This repository has been archived by the owner on Dec 16, 2022. It is now read-only.
/
file_utils.py
1301 lines (1085 loc) · 46.5 KB
/
file_utils.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
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
576
577
578
579
580
581
582
583
584
585
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
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
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
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
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
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
"""
Utilities for working with the local dataset cache.
"""
import string
import weakref
from contextlib import contextmanager
import glob
import io
import os
import logging
import tempfile
import json
from abc import ABC
from collections import defaultdict
from dataclasses import dataclass, asdict
from datetime import timedelta
from fnmatch import fnmatch
from os import PathLike
from urllib.parse import urlparse
from pathlib import Path
from typing import (
Optional,
Tuple,
Union,
IO,
Callable,
Set,
List,
Iterator,
Iterable,
Dict,
NamedTuple,
MutableMapping,
)
from hashlib import sha256
from functools import wraps
from weakref import WeakValueDictionary
from zipfile import ZipFile, is_zipfile
import tarfile
import shutil
import pickle
import time
import warnings
import boto3
import botocore
import torch
from filelock import FileLock as _FileLock
from google.cloud import storage
from google.api_core.exceptions import NotFound
import numpy as np
from overrides import overrides
import requests
from requests.adapters import HTTPAdapter
from requests.packages.urllib3.util.retry import Retry
import lmdb
from torch import Tensor
import huggingface_hub as hf_hub
from allennlp.version import VERSION
from allennlp.common.tqdm import Tqdm
logger = logging.getLogger(__name__)
CACHE_ROOT = Path(os.getenv("ALLENNLP_CACHE_ROOT", Path.home() / ".allennlp"))
CACHE_DIRECTORY = str(CACHE_ROOT / "cache")
DEPRECATED_CACHE_DIRECTORY = str(CACHE_ROOT / "datasets")
# This variable was deprecated in 0.7.2 since we use a single folder for caching
# all types of files (datasets, models, etc.)
DATASET_CACHE = CACHE_DIRECTORY
# Warn if the user is still using the deprecated cache directory.
if os.path.exists(DEPRECATED_CACHE_DIRECTORY):
logger.warning(
f"Deprecated cache directory found ({DEPRECATED_CACHE_DIRECTORY}). "
f"Please remove this directory from your system to free up space."
)
class FileLock(_FileLock):
"""
This is just a subclass of the `FileLock` class from the `filelock` library, except that
it adds an additional argument to the `__init__` method: `read_only_ok`.
By default this flag is `False`, which an exception will be thrown when a lock
can't be acquired due to lack of write permissions.
But if this flag is set to `True`, a warning will be emitted instead of an error when
the lock already exists but the lock can't be acquired because write access is blocked.
"""
def __init__(
self, lock_file: Union[str, PathLike], timeout=-1, read_only_ok: bool = False
) -> None:
super().__init__(str(lock_file), timeout=timeout)
self._read_only_ok = read_only_ok
@overrides
def acquire(self, timeout=None, poll_interval=0.05):
try:
super().acquire(timeout=timeout, poll_intervall=poll_interval)
except OSError as err:
# OSError could be a lot of different things, but what we're looking
# for in particular are permission errors, such as:
# - errno 1 - EPERM - "Operation not permitted"
# - errno 13 - EACCES - "Permission denied"
# - errno 30 - EROFS - "Read-only file system"
if err.errno not in (1, 13, 30):
raise
if os.path.isfile(self._lock_file) and self._read_only_ok:
warnings.warn(
f"Lacking permissions required to obtain lock '{self._lock_file}'. "
"Race conditions are possible if other processes are writing to the same resource.",
UserWarning,
)
else:
raise
def _resource_to_filename(resource: str, etag: str = None) -> str:
"""
Convert a `resource` into a hashed filename in a repeatable way.
If `etag` is specified, append its hash to the resources's, delimited
by a period.
"""
resource_bytes = resource.encode("utf-8")
resource_hash = sha256(resource_bytes)
filename = resource_hash.hexdigest()
if etag:
etag_bytes = etag.encode("utf-8")
etag_hash = sha256(etag_bytes)
filename += "." + etag_hash.hexdigest()
return filename
def filename_to_url(filename: str, cache_dir: Union[str, Path] = None) -> Tuple[str, str]:
"""
Return the url and etag (which may be `None`) stored for `filename`.
