feat: added encode method
Browse files- configuration_bert.py +2 -0
- modeling_bert.py +165 -0
configuration_bert.py
CHANGED
@@ -84,6 +84,7 @@ class JinaBertConfig(PretrainedConfig):
|
|
84 |
num_tasks=0,
|
85 |
use_flash_attn=True,
|
86 |
use_qk_norm=True,
|
|
|
87 |
**kwargs,
|
88 |
):
|
89 |
assert 'position_embedding_type' not in kwargs
|
@@ -112,3 +113,4 @@ class JinaBertConfig(PretrainedConfig):
|
|
112 |
self.num_tasks = num_tasks
|
113 |
self.use_flash_attn = use_flash_attn
|
114 |
self.use_qk_norm = use_qk_norm
|
|
|
|
84 |
num_tasks=0,
|
85 |
use_flash_attn=True,
|
86 |
use_qk_norm=True,
|
87 |
+
emb_pooler=None,
|
88 |
**kwargs,
|
89 |
):
|
90 |
assert 'position_embedding_type' not in kwargs
|
|
|
113 |
self.num_tasks = num_tasks
|
114 |
self.use_flash_attn = use_flash_attn
|
115 |
self.use_qk_norm = use_qk_norm
|
116 |
+
self.emb_pooler = emb_pooler
|
modeling_bert.py
CHANGED
@@ -15,7 +15,10 @@ and made modifications to use ALiBi.
|
|
15 |
import logging
|
16 |
from collections.abc import Sequence
|
17 |
from functools import partial
|
|
|
|
|
18 |
|
|
|
19 |
import torch
|
20 |
import torch.nn as nn
|
21 |
import torch.nn.functional as F
|
@@ -54,6 +57,10 @@ try:
|
|
54 |
except ImportError:
|
55 |
CrossEntropyLoss = None
|
56 |
|
|
|
|
|
|
|
|
|
57 |
|
58 |
logger = logging.getLogger(__name__)
|
59 |
|
@@ -346,6 +353,15 @@ class BertModel(BertPreTrainedModel):
|
|
346 |
self.pooler = BertPooler(config) if add_pooling_layer else None
|
347 |
self.task_type_embeddings = nn.Embedding(config.num_tasks, config.hidden_size)
|
348 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
349 |
# We now initialize the task embeddings to 0; We do not use task types during
|
350 |
# pretraining. When we start using task types during embedding training,
|
351 |
# we want the model to behave exactly as in pretraining (i.e. task types
|
@@ -419,6 +435,155 @@ class BertModel(BertPreTrainedModel):
|
|
419 |
)
|
420 |
|
421 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
422 |
class BertForPreTraining(BertPreTrainedModel):
|
423 |
def __init__(self, config: JinaBertConfig):
|
424 |
super().__init__(config)
|
|
|
15 |
import logging
|
16 |
from collections.abc import Sequence
|
17 |
from functools import partial
|
18 |
+
from typing import Union, List, Optional
|
19 |
+
import warnings
|
20 |
|
21 |
+
import numpy as np
|
22 |
import torch
|
23 |
import torch.nn as nn
|
24 |
import torch.nn.functional as F
|
|
|
57 |
except ImportError:
|
58 |
CrossEntropyLoss = None
|
59 |
|
60 |
+
try:
|
61 |
+
from tqdm.autonotebook import trange
|
62 |
+
except ImportError:
|
63 |
+
trange = None
|
64 |
|
65 |
logger = logging.getLogger(__name__)
|
66 |
|
|
|
353 |
self.pooler = BertPooler(config) if add_pooling_layer else None
|
354 |
self.task_type_embeddings = nn.Embedding(config.num_tasks, config.hidden_size)
|
355 |
|
356 |
+
self.emb_pooler = config.emb_pooler
|
357 |
+
self._name_or_path = config._name_or_path
|
358 |
+
if self.emb_pooler is not None:
|
359 |
+
from transformers import AutoTokenizer
|
360 |
+
|
361 |
+
self.tokenizer = AutoTokenizer.from_pretrained(config._name_or_path)
|
362 |
+
else:
|
363 |
+
self.tokenizer = None
|
364 |
+
|
365 |
# We now initialize the task embeddings to 0; We do not use task types during
|
366 |
# pretraining. When we start using task types during embedding training,
|
367 |
# we want the model to behave exactly as in pretraining (i.e. task types
|
|
|
435 |
)
|
