-
Notifications
You must be signed in to change notification settings - Fork 387
/
bert.py
447 lines (386 loc) · 20.5 KB
/
bert.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
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
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
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
# Copyright The Lightning team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import csv
import urllib
from typing import Any, Callable, Dict, List, Optional, Sequence, Tuple, Union
import torch
from torch import Tensor
from torch.nn import Module
from torch.utils.data import DataLoader
from torchmetrics.functional.text.helper_embedding_metric import (
TextDataset,
TokenizedDataset,
_check_shape_of_model_output,
_get_progress_bar,
_input_data_collator,
_output_data_collator,
_process_attention_mask_for_special_tokens,
)
from torchmetrics.utilities import rank_zero_warn
from torchmetrics.utilities.checks import _SKIP_SLOW_DOCTEST, _try_proceed_with_timeout
from torchmetrics.utilities.imports import _TQDM_AVAILABLE, _TRANSFORMERS_GREATER_EQUAL_4_4
# Default model recommended in the original implementation.
_DEFAULT_MODEL = "roberta-large"
if _TRANSFORMERS_GREATER_EQUAL_4_4:
from transformers import AutoModel, AutoTokenizer
def _download_model() -> None:
"""Download intensive operations."""
AutoTokenizer.from_pretrained(_DEFAULT_MODEL)
AutoModel.from_pretrained(_DEFAULT_MODEL)
if _SKIP_SLOW_DOCTEST and not _try_proceed_with_timeout(_download_model):
__doctest_skip__ = ["bert_score"]
else:
__doctest_skip__ = ["bert_score"]
def _get_embeddings_and_idf_scale(
dataloader: DataLoader,
target_len: int,
model: Module,
device: Optional[Union[str, torch.device]] = None,
num_layers: Optional[int] = None,
all_layers: bool = False,
idf: bool = False,
verbose: bool = False,
user_forward_fn: Optional[Callable[[Module, Dict[str, Tensor]], Tensor]] = None,
) -> Tuple[Tensor, Tensor]:
"""Calculate sentence embeddings and the inverse-document-frequency scaling factor.
Args:
dataloader: dataloader instance.
target_len: A length of the longest sequence in the data. Used for padding the model output.
model: BERT model.
device: A device to be used for calculation.
num_layers: The layer of representation to use.
all_layers: An indication whether representation from all model layers should be used for BERTScore.
idf: An Indication whether normalization using inverse document frequencies should be used.
verbose: An indication of whether a progress bar to be displayed during the embeddings' calculation.
user_forward_fn:
A user's own forward function used in a combination with ``user_model``. This function must
take ``user_model`` and a python dictionary of containing ``"input_ids"`` and ``"attention_mask"``
represented by :class:`~torch.Tensor` as an input and return the model's output represented by the single
:class:`~torch.Tensor`.
Return:
A tuple of :class:`~torch.Tensor`s containing the model's embeddings and the normalized tokens IDF.
When ``idf = False``, tokens IDF is not calculated, and a matrix of mean weights is returned instead.
For a single sentence, ``mean_weight = 1/seq_len``, where ``seq_len`` is a sum over the corresponding
``attention_mask``.
Raises:
ValueError:
If ``all_layers = True`` and a model, which is not from the ``transformers`` package, is used.
"""
embeddings_list: List[Tensor] = []
idf_scale_list: List[Tensor] = []
for batch in _get_progress_bar(dataloader, verbose):
with torch.no_grad():
batch = _input_data_collator(batch, device)
# Output shape: batch_size x num_layers OR 1 x sequence_length x bert_dim
if not all_layers:
if not user_forward_fn:
out = model(batch["input_ids"], batch["attention_mask"], output_hidden_states=True)
out = out.hidden_states[num_layers if num_layers is not None else -1]
else:
out = user_forward_fn(model, batch)
_check_shape_of_model_output(out, batch["input_ids"])
out = out.unsqueeze(1)
else:
if user_forward_fn:
raise ValueError(
"The option `all_layers=True` can be used only with default `transformers` models."
)
out = model(batch["input_ids"], batch["attention_mask"], output_hidden_states=True)
out = torch.cat([o.unsqueeze(1) for o in out.hidden_states], dim=1)
out /= out.norm(dim=-1).unsqueeze(-1) # normalize embeddings
out, attention_mask = _output_data_collator(out, batch["attention_mask"], target_len)
processed_attention_mask = _process_attention_mask_for_special_tokens(attention_mask)
# Multiply embeddings with attention_mask (b=batch_size, l=num_layers, s=seq_len, d=emb_dim)
out = torch.einsum("blsd, bs -> blsd", out, processed_attention_mask)
embeddings_list.append(out.cpu())
# Calculate weighted (w.r.t. sentence length) input_ids IDF matrix
input_ids_idf = (
batch["input_ids_idf"] * processed_attention_mask if idf else processed_attention_mask.type(out.dtype)
)
input_ids_idf /= input_ids_idf.sum(-1, keepdim=True)
idf_scale_list.append(input_ids_idf.cpu())
embeddings = torch.cat(embeddings_list)
idf_scale = torch.cat(idf_scale_list)
return embeddings, idf_scale
def _get_scaled_precision_or_recall(cos_sim: Tensor, metric: str, idf_scale: Tensor) -> Tensor:
"""Calculate precision or recall, transpose it and scale it with idf_scale factor."""