Raise `FileNotFoundError` if `filename` or its stored metadata do not exist.
"""
if cache_dir is None:
cache_dir = CACHE_DIRECTORY
cache_path = os.path.join(cache_dir, filename)
if not os.path.exists(cache_path):
raise FileNotFoundError("file {} not found".format(cache_path))
meta_path = cache_path + ".json"
if not os.path.exists(meta_path):
raise FileNotFoundError("file {} not found".format(meta_path))
with open(meta_path) as meta_file:
metadata = json.load(meta_file)
url = metadata["url"]
etag = metadata["etag"]
return url, etag
def check_tarfile(tar_file: tarfile.TarFile):
"""Tar files can contain files outside of the extraction directory, or symlinks that point
outside the extraction directory. We also don't want any block devices fifos, or other
weird file types extracted. This checks for those issues and throws an exception if there
is a problem."""
base_path = os.path.join("tmp", "pathtest")
base_path = os.path.normpath(base_path)
def normalize_path(path: str) -> str:
path = path.rstrip("/")
path = path.replace("/", os.sep)
path = os.path.join(base_path, path)
path = os.path.normpath(path)
return path
for tarinfo in tar_file:
if not (
tarinfo.isreg()
or tarinfo.isdir()
or tarinfo.isfile()
or tarinfo.islnk()
or tarinfo.issym()
):
raise ValueError(
f"Tar file {str(tar_file.name)} contains invalid member {tarinfo.name}."
)
target_path = normalize_path(tarinfo.name)
if os.path.commonprefix([base_path, target_path]) != base_path:
raise ValueError(
f"Tar file {str(tar_file.name)} is trying to create a file outside of its extraction directory."
)
if tarinfo.islnk() or tarinfo.issym():
target_path = normalize_path(tarinfo.linkname)
if os.path.commonprefix([base_path, target_path]) != base_path:
raise ValueError(
f"Tar file {str(tar_file.name)} is trying to link to a file "
"outside of its extraction directory."
)
def cached_path(
url_or_filename: Union[str, PathLike],
cache_dir: Union[str, Path] = None,
extract_archive: bool = False,
force_extract: bool = False,
) -> str:
"""
Given something that might be a URL or local path, determine which.
If it's a remote resource, download the file and cache it, and
then return the path to the cached file. If it's already a local path,
make sure the file exists and return the path.
For URLs, "http://", "https://", "s3://", "gs://", and "hf://" are all supported.
The latter corresponds to the HuggingFace Hub.
For example, to download the PyTorch weights for the model `epwalsh/bert-xsmall-dummy`
on HuggingFace, you could do:
```python
cached_path("hf://epwalsh/bert-xsmall-dummy/pytorch_model.bin")
```
For paths or URLs that point to a tarfile or zipfile, you can also add a path
to a specific file to the `url_or_filename` preceeded by a "!", and the archive will
be automatically extracted (provided you set `extract_archive` to `True`),
returning the local path to the specific file. For example:
```python
cached_path("model.tar.gz!weights.th", extract_archive=True)
```
# Parameters
url_or_filename : `Union[str, Path]`
A URL or path to parse and possibly download.
cache_dir : `Union[str, Path]`, optional (default = `None`)
The directory to cache downloads.
extract_archive : `bool`, optional (default = `False`)
If `True`, then zip or tar.gz archives will be automatically extracted.
In which case the directory is returned.
force_extract : `bool`, optional (default = `False`)
If `True` and the file is an archive file, it will be extracted regardless
of whether or not the extracted directory already exists.
!!! Warning
Use this flag with caution! This can lead to race conditions if used
from multiple processes on the same file.
"""
if cache_dir is None:
cache_dir = CACHE_DIRECTORY
cache_dir = os.path.expanduser(cache_dir)
os.makedirs(cache_dir, exist_ok=True)
if not isinstance(url_or_filename, str):
url_or_filename = str(url_or_filename)
file_path: str
extraction_path: Optional[str] = None
# If we're using the /a/b/foo.zip!c/d/file.txt syntax, handle it here.
exclamation_index = url_or_filename.find("!")
if extract_archive and exclamation_index >= 0:
archive_path = url_or_filename[:exclamation_index]
file_name = url_or_filename[exclamation_index + 1 :]
# Call 'cached_path' recursively now to get the local path to the archive itself.
cached_archive_path = cached_path(archive_path, cache_dir, True, force_extract)
if not os.path.isdir(cached_archive_path):
raise ValueError(
f"{url_or_filename} uses the ! syntax, but does not specify an archive file."
