-
Notifications
You must be signed in to change notification settings - Fork 615
/
integrations.py
625 lines (534 loc) · 21.8 KB
/
integrations.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
"""
License:
This Source Code Form is subject to the terms of the Mozilla Public License, v. 2.0.
If a copy of the MPL was not distributed with this file, You can obtain one at https://mozilla.org/MPL/2.0/.
"""
import sys
from collections import defaultdict
from hub.exceptions import ModuleNotInstalledException, OutOfBoundsError
from hub.schema.features import Primitive, Tensor, SchemaDict
from hub.schema import Audio, BBox, ClassLabel, Image, Sequence, Text, Video
from .dataset import Dataset
import hub.store.pickle_s3_storage
import hub.schema.serialize
import hub.schema.deserialize
def _to_pytorch(
dataset,
transform=None,
inplace=True,
output_type=dict,
indexes=None,
):
"""| Converts the dataset into a pytorch compatible format.
Parameters
----------
transform: function that transforms data in a dict format
inplace: bool, optional
Defines if data should be converted to torch.Tensor before or after Transforms applied (depends on what data
type you need for Transforms). Default is True.
output_type: one of list, tuple, dict, optional
Defines the output type. Default is dict - same as in original Hub Dataset.
indexes: list or int, optional
The samples to be converted into tensorflow format. Takes all samples in dataset by default.
"""
try:
import torch
except ModuleNotFoundError:
raise ModuleNotInstalledException("torch")
global torch
indexes = indexes or dataset.indexes
if "r" not in dataset.mode:
dataset.flush() # FIXME Without this some tests in test_converters.py fails, not clear why
return TorchDataset(
dataset, transform, inplace=inplace, output_type=output_type, indexes=indexes
)
def _from_pytorch(dataset, scheduler: str = "single", workers: int = 1):
"""| Converts a pytorch dataset object into hub format
Parameters
----------
dataset:
The pytorch dataset object that needs to be converted into hub format
scheduler: str
choice between "single", "threaded", "processed"
workers: int
how many threads or processes to use
"""
if "torch" not in sys.modules:
raise ModuleNotInstalledException("torch")
else:
import torch
global torch
max_dict = defaultdict(lambda: None)
def sampling(ds):
for sample in ds:
dict_sampling(sample)
def dict_sampling(d, path=""):
for k, v in d.items():
k = k.replace("/", "_")
cur_path = path + "/" + k
if isinstance(v, dict):
dict_sampling(v, path=cur_path)
elif isinstance(v, str):
if cur_path not in max_dict.keys():
max_dict[cur_path] = (len(v),)
else:
max_dict[cur_path] = max(((len(v)),), max_dict[cur_path])
elif hasattr(v, "shape"):
if cur_path not in max_dict.keys():
max_dict[cur_path] = v.shape
else:
max_dict[cur_path] = tuple(
[max(value) for value in zip(max_dict[cur_path], v.shape)]
)
sampling(dataset)
def generate_schema(dataset):
sample = dataset[0]
return dict_to_hub(sample).dict_
def dict_to_hub(dic, path=""):
d = {}
for k, v in dic.items():
k = k.replace("/", "_")
cur_path = path + "/" + k
if isinstance(v, dict):
d[k] = dict_to_hub(v, path=cur_path)
else:
value_shape = v.shape if hasattr(v, "shape") else ()
if isinstance(v, torch.Tensor):
v = v.numpy()
shape = tuple(None for it in value_shape)
max_shape = (
max_dict[cur_path] or tuple(10000 for it in value_shape)
if not isinstance(v, str)
else (10000,)
)
dtype = v.dtype.name if hasattr(v, "dtype") else type(v)
dtype = "int64" if isinstance(v, str) else dtype
d[k] = (
Tensor(shape=shape, dtype=dtype, max_shape=max_shape)
if not isinstance(v, str)
else Text(shape=(None,), dtype=dtype, max_shape=max_shape)
)
return SchemaDict(d)
my_schema = generate_schema(dataset)
def transform_numpy(sample):
d = {}
for k, v in sample.items():
k = k.replace("/", "_")
d[k] = transform_numpy(v) if isinstance(v, dict) else v
return d
@hub.transform(schema=my_schema, scheduler=scheduler, workers=workers)
def my_transform(sample):
return transform_numpy(sample)
return my_transform(dataset)
def _to_tensorflow(dataset, indexes=None, include_shapes=False):
"""| Converts the dataset into a tensorflow compatible format
Parameters
----------
indexes: list or int, optional
The samples to be converted into tensorflow format. Takes all samples in dataset by default.
include_shapes: boolean, optional
False by default. Setting it to True passes the shapes to tf.data.Dataset.from_generator.