dim = 3 if metric == "precision" else 2
res = cos_sim.max(dim=dim).values
res = torch.einsum("bls, bs -> bls", res, idf_scale).sum(-1)
# We transpose the results and squeeze if possible to match the format of the original BERTScore implementation
return res.transpose(0, 1).squeeze()
def _get_precision_recall_f1(
preds_embeddings: Tensor, target_embeddings: Tensor, preds_idf_scale: Tensor, target_idf_scale: Tensor
) -> Tuple[Tensor, Tensor, Tensor]:
"""Calculate precision, recall and F1 score over candidate and reference sentences.
Args:
preds_embeddings: Embeddings of candidate sentences.
target_embeddings: Embeddings of reference sentences.
preds_idf_scale: An IDF scale factor for candidate sentences.
target_idf_scale: An IDF scale factor for reference sentences.
Return:
Tensors containing precision, recall and F1 score, respectively.
"""
# Dimensions: b = batch_size, l = num_layers, p = predictions_seq_len, r = references_seq_len, d = bert_dim
cos_sim = torch.einsum("blpd, blrd -> blpr", preds_embeddings, target_embeddings)
# Final metrics shape = (batch_size * num_layers | batch_size)
precision = _get_scaled_precision_or_recall(cos_sim, "precision", preds_idf_scale)
recall = _get_scaled_precision_or_recall(cos_sim, "recall", target_idf_scale)
f1_score = 2 * precision * recall / (precision + recall)
f1_score = f1_score.masked_fill(torch.isnan(f1_score), 0.0)
return precision, recall, f1_score
def _get_hash(model_name_or_path: Optional[str] = None, num_layers: Optional[int] = None, idf: bool = False) -> str:
"""Compute `BERT_score`_ (copied and adjusted)."""
return f"{model_name_or_path}_L{num_layers}{'_idf' if idf else '_no-idf'}"
def _read_csv_from_local_file(baseline_path: str) -> Tensor:
"""Read baseline from csv file from the local file.
This method implemented to avoid `pandas` dependency.
"""
with open(baseline_path) as fname:
csv_file = csv.reader(fname)
baseline_list = [[float(item) for item in row] for idx, row in enumerate(csv_file) if idx > 0]
return torch.tensor(baseline_list)[:, 1:]
def _read_csv_from_url(baseline_url: str) -> Tensor:
"""Read baseline from csv file from URL.
This method is implemented to avoid `pandas` dependency.
"""
with urllib.request.urlopen(baseline_url) as http_request:
baseline_list = [
[float(item) for item in row.strip().decode("utf-8").split(",")]
for idx, row in enumerate(http_request)
if idx > 0
]
return torch.tensor(baseline_list)[:, 1:]
def _load_baseline(
lang: str = "en",
model_name_or_path: Optional[str] = None,
baseline_path: Optional[str] = None,
baseline_url: Optional[str] = None,
) -> Optional[Tensor]:
"""Load a CSV file with the baseline values used for rescaling."""
if baseline_path:
baseline: Optional[Tensor] = _read_csv_from_local_file(baseline_path)
elif baseline_url:
baseline = _read_csv_from_url(baseline_url)
# Read default baseline from the original `bert-score` package https://github.com/Tiiiger/bert_score
elif lang and model_name_or_path:
url_base = "https://raw.githubusercontent.com/Tiiiger/bert_score/master/bert_score/rescale_baseline"
baseline_url = f"{url_base}/{lang}/{model_name_or_path}.tsv"
baseline = _read_csv_from_url(baseline_url)
else:
rank_zero_warn("Baseline was not successfully loaded. No baseline is going to be used.")
return None
return baseline
def _rescale_metrics_with_baseline(
precision: Tensor,
recall: Tensor,
f1_score: Tensor,
baseline: Tensor,
num_layers: Optional[int] = None,
all_layers: bool = False,
) -> Tuple[Tensor, Tensor, Tensor]:
"""Rescale the computed metrics with the pre-computed baseline."""