)
# Now return the full path to the desired file within the extracted archive,
# provided it exists.
file_path = os.path.join(cached_archive_path, file_name)
if not os.path.exists(file_path):
raise FileNotFoundError(f"file {file_name} not found within {archive_path}")
return file_path
parsed = urlparse(url_or_filename)
if parsed.scheme in ("http", "https", "s3", "hf", "gs"):
# URL, so get it from the cache (downloading if necessary)
file_path = get_from_cache(url_or_filename, cache_dir)
if extract_archive and (is_zipfile(file_path) or tarfile.is_tarfile(file_path)):
# This is the path the file should be extracted to.
# For example ~/.allennlp/cache/234234.21341 -> ~/.allennlp/cache/234234.21341-extracted
extraction_path = file_path + "-extracted"
else:
url_or_filename = os.path.expanduser(url_or_filename)
if os.path.exists(url_or_filename):
# File, and it exists.
file_path = url_or_filename
# Normalize the path.
url_or_filename = os.path.abspath(url_or_filename)
if (
extract_archive
and os.path.isfile(file_path)
and (is_zipfile(file_path) or tarfile.is_tarfile(file_path))
):
# We'll use a unique directory within the cache to root to extract the archive to.
# The name of the directory is a hash of the resource file path and it's modification
# time. That way, if the file changes, we'll know when to extract it again.
extraction_name = (
_resource_to_filename(url_or_filename, str(os.path.getmtime(file_path)))
+ "-extracted"
)
extraction_path = os.path.join(cache_dir, extraction_name)
elif parsed.scheme == "":
# File, but it doesn't exist.
raise FileNotFoundError(f"file {url_or_filename} not found")
else:
# Something unknown
raise ValueError(f"unable to parse {url_or_filename} as a URL or as a local path")
if extraction_path is not None:
# If the extracted directory already exists (and is non-empty), then no
# need to create a lock file and extract again unless `force_extract=True`.
if os.path.isdir(extraction_path) and os.listdir(extraction_path) and not force_extract:
return extraction_path
# Extract it.
with FileLock(extraction_path + ".lock"):
# Check again if the directory exists now that we've acquired the lock.
if os.path.isdir(extraction_path) and os.listdir(extraction_path):
if force_extract:
logger.warning(
"Extraction directory for %s (%s) already exists, "
"overwriting it since 'force_extract' is 'True'",
url_or_filename,
extraction_path,
)
else:
return extraction_path
logger.info("Extracting %s to %s", url_or_filename, extraction_path)
shutil.rmtree(extraction_path, ignore_errors=True)
# We extract first to a temporary directory in case something goes wrong
# during the extraction process so we don't end up with a corrupted cache.
tmp_extraction_dir = tempfile.mkdtemp(dir=os.path.split(extraction_path)[0])
try:
if is_zipfile(file_path):
with ZipFile(file_path, "r") as zip_file:
zip_file.extractall(tmp_extraction_dir)
zip_file.close()
else:
tar_file = tarfile.open(file_path)
check_tarfile(tar_file)
tar_file.extractall(tmp_extraction_dir)
tar_file.close()
# Extraction was successful, rename temp directory to final
# cache directory and dump the meta data.
os.replace(tmp_extraction_dir, extraction_path)
meta = _Meta(
resource=url_or_filename,
cached_path=extraction_path,
creation_time=time.time(),
extraction_dir=True,
size=_get_resource_size(extraction_path),
)
meta.to_file()
finally:
shutil.rmtree(tmp_extraction_dir, ignore_errors=True)
return extraction_path
return file_path
def is_url_or_existing_file(url_or_filename: Union[str, Path, None]) -> bool:
"""
Given something that might be a URL (or might be a local path),
determine check if it's url or an existing file path.