Setting to True could lead to issues with dictionaries inside Tensors.
"""
try:
import tensorflow as tf
global tf
except ModuleNotFoundError:
raise ModuleNotInstalledException("tensorflow")
indexes = indexes or dataset.indexes
indexes = [indexes] if isinstance(indexes, int) else indexes
_samples_in_chunks = {
key: (None in value.shape) and 1 or value.chunks[0]
for key, value in dataset._tensors.items()
}
_active_chunks = {}
_active_chunks_range = {}
def _get_active_item(key, index):
active_range = _active_chunks_range.get(key)
samples_per_chunk = _samples_in_chunks[key]
if active_range is None or index not in active_range:
active_range_start = index - index % samples_per_chunk
active_range = range(
active_range_start, active_range_start + samples_per_chunk
)
_active_chunks_range[key] = active_range
_active_chunks[key] = dataset._tensors[key][
active_range.start : active_range.stop
]
return _active_chunks[key][index % samples_per_chunk]
def tf_gen():
key_dtype_map = {key: dataset[key, indexes[0]].dtype for key in dataset.keys}
for index in indexes:
d = {}
for key in dataset.keys:
split_key, cur = key.split("/"), d
for i in range(1, len(split_key) - 1):
if split_key[i] in cur.keys():
cur = cur[split_key[i]]
else:
cur[split_key[i]] = {}
cur = cur[split_key[i]]
cur[split_key[-1]] = _get_active_item(key, index)
if isinstance(key_dtype_map[key], Text):
value = cur[split_key[-1]]
cur[split_key[-1]] = (
"".join(chr(it) for it in value.tolist())
if value.ndim == 1
else ["".join(chr(it) for it in val.tolist()) for val in value]
)
yield (d)
def dict_to_tf(my_dtype):
d = {}
for k, v in my_dtype.dict_.items():
d[k] = dtype_to_tf(v)
return d
def tensor_to_tf(my_dtype):
return dtype_to_tf(my_dtype.dtype)
def text_to_tf(my_dtype):
return "string"
def dtype_to_tf(my_dtype):
if isinstance(my_dtype, SchemaDict):
return dict_to_tf(my_dtype)
elif isinstance(my_dtype, Text):
return text_to_tf(my_dtype)
elif isinstance(my_dtype, Tensor):
return tensor_to_tf(my_dtype)
elif isinstance(my_dtype, Primitive):
if str(my_dtype._dtype) == "object":
return "string"
return str(my_dtype._dtype)
def get_output_shapes(my_dtype):
if isinstance(my_dtype, SchemaDict):
return output_shapes_from_dict(my_dtype)
elif isinstance(my_dtype, (Text, Primitive)):
return ()
elif isinstance(my_dtype, Tensor):
return my_dtype.shape
def output_shapes_from_dict(my_dtype):
d = {}
for k, v in my_dtype.dict_.items():
d[k] = get_output_shapes(v)
return d
output_types = dtype_to_tf(dataset._schema)
if include_shapes:
output_shapes = get_output_shapes(dataset._schema)
return tf.data.Dataset.from_generator(
tf_gen, output_types=output_types, output_shapes=output_shapes
)
else:
return tf.data.Dataset.from_generator(tf_gen, output_types=output_types)
def _from_tensorflow(ds, scheduler: str = "single", workers: int = 1):
"""Converts a tensorflow dataset into hub format.
