Truong-Phuc Nguyen
commited on
Commit
•
e7fcbb8
1
Parent(s):
a4dc1dd
Update plms/language_model.py
Browse files- plms/language_model.py +761 -613
plms/language_model.py
CHANGED
@@ -1,613 +1,761 @@
|
|
1 |
-
import os
|
2 |
-
import logging
|
3 |
-
import pickle
|
4 |
-
import re
|
5 |
-
import urllib
|
6 |
-
from itertools import chain
|
7 |
-
from typing import List, Dict
|
8 |
-
from multiprocessing import Pool
|
9 |
-
import numpy as np
|
10 |
-
from tqdm import tqdm
|
11 |
-
import torch
|
12 |
-
from torch.nn import functional
|
13 |
-
import transformers
|
14 |
-
from .exceptions import ExceedMaxLengthError, HighlightNotFoundError, AnswerNotFoundError
|
15 |
-
from .spacy_module import SpacyPipeline, VALID_METHODS
|
16 |
-
|
17 |
-
__all__ = ('TransformersQG', 'ADDITIONAL_SP_TOKENS', 'TASK_PREFIX', 'clean', 'internet_connection')
|
18 |
-
|
19 |
-
os.environ["TOKENIZERS_PARALLELISM"] = "false" # to turn off warning message
|
20 |
-
TASK_PREFIX = {
|
21 |
-
"ae": "extract answers",
|
22 |
-
"qg": "generate question",
|
23 |
-
"qag": "generate question and answer",
|
24 |
-
"qa": "answer question"
|
25 |
-
}
|
26 |
-
CE_IGNORE_INDEX = -100
|
27 |
-
ADDITIONAL_SP_TOKENS = {'hl': '<hl>'}
|
28 |
-
NUM_WORKERS = int(os.getenv('NUM_WORKERS', '0'))
|
29 |
-
PARALLEL_PROCESSING = bool(int(os.getenv('PARALLEL_PROCESSING', '0')))
|
30 |
-
DEFAULT_MODELS = {
|
31 |
-
'vi': 'VietAI/vit5-base'
|
32 |
-
}
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
string = re.sub(r'\s
|
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 |
-
tokenizer
|
96 |
-
|
97 |
-
|
98 |
-
|
99 |
-
|
100 |
-
|
101 |
-
|
102 |
-
|
103 |
-
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
-
# will
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
nll_loss.
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
smoothed_loss.
|
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 |
-
@param
|
160 |
-
@param
|
161 |
-
@param
|
162 |
-
@param
|
163 |
-
@param
|
164 |
-
@param
|
165 |
-
@param
|
166 |
-
|
167 |
-
|
168 |
-
self.
|
169 |
-
self.
|
170 |
-
self.
|
171 |
-
|
172 |
-
|
173 |
-
self.
|
174 |
-
self.
|
175 |
-
|
176 |
-
|
177 |
-
self.
|
178 |
-
|
179 |
-
|
180 |
-
self.
|
181 |
-
|
182 |
-
|
183 |
-
|
184 |
-
|
185 |
-
|
186 |
-
|
187 |
-
|
188 |
-
|
189 |
-
@param
|
190 |
-
@param
|
191 |
-
@
|
192 |
-
|
193 |
-
|
194 |
-
|
195 |
-
|
196 |
-
|
197 |
-
|
198 |
-
|
199 |
-
|
200 |
-
|
201 |
-
|
202 |
-
input_sequence[position
|
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 |
-
@param
|
259 |
-
@param
|
260 |
-
@param
|
261 |
-
@param
|
262 |
-
@param
|
263 |
-
@param
|
264 |
-
@param
|
265 |
-
@param
|
266 |
-
@param
|
267 |
-
@param
|
268 |
-
|
269 |
-
|
270 |
-
|
271 |
-
|
272 |
-
|
273 |
-
|
274 |
-
f"'{
|
275 |
-
|
276 |
-
|
277 |
-
|
278 |
-
|
279 |
-
self.
|
280 |
-
self.
|
281 |
-
self.
|
282 |
-
|
283 |
-
|
284 |
-
self.
|
285 |
-
self.
|
286 |
-
self.
|
287 |
-
self.
|
288 |
-
self.
|
289 |
-
self.
|
290 |
-
self.
|
291 |
-
self.
|
292 |
-
self.
|
293 |
-
self.
|
294 |
-
|
295 |
-
|
296 |
-
|
297 |
-
|
298 |
-
|
299 |
-
|
300 |
-
#
|
301 |
-
|
302 |
-
|
303 |
-
|
304 |
-
|
305 |
-
|
306 |
-
|
307 |
-
|
308 |
-
|
309 |
-
|
310 |
-
|
311 |
-
|
312 |
-
|
313 |
-
self.
|
314 |
-
self.
|
315 |
-
|
316 |
-
|
317 |
-
|
318 |
-
|
319 |
-
|
320 |
-
self.
|
321 |
-
self.
|
322 |
-
|
323 |
-
|
324 |
-
|
325 |
-
self.
|
326 |
-
self.
|
327 |
-
|
328 |
-
|
329 |
-
|
330 |
-
|
331 |
-
self.
|
332 |
-
|
333 |
-
|
334 |
-
self.
