file_utils.py 37.4 KB
Newer Older
thomwolf's avatar
thomwolf committed
1
2
3
4
5
"""
Utilities for working with the local dataset cache.
This file is adapted from the AllenNLP library at https://github.com/allenai/allennlp
Copyright by the AllenNLP authors.
"""
Aymeric Augustin's avatar
Aymeric Augustin committed
6

Aymeric Augustin's avatar
Aymeric Augustin committed
7
import fnmatch
thomwolf's avatar
thomwolf committed
8
9
import json
import os
10
import re
11
import shutil
Aymeric Augustin's avatar
Aymeric Augustin committed
12
import sys
13
import tarfile
thomwolf's avatar
thomwolf committed
14
import tempfile
15
from collections import OrderedDict
Aymeric Augustin's avatar
Aymeric Augustin committed
16
from contextlib import contextmanager
17
from dataclasses import fields
18
from functools import partial, wraps
thomwolf's avatar
thomwolf committed
19
from hashlib import sha256
20
from pathlib import Path
21
from typing import Any, Dict, Optional, Tuple, Union
Aymeric Augustin's avatar
Aymeric Augustin committed
22
from urllib.parse import urlparse
23
from zipfile import ZipFile, is_zipfile
thomwolf's avatar
thomwolf committed
24

25
import numpy as np
26
27
from tqdm.auto import tqdm

Aymeric Augustin's avatar
Aymeric Augustin committed
28
29
30
import requests
from filelock import FileLock

31
from . import __version__
Lysandre Debut's avatar
Lysandre Debut committed
32
from .utils import logging
thomwolf's avatar
thomwolf committed
33

Lysandre's avatar
Lysandre committed
34

Lysandre Debut's avatar
Lysandre Debut committed
35
logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
thomwolf's avatar
thomwolf committed
36

thomwolf's avatar
thomwolf committed
37
try:
38
39
40
    USE_TF = os.environ.get("USE_TF", "AUTO").upper()
    USE_TORCH = os.environ.get("USE_TORCH", "AUTO").upper()
    if USE_TORCH in ("1", "ON", "YES", "AUTO") and USE_TF not in ("1", "ON", "YES"):
41
        import torch
42

43
44
        _torch_available = True  # pylint: disable=invalid-name
        logger.info("PyTorch version {} available.".format(torch.__version__))
45
    else:
46
        logger.info("Disabling PyTorch because USE_TF is set")
47
        _torch_available = False
thomwolf's avatar
thomwolf committed
48
49
50
except ImportError:
    _torch_available = False  # pylint: disable=invalid-name

Lysandre's avatar
Lysandre committed
51
try:
52
53
54
55
    USE_TF = os.environ.get("USE_TF", "AUTO").upper()
    USE_TORCH = os.environ.get("USE_TORCH", "AUTO").upper()

    if USE_TF in ("1", "ON", "YES", "AUTO") and USE_TORCH not in ("1", "ON", "YES"):
thomwolf's avatar
thomwolf committed
56
        import tensorflow as tf
57
58

        assert hasattr(tf, "__version__") and int(tf.__version__[0]) >= 2
thomwolf's avatar
thomwolf committed
59
60
61
        _tf_available = True  # pylint: disable=invalid-name
        logger.info("TensorFlow version {} available.".format(tf.__version__))
    else:
62
        logger.info("Disabling Tensorflow because USE_TORCH is set")
thomwolf's avatar
thomwolf committed
63
        _tf_available = False
Lysandre's avatar
Lysandre committed
64
65
except (ImportError, AssertionError):
    _tf_available = False  # pylint: disable=invalid-name
thomwolf's avatar
thomwolf committed
66

67

68
try:
69
    import datasets  # noqa: F401
70

71
72
73
74
75
76
    # Check we're not importing a "datasets" directory somewhere
    _datasets_available = hasattr(datasets, "__version__") and hasattr(datasets, "load_dataset")
    if _datasets_available:
        logger.debug(f"Succesfully imported datasets version {datasets.__version__}")
    else:
        logger.debug("Imported a datasets object but this doesn't seem to be the 馃 datasets library.")
77
78

except ImportError:
79
    _datasets_available = False
80

81
82
try:
    from torch.hub import _get_torch_home
83

84
85
86
    torch_cache_home = _get_torch_home()
except ImportError:
    torch_cache_home = os.path.expanduser(
87
88
        os.getenv("TORCH_HOME", os.path.join(os.getenv("XDG_CACHE_HOME", "~/.cache"), "torch"))
    )
89
90
91


try:
92
    import torch_xla.core.xla_model as xm  # noqa: F401
93
94
95
96
97
98
99
100
101

    if _torch_available:
        _torch_tpu_available = True  # pylint: disable=
    else:
        _torch_tpu_available = False
except ImportError:
    _torch_tpu_available = False


Patrick von Platen's avatar
Patrick von Platen committed
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
try:
    import psutil  # noqa: F401

    _psutil_available = True

except ImportError:
    _psutil_available = False


try:
    import py3nvml  # noqa: F401

    _py3nvml_available = True

except ImportError:
    _py3nvml_available = False


try:
    from apex import amp  # noqa: F401

    _has_apex = True
except ImportError:
    _has_apex = False

Ola Piktus's avatar
Ola Piktus committed
127
128
129
130
131
132
133
134
135
136

try:
    import faiss  # noqa: F401

    _faiss_available = True
    logger.debug(f"Succesfully imported faiss version {faiss.__version__}")
except ImportError:
    _faiss_available = False