if num_layers is None and all_layers is False:
num_layers = -1
all_metrics = torch.stack([precision, recall, f1_score], dim=-1)
baseline_scale = baseline.unsqueeze(1) if all_layers else baseline[num_layers]
all_metrics = (all_metrics - baseline_scale) / (1 - baseline_scale)
return all_metrics[..., 0], all_metrics[..., 1], all_metrics[..., 2]
def bert_score(
preds: Union[str, Sequence[str], Dict[str, Tensor]],
target: Union[str, Sequence[str], Dict[str, Tensor]],
model_name_or_path: Optional[str] = None,
num_layers: Optional[int] = None,
all_layers: bool = False,
model: Optional[Module] = None,
user_tokenizer: Any = None,
user_forward_fn: Optional[Callable[[Module, Dict[str, Tensor]], Tensor]] = None,
verbose: bool = False,
idf: bool = False,
device: Optional[Union[str, torch.device]] = None,
max_length: int = 512,
batch_size: int = 64,
num_threads: int = 0,
return_hash: bool = False,
lang: str = "en",
rescale_with_baseline: bool = False,
baseline_path: Optional[str] = None,
baseline_url: Optional[str] = None,
) -> Dict[str, Union[Tensor, List[float], str]]:
"""`Bert_score Evaluating Text Generation`_ for text similirity matching.
This metric leverages the pre-trained contextual embeddings from BERT and matches words in candidate and reference
sentences by cosine similarity. It has been shown to correlate with human judgment on sentence-level and
system-level evaluation. Moreover, BERTScore computes precision, recall, and F1 measure, which can be useful for
evaluating different language generation tasks.
This implementation follows the original implementation from `BERT_score`_.
Args:
preds: Either an iterable of predicted sentences or a ``Dict[input_ids, attention_mask]``.
target: Either an iterable of target sentences or a ``Dict[input_ids, attention_mask]``.
model_name_or_path: A name or a model path used to load ``transformers`` pretrained model.
num_layers: A layer of representation to use.
all_layers:
An indication of whether the representation from all model's layers should be used.
If ``all_layers = True``, the argument ``num_layers`` is ignored.
model: A user's own model.
user_tokenizer:
A user's own tokenizer used with the own model. This must be an instance with the ``__call__`` method.
This method must take an iterable of sentences (``List[str]``) and must return a python dictionary
containing ``"input_ids"`` and ``"attention_mask"`` represented by :class:`~torch.Tensor`.
It is up to the user's model of whether ``"input_ids"`` is a :class:`~torch.Tensor` of input ids
or embedding vectors. his tokenizer must prepend an equivalent of ``[CLS]`` token and append an equivalent
of ``[SEP]`` token as `transformers` tokenizer does.
user_forward_fn:
A user's own forward function used in a combination with ``user_model``.
This function must take ``user_model`` and a python dictionary of containing ``"input_ids"``
and ``"attention_mask"`` represented by :class:`~torch.Tensor` as an input and return the model's output
represented by the single :class:`~torch.Tensor`.
verbose: An indication of whether a progress bar to be displayed during the embeddings' calculation.
idf: An indication of whether normalization using inverse document frequencies should be used.
device: A device to be used for calculation.
max_length: A maximum length of input sequences. Sequences longer than ``max_length`` are to be trimmed.
batch_size: A batch size used for model processing.
num_threads: A number of threads to use for a dataloader.
return_hash: An indication of whether the correspodning ``hash_code`` should be returned.
lang: A language of input sentences. It is used when the scores are rescaled with a baseline.
rescale_with_baseline:
An indication of whether bertscore should be rescaled with a pre-computed baseline.
When a pretrained model from ``transformers`` model is used, the corresponding baseline is downloaded
from the original ``bert-score`` package from `BERT_score`_ if available.
In other cases, please specify a path to the baseline csv/tsv file, which must follow the formatting
of the files from `BERT_score`_
baseline_path: A path to the user's own local csv/tsv file with the baseline scale.
baseline_url: A url path to the user's own csv/tsv file with the baseline scale.
Returns:
Python dictionary containing the keys ``precision``, ``recall`` and ``f1`` with corresponding values.
Raises:
ValueError:
If ``len(preds) != len(target)``.
ModuleNotFoundError:
If `tqdm` package is required and not installed.
ModuleNotFoundError:
If ``transformers`` package is required and not installed.
ValueError:
If ``num_layer`` is larger than the number of the model layers.
ValueError:
If invalid input is provided.
