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
|
# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
#
# 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.
# ==============================================================================
"""Experimental shuffle ops."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from tensorflow.python.data.ops import dataset_ops
from tensorflow.python.data.util import random_seed
from tensorflow.python.framework import constant_op
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import ops
from tensorflow.python.ops import gen_dataset_ops
from tensorflow.python.util.tf_export import tf_export
class _ShuffleAndRepeatDataset(dataset_ops.UnaryDataset):
"""A `Dataset` that fuses `shuffle` and `repeat`."""
def __init__(self, input_dataset, buffer_size, count=None, seed=None):
super(_ShuffleAndRepeatDataset, self).__init__(input_dataset)
self._input_dataset = input_dataset
self._buffer_size = ops.convert_to_tensor(
buffer_size, dtype=dtypes.int64, name="buffer_size")
if count is None:
self._count = constant_op.constant(-1, dtype=dtypes.int64, name="count")
else:
self._count = ops.convert_to_tensor(
count, dtype=dtypes.int64, name="count")
# NOTE(mrry): We generate the seed-pair once per graph in which the dataset
# is iterated over, and cache it in `self._graph_seed_map`. This supports
# two features: iterating over the same `ShuffleDataset` twice in the same
# pipeline and observing the same order (by tying the seeds together with
# a randomly-generated seed), and using `Dataset.make_one_shot_iterator()`,
# which requires the stateful RNG op to be created inside the same graph as
# the dataset.
self._original_seed = seed
self._graph_seed_map = {}
def _as_variant_tensor(self):
try:
seed, seed2 = self._graph_seed_map[ops.get_default_graph()]
except KeyError:
seed, seed2 = random_seed.get_seed(self._original_seed)
self._graph_seed_map[ops.get_default_graph()] = (seed, seed2)
# pylint: disable=protected-access
input_resource = self._input_dataset._as_variant_tensor()
return gen_dataset_ops.shuffle_and_repeat_dataset(
input_resource,
buffer_size=self._buffer_size,
count=self._count,
seed=seed,
seed2=seed2,
**dataset_ops.flat_structure(self))
# pylint: enable=protected-access
@property
def output_classes(self):
return self._input_dataset.output_classes
@property
def output_shapes(self):
return self._input_dataset.output_shapes
@property
def output_types(self):
return self._input_dataset.output_types
@tf_export("data.experimental.shuffle_and_repeat")
def shuffle_and_repeat(buffer_size, count=None, seed=None):
"""Shuffles and repeats a Dataset returning a new permutation for each epoch.
`dataset.apply(tf.data.experimental.shuffle_and_repeat(buffer_size, count))`
is equivalent to
`dataset.shuffle(buffer_size, reshuffle_each_iteration=True).repeat(count)`
The difference is that the latter dataset is not serializable. So,
if you need to checkpoint an input pipeline with reshuffling you must use
this implementation.
Args:
buffer_size: A `tf.int64` scalar `tf.Tensor`, representing the
maximum number elements that will be buffered when prefetching.
count: (Optional.) A `tf.int64` scalar `tf.Tensor`, representing the
number of times the dataset should be repeated. The default behavior
(if `count` is `None` or `-1`) is for the dataset be repeated
indefinitely.
seed: (Optional.) A `tf.int64` scalar `tf.Tensor`, representing the
random seed that will be used to create the distribution. See
`tf.set_random_seed` for behavior.
Returns:
A `Dataset` transformation function, which can be passed to
`tf.data.Dataset.apply`.
"""
def _apply_fn(dataset): # pylint: disable=missing-docstring
return _ShuffleAndRepeatDataset(dataset, buffer_size, count, seed)
return _apply_fn
|