137
default_cache_path = os.path.join(torch_cache_home, "transformers")
138

139

140
141
142
PYTORCH_PRETRAINED_BERT_CACHE = os.getenv("PYTORCH_PRETRAINED_BERT_CACHE", default_cache_path)
PYTORCH_TRANSFORMERS_CACHE = os.getenv("PYTORCH_TRANSFORMERS_CACHE", PYTORCH_PRETRAINED_BERT_CACHE)
TRANSFORMERS_CACHE = os.getenv("TRANSFORMERS_CACHE", PYTORCH_TRANSFORMERS_CACHE)
thomwolf's avatar
thomwolf committed
143

144
WEIGHTS_NAME = "pytorch_model.bin"
145
146
TF2_WEIGHTS_NAME = "tf_model.h5"
TF_WEIGHTS_NAME = "model.ckpt"
147
CONFIG_NAME = "config.json"
148
MODEL_CARD_NAME = "modelcard.json"
Thomas Wolf's avatar
Thomas Wolf committed
149

Lysandre's avatar
Lysandre committed
150

151
152
153
MULTIPLE_CHOICE_DUMMY_INPUTS = [
    [[0, 1, 0, 1], [1, 0, 0, 1]]
] * 2  # Needs to have 0s and 1s only since XLM uses it for langs too.
154
155
156
DUMMY_INPUTS = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
DUMMY_MASK = [[1, 1, 1, 1, 1], [1, 1, 1, 0, 0], [0, 0, 0, 1, 1]]

157
S3_BUCKET_PREFIX = "https://s3.amazonaws.com/models.huggingface.co/bert"
Julien Chaumond's avatar
Julien Chaumond committed
158
CLOUDFRONT_DISTRIB_PREFIX = "https://cdn.huggingface.co"
159
160
161
162
PRESET_MIRROR_DICT = {
    "tuna": "https://mirrors.tuna.tsinghua.edu.cn/hugging-face-models",
    "bfsu": "https://mirrors.bfsu.edu.cn/hugging-face-models",
}
163

Thomas Wolf's avatar
Thomas Wolf committed
164

thomwolf's avatar
thomwolf committed
165
166
167
def is_torch_available():
    return _torch_available

168

thomwolf's avatar
thomwolf committed
169
170
171
def is_tf_available():
    return _tf_available

172

173
174
175
176
def is_torch_tpu_available():
    return _torch_tpu_available


177
178
def is_datasets_available():
    return _datasets_available
179
180


Patrick von Platen's avatar
Patrick von Platen committed
181
182
183
184
185
186
187
188
189
190
191
192
def is_psutil_available():
    return _psutil_available


def is_py3nvml_available():
    return _py3nvml_available


def is_apex_available():
    return _has_apex


Ola Piktus's avatar
Ola Piktus committed
193
194
195
196
def is_faiss_available():
    return _faiss_available


Aymeric Augustin's avatar
Aymeric Augustin committed
197
198
def add_start_docstrings(*docstr):
    def docstring_decorator(fn):
199
200
201
202
203
204
205
206
207
208
        fn.__doc__ = "".join(docstr) + (fn.__doc__ if fn.__doc__ is not None else "")
        return fn

    return docstring_decorator


def add_start_docstrings_to_callable(*docstr):
    def docstring_decorator(fn):
        class_name = ":class:`~transformers.{}`".format(fn.__qualname__.split(".")[0])
        intro = "   The {} forward method, overrides the :func:`__call__` special method.".format(class_name)
Lysandre's avatar
Lysandre committed
209
210
        note = r"""

211
212
213
214
    .. note::
        Although the recipe for forward pass needs to be defined within
        this function, one should call the :class:`Module` instance afterwards
        instead of this since the former takes care of running the
215
        pre and post processing steps while the latter silently ignores them.
216
217
        """
        fn.__doc__ = intro + note + "".join(docstr) + (fn.__doc__ if fn.__doc__ is not None else "")
Aymeric Augustin's avatar
Aymeric Augustin committed
218
        return fn
219

Aymeric Augustin's avatar
Aymeric Augustin committed
220
    return docstring_decorator
221

222

Aymeric Augustin's avatar
Aymeric Augustin committed
223
224
225
226
def add_end_docstrings(*docstr):
    def docstring_decorator(fn):
        fn.__doc__ = fn.__doc__ + "".join(docstr)
        return fn
227

Aymeric Augustin's avatar
Aymeric Augustin committed
228
    return docstring_decorator
thomwolf's avatar
thomwolf committed
229

230

Sylvain Gugger's avatar
Sylvain Gugger committed
231
PT_RETURN_INTRODUCTION = r"""
232
    Returns:
233
        :class:`~{full_output_type}` or :obj:`tuple(torch.FloatTensor)`:
234
235
        A :class:`~{full_output_type}` (if ``return_dict=True`` is passed or when ``config.return_dict=True``) or a
        tuple of :obj:`torch.FloatTensor` comprising various elements depending on the configuration
236
237
        (:class:`~transformers.{config_class}`) and inputs.

238
239
240
"""


Sylvain Gugger's avatar
Sylvain Gugger committed
241
242
243
244
245
246
247
248
249
250
TF_RETURN_INTRODUCTION = r"""
    Returns:
        :class:`~{full_output_type}` or :obj:`tuple(tf.Tensor)`:
        A :class:`~{full_output_type}` (if ``return_dict=True`` is passed or when ``config.return_dict=True``) or a
        tuple of :obj:`tf.Tensor` comprising various elements depending on the configuration
        (:class:`~transformers.{config_class}`) and inputs.