Parameters
----------
dataset:
The tensorflow dataset object that needs to be converted into hub format
scheduler: str
choice between "single", "threaded", "processed"
workers: int
how many threads or processes to use
Examples
--------
>>> ds = tf.data.Dataset.from_tensor_slices(tf.range(10))
>>> out_ds = hub.Dataset.from_tensorflow(ds)
>>> res_ds = out_ds.store("username/new_dataset") # res_ds is now a usable hub dataset
>>> ds = tf.data.Dataset.from_tensor_slices({'a': [1, 2], 'b': [5, 6]})
>>> out_ds = hub.Dataset.from_tensorflow(ds)
>>> res_ds = out_ds.store("username/new_dataset") # res_ds is now a usable hub dataset
>>> ds = hub.Dataset(schema=my_schema, shape=(1000,), url="username/dataset_name", mode="w")
>>> ds = ds.to_tensorflow()
>>> out_ds = hub.Dataset.from_tensorflow(ds)
>>> res_ds = out_ds.store("username/new_dataset") # res_ds is now a usable hub dataset
"""
if "tensorflow" not in sys.modules:
raise ModuleNotInstalledException("tensorflow")
else:
import tensorflow as tf
global tf
def generate_schema(ds):
if isinstance(ds._structure, tf.TensorSpec):
return tf_to_hub({"data": ds._structure}).dict_
return tf_to_hub(ds._structure).dict_
def tf_to_hub(tf_dt):
if isinstance(tf_dt, dict):
return dict_to_hub(tf_dt)
elif isinstance(tf_dt, tf.TensorSpec):
return TensorSpec_to_hub(tf_dt)
def TensorSpec_to_hub(tf_dt):
dt = tf_dt.dtype.name if tf_dt.dtype.name != "string" else "object"
shape = tuple(tf_dt.shape) if tf_dt.shape.rank is not None else (None,)
return Tensor(shape=shape, dtype=dt)
def dict_to_hub(tf_dt):
d = {key.replace("/", "_"): tf_to_hub(value) for key, value in tf_dt.items()}
return SchemaDict(d)
my_schema = generate_schema(ds)
def transform_numpy(sample):
d = {}
for k, v in sample.items():
k = k.replace("/", "_")
if not isinstance(v, dict):
if isinstance(v, (tuple, list)):
new_v = list(v)
for i in range(len(new_v)):
new_v[i] = new_v[i].numpy()
d[k] = tuple(new_v) if isinstance(v, tuple) else new_v
else:
d[k] = v.numpy()
else:
d[k] = transform_numpy(v)
return d
@hub.transform(schema=my_schema, scheduler=scheduler, workers=workers)
def my_transform(sample):
sample = sample if isinstance(sample, dict) else {"data": sample}
return transform_numpy(sample)
return my_transform(ds)
def _from_tfds(
dataset,
split=None,
num: int = -1,
sampling_amount: int = 1,
scheduler: str = "single",
workers: int = 1,
):
"""| Converts a TFDS Dataset into hub format.
Parameters
----------
dataset: str
The name of the tfds dataset that needs to be converted into hub format
split: str, optional
A string representing the splits of the dataset that are required such as "train" or "test+train"
If not present, all the splits of the dataset are used.
num: int, optional
The number of samples required. If not present, all the samples are taken.
If count is -1, or if count is greater than the size of this dataset, the new dataset will contain all elements of this dataset.
