/
feature.py
568 lines (452 loc) · 16.6 KB
/
feature.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
# coding=utf-8
# Copyright 2020 The TensorFlow Datasets Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Feature connector.
FeatureConnector is a way of abstracting what data is returned by the
tensorflow/datasets builders from how they are encoded/decoded from file.
# Use FeatureConnector in `GeneratorBasedBuilder`
1) In the _build_info() function, define the features as you would like them
to be returned by the tf.data.Dataset() object.
Ex:
```
features=features.FeaturesDict({
'input': features.Image(),
'target': features.Text(encoder=SubWordEncoder()),
'extra_data': {
'label_id': tf.int64,
'language': tf.string,
}
})
```
The tf.data.Dataset will return each examples as a dict:
```
{
'input': tf.Tensor(shape=(batch, height, width, channel), tf.uint8),
'target': tf.Tensor(shape=(batch, sequence_length), tf.int64),
'extra_data': {
'label_id': tf.Tensor(shape=(batch,), tf.int64),
'language': tf.Tensor(shape=(batch,), tf.string),
}
}
```
2) In the generator function, yield the examples to match what you have defined
in the spec. The values will automatically be encoded.
```
yield {
'input': np_image,
'target': 'This is some text',
'extra_data': {
'label_id': 43,
'language': 'en',
}
}
```
# Create your own FeatureConnector
To create your own feature connector, you need to inherit from FeatureConnector
and implement the abstract methods.
1. If your connector only contains one value, then the get_serialized_info,
get_tensor_info, encode_example, and decode_example can directly process
single value, without wrapping it in a dict.
2. If your connector is a container of multiple sub-connectors, the easiest
way is to inherit from features.FeaturesDict and use the super() methods to
automatically encode/decode the sub-connectors.
This file contains the following FeatureConnector:
* FeatureConnector: The abstract base class defining the interface
* FeaturesDict: Container of FeatureConnector
* Tensor: Simple tensor value with static or dynamic shape
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import abc
import collections
import numpy as np
import six
import tensorflow.compat.v2 as tf
from tensorflow_datasets.core import api_utils
from tensorflow_datasets.core import utils
class TensorInfo(object):
"""Structure containing info on the `tf.Tensor` shape/dtype."""
def __init__(self, shape, dtype, default_value=None, sequence_rank=None):
"""Constructor.
Args:
shape: `tuple[int]`, shape of the tensor
dtype: Tensor dtype
default_value: Used for retrocompatibility with previous files if a new
field is added to provide a default value when reading the file.
sequence_rank: `int`, Number of `tfds.features.Sequence` dimension.
"""
self.shape = shape
self.dtype = dtype
self.default_value = default_value
self.sequence_rank = sequence_rank or 0
@classmethod
def copy_from(cls, tensor_info):
"""Copy constructor."""
return cls(
shape=tensor_info.shape,
dtype=tensor_info.dtype,
default_value=tensor_info.default_value,
sequence_rank=tensor_info.sequence_rank,
)
def __eq__(self, other):
"""Equality."""
return (
self.shape == other.shape and
self.dtype == other.dtype and
self.default_value == other.default_value
)
def __repr__(self):
return '{}(shape={}, dtype={})'.format(
type(self).__name__,
self.shape,
repr(self.dtype),
)
@six.add_metaclass(abc.ABCMeta)
class FeatureConnector(object):
"""Abstract base class for feature types.
This class provides an interface between the way the information is stored
on disk, and the way it is presented to the user.
Here is a diagram on how FeatureConnector methods fit into the data
generation/reading:
```
generator => encode_example() => tf_example => decode_example() => data dict
```
The connector can either get raw or dictionary values as input, depending on
the connector type.
"""
@abc.abstractmethod
def get_tensor_info(self):
"""Return the tf.Tensor dtype/shape of the feature.
This returns the tensor dtype/shape, as returned by .as_dataset by the
`tf.data.Dataset` object.
Ex:
```
return {
'image': tfds.features.TensorInfo(shape=(None,), dtype=tf.uint8),
'height': tfds.features.TensorInfo(shape=(), dtype=tf.int32),
'width': tfds.features.TensorInfo(shape=(), dtype=tf.int32),
}
```
FeatureConnector which are not containers should return the feature proto
directly:
```
return tfds.features.TensorInfo(shape=(256, 256), dtype=tf.uint8)
```
Returns:
tensor_info: Either a dict of `tfds.features.TensorInfo` object, or a
`tfds.features.TensorInfo`
"""
raise NotImplementedError
@property
def shape(self):
"""Return the shape (or dict of shape) of this FeatureConnector."""
return utils.map_nested(lambda t: t.shape, self.get_tensor_info())
@property
def dtype(self):
"""Return the dtype (or dict of dtype) of this FeatureConnector."""
return utils.map_nested(lambda t: t.dtype, self.get_tensor_info())
def get_serialized_info(self):
"""Return the shape/dtype of features after encoding (for the adapter).
