• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
2#
3# Licensed under the Apache License, Version 2.0 (the "License");
4# you may not use this file except in compliance with the License.
5# You may obtain a copy of the License at
6#
7#     http://www.apache.org/licenses/LICENSE-2.0
8#
9# Unless required by applicable law or agreed to in writing, software
10# distributed under the License is distributed on an "AS IS" BASIS,
11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12# See the License for the specific language governing permissions and
13# limitations under the License.
14# ==============================================================================
15"""Experimental shuffle ops."""
16from tensorflow.python.data.ops import dataset_ops
17from tensorflow.python.data.util import random_seed
18from tensorflow.python.framework import constant_op
19from tensorflow.python.framework import dtypes
20from tensorflow.python.framework import ops
21from tensorflow.python.ops import gen_dataset_ops
22from tensorflow.python.util import deprecation
23from tensorflow.python.util.tf_export import tf_export
24
25
26class _ShuffleAndRepeatDataset(dataset_ops.UnaryUnchangedStructureDataset):
27  """A `Dataset` that fuses `shuffle` and `repeat`."""
28
29  def __init__(self, input_dataset, buffer_size, count=None, seed=None):
30    self._input_dataset = input_dataset
31    self._buffer_size = ops.convert_to_tensor(
32        buffer_size, dtype=dtypes.int64, name="buffer_size")
33    if count is None:
34      self._count = constant_op.constant(-1, dtype=dtypes.int64, name="count")
35    else:
36      self._count = ops.convert_to_tensor(
37          count, dtype=dtypes.int64, name="count")
38    self._seed, self._seed2 = random_seed.get_seed(seed)
39    variant_tensor = gen_dataset_ops.shuffle_and_repeat_dataset(
40        self._input_dataset._variant_tensor,  # pylint: disable=protected-access
41        buffer_size=self._buffer_size,
42        count=self._count,
43        seed=self._seed,
44        seed2=self._seed2,
45        **self._flat_structure)
46    super(_ShuffleAndRepeatDataset, self).__init__(input_dataset,
47                                                   variant_tensor)
48
49
50@deprecation.deprecated(
51    None,
52    "Use `tf.data.Dataset.shuffle(buffer_size, seed)` followed by "
53    "`tf.data.Dataset.repeat(count)`. Static tf.data optimizations will take "
54    "care of using the fused implementation.")
55@tf_export("data.experimental.shuffle_and_repeat")
56def shuffle_and_repeat(buffer_size, count=None, seed=None):
57  """Shuffles and repeats a Dataset, reshuffling with each repetition.
58
59  >>> d = tf.data.Dataset.from_tensor_slices([1, 2, 3])
60  >>> d = d.apply(tf.data.experimental.shuffle_and_repeat(2, count=2))
61  >>> [elem.numpy() for elem in d] # doctest: +SKIP
62  [2, 3, 1, 1, 3, 2]
63
64  ```python
65  dataset.apply(
66    tf.data.experimental.shuffle_and_repeat(buffer_size, count, seed))
67  ```
68
69  produces the same output as
70
71  ```python
72  dataset.shuffle(
73    buffer_size, seed=seed, reshuffle_each_iteration=True).repeat(count)
74  ```
75
76  In each repetition, this dataset fills a buffer with `buffer_size` elements,
77  then randomly samples elements from this buffer, replacing the selected
78  elements with new elements. For perfect shuffling, set the buffer size equal
79  to the full size of the dataset.
80
81  For instance, if your dataset contains 10,000 elements but `buffer_size` is
82  set to 1,000, then `shuffle` will initially select a random element from
83  only the first 1,000 elements in the buffer. Once an element is selected,
84  its space in the buffer is replaced by the next (i.e. 1,001-st) element,
85  maintaining the 1,000 element buffer.
86
87  Args:
88    buffer_size: A `tf.int64` scalar `tf.Tensor`, representing the maximum
89      number elements that will be buffered when prefetching.
90    count: (Optional.) A `tf.int64` scalar `tf.Tensor`, representing the number
91      of times the dataset should be repeated. The default behavior (if `count`
92      is `None` or `-1`) is for the dataset be repeated indefinitely.
93    seed: (Optional.) A `tf.int64` scalar `tf.Tensor`, representing the random
94      seed that will be used to create the distribution. See
95      `tf.random.set_seed` for behavior.
96
97  Returns:
98    A `Dataset` transformation function, which can be passed to
99    `tf.data.Dataset.apply`.
100  """
101
102  def _apply_fn(dataset):  # pylint: disable=missing-docstring
103    return _ShuffleAndRepeatDataset(dataset, buffer_size, count, seed)
104
105  return _apply_fn
106