436 |
|
437 |
|
438 |
+
@torch.inference_mode()
|
439 |
+
def encode(
|
440 |
+
self: 'BertModel',
|
441 |
+
sentences: Union[str, List[str]],
|
442 |
+
batch_size: int = 32,
|
443 |
+
show_progress_bar: Optional[bool] = None,
|
444 |
+
output_value: str = 'sentence_embedding',
|
445 |
+
convert_to_numpy: bool = True,
|
446 |
+
convert_to_tensor: bool = False,
|
447 |
+
device: Optional[torch.device] = None,
|
448 |
+
normalize_embeddings: bool = False,
|
449 |
+
**tokenizer_kwargs,
|
450 |
+
) -> Union[List[torch.Tensor], np.ndarray, torch.Tensor]:
|
451 |
+
"""
|
452 |
+
Computes sentence embeddings
|
453 |
+
Args:
|
454 |
+
sentences(`str` or `List[str]`):
|
455 |
+
Sentence or sentences to be encoded
|
456 |
+
batch_size(`int`, *optional*, defaults to 32):
|
457 |
+
Batch size for the computation
|
458 |
+
show_progress_bar(`bool`, *optional*, defaults to None):
|
459 |
+
Show a progress bar when encoding sentences.
|
460 |
+
If set to None, progress bar is only shown when `logger.level == logging.INFO` or `logger.level == logging.DEBUG`.
|
461 |
+
output_value(`str`, *optional*, defaults to 'sentence_embedding'):
|
462 |
+
Default sentence_embedding, to get sentence embeddings.
|
463 |
+
Can be set to token_embeddings to get wordpiece token embeddings.
|
464 |
+
Set to None, to get all output values
|
465 |
+
convert_to_numpy(`bool`, *optional*, defaults to True):
|
466 |
+
If true, the output is a list of numpy vectors.
|
467 |
+
Else, it is a list of pytorch tensors.
|
468 |
+
convert_to_tensor(`bool`, *optional*, defaults to False):
|
469 |
+
If true, you get one large tensor as return.
|
470 |
+
Overwrites any setting from convert_to_numpy
|
471 |
+
device(`torch.device`, *optional*, defaults to None):
|
472 |
+
Which torch.device to use for the computation
|
473 |
+
normalize_embeddings(`bool`, *optional*, defaults to False):
|
474 |
+
If set to true, returned vectors will have length 1. In that case, the faster dot-product (util.dot_score) instead of cosine similarity can be used.
|
475 |
+
tokenizer_kwargs(`Dict[str, Any]`, *optional*, defaults to {}):
|
476 |
+
Keyword arguments for the tokenizer
|
477 |
+
Returns:
|
478 |
+
By default, a list of tensors is returned.
|
479 |
+
If convert_to_tensor, a stacked tensor is returned.
|
480 |
+
If convert_to_numpy, a numpy matrix is returned.
|
481 |
+
"""
|
482 |
+
if self.emb_pooler is None:
|
483 |
+
warnings.warn("No emb_pooler specified, defaulting to mean pooling.")
|
484 |
+
self.emb_pooler = 'mean'
|
485 |
+
from transformers import AutoTokenizer
|
486 |
+
|
487 |
+
self.tokenizer = AutoTokenizer.from_pretrained(self._name_or_path)
|
488 |
+
if self.emb_pooler != 'mean':
|
489 |
+
raise NotImplementedError
|
490 |
+
|
491 |
+
is_training = self.training
|
492 |
+
self.eval()
|
493 |
+
|
494 |
+
if show_progress_bar is None:
|
495 |
+
show_progress_bar = (
|
496 |
+
logger.getEffectiveLevel() == logging.INFO
|
497 |
+
or logger.getEffectiveLevel() == logging.DEBUG
|
498 |
+
)
|
499 |
+
|
500 |
+
if convert_to_tensor:
|
501 |
+
convert_to_numpy = False
|
502 |
+
|
503 |
+
if output_value != 'sentence_embedding':
|
504 |
+
convert_to_tensor = False
|
505 |
+
convert_to_numpy = False
|
506 |
+
|
507 |
+
input_was_string = False
|
508 |
+
if isinstance(sentences, str) or not hasattr(sentences, '__len__'):
|
509 |
+
sentences = [sentences]
|
510 |
+
input_was_string = True
|
511 |
+
|
512 |
+
if device is not None:
|
513 |
+
self.to(device)