The `FileAdapter` then use those information to write data on disk.
This function indicates how this feature is encoded on file internally.
The DatasetBuilder are written on disk as tf.train.Example proto.
Ex:
```
return {
'image': tfds.features.TensorInfo(shape=(None,), dtype=tf.uint8),
'height': tfds.features.TensorInfo(shape=(), dtype=tf.int32),
'width': tfds.features.TensorInfo(shape=(), dtype=tf.int32),
}
```
FeatureConnector which are not containers should return the feature proto
directly:
```
return tfds.features.TensorInfo(shape=(64, 64), tf.uint8)
```
If not defined, the retuned values are automatically deduced from the
`get_tensor_info` function.
Returns:
features: Either a dict of feature proto object, or a feature proto object
"""
return self.get_tensor_info()
@abc.abstractmethod
def encode_example(self, example_data):
"""Encode the feature dict into tf-example compatible input.
The input example_data can be anything that the user passed at data
generation. For example:
For features:
```
features={
'image': tfds.features.Image(),
'custom_feature': tfds.features.CustomFeature(),
}
```
At data generation (in `_generate_examples`), if the user yields:
```
yield {
'image': 'path/to/img.png',
'custom_feature': [123, 'str', lambda x: x+1]
}
```
Then:
* `tfds.features.Image.encode_example` will get `'path/to/img.png'` as
input
* `tfds.features.CustomFeature.encode_example` will get `[123, 'str',
lambda x: x+1] as input
Args:
example_data: Value or dictionary of values to convert into tf-example
compatible data.
Returns:
tfexample_data: Data or dictionary of data to write as tf-example. Data
can be a list or numpy array.
Note that numpy arrays are flattened so it's the feature connector
responsibility to reshape them in `decode_example()`.
Note that tf.train.Example only supports int64, float32 and string so
the data returned here should be integer, float or string. User type
can be restored in `decode_example()`.
"""
raise NotImplementedError
def decode_example(self, tfexample_data):
"""Decode the feature dict to TF compatible input.
Note: If eager is not enabled, this function will be executed as a
tensorflow graph (in `tf.data.Dataset.map(features.decode_example)`).
Args:
tfexample_data: Data or dictionary of data, as read by the tf-example
reader. It correspond to the `tf.Tensor()` (or dict of `tf.Tensor()`)
extracted from the `tf.train.Example`, matching the info defined in
`get_serialized_info()`.
Returns:
tensor_data: Tensor or dictionary of tensor, output of the tf.data.Dataset
object
"""
return tfexample_data
def decode_batch_example(self, tfexample_data):
"""Decode multiple features batched in a single tf.Tensor.
This function is used to decode features wrapped in
`tfds.features.Sequence()`.
By default, this function apply `decode_example` on each individual
elements using `tf.map_fn`. However, for optimization, features can
overwrite this method to apply a custom batch decoding.
Args:
tfexample_data: Same `tf.Tensor` inputs as `decode_example`, but with
and additional first dimension for the sequence length.
Returns:
tensor_data: Tensor or dictionary of tensor, output of the tf.data.Dataset
object
"""
# Note: This all works fine in Eager mode (without tf.function) because
# tf.data pipelines are always executed in Graph mode.
# Apply the decoding to each of the individual distributed features.
return tf.map_fn(
self.decode_example,
tfexample_data,
dtype=self.dtype,
parallel_iterations=10,
back_prop=False,
name='sequence_decode',
)
def decode_ragged_example(self, tfexample_data):
"""Decode nested features from a tf.RaggedTensor.
This function is used to decode features wrapped in nested
`tfds.features.Sequence()`.
By default, this function apply `decode_batch_example` on the flat values
of the ragged tensor. For optimization, features can
overwrite this method to apply a custom batch decoding.
Args:
tfexample_data: `tf.RaggedTensor` inputs containing the nested encoded
examples.
Returns:
tensor_data: The decoded `tf.RaggedTensor` or dictionary of tensor,
output of the tf.data.Dataset object
"""
return tf.ragged.map_flat_values(self.decode_batch_example, tfexample_data)
def _flatten(self, x):
"""Flatten the input dict into a list of values.
For instance, the following feature:
```
feature = FeatureDict({
'a': w,
'b': x,
'c': {
'd': y,
'e': z,
},
})
```
Applied to the following `dict`:
```
feature._flatten({
'b': X,
'c': {
'd': Y,
},
})
```
Will produce the following flattened output:
```
[
None,
X,
Y,
None,
]
```
Args:
x: A nested `dict` like structure matching the structure of the
`FeatureConnector`. Note that some elements may be missing.
Returns:
`list`: The flattened list of element of `x`. Order is guaranteed to be
deterministic. Missing elements will be filled with `None`.