"""
if url_or_filename is None:
return False
url_or_filename = os.path.expanduser(str(url_or_filename))
parsed = urlparse(url_or_filename)
return parsed.scheme in ("http", "https", "s3", "gs") or os.path.exists(url_or_filename)
def _split_s3_path(url: str) -> Tuple[str, str]:
return _split_cloud_path(url, "s3")
def _split_gcs_path(url: str) -> Tuple[str, str]:
return _split_cloud_path(url, "gs")
def _split_cloud_path(url: str, provider: str) -> Tuple[str, str]:
"""Split a full s3 path into the bucket name and path."""
parsed = urlparse(url)
if not parsed.netloc or not parsed.path:
raise ValueError("bad {} path {}".format(provider, url))
bucket_name = parsed.netloc
provider_path = parsed.path
# Remove '/' at beginning of path.
if provider_path.startswith("/"):
provider_path = provider_path[1:]
return bucket_name, provider_path
def _s3_request(func: Callable):
"""
Wrapper function for s3 requests in order to create more helpful error
messages.
"""
@wraps(func)
def wrapper(url: str, *args, **kwargs):
try:
return func(url, *args, **kwargs)
except botocore.exceptions.ClientError as exc:
if int(exc.response["Error"]["Code"]) == 404:
raise FileNotFoundError("file {} not found".format(url))
else:
raise
return wrapper
def _get_s3_resource():
session = boto3.session.Session()
if session.get_credentials() is None:
# Use unsigned requests.
s3_resource = session.resource(
"s3", config=botocore.client.Config(signature_version=botocore.UNSIGNED)
)
else:
s3_resource = session.resource("s3")
return s3_resource
@_s3_request
def _s3_etag(url: str) -> Optional[str]:
"""Check ETag on S3 object."""
s3_resource = _get_s3_resource()
bucket_name, s3_path = _split_s3_path(url)
s3_object = s3_resource.Object(bucket_name, s3_path)
return s3_object.e_tag
@_s3_request
def _s3_get(url: str, temp_file: IO) -> None:
"""Pull a file directly from S3."""
s3_resource = _get_s3_resource()
bucket_name, s3_path = _split_s3_path(url)
s3_resource.Bucket(bucket_name).download_fileobj(s3_path, temp_file)
def _gcs_request(func: Callable):
"""
Wrapper function for gcs requests in order to create more helpful error
messages.
"""
@wraps(func)
def wrapper(url: str, *args, **kwargs):
try:
return func(url, *args, **kwargs)
except NotFound:
raise FileNotFoundError("file {} not found".format(url))
return wrapper
def _get_gcs_client():
storage_client = storage.Client()
return storage_client
def _get_gcs_blob(url: str) -> storage.blob.Blob:
gcs_resource = _get_gcs_client()
bucket_name, gcs_path = _split_gcs_path(url)
bucket = gcs_resource.bucket(bucket_name)
blob = bucket.blob(gcs_path)
return blob
@_gcs_request
def _gcs_md5(url: str) -> Optional[str]:
"""Get GCS object's md5."""
blob = _get_gcs_blob(url)
return blob.md5_hash
@_gcs_request
def _gcs_get(url: str, temp_filename: str) -> None:
"""Pull a file directly from GCS."""
blob = _get_gcs_blob(url)
blob.download_to_filename(temp_filename)
def _session_with_backoff() -> requests.Session:
"""
We ran into an issue where http requests to s3 were timing out,
possibly because we were making too many requests too quickly.
This helper function returns a requests session that has retry-with-backoff
built in. See
<https://stackoverflow.com/questions/23267409/how-to-implement-retry-mechanism-into-python-requests-library>.
"""
session = requests.Session()
retries = Retry(total=5, backoff_factor=1, status_forcelist=[502, 503, 504])
session.mount("http://", HTTPAdapter(max_retries=retries))
session.mount("https://", HTTPAdapter(max_retries=retries))
return session
def _http_etag(url: str) -> Optional[str]:
with _session_with_backoff() as session:
response = session.head(url, allow_redirects=True)
if response.status_code != 200:
raise OSError(
"HEAD request failed for url {} with status code {}".format(url, response.status_code)
)
return response.headers.get("ETag")
def _http_get(url: str, temp_file: IO) -> None:
with _session_with_backoff() as session:
req = session.get(url, stream=True)
req.raise_for_status()
content_length = req.headers.get("Content-Length")
total = int(content_length) if content_length is not None else None
progress = Tqdm.tqdm(unit="B", total=total, desc="downloading")
for chunk in req.iter_content(chunk_size=1024):
if chunk: # filter out keep-alive new chunks
progress.update(len(chunk))
temp_file.write(chunk)
progress.close()
def _find_latest_cached(url: str, cache_dir: Union[str, Path]) -> Optional[str]:
filename = _resource_to_filename(url)
cache_path = os.path.join(cache_dir, filename)
candidates: List[Tuple[str, float]] = []
for path in glob.glob(cache_path + "*"):
if path.endswith(".json") or path.endswith("-extracted") or path.endswith(".lock"):
continue
mtime = os.path.getmtime(path)
candidates.append((path, mtime))
