• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1# Copyright 2019 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"""Code for creating a dataset out of a NumPy array."""
16
17from __future__ import absolute_import
18from __future__ import division
19from __future__ import print_function
20
21import numpy as np
22
23from tensorflow.python.data.ops import dataset_ops
24from tensorflow.python.eager import context
25from tensorflow.python.framework import dtypes
26from tensorflow.python.framework import ops
27from tensorflow.python.ops import array_ops
28from tensorflow.python.ops import variable_scope
29from tensorflow.python.util import nest
30
31
32def init_var_from_numpy(input_var, numpy_input, session):
33  """Initialize `input_var` to `numpy_input` using `session` in graph mode."""
34  with ops.init_scope():
35    if context.executing_eagerly():
36      input_var.assign(numpy_input)
37      return
38
39    assert session is not None
40    session.run(input_var.initializer)
41
42    start_placeholder = array_ops.placeholder(dtypes.int64, ())
43    end_placeholder = array_ops.placeholder(dtypes.int64, ())
44    slice_placeholder = array_ops.placeholder(input_var.dtype)
45    assign_slice_op = input_var[start_placeholder:end_placeholder].assign(
46        slice_placeholder)
47
48    # If each batch element is > 64 MB, then we copy each batch element
49    # individually. Otherwise, the slices will be < 128 MB. There might be
50    # padding which might mean that the slices are 128 MB even if the size of
51    # the tensor allocated is less than 128 MB.  This formula gives slices with
52    # size: ceil(64 MB / byte size per batch element) bytes.  Using ceil()
53    # guarantees we get a number >= 1.
54
55    # Calculate the size of each batch element.
56    byte_size_per_batch_element = (
57        np.prod(numpy_input.shape[1:]) * input_var.dtype.size)
58
59    # Calculate number of elements we want to copy per slice.
60    batch_size_per_slice = int(
61        np.ceil((64 << 20) / byte_size_per_batch_element))
62
63    # Copy slices of the above size starting at 0, except the last slice will be
64    # smaller.
65    start = 0
66    limit = numpy_input.shape[0]
67    while start < limit:
68      end = min(start + batch_size_per_slice, limit)
69      session.run(assign_slice_op, feed_dict={
70          start_placeholder: start,
71          end_placeholder: end,
72          slice_placeholder: numpy_input[start:end]})
73      start = end
74
75
76def one_host_numpy_dataset(numpy_input, colocate_with, session):
77  """Create a dataset on `colocate_with` from `numpy_input`."""
78  def create_colocated_variable(next_creator, *args, **kwargs):
79    kwargs["colocate_with"] = colocate_with
80    return next_creator(*args, **kwargs)
81
82  numpy_flat = nest.flatten(numpy_input)
83  with variable_scope.variable_creator_scope(create_colocated_variable):
84    vars_flat = tuple(variable_scope.variable(array_ops.zeros(i.shape, i.dtype),
85                                              trainable=False)
86                      for i in numpy_flat)
87  for v, i in zip(vars_flat, numpy_flat):
88    init_var_from_numpy(v, i, session)
89  vars_nested = nest.pack_sequence_as(numpy_input, vars_flat)
90  return dataset_ops.Dataset.from_tensor_slices(vars_nested)
91
92
93class SingleDevice(object):
94  """Used with `colocate_with` to create a non-mirrored variable."""
95
96  def __init__(self, device):
97    self.device = device
98