|
335 |
-
|
336 |
-
|
337 |
-
|
338 |
-
|
339 |
-
|
340 |
-
|
341 |
-
logging.info(f'
|
342 |
-
logging.info(f'\t *
|
343 |
-
logging.info(f'\t *
|
344 |
-
|
345 |
-
|
346 |
-
|
347 |
-
|
348 |
-
|
349 |
-
|
350 |
-
|
351 |
-
|
352 |
-
|
353 |
-
|
354 |
-
|
355 |
-
|
356 |
-
|
357 |
-
|
358 |
-
|
359 |
-
|
360 |
-
|
361 |
-
"""
|
362 |
-
|
363 |
-
|
364 |
-
|
365 |
-
|
366 |
-
|
367 |
-
|
368 |
-
|
369 |
-
|
370 |
-
|
371 |
-
|
372 |
-
|
373 |
-
|
374 |
-
|
375 |
-
|
376 |
-
|
377 |
-
|
378 |
-
|
379 |
-
|
380 |
-
|
381 |
-
|
382 |
-
|
383 |
-
|
384 |
-
|
385 |
-
|
386 |
-
|
387 |
-
|
388 |
-
|
389 |
-
|
390 |
-
|
391 |
-
|
392 |
-
|
393 |
-
|
394 |
-
|
395 |
-
|
396 |
-
|
397 |
-
|
398 |
-
|
399 |
-
|
400 |
-
|
401 |
-
|
402 |
-
|
403 |
-
|
404 |
-
|
405 |
-
|
406 |
-
|
407 |
-
|
408 |
-
|
409 |
-
|
410 |
-
|
411 |
-
|
412 |
-
|
413 |
-
|
414 |
-
|
415 |
-
batch_size
|
416 |
-
|
417 |
-
|
418 |
-
|
419 |
-
|
420 |
-
)
|
421 |
-
|
422 |
-
|
423 |
-
|
424 |
-
|
425 |
-
|
426 |
-
|
427 |
-
|
428 |
-
|
429 |
-
|
430 |
-
|
431 |
-
|
432 |
-
|
433 |
-
|
434 |
-
|
435 |
-
|
436 |
-
|
437 |
-
)
|
438 |
-
|
439 |
-
|
440 |
-
|
441 |
-
|
442 |
-
|
443 |
-
|
444 |
-
|
445 |
-
|
446 |
-
|
447 |
-
|
448 |
-
|
449 |
-
|
450 |
-
|
451 |
-
|
452 |
-
|
453 |
-
|
454 |
-
|
455 |
-
|
456 |
-
|
457 |
-
|
458 |
-
|
459 |
-
|
460 |
-
|
461 |
-
|
462 |
-
|
463 |
-
|
464 |
-
|
465 |
-
|
466 |
-
|
467 |
-
|
468 |
-
|
469 |
-
|
470 |
-
|
471 |
-
|
472 |
-
|
473 |
-
|
474 |
-
|
475 |
-
|
476 |
-
|
477 |
-
|
478 |
-
|
479 |
-
|
480 |
-
|
481 |
-
|
482 |
-
|
483 |
-
|
484 |
-
|
485 |
-
|
486 |
-
|
487 |
-
|
488 |
-
|
489 |
-
|
490 |
-
|
491 |
-
|
492 |
-
|
493 |
-
|
494 |
-
|
495 |
-
|
496 |
-
|
497 |
-
|
498 |
-
|
499 |
-
|
500 |
-
|
501 |
-
|
502 |
-
)
|
503 |
-
|
504 |
-
|
505 |
-
|
506 |
-
|
507 |
-
|
508 |
-
|
509 |
-
|
510 |
-
|
511 |
-
|
512 |
-
|
513 |
-
|
514 |
-
|
515 |
-
|
516 |
-
|
517 |
-
|
518 |
-
|
519 |
-
|
520 |
-
|
521 |
-
|
522 |
-
|
523 |
-
|
524 |
-
|
525 |
-
|
526 |
-
|
527 |
-
|
528 |
-
|
529 |
-
|
530 |
-
|
531 |
-
|
532 |
-
|
533 |
-
"""
|
534 |
-
|
535 |
-
|
536 |
-
|
537 |
-
|
538 |
-
|
539 |
-
|
540 |
-
|
541 |
-
|
542 |
-
|
543 |
-
|
544 |
-
|
545 |
-
|
546 |
-
|
547 |
-
|
548 |
-
|
549 |
-
|
550 |
-
|
551 |
-
|
552 |
-
|
553 |
-
|
554 |
-
|
555 |
-
|
556 |
-
|
557 |
-
|
558 |
-
|
559 |
-
|
560 |
-
|
561 |
-
|
562 |
-
|
563 |
-
|
564 |
-
|
565 |
-
|
566 |
-
|
567 |
-
|
568 |
-
|
569 |
-
|
570 |
-
|
571 |
-
|
572 |
-
|
573 |
-
""
|
574 |
-
|
575 |
-
if
|
576 |
-
|
577 |
-
|
578 |
-
|
579 |
-
|
580 |
-
|
581 |
-
|
582 |
-
|
583 |
-
|
584 |
-
|
585 |
-
if
|
586 |
-
|
587 |
-
|
588 |
-
|
589 |
-
|
590 |
-
|
591 |
-
|
592 |
-
|
593 |
-
|
594 |
-
|
595 |
-
|
596 |
-
|
597 |
-
|
598 |
-
|
599 |
-
|
600 |
-
|
601 |
-
|
602 |
-
|
603 |
-
|
604 |
-
|
605 |
-
|
606 |
-
|
607 |
-
|
608 |
-
|
609 |
-
|
610 |
-
|
611 |
-
|
612 |
-
|
613 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import logging
|
3 |
+
import pickle
|
4 |
+
import re
|
5 |
+
import urllib
|
6 |
+
from itertools import chain
|
7 |
+
from typing import List, Dict
|
8 |
+
from multiprocessing import Pool
|
9 |
+
import numpy as np
|
10 |
+
from tqdm import tqdm
|
11 |
+
import torch
|
12 |
+
from torch.nn import functional
|
13 |
+
import transformers
|
14 |
+
from .exceptions import ExceedMaxLengthError, HighlightNotFoundError, AnswerNotFoundError
|
15 |
+
from .spacy_module import SpacyPipeline, VALID_METHODS
|
16 |
+
|
17 |
+
__all__ = ('TransformersQG', 'ADDITIONAL_SP_TOKENS', 'TASK_PREFIX', 'clean', 'internet_connection')
|
18 |
+
|
19 |
+
os.environ["TOKENIZERS_PARALLELISM"] = "false" # to turn off warning message
|
20 |
+
TASK_PREFIX = {
|
21 |
+
"ae": "extract answers",
|
22 |
+
"qg": "generate question",
|
23 |
+
"qag": "generate question and answer",
|
24 |
+
"qa": "answer question"
|
25 |
+
}
|
26 |
+
CE_IGNORE_INDEX = -100
|
27 |
+
ADDITIONAL_SP_TOKENS = {'hl': '<hl>'}
|
28 |
+
NUM_WORKERS = int(os.getenv('NUM_WORKERS', '0'))
|
29 |
+
PARALLEL_PROCESSING = bool(int(os.getenv('PARALLEL_PROCESSING', '0')))
|
30 |
+
DEFAULT_MODELS = {
|
31 |
+
'vi': 'VietAI/vit5-base'
|
32 |
+
}
|
33 |
+
|
34 |
+
|
35 |
+
def pickle_save(obj, path: str):
|
36 |
+
with open(path, "wb") as fp:
|
37 |
+
pickle.dump(obj, fp)
|
38 |
+
|
39 |
+
|
40 |
+
def pickle_load(path: str):
|
41 |
+
with open(path, "rb") as fp: # Unpickling
|
42 |
+
return pickle.load(fp)
|
43 |
+
|
44 |
+
|
45 |
+
def clean(string):
|
46 |
+
string = re.sub(r'\A\s*', '', string)
|
47 |
+
string = re.sub(r'\s*\Z', '', string)
|
48 |
+
if len(string) > 0:
|
49 |
+
return string
|
50 |
+
return None
|
51 |
+
|
52 |
+
|
53 |
+
def internet_connection(host='http://google.com'):
|
54 |
+
try:
|
55 |
+
urllib.request.urlopen(host)
|
56 |
+
return True
|
57 |
+
except:
|
58 |
+
return False
|
59 |
+
|
60 |
+
|
61 |
+
def load_language_model(model_name,
|
62 |
+
cache_dir: str = None,
|
63 |
+
use_auth_token: bool = False,
|
64 |
+
torch_dtype=None,
|
65 |
+
device_map: str = None,
|
66 |
+
low_cpu_mem_usage: bool = False):
|
67 |
+
""" load language model from huggingface model hub """
|
68 |
+
# tokenizer
|
69 |
+