"""


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
def _get_indent(t):
    """Returns the indentation in the first line of t"""
    search = re.search(r"^(\s*)\S", t)
    return "" if search is None else search.groups()[0]


def _convert_output_args_doc(output_args_doc):
    """Convert output_args_doc to display properly."""
    # Split output_arg_doc in blocks argument/description
    indent = _get_indent(output_args_doc)
    blocks = []
    current_block = ""
    for line in output_args_doc.split("\n"):
        # If the indent is the same as the beginning, the line is the name of new arg.
        if _get_indent(line) == indent:
            if len(current_block) > 0:
                blocks.append(current_block[:-1])
            current_block = f"{line}\n"
        else:
            # Otherwise it's part of the description of the current arg.
            # We need to remove 2 spaces to the indentation.
            current_block += f"{line[2:]}\n"
    blocks.append(current_block[:-1])

    # Format each block for proper rendering
    for i in range(len(blocks)):
        blocks[i] = re.sub(r"^(\s+)(\S+)(\s+)", r"\1- **\2**\3", blocks[i])
        blocks[i] = re.sub(r":\s*\n\s*(\S)", r" -- \1", blocks[i])

    return "\n".join(blocks)


283
284
285
286
287
288
289
290
291
292
293
294
295
def _prepare_output_docstrings(output_type, config_class):
    """
    Prepares the return part of the docstring using `output_type`.
    """
    docstrings = output_type.__doc__

    # Remove the head of the docstring to keep the list of args only
    lines = docstrings.split("\n")
    i = 0
    while i < len(lines) and re.search(r"^\s*(Args|Parameters):\s*$", lines[i]) is None:
        i += 1
    if i < len(lines):
        docstrings = "\n".join(lines[(i + 1) :])
296
        docstrings = _convert_output_args_doc(docstrings)
297
298

    # Add the return introduction
299
    full_output_type = f"{output_type.__module__}.{output_type.__name__}"
Sylvain Gugger's avatar
Sylvain Gugger committed
300
301
    intro = TF_RETURN_INTRODUCTION if output_type.__name__.startswith("TF") else PT_RETURN_INTRODUCTION
    intro = intro.format(full_output_type=full_output_type, config_class=config_class)
302
303
304
    return intro + docstrings


305
306
307
308
309
310
311
PT_TOKEN_CLASSIFICATION_SAMPLE = r"""
    Example::

        >>> from transformers import {tokenizer_class}, {model_class}
        >>> import torch

        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
312
        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True)
313
314
315
316
317

        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
        >>> labels = torch.tensor([1] * inputs["input_ids"].size(1)).unsqueeze(0)  # Batch size 1

        >>> outputs = model(**inputs, labels=labels)
318
319
        >>> loss = outputs.loss
        >>> logits = outputs.logits
320
321
322
323
324
325
326
327
328
"""

PT_QUESTION_ANSWERING_SAMPLE = r"""
    Example::

        >>> from transformers import {tokenizer_class}, {model_class}
        >>> import torch

        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
329
        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True)
330

331
332
        >>> question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
        >>> inputs = tokenizer(question, text, return_tensors='pt')
333
334
335
336
        >>> start_positions = torch.tensor([1])
        >>> end_positions = torch.tensor([3])

        >>> outputs = model(**inputs, start_positions=start_positions, end_positions=end_positions)
337
        >>> loss = outputs.loss
338
339
        >>> start_scores = outputs.start_logits
        >>> end_scores = outputs.end_logits
340
341
342
343
344
345
346
347
348
"""

PT_SEQUENCE_CLASSIFICATION_SAMPLE = r"""
    Example::

        >>> from transformers import {tokenizer_class}, {model_class}
        >>> import torch

        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
349
        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True)
350
351
352
353

        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
        >>> labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
        >>> outputs = model(**inputs, labels=labels)
354
355
        >>> loss = outputs.loss
        >>> logits = outputs.logits
356
357
358
359
360
361
362
363
364
"""

PT_MASKED_LM_SAMPLE = r"""
    Example::

        >>> from transformers import {tokenizer_class}, {model_class}
        >>> import torch

        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
365
        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True)
366
367
368
369

        >>> input_ids = tokenizer("Hello, my dog is cute", return_tensors="pt")["input_ids"]

        >>> outputs = model(input_ids, labels=input_ids)
370
371
        >>> loss = outputs.loss
        >>> prediction_logits = outputs.logits
372
373
374
375
376
377
378
379
380
"""

PT_BASE_MODEL_SAMPLE = r"""
    Example::

        >>> from transformers import {tokenizer_class}, {model_class}
        >>> import torch

        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
381
        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True)
382
383
384
385

        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
        >>> outputs = model(**inputs)

386
        >>> last_hidden_states = outputs.last_hidden_state
387
388
389
390
391
392
393
394
395
"""

PT_MULTIPLE_CHOICE_SAMPLE = r"""
    Example::

        >>> from transformers import {tokenizer_class}, {model_class}
        >>> import torch

        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
396
        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True)
397
398
399
400
401
402

        >>> prompt = "In Italy, pizza served in formal settings, such as at a restaurant, is presented unsliced."
        >>> choice0 = "It is eaten with a fork and a knife."
        >>> choice1 = "It is eaten while held in the hand."
        >>> labels = torch.tensor(0).unsqueeze(0)  # choice0 is correct (according to Wikipedia ;)), batch size 1

