokenizer = AutoTokenizer.from_pretrained(args.model_dir)
# Build dataloader
test_data, label_list, strip_char = read_data(
os.path.join(args.data_dir, args.test_filename), tokenizer
)
dataset = NerDataset(
tokenizer,
test_data,
label_list,
args.max_length,
args.batch_size,
shuffle=False,
**kwargs
)
dataloader = dataset.loader
# Run Inference
preds = []
for input_ids, token_type_ids, attention_mask, labels in dataloader:
input_ids, token_type_ids, attention_mask, labels = (
input_ids.to(device),
token_type_ids.to(device),
attention_mask.to(device),
labels.to(device),
)
output = model(
input_ids,
token_type_ids=token_type_ids,
attention_mask=attention_mask,
labels=labels,
)[1]
pred = output.argmax(dim=2)
pred = pred.detach().cpu().numpy()
for p in pred:
preds.append(p.tolist())
# Convert sub-word preds into char preds
output_convertor = OutputConvertor(
tokenizer, label_list, strip_char, args.max_length
)
list_of_character_preds = output_convertor.return_char_pred_output(test_data, preds)
with open(os.path.join(args.output_dir, KLUE_NER_OUTPUT), "w", newline="") as f:
writer = csv.writer(f)
writer.writerows(list_of_character_preds)
f.close()
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument(
"--batch_size",
type=int,
default=16,
metavar="N",
help="input batch size for inference (default: 64)",
)
parser.add_argument(
"--data_dir", type=str, default=os.environ.get("SM_CHANNEL_EVAL", "/data")
)
parser.add_argument(
"--model_dir", type=str, default='./model'
)
parser.add_argument(
"--output_dir",
type=str,
default=os.environ.get("SM_OUTPUT_DATA_DIR", "/output"),
)
parser.add_argument(
"--model_tar_file",
type=str,
default="klue_ner_model.tar.gz",
help="it needs to include all things for loading baseline model & tokenizer, \
only supporting transformers.AutoModelForSequenceClassification as a model \
transformers.XLMRobertaTokenizer or transformers.BertTokenizer as a tokenizer",
)
parser.add_argument(
"--test_filename",
default="klue-ner-v1.1_test.tsv",
type=str,
help="Name of the test file (default: klue-ner-v1.1_test.tsv)",
)
parser.add_argument(
"--max_length",
type=int,
default=510,
help="maximum sequence length (default: 510)",
)
args = parser.parse_args()
inference(args)