local_files_only = not internet_connection()
|
70 |
+
tokenizer = transformers.AutoTokenizer.from_pretrained(
|
71 |
+
model_name, cache_dir=cache_dir, local_files_only=local_files_only, use_auth_token=use_auth_token)
|
72 |
+
config = transformers.AutoConfig.from_pretrained(
|
73 |
+
model_name, local_files_only=local_files_only, cache_dir=cache_dir, use_auth_token=use_auth_token)
|
74 |
+
# model
|
75 |
+
if config.model_type == 't5': # T5 model requires T5ForConditionalGeneration class
|
76 |
+
model_class = transformers.T5ForConditionalGeneration.from_pretrained
|
77 |
+
elif config.model_type == 'mt5':
|
78 |
+
model_class = transformers.MT5ForConditionalGeneration.from_pretrained
|
79 |
+
elif config.model_type == 'bart':
|
80 |
+
model_class = transformers.BartForConditionalGeneration.from_pretrained
|
81 |
+
elif config.model_type == 'mbart':
|
82 |
+
model_class = transformers.MBartForConditionalGeneration.from_pretrained
|
83 |
+
elif config.model_type == 'switch_transformers':
|
84 |
+
model_class = transformers.SwitchTransformersForConditionalGeneration.from_pretrained
|
85 |
+
else:
|
86 |
+
raise ValueError(f'unsupported model type: {config.model_type}')
|
87 |
+
|
88 |
+
param = {'config': config, "local_files_only": local_files_only, "use_auth_token": use_auth_token,
|
89 |
+
"low_cpu_mem_usage": low_cpu_mem_usage, "cache_dir": cache_dir}
|
90 |
+
if torch_dtype is not None:
|
91 |
+
param['torch_dtype'] = torch_dtype
|
92 |
+
if device_map is not None:
|
93 |
+
param['device_map'] = device_map
|
94 |
+
model = model_class(model_name, **param)
|
95 |
+
# add new special tokens to the tokenizer and the model if they don't have it
|
96 |
+
tokenizer.add_special_tokens({'additional_special_tokens': list(ADDITIONAL_SP_TOKENS.values())})
|
97 |
+
model.resize_token_embeddings(len(tokenizer))
|
98 |
+
return tokenizer, model, config
|
99 |
+
|
100 |
+
|
101 |
+
def label_smoothed_loss(logits, labels, epsilon):
|
102 |
+
""" https://github.com/huggingface/transformers/blob/55bb4c06f7be141c6d895dbe1f11018dc8580b2d/src/transformers/trainer_pt_utils.py#L430 """
|
103 |
+
log_probs = - functional.log_softmax(logits, dim=-1)
|
104 |
+
if labels.dim() == log_probs.dim() - 1:
|
105 |
+
labels = labels.unsqueeze(-1)
|
106 |
+
|
107 |
+
padding_mask = labels.eq(CE_IGNORE_INDEX)
|
108 |
+
# In case the ignore_index is -100, the gather will fail, so we replace labels by 0. The padding_mask
|
109 |
+
# will ignore them in any case.
|
110 |
+
labels.clamp_min_(0)
|
111 |
+
|
112 |
+
nll_loss = log_probs.gather(dim=-1, index=labels)
|
113 |
+
nll_loss.masked_fill_(padding_mask, 0.0)
|
114 |
+
|
115 |
+
# works for fp16 input tensor too, by internally upcasting it to fp32
|
116 |
+
smoothed_loss = log_probs.sum(dim=-1, keepdim=True, dtype=torch.float32)
|
117 |
+
smoothed_loss.masked_fill_(padding_mask, 0.0)
|
118 |
+
|
119 |
+
# Take the mean over the label dimensions, then divide by the number of active elements (i.e. not-padded):
|
120 |
+
num_active_elements = padding_mask.numel() - padding_mask.long().sum()
|
121 |
+
nll_loss = nll_loss.sum() / num_active_elements
|
122 |
+
smoothed_loss = smoothed_loss.sum() / (num_active_elements * log_probs.shape[-1])
|
123 |
+
return (1 - epsilon) * nll_loss + epsilon * smoothed_loss
|
124 |
+
|
125 |
+
|
126 |
+
class Dataset(torch.utils.data.Dataset):
|
127 |
+
""" torch.utils.data.Dataset wrapper converting into tensor """
|
128 |
+
float_tensors = ['attention_mask']
|
129 |
+
|
130 |
+
def __init__(self, data: List):
|
131 |
+
self.data = data
|
132 |
+
|
133 |
+
def __len__(self):
|
134 |
+
return len(self.data)
|
135 |
+
|
136 |
+
def to_tensor(self, name, data):
|
137 |
+
if name in self.float_tensors:
|
138 |
+
return torch.tensor(data, dtype=torch.float32)
|
139 |
+
return torch.tensor(data, dtype=torch.long)
|
140 |
+
|
141 |
+
def __getitem__(self, idx):
|
142 |
+
return {k: self.to_tensor(k, v) for k, v in self.data[idx].items()}
|
143 |
+
|
144 |
+
|
145 |
+
class EncodePlus:
|
146 |
+
""" Wrapper of encode_plus for multiprocessing. """
|
147 |
+
|
148 |
+
def __init__(self,
|
149 |
+
tokenizer,
|
150 |
+
max_length: int = 512,
|
151 |
+
max_length_output: int = 34,
|
152 |
+
drop_overflow_error_text: bool = False,
|
153 |
+
skip_overflow_error: bool = False,
|
154 |
+
drop_highlight_error_text: bool = False,
|
155 |
+
prefix_type: str = None,
|
156 |
+
padding: bool = True):
|
157 |
+
""" Wrapper of encode_plus for multiprocessing.
|
158 |
+
|
159 |
+
@param tokenizer: transforms.Tokenizer
|
160 |
+
@param max_length: Max text length of input.
|
161 |
+
@param max_length_output: Max text length of output.
|
162 |
+
@param drop_overflow_error_text: If true, return None when the input exceeds the max length.
|
163 |
+
@param skip_overflow_error: If true, raise an error when the input exceeds the max length.
|
164 |
+
@param drop_highlight_error_text: If true, raise an error when a highlight span is not found in the paragraph.
|
165 |
+
@param prefix_type: Either of `qg` or `answer_extraction`, which is to add at the beginning of the text.
|
166 |
+
@param padding: Pad the sequence to the max length.