403
        >>> encoding = tokenizer([[prompt, prompt], [choice0, choice1]], return_tensors='pt', padding=True)
404
405
406
        >>> outputs = model(**{{k: v.unsqueeze(0) for k,v in encoding.items()}}, labels=labels)  # batch size is 1

        >>> # the linear classifier still needs to be trained
407
408
        >>> loss = outputs.loss
        >>> logits = outputs.logits
409
410
411
412
413
414
415
416
417
"""

PT_CAUSAL_LM_SAMPLE = r"""
    Example::

        >>> import torch
        >>> from transformers import {tokenizer_class}, {model_class}

        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
418
        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True)
419
420
421

        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
        >>> outputs = model(**inputs, labels=inputs["input_ids"])
422
423
        >>> loss = outputs.loss
        >>> logits = outputs.logits
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
"""

TF_TOKEN_CLASSIFICATION_SAMPLE = r"""
    Example::

        >>> from transformers import {tokenizer_class}, {model_class}
        >>> import tensorflow as tf

        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
        >>> model = {model_class}.from_pretrained('{checkpoint}')

        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="tf")
        >>> input_ids = inputs["input_ids"]
        >>> inputs["labels"] = tf.reshape(tf.constant([1] * tf.size(input_ids).numpy()), (-1, tf.size(input_ids))) # Batch size 1

        >>> outputs = model(inputs)
        >>> loss, scores = outputs[:2]
"""

TF_QUESTION_ANSWERING_SAMPLE = r"""
    Example::

        >>> from transformers import {tokenizer_class}, {model_class}
        >>> import tensorflow as tf

        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
        >>> model = {model_class}.from_pretrained('{checkpoint}')

        >>> question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
        >>> input_dict = tokenizer(question, text, return_tensors='tf')
        >>> start_scores, end_scores = model(input_dict)

        >>> all_tokens = tokenizer.convert_ids_to_tokens(input_dict["input_ids"].numpy()[0])
        >>> answer = ' '.join(all_tokens[tf.math.argmax(start_scores, 1)[0] : tf.math.argmax(end_scores, 1)[0]+1])
"""

TF_SEQUENCE_CLASSIFICATION_SAMPLE = r"""
    Example::

        >>> from transformers import {tokenizer_class}, {model_class}
        >>> import tensorflow as tf

        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
        >>> model = {model_class}.from_pretrained('{checkpoint}')

        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="tf")
        >>> inputs["labels"] = tf.reshape(tf.constant(1), (-1, 1)) # Batch size 1

        >>> outputs = model(inputs)
        >>> loss, logits = outputs[:2]
"""

TF_MASKED_LM_SAMPLE = r"""
    Example::
Sylvain Gugger's avatar
Sylvain Gugger committed
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
        >>> from transformers import {tokenizer_class}, {model_class}
        >>> import tensorflow as tf

        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
        >>> model = {model_class}.from_pretrained('{checkpoint}')

        >>> input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1

        >>> outputs = model(input_ids)
        >>> prediction_scores = outputs[0]
"""

TF_BASE_MODEL_SAMPLE = r"""
    Example::

        >>> from transformers import {tokenizer_class}, {model_class}
        >>> import tensorflow as tf

        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
        >>> model = {model_class}.from_pretrained('{checkpoint}')

        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="tf")
        >>> outputs = model(inputs)

        >>> last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
"""

TF_MULTIPLE_CHOICE_SAMPLE = r"""
    Example::

        >>> from transformers import {tokenizer_class}, {model_class}
        >>> import tensorflow as tf

        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
        >>> model = {model_class}.from_pretrained('{checkpoint}')

        >>> prompt = "In Italy, pizza served in formal settings, such as at a restaurant, is presented unsliced."
        >>> choice0 = "It is eaten with a fork and a knife."
        >>> choice1 = "It is eaten while held in the hand."

519
        >>> encoding = tokenizer([[prompt, prompt], [choice0, choice1]], return_tensors='tf', padding=True)
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
        >>> inputs = {{k: tf.expand_dims(v, 0) for k, v in encoding.items()}}
        >>> outputs = model(inputs)  # batch size is 1

        >>> # the linear classifier still needs to be trained
        >>> logits = outputs[0]
"""

TF_CAUSAL_LM_SAMPLE = r"""
    Example::

        >>> from transformers import {tokenizer_class}, {model_class}
        >>> import tensorflow as tf

        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
        >>> model = {model_class}.from_pretrained('{checkpoint}')

        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="tf")
        >>> outputs = model(inputs)
        >>> logits = outputs[0]
"""


542
def add_code_sample_docstrings(*docstr, tokenizer_class=None, checkpoint=None, output_type=None, config_class=None):
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
    def docstring_decorator(fn):
        model_class = fn.__qualname__.split(".")[0]
        is_tf_class = model_class[:2] == "TF"

        if "SequenceClassification" in model_class:
            code_sample = TF_SEQUENCE_CLASSIFICATION_SAMPLE if is_tf_class else PT_SEQUENCE_CLASSIFICATION_SAMPLE
        elif "QuestionAnswering" in model_class:
            code_sample = TF_QUESTION_ANSWERING_SAMPLE if is_tf_class else PT_QUESTION_ANSWERING_SAMPLE
        elif "TokenClassification" in model_class:
            code_sample = TF_TOKEN_CLASSIFICATION_SAMPLE if is_tf_class else PT_TOKEN_CLASSIFICATION_SAMPLE
        elif "MultipleChoice" in model_class:
            code_sample = TF_MULTIPLE_CHOICE_SAMPLE if is_tf_class else PT_MULTIPLE_CHOICE_SAMPLE
        elif "MaskedLM" in model_class:
            code_sample = TF_MASKED_LM_SAMPLE if is_tf_class else PT_MASKED_LM_SAMPLE
        elif "LMHead" in model_class:
            code_sample = TF_CAUSAL_LM_SAMPLE if is_tf_class else PT_CAUSAL_LM_SAMPLE
559
        elif "Model" in model_class or "Encoder" in model_class:
560
561
562
563
            code_sample = TF_BASE_MODEL_SAMPLE if is_tf_class else PT_BASE_MODEL_SAMPLE
        else:
            raise ValueError(f"Docstring can't be built for model {model_class}")