Example:
>>> from pprint import pprint
>>> from torchmetrics.functional.text.bert import bert_score
>>> preds = ["hello there", "general kenobi"]
>>> target = ["hello there", "master kenobi"]
>>> pprint(bert_score(preds, target))
{'f1': tensor([1.0000, 0.9961]), 'precision': tensor([1.0000, 0.9961]), 'recall': tensor([1.0000, 0.9961])}
"""
if len(preds) != len(target):
raise ValueError("Number of predicted and reference sententes must be the same!")
if not isinstance(preds, (str, list, dict)): # dict for BERTScore class compute call
preds = list(preds)
if not isinstance(target, (str, list, dict)): # dict for BERTScore class compute call
target = list(target)
if verbose and (not _TQDM_AVAILABLE):
raise ModuleNotFoundError(
"An argument `verbose = True` requires `tqdm` package be installed. Install with `pip install tqdm`."
)
if model is None:
if not _TRANSFORMERS_GREATER_EQUAL_4_4:
raise ModuleNotFoundError(
"`bert_score` metric with default models requires `transformers` package be installed."
" Either install with `pip install transformers>=4.4` or `pip install torchmetrics[text]`."
)
if model_name_or_path is None:
rank_zero_warn(
"The argument `model_name_or_path` was not specified while it is required when default"
" `transformers` model are used."
f"It is, therefore, used the default recommended model - {_DEFAULT_MODEL}."
)
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path or _DEFAULT_MODEL)
model = AutoModel.from_pretrained(model_name_or_path or _DEFAULT_MODEL)
else:
tokenizer = user_tokenizer
model.eval()
model.to(device)
try:
if num_layers and num_layers > model.config.num_hidden_layers:
raise ValueError(
f"num_layers={num_layers} is forbidden for {model_name_or_path}."
f" Please use num_layers <= {model.config.num_hidden_layers}"
)
except AttributeError:
rank_zero_warn("It was not possible to retrieve the parameter `num_layers` from the model specification.")
_are_empty_lists = all(isinstance(text, list) and len(text) == 0 for text in (preds, target))
_are_valid_lists = all(
isinstance(text, list) and len(text) > 0 and isinstance(text[0], str) for text in (preds, target)
)
_are_valid_tensors = all(
isinstance(text, dict) and isinstance(text["input_ids"], Tensor) for text in (preds, target)
)
if _are_empty_lists:
rank_zero_warn("Predictions and references are empty.")
output_dict: Dict[str, Union[Tensor, List[float], str]] = {
"precision": [0.0],
"recall": [0.0],
"f1": [0.0],
}
if return_hash:
output_dict.update({"hash": _get_hash(model_name_or_path, num_layers, idf)})
return output_dict
# Load baselines if needed
baseline = _load_baseline(lang, model_name_or_path, baseline_path, baseline_url) if rescale_with_baseline else None
# We ignore mypy typing below as the proper typing is ensured by conditions above, only mypy cannot infer that.
if _are_valid_lists:
target_dataset = TextDataset(target, tokenizer, max_length, idf=idf) # type: ignore
preds_dataset = TextDataset(
preds, # type: ignore
tokenizer,
max_length,
idf=idf,
tokens_idf=target_dataset.tokens_idf,
)
elif _are_valid_tensors:
target_dataset = TokenizedDataset(**target, idf=idf) # type: ignore
preds_dataset = TokenizedDataset(**preds, idf=idf, tokens_idf=target_dataset.tokens_idf) # type: ignore
else:
raise ValueError("Invalid input provided.")
target_loader = DataLoader(target_dataset, batch_size=batch_size, num_workers=num_threads)
preds_loader = DataLoader(preds_dataset, batch_size=batch_size, num_workers=num_threads)
target_embeddings, target_idf_scale = _get_embeddings_and_idf_scale(
target_loader, target_dataset.max_length, model, device, num_layers, all_layers, idf, verbose, user_forward_fn
)
preds_embeddings, preds_idf_scale = _get_embeddings_and_idf_scale(
preds_loader, preds_dataset.max_length, model, device, num_layers, all_layers, idf, verbose, user_forward_fn
)
precision, recall, f1_score = _get_precision_recall_f1(
preds_embeddings, target_embeddings, preds_idf_scale, target_idf_scale
)
# Sort predictions
if len(precision.shape) == 1: # i.e. when all_layers = False
precision = precision[preds_loader.dataset.sorting_indices]
recall = recall[preds_loader.dataset.sorting_indices]
f1_score = f1_score[preds_loader.dataset.sorting_indices]
elif len(precision.shape) == 2: # i.e. when all_layers = True
precision = precision[:, preds_loader.dataset.sorting_indices]
recall = recall[:, preds_loader.dataset.sorting_indices]
f1_score = f1_score[:, preds_loader.dataset.sorting_indices]
if baseline is not None:
precision, recall, f1_score = _rescale_metrics_with_baseline(
precision, recall, f1_score, baseline, num_layers, all_layers
)
output_dict = {
"precision": precision,
"recall": recall,
"f1": f1_score,
}
if return_hash:
output_dict.update({"hash": _get_hash(model_name_or_path, num_layers, idf)})
return output_dict