|
167 |
+
"""
|
168 |
+
self.prefix = TASK_PREFIX[prefix_type] if prefix_type is not None else None
|
169 |
+
self.tokenizer = tokenizer
|
170 |
+
self.max_length = max_length
|
171 |
+
self.max_length_output = max_length_output
|
172 |
+
# NOTE: for model training, we should drop the exceeded input but not for the evaluator
|
173 |
+
self.drop_overflow_error_text = drop_overflow_error_text
|
174 |
+
self.skip_overflow_error = skip_overflow_error
|
175 |
+
self.drop_highlight_error_text = drop_highlight_error_text
|
176 |
+
# truncation should be true for the batch process, but not necessary to process single input
|
177 |
+
self.param_in = {'truncation': True, 'max_length': self.max_length}
|
178 |
+
self.param_out = {'truncation': True, 'max_length': self.max_length_output}
|
179 |
+
if padding:
|
180 |
+
self.param_in['padding'] = 'max_length'
|
181 |
+
self.param_out['padding'] = 'max_length'
|
182 |
+
|
183 |
+
def __call__(self, inputs):
|
184 |
+
return self.encode_plus(*inputs)
|
185 |
+
|
186 |
+
def encode_plus(self, input_sequence: str, output_sequence: str = None, input_highlight: str = None):
|
187 |
+
""" encode_plus
|
188 |
+
|
189 |
+
@param input_sequence: Input sequence.
|
190 |
+
@param output_sequence: Output sequence.
|
191 |
+
@param input_highlight: Sub-sequence of `input_sequence` to be surrounded by <hl>.
|
192 |
+
@return: The output of `encode_plus`.
|
193 |
+
"""
|
194 |
+
# add highlight to the input
|
195 |
+
if input_highlight is not None:
|
196 |
+
position = input_sequence.find(input_highlight)
|
197 |
+
if position == -1:
|
198 |
+
if self.drop_highlight_error_text:
|
199 |
+
return None
|
200 |
+
raise HighlightNotFoundError(input_highlight, input_sequence)
|
201 |
+
input_sequence = '{0}{1} {2} {1}{3}'.format(
|
202 |
+
input_sequence[:position], ADDITIONAL_SP_TOKENS['hl'], input_highlight,
|
203 |
+
input_sequence[position+len(input_highlight):])
|
204 |
+
if self.prefix is not None:
|
205 |
+
input_sequence = f'{self.prefix}: {input_sequence}'
|
206 |
+
|
207 |
+
# handling overflow text
|
208 |
+
# drop_overflow_error_text ==> remove the overflow sentence from input
|
209 |
+
# skip_overflow_error ==> keep the overflow sentence
|
210 |
+
# none of them ==> raise error
|
211 |
+
if self.drop_overflow_error_text or not self.skip_overflow_error:
|
212 |
+
if len(self.tokenizer.encode(input_sequence)) > self.max_length:
|
213 |
+
if not self.drop_overflow_error_text: # raise error for overflow text
|
214 |
+
raise ExceedMaxLengthError(self.max_length)
|
215 |
+
return None # remove overflow text
|
216 |
+
if output_sequence is not None:
|
217 |
+
if len(self.tokenizer.encode(output_sequence)) > self.max_length_output:
|
218 |
+
if not self.drop_overflow_error_text: # raise error for overflow text
|
219 |
+
raise ExceedMaxLengthError(self.max_length)
|
220 |
+
return None # remove overflow text
|
221 |
+
if type(self.tokenizer) is transformers.models.mbart.tokenization_mbart_fast.MBartTokenizerFast:
|
222 |
+
encode = self.tokenizer(input_sequence, **self.param_in)
|
223 |
+
else:
|
224 |
+
encode = self.tokenizer(text_target=input_sequence, **self.param_in)
|
225 |
+
if output_sequence is not None:
|
226 |
+
encode['labels'] = self.tokenizer.encode(output_sequence, **self.param_out)
|
227 |
+
return encode
|
228 |
+
|
229 |
+
|
230 |
+
class TransformersQG:
|
231 |
+
""" Transformers Language Model for Question Generation. """
|
232 |
+
|
233 |
+
def __init__(self,
|
234 |
+
model: str = None,
|
235 |
+
max_length: int = 512,
|
236 |
+
max_length_output: int = 256,
|
237 |
+
model_ae: str = None,
|
238 |
+
max_length_ae: int = 512,
|
239 |
+
max_length_output_ae: int = 64,
|
240 |
+
cache_dir: str = None,
|
241 |
+
add_prefix: bool = None,
|
242 |
+
language: str = 'vi',
|
243 |
+
label_smoothing: float = None,
|
244 |
+
skip_overflow_error: bool = False,
|
245 |
+
drop_overflow_error_text: bool = False,
|
246 |
+
drop_highlight_error_text: bool = False,
|
247 |
+
drop_answer_error_text: bool = False,
|
248 |
+
use_auth_token: bool = False,
|
249 |
+
torch_dtype=None,
|
250 |
+
device_map: str = None,
|
251 |
+
low_cpu_mem_usage: bool = False,
|
252 |
+
is_qg: bool = None,
|
253 |
+
is_qag: bool = None,
|
254 |
+
is_qa: bool = None,
|
255 |
+
is_ae: bool = None):
|
256 |
+
""" Transformers Language Model for Question Generation.
|
257 |
+
|
258 |
+
@param model: Model alias or path to local model file.
|
259 |
+
@param max_length: Max text length of input.
|
260 |
+
@param max_length_output: Max text length of output.
|
261 |
+
@param cache_dir: Directory to cache transformers model files.
|
262 |
+
@param add_prefix: Whether model uses task-specific prefix (eg. True for T5 but False for BART models).
|
263 |
+
@param language: Language alias for SpaCy language-specific pipelines (sentencizer/keyword extraction).
|
264 |
+
@param label_smoothing: [Fine-tuning parameter] Label smoothing.
|
265 |
+
@param drop_overflow_error_text: If true, return None when the input exceeds the max length.
|
266 |
+
@param skip_overflow_error: If true, raise an error when the input exceeds the max length.
|
267 |
+
@param drop_highlight_error_text: If true, raise an error when a highlight span is not found in the paragraph.
|
268 |
+
@param use_auth_token: [optional] Huggingface transformers argument of `use_auth_token`
|
269 |
+
"""
|
270 |
+
|
271 |
+
# take default model given the language
|
272 |
+
if model is None:
|
273 |
+
assert language in DEFAULT_MODELS.keys(),\
|
274 |
+
f"Model with language '{language}' is not available. Please choose language from " \
|
275 |
+
f"'{DEFAULT_MODELS.keys()}' or specify 'model'."