564
        output_doc = _prepare_output_docstrings(output_type, config_class) if output_type is not None else ""
565
        built_doc = code_sample.format(model_class=model_class, tokenizer_class=tokenizer_class, checkpoint=checkpoint)
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
        fn.__doc__ = (fn.__doc__ or "") + "".join(docstr) + output_doc + built_doc
        return fn

    return docstring_decorator


def replace_return_docstrings(output_type=None, config_class=None):
    def docstring_decorator(fn):
        docstrings = fn.__doc__
        lines = docstrings.split("\n")
        i = 0
        while i < len(lines) and re.search(r"^\s*Returns?:\s*$", lines[i]) is None:
            i += 1
        if i < len(lines):
            lines[i] = _prepare_output_docstrings(output_type, config_class)
            docstrings = "\n".join(lines)
        else:
            raise ValueError(
                f"The function {fn} should have an empty 'Return:' or 'Returns:' in its docstring as placeholder, current docstring is:\n{docstrings}"
            )
        fn.__doc__ = docstrings
587
588
589
590
591
        return fn

    return docstring_decorator


592
593
def is_remote_url(url_or_filename):
    parsed = urlparse(url_or_filename)
Julien Chaumond's avatar
Julien Chaumond committed
594
    return parsed.scheme in ("http", "https")
595

596

597
def hf_bucket_url(model_id: str, filename: str, use_cdn=True, mirror=None) -> str:
Julien Chaumond's avatar
Julien Chaumond committed
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
    """
    Resolve a model identifier, and a file name, to a HF-hosted url
    on either S3 or Cloudfront (a Content Delivery Network, or CDN).

    Cloudfront is replicated over the globe so downloads are way faster
    for the end user (and it also lowers our bandwidth costs). However, it
    is more aggressively cached by default, so may not always reflect the
    latest changes to the underlying file (default TTL is 24 hours).

    In terms of client-side caching from this library, even though
    Cloudfront relays the ETags from S3, using one or the other
    (or switching from one to the other) will affect caching: cached files
    are not shared between the two because the cached file's name contains
    a hash of the url.
    """
613
614
615
616
617
618
619
    endpoint = (
        PRESET_MIRROR_DICT.get(mirror, mirror)
        if mirror
        else CLOUDFRONT_DISTRIB_PREFIX
        if use_cdn
        else S3_BUCKET_PREFIX
    )
Julien Chaumond's avatar
Julien Chaumond committed
620
621
622
    legacy_format = "/" not in model_id
    if legacy_format:
        return f"{endpoint}/{model_id}-{filename}"
623
    else:
Julien Chaumond's avatar
Julien Chaumond committed
624
        return f"{endpoint}/{model_id}/{filename}"
625
626


thomwolf's avatar
thomwolf committed
627
def url_to_filename(url, etag=None):
thomwolf's avatar
thomwolf committed
628
629
630
631
    """
    Convert `url` into a hashed filename in a repeatable way.
    If `etag` is specified, append its hash to the url's, delimited
    by a period.
632
    If the url ends with .h5 (Keras HDF5 weights) adds '.h5' to the name
thomwolf's avatar
thomwolf committed
633
634
    so that TF 2.0 can identify it as a HDF5 file
    (see https://github.com/tensorflow/tensorflow/blob/00fad90125b18b80fe054de1055770cfb8fe4ba3/tensorflow/python/keras/engine/network.py#L1380)
thomwolf's avatar
thomwolf committed
635
    """
636
    url_bytes = url.encode("utf-8")
thomwolf's avatar
thomwolf committed
637
638
639
640
    url_hash = sha256(url_bytes)
    filename = url_hash.hexdigest()

    if etag:
641
        etag_bytes = etag.encode("utf-8")
thomwolf's avatar
thomwolf committed
642
        etag_hash = sha256(etag_bytes)
643
        filename += "." + etag_hash.hexdigest()
thomwolf's avatar
thomwolf committed
644

645
646
    if url.endswith(".h5"):
        filename += ".h5"
thomwolf's avatar
thomwolf committed
647

thomwolf's avatar
thomwolf committed
648
649
650
    return filename


thomwolf's avatar
thomwolf committed
651
def filename_to_url(filename, cache_dir=None):
thomwolf's avatar
thomwolf committed
652
653
    """
    Return the url and etag (which may be ``None``) stored for `filename`.
thomwolf's avatar
thomwolf committed
654
    Raise ``EnvironmentError`` if `filename` or its stored metadata do not exist.
thomwolf's avatar
thomwolf committed
655
656
    """
    if cache_dir is None:
657
        cache_dir = TRANSFORMERS_CACHE
658
    if isinstance(cache_dir, Path):
659
        cache_dir = str(cache_dir)
thomwolf's avatar
thomwolf committed
660
661
662

    cache_path = os.path.join(cache_dir, filename)
    if not os.path.exists(cache_path):
thomwolf's avatar
thomwolf committed
663
        raise EnvironmentError("file {} not found".format(cache_path))
thomwolf's avatar
thomwolf committed
664