# Sort candidates by modification time, newest first.
candidates.sort(key=lambda x: x[1], reverse=True)
if candidates:
return candidates[0][0]
return None
def _serialize(data):
buffer = pickle.dumps(data, protocol=-1)
return np.frombuffer(buffer, dtype=np.uint8)
_active_tensor_caches: MutableMapping[int, "TensorCache"] = weakref.WeakValueDictionary()
def _unique_file_id(path: Union[str, PathLike]) -> int:
result = os.stat(path).st_ino
assert result != 0
return result
class TensorCache(MutableMapping[str, Tensor], ABC):
"""
This is a key-value store, mapping strings to tensors. The data is kept on disk,
making this class useful as a cache for storing tensors.
`TensorCache` is also safe to access from multiple processes at the same time, so
you can use it in distributed training situations, or from multiple training
runs at the same time.
"""
def __new__(cls, filename: Union[str, PathLike], *, read_only: bool = False, **kwargs):
# This mechanism makes sure we re-use open lmdb file handles. Lmdb has a problem when the same file is
# opened by the same process multiple times. This is our workaround.
filename = str(filename)
try:
result = _active_tensor_caches.get(_unique_file_id(filename))
except FileNotFoundError:
result = None
if result is None:
result = super(TensorCache, cls).__new__(
cls, filename, read_only=read_only, **kwargs
) # type: ignore
return result
def __init__(
self,
filename: Union[str, PathLike],
*,
map_size: int = 1024 * 1024 * 1024 * 1024,
read_only: bool = False,
) -> None:
"""
Creates a `TensorCache` by either opening an existing one on disk, or creating
a new one. Its interface is almost exactly like a Python dictionary, where the
keys are strings and the values are `torch.Tensor`.
Parameters
----------
filename: `str`
Path to the location of the cache
map_size: `int`, optional, defaults to 1TB
This is the maximum size the cache will ever grow to. On reasonable operating
systems, there is no penalty to making this a large value.
`TensorCache` uses a memory-mapped file to store the data. When the file is
first opened, we have to give the maximum size it can ever grow to. This is
that number. Reasonable operating systems don't actually allocate that space
until it is really needed.
"""
self.lmdb_env: lmdb.Environment
if hasattr(self, "lmdb_env"):
# We're being initialized again after a cache hit in _active_tensor_caches, thanks
# to __new__. In this case, we may have to upgrade to read/write, but other than
# that we are good to go.
if read_only:
return
if not self.read_only:
return
# Upgrade a read-only lmdb env to a read/write lmdb env.
filename = self.lmdb_env.path()
old_info = self.lmdb_env.info()
self.lmdb_env.close()
self.lmdb_env = lmdb.open(
filename,
map_size=old_info["map_size"],
subdir=False,
metasync=False,
sync=True,
readahead=False,
meminit=False,
readonly=False,
lock=True,
)
else:
filename = str(filename)
cpu_count = os.cpu_count() or 1
if os.path.exists(filename):
if os.path.isfile(filename):
# If the file is not writable, set read_only to True, but issue a warning.
if not os.access(filename, os.W_OK):
if not read_only:
warnings.warn(
f"File '{filename}' is read-only, so cache will be read-only",
UserWarning,
)
read_only = True
else:
# If it's not a file, raise an error.
raise ValueError("Expect a file, found a directory instead")
use_lock = True
if read_only:
# Check if the lock file is writable. If it's not, then we won't be able to use the lock.
# This is always how lmdb names the lock file.
lock_filename = filename + "-lock"
if os.path.isfile(lock_filename):
use_lock = os.access(lock_filename, os.W_OK)
else:
# If the lock file doesn't exist yet, then the directory needs to be writable in
# order to create and use the lock file.
use_lock = os.access(os.path.dirname(lock_filename), os.W_OK)
if not use_lock:
warnings.warn(
f"Lacking permissions to use lock file on cache '{filename}'.\nUse at your own risk!",
UserWarning,
)
self.lmdb_env = lmdb.open(
filename,
subdir=False,
map_size=map_size,
max_readers=cpu_count * 4,
max_spare_txns=cpu_count * 4,
metasync=False,
sync=True,
readahead=False,
meminit=False,
readonly=read_only,
lock=use_lock,
)