|
276 |
+
model = DEFAULT_MODELS[language]
|
277 |
+
|
278 |
+
# classify model type
|
279 |
+
self.is_qg = 'qg' in model.split('-') if is_qg is None else is_qg
|
280 |
+
self.is_ae = 'ae' in model.split('-') if is_ae is None else is_ae
|
281 |
+
self.is_qa = 'qa' in model.split('-') if is_qa is None else is_qa
|
282 |
+
self.is_qag = 'qag' in model.split('-') if is_qag is None else is_qag
|
283 |
+
# configs
|
284 |
+
self.model_name = model
|
285 |
+
self.max_length = max_length
|
286 |
+
self.max_length_output = max_length_output
|
287 |
+
self.label_smoothing = label_smoothing
|
288 |
+
self.drop_overflow_error_text = drop_overflow_error_text
|
289 |
+
self.skip_overflow_error = skip_overflow_error
|
290 |
+
self.drop_highlight_error_text = drop_highlight_error_text
|
291 |
+
self.drop_answer_error_text = drop_answer_error_text
|
292 |
+
self.model_name_ae = model_ae
|
293 |
+
self.max_length_ae = max_length_ae
|
294 |
+
self.max_length_output_ae = max_length_output_ae
|
295 |
+
# load model
|
296 |
+
self.tokenizer, self.model, config = load_language_model(
|
297 |
+
self.model_name, cache_dir=cache_dir, use_auth_token=use_auth_token, device_map=device_map,
|
298 |
+
torch_dtype=torch_dtype, low_cpu_mem_usage=low_cpu_mem_usage)
|
299 |
+
if 'add_prefix' not in config.to_dict().keys():
|
300 |
+
# this means the model is not fine-tuned
|
301 |
+
# assert add_prefix, '`add_prefix` is required for non-fine-tuned models'
|
302 |
+
self.add_prefix = add_prefix
|
303 |
+
else:
|
304 |
+
self.add_prefix = config.add_prefix
|
305 |
+
|
306 |
+
# set default behaviour for answer extraction
|
307 |
+
if self.model_name_ae is None:
|
308 |
+
self.model_name_ae = self.model_name if self.is_ae else "positionrank"
|
309 |
+
# load answer extraction model
|
310 |
+
self.answer_model_type = None
|
311 |
+
if self.model_name_ae in VALID_METHODS:
|
312 |
+
logging.info(f'use spaCy answer extraction model: {self.model_name_ae}')
|
313 |
+
self.tokenizer_ae = self.model_ae = self.add_prefix_ae = None
|
314 |
+
self.spacy_module = SpacyPipeline(language, self.model_name_ae)
|
315 |
+
self.answer_model_type = 'spacy'
|
316 |
+
else:
|
317 |
+
logging.info(f'use LMQG fine-tuned answer extraction model: {self.model_name_ae}')
|
318 |
+
if self.model_name == self.model_name_ae:
|
319 |
+
logging.info("the same model as QG is used as AE")
|
320 |
+
assert self.is_ae, f"the model ({self.model_name_ae}) is not fine-tuned for AE"
|
321 |
+
self.tokenizer_ae = self.model_ae = self.add_prefix_ae = None
|
322 |
+
self.answer_model_type = 'multitask'
|
323 |
+
else:
|
324 |
+
logging.info(f"loading 2nd model for AE: {self.model_name_ae}")
|
325 |
+
self.tokenizer_ae, self.model_ae, config_ae = load_language_model(model_ae, cache_dir=cache_dir, use_auth_token=use_auth_token)
|
326 |
+
self.add_prefix_ae = config_ae.add_prefix
|
327 |
+
self.answer_model_type = 'pipeline'
|
328 |
+
self.spacy_module = SpacyPipeline(language)
|
329 |
+
|
330 |
+
# GPU setup
|
331 |
+
self.device = 'cuda' if torch.cuda.device_count() > 0 else 'cpu'
|
332 |
+
self.parallel = False
|
333 |
+
if torch.cuda.device_count() > 1:
|
334 |
+
self.parallel = True
|
335 |
+
self.model = torch.nn.DataParallel(self.model)
|
336 |
+
if self.model_ae is not None:
|
337 |
+
self.model_ae = torch.nn.DataParallel(self.model_ae)
|
338 |
+
self.model.to(self.device)
|
339 |
+
if self.model_ae is not None:
|
340 |
+
self.model_ae.to(self.device)
|
341 |
+
logging.info(f'Model `{self.model_name}`')
|
342 |
+
logging.info(f'\t * Num of GPU in use: {torch.cuda.device_count()}')
|
343 |
+
logging.info(f'\t * Prefix: {self.add_prefix}')
|
344 |
+
logging.info(f'\t * Language: {language} (ignore at the training phase)')
|
345 |
+
|
346 |
+
def push_to_hub(self, repo_id):
|
347 |
+
if self.parallel:
|
348 |
+
self.model.module.push_to_hub(repo_id)
|
349 |
+
else:
|
350 |
+
self.model.push_to_hub(repo_id)
|
351 |
+
self.tokenizer.push_to_hub(repo_id)
|
352 |
+
|
353 |
+
def generate_qa_end2end(self,
|
354 |
+
list_context: str or List,
|
355 |
+
batch_size: int = None,
|
356 |
+
num_beams: int = 4,
|
357 |
+
cache_path: str = None,
|
358 |
+
splitting_symbol: str = ' [SEP] ',
|
359 |
+
question_prefix: str = "question: ",
|
360 |
+
answer_prefix: str = ", answer: "):
|
361 |
+
""" Generate question from paragraph and answer. Note that `list_answer` is needed unless they are already
|
362 |
+
highlighted in the `list_context`. eg) "I live in <hl> Tokyo <hl>."
|
363 |
+
|
364 |
+
@param list_context: List of input texts.
|
365 |
+
@param batch_size: Batch size.
|
366 |
+
@param num_beams: Number of beam for model generation.
|
367 |
+
@param cache_path: Path to pre-compute features.
|
368 |
+
@return: List of generated sentences.
|
369 |
+
"""
|
370 |
+
logging.info(f'running model for `question_answer_pair_generation`')
|
371 |
+
assert self.is_qag, "`generate_qa_end2end` is available for end2end_qag_model"
|
372 |
+
prefix_type = 'qag' if self.add_prefix else None
|
373 |
+
single_input = type(list_context) is str
|
374 |
+
list_context = [list_context] if single_input else list_context
|
375 |
+
output = self.generate_prediction(
|
376 |
+
list_context, prefix_type=prefix_type, cache_path=cache_path, num_beams=num_beams, batch_size=batch_size
|
377 |
+
)
|
378 |
+
|
379 |
+
def format_qa(list_raw_string):
|
380 |
+
tmp = []
|
381 |
+
for raw_string in list_raw_string:
|
382 |
+
if len(raw_string.split(answer_prefix)) != 2 or question_prefix not in raw_string:
|
383 |
+
logging.info(f"invalid prediction: {raw_string}")
|
384 |
+
else:
|
385 |
+
q, a = raw_string.split(answer_prefix)
|
386 |
+
a = re.sub(r'\A\s+', '', a)
|
387 |
+
a = re.sub(r'\s+\Z', '', a)
|
388 |
+
q = q.replace(question_prefix, "")
|
389 |
+
q = re.sub(r'\A\s+', '', q)
|
390 |
+
q = re.sub(r'\s+\Z', '', q)
|
391 |
+
tmp.append((q, a))
|
392 |
+
return tmp
|
393 |
+
|
394 |
+
output = [format_qa(o.split(splitting_symbol)) for o in output]
|
395 |
+
return output[0] if single_input else output
|
396 |
+
|
397 |
+
def generate_qa(self,
|
398 |
+
list_context: str or List,
|
399 |
+
batch_size: int = None,
|
400 |
+
num_beams: int = 4,
|
401 |
+
cache_path: str = None,
|
402 |
+
num_questions: int = None,
|
403 |
+
sentence_level: bool = False):
|
404 |
+
""" Generate question given context.