sampling_amount: float, optional
a value from 0 to 1, that specifies how much of the dataset would be sampled to determinte feature shapes
value of 0 would mean no sampling and 1 would imply that entire dataset would be sampled
scheduler: str
choice between "single", "threaded", "processed"
workers: int
how many threads or processes to use
Examples
--------
>>> out_ds = hub.Dataset.from_tfds('mnist', split='test+train', num=1000)
>>> res_ds = out_ds.store("username/mnist") # res_ds is now a usable hub dataset
"""
try:
import tensorflow_datasets as tfds
global tfds
except Exception:
raise ModuleNotInstalledException("tensorflow_datasets")
ds_info = tfds.load(dataset, with_info=True)
if split is None:
all_splits = ds_info[1].splits.keys()
split = "+".join(all_splits)
ds = tfds.load(dataset, split=split)
ds = ds.take(num)
max_dict = defaultdict(lambda: None)
def sampling(ds):
try:
subset_len = len(ds) if hasattr(ds, "__len__") else num
except Exception:
subset_len = max(num, 5)
subset_len = int(max(subset_len * sampling_amount, 5))
samples = ds.take(subset_len)
for smp in samples:
dict_sampling(smp)
def dict_sampling(d, path=""):
for k, v in d.items():
k = k.replace("/", "_")
cur_path = path + "/" + k
if isinstance(v, dict):
dict_sampling(v)
elif hasattr(v, "shape") and v.dtype != "string":
if cur_path not in max_dict.keys():
max_dict[cur_path] = v.shape
else:
max_dict[cur_path] = tuple(
[max(value) for value in zip(max_dict[cur_path], v.shape)]
)
elif hasattr(v, "shape") and v.dtype == "string":
if cur_path not in max_dict.keys():
max_dict[cur_path] = (len(v.numpy()),)
else:
max_dict[cur_path] = max(((len(v.numpy()),), max_dict[cur_path]))
if sampling_amount > 0:
sampling(ds)
def generate_schema(ds):
tf_schema = ds[1].features
return to_hub(tf_schema).dict_
def to_hub(tf_dt, max_shape=None, path=""):
if isinstance(tf_dt, tfds.features.FeaturesDict):
return sdict_to_hub(tf_dt, path=path)
elif isinstance(tf_dt, tfds.features.Image):
return image_to_hub(tf_dt, max_shape=max_shape)
elif isinstance(tf_dt, tfds.features.ClassLabel):
return class_label_to_hub(tf_dt, max_shape=max_shape)
elif isinstance(tf_dt, tfds.features.Video):
return video_to_hub(tf_dt, max_shape=max_shape)
elif isinstance(tf_dt, tfds.features.Text):
return text_to_hub(tf_dt, max_shape=max_shape)
elif isinstance(tf_dt, tfds.features.Sequence):
return sequence_to_hub(tf_dt, max_shape=max_shape)
elif isinstance(tf_dt, tfds.features.BBoxFeature):
return bbox_to_hub(tf_dt, max_shape=max_shape)
elif isinstance(tf_dt, tfds.features.Audio):
return audio_to_hub(tf_dt, max_shape=max_shape)
elif isinstance(tf_dt, tfds.features.Tensor):
return tensor_to_hub(tf_dt, max_shape=max_shape)
else:
if tf_dt.dtype.name != "string":
return tf_dt.dtype.name
def sdict_to_hub(tf_dt, path=""):
d = {}
for key, value in tf_dt.items():
key = key.replace("/", "_")
cur_path = path + "/" + key
d[key] = to_hub(value, max_dict[cur_path], cur_path)
return SchemaDict(d)
def tensor_to_hub(tf_dt, max_shape=None):
if tf_dt.dtype.name == "string":
max_shape = max_shape or (100000,)
return Text(shape=(None,), dtype="int64", max_shape=(100000,))
dt = tf_dt.dtype.name
if max_shape and len(max_shape) > len(tf_dt.shape):
max_shape = max_shape[(len(max_shape) - len(tf_dt.shape)) :]
max_shape = max_shape or tuple(
10000 if dim is None else dim for dim in tf_dt.shape
)
return Tensor(shape=tf_dt.shape, dtype=dt, max_shape=max_shape)
def image_to_hub(tf_dt, max_shape=None):
dt = tf_dt.dtype.name
if max_shape and len(max_shape) > len(tf_dt.shape):
max_shape = max_shape[(len(max_shape) - len(tf_dt.shape)) :]
max_shape = max_shape or tuple(
10000 if dim is None else dim for dim in tf_dt.shape
)
return Image(
shape=tf_dt.shape,
dtype=dt,
max_shape=max_shape, # compressor="png"
)