|
514 |
+
|
515 |
+
# TODO: Maybe use better length heuristic?
|
516 |
+
permutation = np.argsort([-len(i) for i in sentences])
|
517 |
+
inverse_permutation = np.argsort(permutation)
|
518 |
+
sentences = [sentences[idx] for idx in permutation]
|
519 |
+
|
520 |
+
tokenizer_kwargs['padding'] = tokenizer_kwargs.get('padding', True)
|
521 |
+
tokenizer_kwargs['max_length'] = tokenizer_kwargs.get('max_length', 8192)
|
522 |
+
tokenizer_kwargs['truncation'] = tokenizer_kwargs.get('truncation', True)
|
523 |
+
|
524 |
+
all_embeddings = []
|
525 |
+
|
526 |
+
if trange is not None:
|
527 |
+
range_iter = trange(
|
528 |
+
0,
|
529 |
+
len(sentences),
|
530 |
+
batch_size,
|
531 |
+
desc="Encoding",
|
532 |
+
disable=not show_progress_bar,
|
533 |
+
)
|
534 |
+
else:
|
535 |
+
range_iter = range(0, len(sentences), batch_size)
|
536 |
+
|
537 |
+
for i in range_iter:
|
538 |
+
encoded_input = self.tokenizer(
|
539 |
+
sentences[i : i + batch_size],
|
540 |
+
return_tensors='pt',
|
541 |
+
**tokenizer_kwargs,
|
542 |
+
).to(self.device)
|
543 |
+
token_embs = self.forward(**encoded_input)[0]
|
544 |
+
|
545 |
+
# Accumulate in fp32 to avoid overflow
|
546 |
+
token_embs = token_embs.float()
|
547 |
+
|
548 |
+
if output_value == 'token_embeddings':
|
549 |
+
raise NotImplementedError
|
550 |
+
elif output_value is None:
|
551 |
+
raise NotImplementedError
|
552 |
+
else:
|
553 |
+
embeddings = self.mean_pooling(
|
554 |
+
token_embs, encoded_input['attention_mask']
|
555 |
+
)
|
556 |
+
|
557 |
+
if normalize_embeddings:
|
558 |
+
embeddings = torch.nn.functional.normalize(embeddings, p=2, dim=1)
|
559 |
+
|
560 |
+
if convert_to_numpy:
|
561 |
+
embeddings = embeddings.cpu()
|
562 |
+
all_embeddings.extend(embeddings)
|
563 |
+
|
564 |
+
all_embeddings = [all_embeddings[idx] for idx in inverse_permutation]
|
565 |
+
|
566 |
+
if convert_to_tensor:
|
567 |
+
all_embeddings = torch.stack(all_embeddings)
|
568 |
+
elif convert_to_numpy:
|
569 |
+
all_embeddings = np.asarray([emb.numpy() for emb in all_embeddings])
|
570 |
+
|
571 |
+
if input_was_string:
|
572 |
+
all_embeddings = all_embeddings[0]
|
573 |
+
|
574 |
+
self.train(is_training)
|
575 |
+
return all_embeddings
|
576 |
+
|
577 |
+
def mean_pooling(
|
578 |
+
self, token_embeddings: torch.Tensor, attention_mask: torch.Tensor
|
579 |
+
):
|
580 |
+
input_mask_expanded = (
|
581 |
+
attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
|
582 |
+
)
|
583 |
+
return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(
|
584 |
+
input_mask_expanded.sum(1), min=1e-9
|
585 |
+
)
|
586 |
+
|
587 |
class BertForPreTraining(BertPreTrainedModel):
|
588 |
def __init__(self, config: JinaBertConfig):
|
589 |
super().__init__(config)
|