|
405 |
+
|
406 |
+
@param list_context: Input text.
|
407 |
+
@param batch_size: Batch size.
|
408 |
+
@param num_beams: Number of beam for model generation.
|
409 |
+
@param cache_path: Path to pre-compute features.
|
410 |
+
@param num_questions: Max number of questions.
|
411 |
+
@param sentence_level: Run prediction on each sentence of the context independently to reduce complexity.
|
412 |
+
@return: List of generated sentences.
|
413 |
+
"""
|
414 |
+
if self.is_qag:
|
415 |
+
return self.generate_qa_end2end(list_context, batch_size, num_beams, cache_path)
|
416 |
+
single_input = type(list_context) is str
|
417 |
+
list_context = [list_context] if single_input else list_context
|
418 |
+
original_input_length = len(list_context)
|
419 |
+
|
420 |
+
logging.info('running model for `ae`')
|
421 |
+
list_answer = self.generate_a(
|
422 |
+
list_context,
|
423 |
+
batch_size=batch_size,
|
424 |
+
num_beams=num_beams,
|
425 |
+
cache_path=cache_path,
|
426 |
+
sentence_level=sentence_level,
|
427 |
+
num_questions=num_questions
|
428 |
+
)
|
429 |
+
valid_context_id = [n for n, a in enumerate(list_answer) if a is not None]
|
430 |
+
list_context = [list_context[n] for n in valid_context_id]
|
431 |
+
list_answer = [list_answer[n] for n in valid_context_id]
|
432 |
+
qg_input, qg_hl, list_length = [], [], [0]
|
433 |
+
for c, a in zip(list_context, list_answer):
|
434 |
+
qg_hl += a
|
435 |
+
qg_input += [c] * len(a)
|
436 |
+
list_length.append(list_length[-1] + len(a))
|
437 |
+
logging.info('running model for `qg`')
|
438 |
+
list_question = self.generate_q(
|
439 |
+
qg_input,
|
440 |
+
list_answer=qg_hl,
|
441 |
+
batch_size=batch_size,
|
442 |
+
cache_path=cache_path,
|
443 |
+
num_beams=num_beams,
|
444 |
+
sentence_level=sentence_level
|
445 |
+
)
|
446 |
+
|
447 |
+
assert len(qg_hl) == len(list_question), f"{len(qg_input)} != {len(list_question)}"
|
448 |
+
|
449 |
+
# return to nested list
|
450 |
+
list_question = [list_question[list_length[n - 1]:list_length[n]] for n in range(1, len(list_length))]
|
451 |
+
list_answer = [qg_hl[list_length[n - 1]:list_length[n]] for n in range(1, len(list_length))]
|
452 |
+
output_list = [None] * original_input_length
|
453 |
+
# print(len(valid_context_id), valid_context_id[:10], valid_context_id[-10:0])
|
454 |
+
# print(original_input_length)
|
455 |
+
# print(len(list_question), len(list_answer))
|
456 |
+
for n, _id in enumerate(valid_context_id):
|
457 |
+
output_list[_id] = [(q, a) for q, a in zip(list_question[n], list_answer[n])]
|
458 |
+
return output_list[0] if single_input else output_list
|
459 |
+
|
460 |
+
def generate_a(self,
|
461 |
+
context: str or List,
|
462 |
+
batch_size: int = None,
|
463 |
+
num_beams: int = 4,
|
464 |
+
cache_path: str = None,
|
465 |
+
sentence_level: bool = False,
|
466 |
+
num_questions: int = None):
|
467 |
+
""" Generate answers from each sentence.
|
468 |
+
|
469 |
+
@param context: Input text.
|
470 |
+
@param batch_size: Batch size.
|
471 |
+
@param num_beams: Number of beam for model generation.
|
472 |
+
@param cache_path: Path to pre-compute features.
|
473 |
+
@param sentence_level: Run prediction on each sentence of the context independently to reduce complexity.
|
474 |
+
@param num_questions: Max number of questions.
|
475 |
+
@return: List of generated answers.
|
476 |
+
"""
|
477 |
+
logging.info(f'running model for `answer_extraction`')
|
478 |
+
if self.answer_model_type == 'spacy':
|
479 |
+
num_questions = 10 if num_questions is None else num_questions
|
480 |
+
if type(context) is str:
|
481 |
+
return self.spacy_module.keyword(context, num_questions)
|
482 |
+
else:
|
483 |
+
return [self.spacy_module.keyword(c, num_questions) for c in context]
|
484 |
+
single_input = type(context) is str
|
485 |
+
context = [context] if single_input else context
|
486 |
+
list_sentences = [self.spacy_module.sentence(c) for c in context] # split into sentence
|
487 |
+
list_inputs = [[c] * len(s) for c, s in zip(context, list_sentences)]
|
488 |
+
list_length = [0] + np.cumsum([len(s) for s in list_sentences]).tolist()
|
489 |
+
if sentence_level:
|
490 |
+
list_inputs = list_sentences
|
491 |
+
# flatten inputs
|
492 |
+
flat_sentences = list(chain(*list_sentences))
|
493 |
+
flat_inputs = list(chain(*list_inputs))
|
494 |
+
if self.answer_model_type == 'multitask':
|
495 |
+
answer = self.generate_prediction(
|
496 |
+
flat_inputs, # list_input,
|
497 |
+
highlights=flat_sentences, # highlights=list_sentence,
|
498 |
+
prefix_type='ae' if self.add_prefix else None,
|
499 |
+
cache_path=cache_path,
|
500 |
+
num_beams=num_beams,
|
501 |
+
batch_size=batch_size
|
502 |
+
)
|
503 |
+
elif self.answer_model_type == 'pipeline':
|
504 |
+
answer = self.generate_prediction(
|
505 |
+
flat_inputs, # list_input,
|
506 |
+
highlights=flat_sentences, # highlights=list_sentence,
|
507 |
+
prefix_type='ae' if self.add_prefix_ae else None,
|
508 |
+
cache_path=cache_path,
|
509 |
+
num_beams=num_beams,
|
510 |
+
batch_size=batch_size,
|
511 |
+
switch_to_model_ae=True
|
512 |
+
)
|
513 |
+
else:
|
514 |
+
raise ValueError(f"unknown answer model type: {self.answer_model_type}")
|
515 |
+
# return to nested list
|
516 |
+
answer = [clean(a) for a in answer]
|
517 |
+
list_answer = [answer[list_length[n - 1]:list_length[n]] for n in range(1, len(list_length))]
|
518 |
+
list_answer = [[a for a, c in zip(a_sent, c_sent) if a is not None and a in c]
|
519 |
+
for a_sent, c_sent in zip(list_answer, list_inputs)]
|
520 |
+
list_answer = [None if len(a) == 0 else a for a in list_answer]
|
521 |
+
if not self.drop_answer_error_text:
|
522 |
+
if any(a is None for a in list_answer):
|
523 |
+
raise AnswerNotFoundError([context[n] for n, a in enumerate(list_answer) if a is None][0])
|
524 |
+
return list_answer[0] if single_input else list_answer
|
525 |
+
|
526 |
+
def generate_q(self,
|
527 |
+
list_context: str or List,
|
528 |
+
list_answer: List = None,
|
529 |
+
batch_size: int = None,
|
530 |
+
num_beams: int = 4,
|
531 |
+
cache_path: str = None,
|
532 |
+
sentence_level: bool = False):
|
533 |
+
""" Generate question from paragraph and answer. Note that `list_answer` is needed unless they are already
|
534 |
+
highlighted in the `list_context`. eg) "I live in <hl> Tokyo <hl>."