def class_label_to_hub(tf_dt, max_shape=None):
if hasattr(tf_dt, "_num_classes"):
return ClassLabel(
num_classes=tf_dt.num_classes,
)
else:
return ClassLabel(names=tf_dt.names)
def text_to_hub(tf_dt, max_shape=None):
max_shape = max_shape or (100000,)
dt = "int64"
return Text(shape=(None,), dtype=dt, max_shape=max_shape)
def bbox_to_hub(tf_dt, max_shape=None):
dt = tf_dt.dtype.name
return BBox(dtype=dt)
def sequence_to_hub(tf_dt, max_shape=None):
return Sequence(dtype=to_hub(tf_dt._feature), shape=())
def audio_to_hub(tf_dt, max_shape=None):
if max_shape and len(max_shape) > len(tf_dt.shape):
max_shape = max_shape[(len(max_shape) - len(tf_dt.shape)) :]
max_shape = max_shape or tuple(
100000 if dim is None else dim for dim in tf_dt.shape
)
dt = tf_dt.dtype.name
return Audio(
shape=tf_dt.shape,
dtype=dt,
max_shape=max_shape,
file_format=tf_dt._file_format,
sample_rate=tf_dt._sample_rate,
)
def video_to_hub(tf_dt, max_shape=None):
if max_shape and len(max_shape) > len(tf_dt.shape):
max_shape = max_shape[(len(max_shape) - len(tf_dt.shape)) :]
max_shape = max_shape or tuple(
10000 if dim is None else dim for dim in tf_dt.shape
)
dt = tf_dt.dtype.name
return Video(shape=tf_dt.shape, dtype=dt, max_shape=max_shape)
my_schema = generate_schema(ds_info)
def transform_numpy(sample):
d = {}
for k, v in sample.items():
k = k.replace("/", "_")
d[k] = transform_numpy(v) if isinstance(v, dict) else v.numpy()
return d
@hub.transform(schema=my_schema, scheduler=scheduler, workers=workers)
def my_transform(sample):
return transform_numpy(sample)
return my_transform(ds)
class TorchDataset:
def __init__(
self, ds, transform=None, inplace=True, output_type=dict, indexes=None
):
self._ds = None
self._url = ds.url
self._token = ds.token
self._transform = transform
self.inplace = inplace
self.output_type = output_type
self.indexes = indexes
self._inited = False
def _do_transform(self, data):
return self._transform(data) if self._transform else data
def _init_ds(self):
"""
For each process, dataset should be independently loaded
"""
if self._ds is None:
self._ds = Dataset(self._url, token=self._token, lock_cache=False)
if not self._inited:
self._inited = True
self._samples_in_chunks = {
key: (None in value.shape) and 1 or value.chunks[0]
for key, value in self._ds._tensors.items()
}
self._active_chunks = {}
self._active_chunks_range = {}
def __len__(self):
self._init_ds()
return len(self.indexes) if isinstance(self.indexes, list) else 1
def _get_active_item(self, key, index):
active_range = self._active_chunks_range.get(key)
samples_per_chunk = self._samples_in_chunks[key]
if active_range is None or index not in active_range:
active_range_start = index - index % samples_per_chunk
active_range = range(
active_range_start, active_range_start + samples_per_chunk
)
self._active_chunks_range[key] = active_range
self._active_chunks[key] = self._ds._tensors[key][
active_range.start : active_range.stop
]
return self._active_chunks[key][index % samples_per_chunk]
def __getitem__(self, ind):
if isinstance(self.indexes, int):
if ind != 0:
raise OutOfBoundsError(f"Got index {ind} for dataset of length 1")
index = self.indexes
else:
index = self.indexes[ind]
self._init_ds()
d = {}
for key in self._ds._tensors.keys():
split_key = key.split("/")
cur = d
for i in range(1, len(split_key) - 1):
if split_key[i] not in cur.keys():
cur[split_key[i]] = {}
cur = cur[split_key[i]]
item = self._get_active_item(key, index)
if not isinstance(item, bytes) and not isinstance(item, str):
t = item
if self.inplace:
t = torch.tensor(t)
cur[split_key[-1]] = t
d = self._do_transform(d)
if self.inplace & (self.output_type != dict) & (isinstance(d, dict)):
d = self.output_type(d.values())
return d
def __iter__(self):
self._init_ds()
for i in range(len(self)):
yield self[i]