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
|
/* 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.
==============================================================================*/
#include "tensorflow/core/kernels/dataset.h"
#include "tensorflow/core/framework/partial_tensor_shape.h"
#include "tensorflow/core/framework/tensor.h"
namespace tensorflow {
namespace {
// See documentation in ../ops/dataset_ops.cc for a high-level
// description of the following op.
class BatchDatasetOp : public UnaryDatasetOpKernel {
public:
explicit BatchDatasetOp(OpKernelConstruction* ctx)
: UnaryDatasetOpKernel(ctx) {}
void MakeDataset(OpKernelContext* ctx, DatasetBase* input,
DatasetBase** output) override {
int64 batch_size = 0;
OP_REQUIRES_OK(ctx,
ParseScalarArgument<int64>(ctx, "batch_size", &batch_size));
OP_REQUIRES(
ctx, batch_size > 0,
errors::InvalidArgument("Batch size must be greater than zero."));
*output = new Dataset(ctx, batch_size, input);
}
private:
class Dataset : public GraphDatasetBase {
public:
Dataset(OpKernelContext* ctx, int64 batch_size, const DatasetBase* input)
: GraphDatasetBase(ctx), batch_size_(batch_size), input_(input) {
input_->Ref();
// NOTE(mrry): Currently we implement "batch up to" semantics. If
// we could tell statically that the input dataset is infinite,
// then we could always report `batch_size` as the 0th dimension.
const auto& input_shapes = input_->output_shapes();
output_shapes_.reserve(input_shapes.size());
for (const auto& input_shape : input_shapes) {
output_shapes_.emplace_back(
PartialTensorShape({-1}).Concatenate(input_shape));
}
}
~Dataset() override { input_->Unref(); }
std::unique_ptr<IteratorBase> MakeIterator(
const string& prefix) const override {
return std::unique_ptr<IteratorBase>(new Iterator(
Iterator::Params{this, strings::StrCat(prefix, "::Batch")}));
}
const DataTypeVector& output_dtypes() const override {
return input_->output_dtypes();
}
const std::vector<PartialTensorShape>& output_shapes() const override {
return output_shapes_;
}
string DebugString() override {
return strings::StrCat("BatchDatasetOp(", batch_size_, ")::Dataset");
}
protected:
Status AsGraphDefInternal(DatasetGraphDefBuilder* b,
Node** output) const override {
Node* input_graph_node = nullptr;
TF_RETURN_IF_ERROR(b->AddParentDataset(input_, &input_graph_node));
Node* batch_size = nullptr;
TF_RETURN_IF_ERROR(b->AddScalar(batch_size_, &batch_size));
TF_RETURN_IF_ERROR(
b->AddDataset(this, {input_graph_node, batch_size}, output));
return Status::OK();
}
private:
// Copies element into the index^th slice of parent (in the 0th dimension).
//
// TODO(mrry): Reconcile this method with the similar method in
// the queue implementation.
template <typename T>
static Status HandleElementToSlice(const Tensor& element, Tensor* parent,
int64 index) {
if (element.NumElements() !=
(parent->NumElements() / parent->dim_size(0))) {
TensorShape chip_shape = parent->shape();
chip_shape.RemoveDim(0);
return errors::InvalidArgument(
"HandleElementToSlice Cannot copy slice: number of elements does "
"not match. Shapes are: [element]: ",
element.shape().DebugString(),
", [parent slice]: ", chip_shape.DebugString());
}
auto parent_as_matrix = parent->flat_outer_dims<T>();
parent_as_matrix.chip(index, 0) = element.flat<T>();
return Status::OK();
}
// Copies element into the index^th slice of parent (in the 0th dimension).
static Status CopyElementToSlice(const Tensor& element, Tensor* parent,
int64 index) {
#define HANDLE_TYPE(T) \
case DataTypeToEnum<T>::value: { \
return HandleElementToSlice<T>(element, parent, index); \
}
switch (element.dtype()) {
TF_CALL_DATASET_TYPES(HANDLE_TYPE);
#undef HANDLE_TYPE
default:
return errors::Unimplemented(
"CopyElementToSlice Unhandled data type: ", element.dtype());
}
}
class Iterator : public DatasetIterator<Dataset> {
public:
explicit Iterator(const Params& params)
: DatasetIterator<Dataset>(params),
input_impl_(params.dataset->input_->MakeIterator(params.prefix)) {}
Status GetNextInternal(IteratorContext* ctx,
std::vector<Tensor>* out_tensors,
bool* end_of_sequence) override {
// Each row of `batch_elements` is a tuple of tensors from the
// input iterator.
std::vector<std::vector<Tensor>> batch_elements;
batch_elements.reserve(dataset()->batch_size_);
{
mutex_lock l(mu_);
*end_of_sequence = false;
for (int i = 0; i < dataset()->batch_size_ && !*end_of_sequence;
++i) {
std::vector<Tensor> batch_element_tuple;
TF_RETURN_IF_ERROR(input_impl_->GetNext(ctx, &batch_element_tuple,
end_of_sequence));
if (!*end_of_sequence) {
batch_elements.emplace_back(std::move(batch_element_tuple));
}
}
}
if (batch_elements.empty()) {
DCHECK(*end_of_sequence);
return Status::OK();
}
// Copy the retrieved batch elements into one output tensor
// per tuple component.
// NOTE(mrry): If the input or output sizes are statically
// known, we could potentially read the input values in-place
// into their respective slice locations. This would require a
// different GetNext() overload that supports zero-copy, and might
// make sense in an optimization pass.
const size_t num_tuple_components = batch_elements[0].size();
const int64 num_batch_elements = batch_elements.size();
for (size_t component_index = 0; component_index < num_tuple_components;
++component_index) {
const Tensor& first_element = batch_elements[0][component_index];
TensorShape batch_component_shape({num_batch_elements});
batch_component_shape.AppendShape(first_element.shape());
Tensor batch_component(cpu_allocator(), first_element.dtype(),
batch_component_shape);
// Build the output tuple component by copying one slice
// from each input element in the batch.
for (size_t i = 0; i < num_batch_elements; ++i) {
TF_RETURN_IF_ERROR(CopyElementToSlice(
batch_elements[i][component_index], &batch_component, i));
}
out_tensors->emplace_back(std::move(batch_component));
}
*end_of_sequence = false;
return Status::OK();
}
protected:
Status SaveInternal(IteratorStateWriter* writer) override {
mutex_lock l(mu_);
TF_RETURN_IF_ERROR(SaveParent(writer, input_impl_));
return Status::OK();
}
Status RestoreInternal(OpKernelContext* ctx,
IteratorStateReader* reader) override {
mutex_lock l(mu_);
TF_RETURN_IF_ERROR(RestoreParent(ctx, reader, input_impl_));
return Status::OK();
}
private:
mutex mu_;
std::unique_ptr<IteratorBase> input_impl_ GUARDED_BY(mu_);
};
const int64 batch_size_;
const DatasetBase* const input_;
std::vector<PartialTensorShape> output_shapes_;
};
};
REGISTER_KERNEL_BUILDER(Name("BatchDataset").Device(DEVICE_CPU),
BatchDatasetOp);
} // namespace
} // namespace tensorflow
|