|
535 |
+
|
536 |
+
@param list_context: List of input texts.
|
537 |
+
@param list_answer: List of answers in the `list_context` that are highlighted by <hl>.
|
538 |
+
@param batch_size: Batch size.
|
539 |
+
@param num_beams: Number of beam for model generation.
|
540 |
+
@param cache_path: Path to pre-compute features.
|
541 |
+
@param sentence_level: Run prediction on each sentence of the context independently to reduce complexity.
|
542 |
+
@return: List of generated sentences.
|
543 |
+
"""
|
544 |
+
assert self.is_qg, "model is not fine-tuned for QG"
|
545 |
+
if list_answer is not None:
|
546 |
+
assert type(list_context) is type(list_answer), f"{type(list_context)} != {type(list_answer)}"
|
547 |
+
single_input = False
|
548 |
+
if type(list_context) is str:
|
549 |
+
list_context = [list_context]
|
550 |
+
list_answer = [list_answer] if list_answer is not None else None
|
551 |
+
single_input = True
|
552 |
+
output = self.generate_prediction(
|
553 |
+
list_context,
|
554 |
+
highlights=list_answer,
|
555 |
+
prefix_type='qg' if self.add_prefix else None,
|
556 |
+
cache_path=cache_path,
|
557 |
+
num_beams=num_beams,
|
558 |
+
batch_size=batch_size,
|
559 |
+
sentence_level=sentence_level
|
560 |
+
)
|
561 |
+
if single_input:
|
562 |
+
return output[0]
|
563 |
+
return output
|
564 |
+
|
565 |
+
def answer_q(self,
|
566 |
+
list_context: str or List,
|
567 |
+
list_question: str or List,
|
568 |
+
batch_size: int = None,
|
569 |
+
num_beams: int = 4,
|
570 |
+
cache_path: str = None):
|
571 |
+
logging.info(f'running model for `question_answering`')
|
572 |
+
assert self.is_qa, "model is not fine-tuned for QA"
|
573 |
+
assert type(list_context) is type(list_question), "invalid input"
|
574 |
+
single_input = type(list_context) is str
|
575 |
+
list_context = [list_context] if single_input else list_context
|
576 |
+
list_question = [list_question] if single_input else list_question
|
577 |
+
assert len(list_context) == len(list_question), f"invalid input: {len(list_context)} != {len(list_question)}"
|
578 |
+
output = self.generate_prediction(
|
579 |
+
[f"question: {q}, context: {c}" for q, c in zip(list_question, list_context)],
|
580 |
+
batch_size=batch_size,
|
581 |
+
prefix_type='qa' if self.add_prefix else None,
|
582 |
+
cache_path=cache_path,
|
583 |
+
num_beams=num_beams
|
584 |
+
)
|
585 |
+
return output[0] if single_input else output
|
586 |
+
|
587 |
+
def generate_prediction(self,
|
588 |
+
inputs: List,
|
589 |
+
highlights: List or None = None,
|
590 |
+
prefix_type: str = None,
|
591 |
+
num_beams: int = 4,
|
592 |
+
batch_size: int = None,
|
593 |
+
cache_path: str = None,
|
594 |
+
sentence_level: bool = False,
|
595 |
+
switch_to_model_ae: bool = False):
|
596 |
+
""" General method to generate model prediction
|
597 |
+
|
598 |
+
@param inputs: List of input sequences.
|
599 |
+
@param highlights: List of sub-sequences from list_context to be highlighted by <hl>.
|
600 |
+
@param batch_size: Batch size.
|
601 |
+
@param num_beams: Number of beam for model generation.
|
602 |
+
@param cache_path: Path to pre-compute features.
|
603 |
+
@param prefix_type: Either of `qg` or `answer_extraction`, which is to add at the beginning of the text.
|
604 |
+
@return: List of generated sequences.
|
605 |
+
"""
|
606 |
+
self.eval()
|
607 |
+
if switch_to_model_ae:
|
608 |
+
assert self.model_ae is not None and self.tokenizer_ae is not None
|
609 |
+
model = self.model_ae
|
610 |
+
tokenizer = self.tokenizer_ae
|
611 |
+
max_length_output = self.max_length_output_ae
|
612 |
+
else:
|
613 |
+
model = self.model
|
614 |
+
tokenizer = self.tokenizer
|
615 |
+
max_length_output = self.max_length_output
|
616 |
+
|
617 |
+
if sentence_level:
|
618 |
+
assert highlights is not None, '`sentence_level` needs `highlights`.'
|
619 |
+
assert len(highlights) == len(inputs), str([len(highlights), len(inputs)])
|
620 |
+
list_sentence = []
|
621 |
+
for context, answer in zip(inputs, highlights):
|
622 |
+
s = [sentence for sentence in self.spacy_module.sentence(context) if answer in sentence]
|
623 |
+
list_sentence.append(s[0] if len(s) != 0 else context)
|
624 |
+
inputs = list_sentence
|
625 |
+
|
626 |
+
assert type(inputs) is list, inputs
|
627 |
+
encode_list = self.text_to_encode(
|
628 |
+
inputs,
|
629 |
+
highlights=highlights,
|
630 |
+
prefix_type=prefix_type,
|
631 |
+
cache_path=cache_path,
|
632 |
+
switch_to_model_ae=switch_to_model_ae
|
633 |
+
)
|
634 |
+
loader = self.get_data_loader(encode_list, batch_size=batch_size)
|
635 |
+
outputs = []
|
636 |
+
for encode in loader:
|
637 |
+
with torch.no_grad():
|
638 |
+
if 'labels' in encode:
|
639 |
+
encode.pop('labels')
|
640 |
+
encode = {k: v.to(self.device) for k, v in encode.items()}
|
641 |
+
encode['max_length'] = max_length_output
|
642 |
+
encode['num_beams'] = num_beams
|
643 |
+
tensor = model.module.generate(**encode) if self.parallel else model.generate(**encode)
|
644 |
+
outputs += tokenizer.batch_decode(tensor, skip_special_tokens=True)
|
645 |
+
return outputs
|
646 |
+
|
647 |
+
def encode_to_loss(self, encode: Dict):
|
648 |
+
""" Transform encoded features to loss value for model finetuning.