665
    meta_path = cache_path + ".json"
thomwolf's avatar
thomwolf committed
666
    if not os.path.exists(meta_path):
thomwolf's avatar
thomwolf committed
667
        raise EnvironmentError("file {} not found".format(meta_path))
thomwolf's avatar
thomwolf committed
668

thomwolf's avatar
thomwolf committed
669
    with open(meta_path, encoding="utf-8") as meta_file:
thomwolf's avatar
thomwolf committed
670
        metadata = json.load(meta_file)
671
672
    url = metadata["url"]
    etag = metadata["etag"]
thomwolf's avatar
thomwolf committed
673
674
675
676

    return url, etag


677
def cached_path(
678
679
680
681
682
    url_or_filename,
    cache_dir=None,
    force_download=False,
    proxies=None,
    resume_download=False,
Julien Chaumond's avatar
Julien Chaumond committed
683
    user_agent: Union[Dict, str, None] = None,
684
685
    extract_compressed_file=False,
    force_extract=False,
686
    local_files_only=False,
687
) -> Optional[str]:
thomwolf's avatar
thomwolf committed
688
689
690
691
692
    """
    Given something that might be a URL (or might be a local path),
    determine which. If it's a URL, download the file and cache it, and
    return the path to the cached file. If it's already a local path,
    make sure the file exists and then return the path.
693
694
695
    Args:
        cache_dir: specify a cache directory to save the file to (overwrite the default cache dir).
        force_download: if True, re-dowload the file even if it's already cached in the cache dir.
696
        resume_download: if True, resume the download if incompletly recieved file is found.
697
        user_agent: Optional string or dict that will be appended to the user-agent on remote requests.
698
699
700
701
        extract_compressed_file: if True and the path point to a zip or tar file, extract the compressed
            file in a folder along the archive.
        force_extract: if True when extract_compressed_file is True and the archive was already extracted,
            re-extract the archive and overide the folder where it was extracted.
702
703
704
705

    Return:
        None in case of non-recoverable file (non-existent or inaccessible url + no cache on disk).
        Local path (string) otherwise
thomwolf's avatar
thomwolf committed
706
707
    """
    if cache_dir is None:
708
        cache_dir = TRANSFORMERS_CACHE
709
    if isinstance(url_or_filename, Path):
710
        url_or_filename = str(url_or_filename)
711
    if isinstance(cache_dir, Path):
712
        cache_dir = str(cache_dir)
thomwolf's avatar
thomwolf committed
713

714
    if is_remote_url(url_or_filename):
thomwolf's avatar
thomwolf committed
715
        # URL, so get it from the cache (downloading if necessary)
716
        output_path = get_from_cache(
717
718
719
720
721
722
            url_or_filename,
            cache_dir=cache_dir,
            force_download=force_download,
            proxies=proxies,
            resume_download=resume_download,
            user_agent=user_agent,
723
            local_files_only=local_files_only,
724
        )
thomwolf's avatar
thomwolf committed
725
726
    elif os.path.exists(url_or_filename):
        # File, and it exists.
727
        output_path = url_or_filename
728
    elif urlparse(url_or_filename).scheme == "":
thomwolf's avatar
thomwolf committed
729
        # File, but it doesn't exist.
thomwolf's avatar
thomwolf committed
730
        raise EnvironmentError("file {} not found".format(url_or_filename))
thomwolf's avatar
thomwolf committed
731
732
733
734
    else:
        # Something unknown
        raise ValueError("unable to parse {} as a URL or as a local path".format(url_or_filename))

735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
    if extract_compressed_file:
        if not is_zipfile(output_path) and not tarfile.is_tarfile(output_path):
            return output_path

        # Path where we extract compressed archives
        # We avoid '.' in dir name and add "-extracted" at the end: "./model.zip" => "./model-zip-extracted/"
        output_dir, output_file = os.path.split(output_path)
        output_extract_dir_name = output_file.replace(".", "-") + "-extracted"
        output_path_extracted = os.path.join(output_dir, output_extract_dir_name)

        if os.path.isdir(output_path_extracted) and os.listdir(output_path_extracted) and not force_extract:
            return output_path_extracted

        # Prevent parallel extractions
        lock_path = output_path + ".lock"
        with FileLock(lock_path):
            shutil.rmtree(output_path_extracted, ignore_errors=True)
            os.makedirs(output_path_extracted)
            if is_zipfile(output_path):
                with ZipFile(output_path, "r") as zip_file:
                    zip_file.extractall(output_path_extracted)
                    zip_file.close()
            elif tarfile.is_tarfile(output_path):
                tar_file = tarfile.open(output_path)
                tar_file.extractall(output_path_extracted)
                tar_file.close()
thomwolf's avatar
cleanup  
thomwolf committed
761
762
            else:
                raise EnvironmentError("Archive format of {} could not be identified".format(output_path))
763
764
765
766
767