"""
return [x]
def _nest(self, list_x):
"""Pack the list into a nested dict.
This is the reverse function of flatten.
For instance, the following feature:
```
feature = FeatureDict({
'a': w,
'b': x,
'c': {
'd': y,
'e': z,
},
})
```
Applied to the following `dict`:
```
feature._nest([
None,
X,
Y,
None,
])
```
Will produce the following flattened output:
```
{
'a': None,
'b': X,
'c': {
'd': Y,
'e': None,
},
}
```
Args:
list_x: List of values matching the flattened `FeatureConnector`
structure. Missing values should be filled with None.
Returns:
nested_x: nested `dict` matching the flattened `FeatureConnector`
structure.
"""
assert len(list_x) == 1
return list_x[0]
def _additional_repr_info(self):
"""Override to return additional info to go into __repr__."""
return {}
def __repr__(self):
"""Display the feature dictionary."""
tensor_info = self.get_tensor_info()
if not isinstance(tensor_info, TensorInfo):
return '{}({})'.format(type(self).__name__, tensor_info)
# Ensure ordering of keys by adding them one-by-one
repr_info = collections.OrderedDict()
repr_info['shape'] = tensor_info.shape
repr_info['dtype'] = repr(tensor_info.dtype)
additional_info = self._additional_repr_info()
for k, v in additional_info.items():
repr_info[k] = v
info_str = ', '.join(['%s=%s' % (k, v) for k, v in repr_info.items()])
return '{}({})'.format(
type(self).__name__,
info_str,
)
def save_metadata(self, data_dir, feature_name):
"""Save the feature metadata on disk.
This function is called after the data has been generated (by
`_download_and_prepare`) to save the feature connector info with the
generated dataset.
Some dataset/features dynamically compute info during
`_download_and_prepare`. For instance:
* Labels are loaded from the downloaded data
* Vocabulary is created from the downloaded data
* ImageLabelFolder compute the image dtypes/shape from the manual_dir
After the info have been added to the feature, this function allow to
save those additional info to be restored the next time the data is loaded.
By default, this function do not save anything, but sub-classes can
overwrite the function.
Args:
data_dir: `str`, path to the dataset folder to which save the info (ex:
`~/datasets/cifar10/1.2.0/`)
feature_name: `str`, the name of the feature (from the FeaturesDict key)
"""
pass
def load_metadata(self, data_dir, feature_name):
"""Restore the feature metadata from disk.
If a dataset is re-loaded and generated files exists on disk, this function
will restore the feature metadata from the saved file.
Args:
data_dir: `str`, path to the dataset folder to which save the info (ex:
`~/datasets/cifar10/1.2.0/`)
feature_name: `str`, the name of the feature (from the FeaturesDict key)
"""
pass
class Tensor(FeatureConnector):
"""`FeatureConnector` for generic data of arbitrary shape and type."""
@api_utils.disallow_positional_args
def __init__(self, shape, dtype):
"""Construct a Tensor feature."""
self._shape = tuple(shape)
self._dtype = dtype
def get_tensor_info(self):
"""See base class for details."""
return TensorInfo(shape=self._shape, dtype=self._dtype)
def decode_batch_example(self, example_data):
"""See base class for details."""
# Overwrite the `tf.map_fn`, decoding is a no-op
return self.decode_example(example_data)
def decode_ragged_example(self, example_data):
"""See base class for details."""
# Overwrite the `tf.map_fn`, decoding is a no-op
return self.decode_example(example_data)
def encode_example(self, example_data):
"""See base class for details."""
np_dtype = np.dtype(self.dtype.as_numpy_dtype)
if not isinstance(example_data, np.ndarray):
example_data = np.array(example_data, dtype=np_dtype)
# Ensure the shape and dtype match
if example_data.dtype != np_dtype:
raise ValueError('Dtype {} do not match {}'.format(
example_data.dtype, np_dtype))
utils.assert_shape_match(example_data.shape, self._shape)
return example_data
def get_inner_feature_repr(feature):
"""Utils which returns the object which should get printed in __repr__.
This is used in container features (Sequence, FeatureDict) to print scalar
Tensor in a less verbose way `Sequence(tf.int32)` rather than
`Sequence(Tensor(shape=(), dtype=tf.in32))`.
Args:
feature: The feature to dispaly
Returns:
Either the feature or it's inner value.
"""
# We only print `tf.int32` rather than `Tensor(shape=(), dtype=tf.int32)`
# * For the base `Tensor` class (and not subclass).
# * When shape is scalar (explicit check to avoid trigger when `shape=None`).
if type(feature) == Tensor and feature.shape == (): # pylint: disable=unidiomatic-typecheck,g-explicit-bool-comparison
return repr(feature.dtype)
else:
return repr(feature)