|
649 |
+
|
650 |
+
@param encode: Encoded feature.
|
651 |
+
@return: Loss value.
|
652 |
+
"""
|
653 |
+
assert 'labels' in encode
|
654 |
+
output = self.model(**{k: v.to(self.device) for k, v in encode.items()})
|
655 |
+
if self.label_smoothing is None or self.label_smoothing == 0.0:
|
656 |
+
return output['loss'].mean() if self.parallel else output['loss']
|
657 |
+
else:
|
658 |
+
return label_smoothed_loss(output['logits'], encode['labels'].to(self.device), self.label_smoothing)
|
659 |
+
|
660 |
+
def text_to_encode(self,
|
661 |
+
inputs,
|
662 |
+
outputs: List = None,
|
663 |
+
highlights: List = None,
|
664 |
+
prefix_type: str = None,
|
665 |
+
cache_path: str = None,
|
666 |
+
switch_to_model_ae: bool = False):
|
667 |
+
""" Transform texts into encoded features.
|
668 |
+
|
669 |
+
@param inputs: List of input sequences.
|
670 |
+
@param outputs: List of output sequences.
|
671 |
+
@param highlights: List of sub-sequences from `inputs` to be highlighted by <hl>.
|
672 |
+
@param prefix_type: Either of `qg` or `answer_extraction`, which is to add at the beginning of the text.
|
673 |
+
@param cache_path: Path to pre-compute features.
|
674 |
+
@return: List of encoded feature.
|
675 |
+
"""
|
676 |
+
if cache_path is not None and os.path.exists(cache_path):
|
677 |
+
logging.info(f'loading preprocessed feature from {cache_path}')
|
678 |
+
return pickle_load(cache_path)
|
679 |
+
outputs = [None] * len(inputs) if outputs is None else outputs
|
680 |
+
highlights = [None] * len(inputs) if highlights is None else highlights
|
681 |
+
assert len(outputs) == len(inputs) == len(highlights), str([len(outputs), len(inputs), len(highlights)])
|
682 |
+
data = list(zip(inputs, outputs, highlights))
|
683 |
+
# process in parallel/single
|
684 |
+
config = {'tokenizer': self.tokenizer, 'max_length': self.max_length, 'prefix_type': prefix_type,
|
685 |
+
'max_length_output': self.max_length_output, 'drop_overflow_error_text': self.drop_overflow_error_text,
|
686 |
+
'skip_overflow_error': self.skip_overflow_error, 'drop_highlight_error_text': self.drop_highlight_error_text,
|
687 |
+
'padding': False if len(data) == 1 else True}
|
688 |
+
if switch_to_model_ae:
|
689 |
+
assert self.model_ae is not None and self.tokenizer_ae is not None
|
690 |
+
config['tokenizer'] = self.tokenizer_ae
|
691 |
+
config['max_length'] = self.max_length_ae
|
692 |
+
config['max_length_output'] = self.max_length_output_ae
|
693 |
+
|
694 |
+
logging.info(f'encode all the data : {len(data)}')
|
695 |
+
if cache_path is not None:
|
696 |
+
os.makedirs(os.path.dirname(cache_path), exist_ok=True)
|
697 |
+
if PARALLEL_PROCESSING:
|
698 |
+
pool = Pool()
|
699 |
+
out = pool.map(EncodePlus(**config), data)
|
700 |
+
pool.close()
|
701 |
+
out = list(filter(None, out)) # remove overflow text
|
702 |
+
else:
|
703 |
+
f = EncodePlus(**config)
|
704 |
+
out = []
|
705 |
+
files = []
|
706 |
+
for i in tqdm(data):
|
707 |
+
e = f(i)
|
708 |
+
if e is not None: # remove overflow text
|
709 |
+
out.append(e)
|
710 |
+
if len(out) > 40000 and cache_path is not None:
|
711 |
+
pickle_save(out, f'{cache_path}.tmp{len(files)}')
|
712 |
+
files.append(f'{cache_path}.tmp{len(files)}')
|
713 |
+
out = []
|
714 |
+
if len(out) > 0 and cache_path is not None:
|
715 |
+
pickle_save(out, f'{cache_path}.tmp{len(files)}')
|
716 |
+
files.append(f'{cache_path}.tmp{len(files)}')
|
717 |
+
if len(files) > 0:
|
718 |
+
out = list(chain(*[pickle_load(i) for i in files]))
|
719 |
+
logging.info(f'after remove the overflow : {len(out)}')
|
720 |
+
# cache the encoded data
|
721 |
+
if cache_path is not None:
|
722 |
+
pickle_save(out, cache_path)
|
723 |
+
logging.info(f'preprocessed feature is saved at {cache_path}')
|
724 |
+
return out
|
725 |
+
|
726 |
+
def save(self, save_dir):
|
727 |
+
""" Save model.
|
728 |
+
|
729 |
+
@param save_dir: Directory to save model related file.
|
730 |
+
"""
|
731 |
+
|
732 |
+
def model_state(model):
|
733 |
+
if self.parallel:
|
734 |
+
return model.module
|
735 |
+
return model
|
736 |
+
|
737 |
+
logging.info('saving model')
|
738 |
+
model_state(self.model).config.update({'add_prefix': self.add_prefix})
|
739 |
+
model_state(self.model).save_pretrained(save_dir)
|
740 |
+
logging.info('saving tokenizer')
|
741 |
+
self.tokenizer.save_pretrained(save_dir)
|
742 |
+
|
743 |
+
@staticmethod
|
744 |
+
def get_data_loader(encode_list, batch_size: int = None, shuffle: bool = False, drop_last: bool = False):
|
745 |
+
""" Get torch.utils.data.DataLoader instance.
|
746 |
+
|
747 |
+
@param encode_list: List of encoded features.
|
748 |
+
@param batch_size: Batch size.
|
749 |
+
@param shuffle: Shuffle data.
|
750 |
+
@param drop_last: Drop residual batch.
|
751 |
+
@return: torch.utils.data.DataLoader
|
752 |
+
"""
|
753 |
+
batch_size = len(encode_list) if batch_size is None else batch_size
|
754 |
+
params = dict(batch_size=batch_size, shuffle=shuffle, drop_last=drop_last, num_workers=NUM_WORKERS)
|
755 |
+
return torch.utils.data.DataLoader(Dataset(encode_list), **params)
|
756 |
+
|
757 |
+
def train(self):
|
758 |
+
self.model.train()
|
759 |
+
|
760 |
+
def eval(self):
|
761 |
+
self.model.eval()
|