        return output_path_extracted

    return output_path

thomwolf's avatar
thomwolf committed
768

Julien Chaumond's avatar
Julien Chaumond committed
769
def http_get(url, temp_file, proxies=None, resume_size=0, user_agent: Union[Dict, str, None] = None):
770
    ua = "transformers/{}; python/{}".format(__version__, sys.version.split()[0])
771
772
773
774
    if is_torch_available():
        ua += "; torch/{}".format(torch.__version__)
    if is_tf_available():
        ua += "; tensorflow/{}".format(tf.__version__)
775
    if isinstance(user_agent, dict):
776
        ua += "; " + "; ".join("{}/{}".format(k, v) for k, v in user_agent.items())
Aymeric Augustin's avatar
Aymeric Augustin committed
777
    elif isinstance(user_agent, str):
778
779
        ua += "; " + user_agent
    headers = {"user-agent": ua}
780
    if resume_size > 0:
781
        headers["Range"] = "bytes=%d-" % (resume_size,)
782
783
784
    response = requests.get(url, stream=True, proxies=proxies, headers=headers)
    if response.status_code == 416:  # Range not satisfiable
        return
785
    content_length = response.headers.get("Content-Length")
786
    total = resume_size + int(content_length) if content_length is not None else None
787
788
789
790
791
792
    progress = tqdm(
        unit="B",
        unit_scale=True,
        total=total,
        initial=resume_size,
        desc="Downloading",
Lysandre's avatar
Lysandre committed
793
        disable=bool(logging.get_verbosity() == logging.NOTSET),
794
    )
795
    for chunk in response.iter_content(chunk_size=1024):
796
        if chunk:  # filter out keep-alive new chunks
thomwolf's avatar
thomwolf committed
797
798
799
800
801
            progress.update(len(chunk))
            temp_file.write(chunk)
    progress.close()


802
def get_from_cache(
803
804
805
806
807
808
    url,
    cache_dir=None,
    force_download=False,
    proxies=None,
    etag_timeout=10,
    resume_download=False,
Julien Chaumond's avatar
Julien Chaumond committed
809
    user_agent: Union[Dict, str, None] = None,
810
    local_files_only=False,
811
) -> Optional[str]:
thomwolf's avatar
thomwolf committed
812
    """
813
    Given a URL, look for the corresponding file in the local cache.
thomwolf's avatar
thomwolf committed
814
    If it's not there, download it. Then return the path to the cached file.
815
816
817
818

    Return:
        None in case of non-recoverable file (non-existent or inaccessible url + no cache on disk).
        Local path (string) otherwise
thomwolf's avatar
thomwolf committed
819
820
    """
    if cache_dir is None:
821
        cache_dir = TRANSFORMERS_CACHE
822
    if isinstance(cache_dir, Path):
823
        cache_dir = str(cache_dir)
thomwolf's avatar
thomwolf committed
824

825
    os.makedirs(cache_dir, exist_ok=True)
thomwolf's avatar
thomwolf committed
826

827
828
    etag = None
    if not local_files_only:
Julien Chaumond's avatar
Julien Chaumond committed
829
830
831
832
833
834
835
        try:
            response = requests.head(url, allow_redirects=True, proxies=proxies, timeout=etag_timeout)
            if response.status_code == 200:
                etag = response.headers.get("ETag")
        except (EnvironmentError, requests.exceptions.Timeout):
            # etag is already None
            pass
thomwolf's avatar
thomwolf committed
836
837
838
839
840
841

    filename = url_to_filename(url, etag)

    # get cache path to put the file
    cache_path = os.path.join(cache_dir, filename)

842
    # etag is None = we don't have a connection, or url doesn't exist, or is otherwise inaccessible.
843
    # try to get the last downloaded one
844
845
846
847
848
849
    if etag is None:
        if os.path.exists(cache_path):
            return cache_path
        else:
            matching_files = [
                file
850
                for file in fnmatch.filter(os.listdir(cache_dir), filename.split(".")[0] + ".*")
851
852
853
854
855
                if not file.endswith(".json") and not file.endswith(".lock")
            ]
            if len(matching_files) > 0:
                return os.path.join(cache_dir, matching_files[-1])
            else:
856
857
858
859
860
861
862
863
864
                # If files cannot be found and local_files_only=True,
                # the models might've been found if local_files_only=False
                # Notify the user about that
                if local_files_only:
                    raise ValueError(
                        "Cannot find the requested files in the cached path and outgoing traffic has been"
                        " disabled. To enable model look-ups and downloads online, set 'local_files_only'"
                        " to False."
                    )
865
866
867
868
869
                return None

    # From now on, etag is not None.
    if os.path.exists(cache_path) and not force_download:
        return cache_path
870

871
    # Prevent parallel downloads of the same file with a lock.
872
    lock_path = cache_path + ".lock"
873
874
    with FileLock(lock_path):

Julien Chaumond's avatar
Julien Chaumond committed
875
876
877
878
879
        # If the download just completed while the lock was activated.
        if os.path.exists(cache_path) and not force_download:
            # Even if returning early like here, the lock will be released.
            return cache_path

880
        if resume_download:
881
882
            incomplete_path = cache_path + ".incomplete"

883
884
            @contextmanager
            def _resumable_file_manager():
885
                with open(incomplete_path, "a+b") as f:
886
                    yield f
887

888
889
890
891
892
            temp_file_manager = _resumable_file_manager
            if os.path.exists(incomplete_path):
                resume_size = os.stat(incomplete_path).st_size
            else:
                resume_size = 0
893
        else:
894
            temp_file_manager = partial(tempfile.NamedTemporaryFile, dir=cache_dir, delete=False)
895
            resume_size = 0
896

897
898
899
900
901
        # Download to temporary file, then copy to cache dir once finished.
        # Otherwise you get corrupt cache entries if the download gets interrupted.
        with temp_file_manager() as temp_file:
            logger.info("%s not found in cache or force_download set to True, downloading to %s", url, temp_file.name)

Julien Chaumond's avatar
Julien Chaumond committed
902
            http_get(url, temp_file, proxies=proxies, resume_size=resume_size, user_agent=user_agent)
903
904

        logger.info("storing %s in cache at %s", url, cache_path)
905
        os.replace(temp_file.name, cache_path)
906
907
908
909
910
911

        logger.info("creating metadata file for %s", cache_path)
        meta = {"url": url, "etag": etag}
        meta_path = cache_path + ".json"
        with open(meta_path, "w") as meta_file:
            json.dump(meta, meta_file)
thomwolf's avatar
thomwolf committed
912
913

    return cache_path
Julien Chaumond's avatar
Julien Chaumond committed
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960


class cached_property(property):
    """
    Descriptor that mimics @property but caches output in member variable.