_active_tensor_caches[_unique_file_id(filename)] = self
# We have another cache here that makes sure we return the same object for the same key. Without it,
# you would get a different tensor, using different memory, every time you call __getitem__(), even
# if you call it with the same key.
# The downside is that we can't keep self.cache_cache up to date when multiple processes modify the
# cache at the same time. We can guarantee though that it is up to date as long as processes either
# write new values, or read existing ones.
self.cache_cache: MutableMapping[str, Tensor] = WeakValueDictionary()
@property
def read_only(self) -> bool:
return self.lmdb_env.flags()["readonly"]
def __contains__(self, key: object):
if not isinstance(key, str):
return False
if key in self.cache_cache:
return True
encoded_key = key.encode()
with self.lmdb_env.begin(write=False) as txn:
result = txn.get(encoded_key)
return result is not None
def __getitem__(self, key: str):
try:
return self.cache_cache[key]
except KeyError:
encoded_key = key.encode()
with self.lmdb_env.begin(write=False) as txn:
buffer = txn.get(encoded_key)
if buffer is None:
raise KeyError()
tensor = torch.load(io.BytesIO(buffer), map_location="cpu")
self.cache_cache[key] = tensor
return tensor
def __setitem__(self, key: str, tensor: torch.Tensor):
if self.read_only:
raise ValueError("cannot write to a read-only cache")
tensor = tensor.cpu()
encoded_key = key.encode()
buffer = io.BytesIO()
if tensor.storage().size() != np.prod(tensor.size()):
tensor = tensor.clone()
assert tensor.storage().size() == np.prod(tensor.size())
torch.save(tensor.detach(), buffer, pickle_protocol=pickle.HIGHEST_PROTOCOL)
with self.lmdb_env.begin(write=True) as txn:
txn.put(encoded_key, buffer.getbuffer())
self.cache_cache[key] = tensor
def __delitem__(self, key: str):
if self.read_only:
raise ValueError("cannot write to a read-only cache")
encoded_key = key.encode()
with self.lmdb_env.begin(write=True) as txn:
txn.delete(encoded_key)
try:
del self.cache_cache[key]
except KeyError:
pass
def __del__(self):
if self.lmdb_env is not None:
self.lmdb_env.close()
self.lmdb_env = None
def __len__(self):
return self.lmdb_env.stat()["entries"]
def __iter__(self):
# It is not hard to implement this, but we have not needed it so far.
raise NotImplementedError()
class CacheFile:
"""
This is a context manager that makes robust caching easier.
On `__enter__`, an IO handle to a temporarily file is returned, which can
be treated as if it's the actual cache file.
On `__exit__`, the temporarily file is renamed to the cache file. If anything
goes wrong while writing to the temporary file, it will be removed.
"""
def __init__(
self, cache_filename: Union[PathLike, str], mode: str = "w+b", suffix: str = ".tmp"
) -> None:
self.cache_filename = (
cache_filename if isinstance(cache_filename, Path) else Path(cache_filename)
)
self.cache_directory = os.path.dirname(self.cache_filename)
self.mode = mode
self.temp_file = tempfile.NamedTemporaryFile(
self.mode, dir=self.cache_directory, delete=False, suffix=suffix
)
def __enter__(self):
return self.temp_file
def __exit__(self, exc_type, exc_value, traceback):
self.temp_file.close()
if exc_value is None:
# Success.
logger.debug(
"Renaming temp file %s to cache at %s", self.temp_file.name, self.cache_filename
)
# Rename the temp file to the actual cache filename.
os.replace(self.temp_file.name, self.cache_filename)
return True
# Something went wrong, remove the temp file.
logger.debug("removing temp file %s", self.temp_file.name)
os.remove(self.temp_file.name)
return False
class LocalCacheResource:
"""
This is a context manager that can be used to fetch and cache arbitrary resources locally
using the same mechanisms that `cached_path` uses for remote resources.
It can be used, for example, when you want to cache the result of an expensive computation.
# Examples
```python
with LocalCacheResource("long-computation", "v1") as cache:
if cache.cached():
with cache.reader() as f:
# read from cache
else:
with cache.writer() as f:
# do the computation
# ...
# write to cache
```
"""
def __init__(self, resource_name: str, version: str, cache_dir: str = CACHE_DIRECTORY) -> None:
self.resource_name = resource_name
self.version = version
self.cache_dir = cache_dir
self.path = os.path.join(self.cache_dir, _resource_to_filename(resource_name, version))
self.file_lock = FileLock(self.path + ".lock")
def cached(self) -> bool:
return os.path.exists(self.path)
@contextmanager
def writer(self, mode="w"):
if self.cached():
raise ValueError(
f"local cache of {self.resource_name} (version '{self.version}') already exists!"
)
with CacheFile(self.path, mode=mode) as f:
yield f
meta = _Meta(
resource=self.resource_name,
cached_path=self.path,
creation_time=time.time(),
etag=self.version,
size=_get_resource_size(self.path),
)
meta.to_file()
@contextmanager
def reader(self, mode="r"):
if not self.cached():
raise ValueError(
f"local cache of {self.resource_name} (version '{self.version}') does not exist yet!"
)
with open(self.path, mode) as f:
yield f
def __enter__(self):
self.file_lock.acquire()
return self
def __exit__(self, exc_type, exc_value, traceback):
self.file_lock.release()
if exc_value is None:
return True
return False
@dataclass
class _Meta:
"""
Any resource that is downloaded to - or extracted in - the cache directory will
have a meta JSON file written next to it, which corresponds to an instance
of this class.
In older versions of AllenNLP, this meta document just had two fields: 'url' and
'etag'. The 'url' field is now the more general 'resource' field, but these old
meta files are still compatible when a `_Meta` is instantiated with the `.from_path()`
class method.
"""
resource: str
"""
URL or normalized path to the resource.
"""
cached_path: str
"""
Path to the corresponding cached version of the resource.
"""
creation_time: float
"""
The unix timestamp of when the corresponding resource was cached or extracted.
"""
size: int = 0
"""
The size of the corresponding resource, in bytes.
"""
etag: Optional[str] = None
"""
Optional ETag associated with the current cached version of the resource.
"""
extraction_dir: bool = False
"""
Does this meta corresponded to an extraction directory?
"""
def to_file(self) -> None:
with open(self.cached_path + ".json", "w") as meta_file:
json.dump(asdict(self), meta_file)
@classmethod
def from_path(cls, path: Union[str, Path]) -> "_Meta":
path = str(path)
with open(path) as meta_file:
data = json.load(meta_file)
# For backwards compat:
if "resource" not in data:
data["resource"] = data.pop("url")
if "creation_time" not in data:
data["creation_time"] = os.path.getmtime(path[:-5])
if "extraction_dir" not in data and path.endswith("-extracted.json"):
data["extraction_dir"] = True
if "cached_path" not in data:
data["cached_path"] = path[:-5]
if "size" not in data:
data["size"] = _get_resource_size(data["cached_path"])
return cls(**data)
def _hf_hub_download(
url, model_identifier: str, filename: Optional[str], cache_dir: Union[str, Path]
) -> str:
revision: Optional[str]
if "@" in model_identifier:
repo_id = model_identifier.split("@")[0]
revision = model_identifier.split("@")[1]
else:
repo_id = model_identifier
revision = None
if filename is not None:
hub_url = hf_hub.hf_hub_url(repo_id=repo_id, filename=filename, revision=revision)
cache_path = str(
hf_hub.cached_download(
url=hub_url,
library_name="allennlp",
library_version=VERSION,
cache_dir=cache_dir,
)
)
# HF writes it's own meta '.json' file which uses the same format we used to use and still
# support, but is missing some fields that we like to have.
# So we overwrite it when it we can.
with FileLock(cache_path + ".lock", read_only_ok=True):
meta = _Meta.from_path(cache_path + ".json")
# The file HF writes will have 'resource' set to the 'http' URL corresponding to the 'hf://' URL,
# but we want 'resource' to be the original 'hf://' URL.
if meta.resource != url:
meta.resource = url
meta.to_file()
else:
cache_path = str(hf_hub.snapshot_download(repo_id, revision=revision, cache_dir=cache_dir))
# Need to write the meta file for snapshot downloads if it doesn't exist.
with FileLock(cache_path + ".lock", read_only_ok=True):
if not os.path.exists(cache_path + ".json"):
meta = _Meta(
resource=url,