    From tensorflow_datasets

    Built-in in functools from Python 3.8.
    """

    def __get__(self, obj, objtype=None):
        # See docs.python.org/3/howto/descriptor.html#properties
        if obj is None:
            return self
        if self.fget is None:
            raise AttributeError("unreadable attribute")
        attr = "__cached_" + self.fget.__name__
        cached = getattr(obj, attr, None)
        if cached is None:
            cached = self.fget(obj)
            setattr(obj, attr, cached)
        return cached


def torch_required(func):
    # Chose a different decorator name than in tests so it's clear they are not the same.
    @wraps(func)
    def wrapper(*args, **kwargs):
        if is_torch_available():
            return func(*args, **kwargs)
        else:
            raise ImportError(f"Method `{func.__name__}` requires PyTorch.")

    return wrapper


def tf_required(func):
    # Chose a different decorator name than in tests so it's clear they are not the same.
    @wraps(func)
    def wrapper(*args, **kwargs):
        if is_tf_available():
            return func(*args, **kwargs)
        else:
            raise ImportError(f"Method `{func.__name__}` requires TF.")

    return wrapper
961
962


963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
def is_tensor(x):
    """ Tests if ``x`` is a :obj:`torch.Tensor`, :obj:`tf.Tensor` or :obj:`np.ndarray`. """
    if is_torch_available():
        import torch

        if isinstance(x, torch.Tensor):
            return True
    if is_tf_available():
        import tensorflow as tf

        if isinstance(x, tf.Tensor):
            return True
    return isinstance(x, np.ndarray)


class ModelOutput(OrderedDict):
979
    """
980
    Base class for all model outputs as dataclass. Has a ``__getitem__`` that allows indexing by integer or slice (like
981
982
983
984
985
986
    a tuple) or strings (like a dictionnary) that will ignore the ``None`` attributes. Otherwise behaves like a
    regular python dictionary.

    .. warning::
        You can't unpack a :obj:`ModelOutput` directly. Use the :meth:`~transformers.file_utils.ModelOutput.to_tuple`
        method to convert it to a tuple before.
987
988
    """

989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
    def __post_init__(self):
        class_fields = fields(self)

        # Safety and consistency checks
        assert len(class_fields), f"{self.__class__.__name__} has no fields."
        assert all(
            field.default is None for field in class_fields[1:]
        ), f"{self.__class__.__name__} should not have more than one required field."

        first_field = getattr(self, class_fields[0].name)
        other_fields_are_none = all(getattr(self, field.name) is None for field in class_fields[1:])

        if other_fields_are_none and not is_tensor(first_field):
            try:
                iterator = iter(first_field)
                first_field_iterator = True
            except TypeError:
                first_field_iterator = False

            # if we provided an iterator as first field and the iterator is a (key, value) iterator
            # set the associated fields
            if first_field_iterator:
                for element in iterator:
                    if (
                        not isinstance(element, (list, tuple))
                        or not len(element) == 2
                        or not isinstance(element[0], str)
                    ):
                        break
                    setattr(self, element[0], element[1])
                    if element[1] is not None:
                        self[element[0]] = element[1]
1021
1022
            elif first_field is not None:
                self[class_fields[0].name] = first_field
1023
1024
1025
1026
1027
        else:
            for field in class_fields:
                v = getattr(self, field.name)
                if v is not None:
                    self[field.name] = v
1028

1029
1030
    def __delitem__(self, *args, **kwargs):
        raise Exception(f"You cannot use ``__delitem__`` on a {self.__class__.__name__} instance.")
1031

1032
1033
    def setdefault(self, *args, **kwargs):
        raise Exception(f"You cannot use ``setdefault`` on a {self.__class__.__name__} instance.")
1034

1035
1036
1037
1038
1039
    def pop(self, *args, **kwargs):
        raise Exception(f"You cannot use ``pop`` on a {self.__class__.__name__} instance.")

    def update(self, *args, **kwargs):
        raise Exception(f"You cannot use ``update`` on a {self.__class__.__name__} instance.")
1040

1041
1042
1043
1044
1045
1046
    def __getitem__(self, k):
        if isinstance(k, str):
            inner_dict = {k: v for (k, v) in self.items()}
            return inner_dict[k]
        else:
            return self.to_tuple()[k]
1047

1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
    def __setattr__(self, name, value):
        if name in self.keys() and value is not None:
            # Don't call self.__setitem__ to avoid recursion errors
            super().__setitem__(name, value)
        super().__setattr__(name, value)

    def __setitem__(self, key, value):
        # Will raise a KeyException if needed
        super().__setitem__(key, value)
        # Don't call self.__setattr__ to avoid recursion errors
        super().__setattr__(key, value)

1060
1061
1062
1063
1064
    def to_tuple(self) -> Tuple[Any]:
        """
        Convert self to a tuple containing all the attributes/keys that are not ``None``.
        """
        return tuple(self[k] for k in self.keys())