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
|
/* 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/framework/partial_tensor_shape.h"
#include "tensorflow/core/framework/tensor.h"
#include "tensorflow/core/kernels/data/dataset.h"
#include "tensorflow/core/util/batch_util.h"
namespace tensorflow {
namespace {
// See documentation in ../ops/dataset_ops.cc for a high-level
// description of the following op.
class UnbatchDatasetOp : public UnaryDatasetOpKernel {
public:
explicit UnbatchDatasetOp(OpKernelConstruction* ctx)
: UnaryDatasetOpKernel(ctx) {}
void MakeDataset(OpKernelContext* ctx, DatasetBase* input,
DatasetBase** output) override {
*output = new Dataset(ctx, input);
}
private:
class Dataset : public DatasetBase {
public:
explicit Dataset(OpKernelContext* ctx, DatasetBase* input)
: DatasetBase(DatasetContext(ctx)), input_(input) {
input_->Ref();
for (const PartialTensorShape& shape : input->output_shapes()) {
gtl::InlinedVector<int64, 4> partial_dim_sizes;
for (int i = 1; i < shape.dims(); ++i) {
partial_dim_sizes.push_back(shape.dim_size(i));
}
shapes_.emplace_back(std::move(partial_dim_sizes));
}
}
std::unique_ptr<IteratorBase> MakeIteratorInternal(
const string& prefix) const override {
return std::unique_ptr<IteratorBase>(
new Iterator({this, strings::StrCat(prefix, "::Unbatch")}));
}
const DataTypeVector& output_dtypes() const override {
return input_->output_dtypes();
}
const std::vector<PartialTensorShape>& output_shapes() const override {
return shapes_;
}
string DebugString() const override { return "UnbatchDatasetOp::Dataset"; }
protected:
Status AsGraphDefInternal(SerializationContext* ctx,
DatasetGraphDefBuilder* b,
Node** output) const override {
Node* input_graph_node = nullptr;
TF_RETURN_IF_ERROR(b->AddInputDataset(ctx, input_, &input_graph_node));
TF_RETURN_IF_ERROR(b->AddDataset(this, {input_graph_node}, output));
return Status::OK();
}
private:
class Iterator : public DatasetIterator<Dataset> {
public:
explicit Iterator(const Params& params)
: DatasetIterator<Dataset>(params),
current_index_(0),
current_batch_size_(0),
shapes_(params.dataset->output_shapes().size()) {}
Status Initialize(IteratorContext* ctx) override {
return dataset()->input_->MakeIterator(ctx, prefix(), &input_impl_);
}
Status GetNextInternal(IteratorContext* ctx,
std::vector<Tensor>* out_tensors,
bool* end_of_sequence) override {
mutex_lock l(mu_);
if (!input_impl_) {
*end_of_sequence = true;
return Status::OK();
}
*end_of_sequence = false;
while (!*end_of_sequence) {
if (current_index_ < current_batch_size_) {
out_tensors->clear();
out_tensors->reserve(tensors_.size());
for (int i = 0; i < tensors_.size(); ++i) {
out_tensors->emplace_back(ctx->allocator({}), tensors_[i].dtype(),
shapes_[i]);
TF_RETURN_IF_ERROR(batch_util::MaybeMoveSliceToElement(
&tensors_[i], &out_tensors->back(), current_index_));
}
++current_index_;
*end_of_sequence = false;
return Status::OK();
}
current_index_ = 0;
current_batch_size_ = 0;
tensors_.clear();
TF_RETURN_IF_ERROR(
input_impl_->GetNext(ctx, &tensors_, end_of_sequence));
if (!*end_of_sequence) {
for (size_t i = 0; i < tensors_.size(); ++i) {
if (tensors_[i].dims() == 0) {
return errors::InvalidArgument(
"Input element must have a non-scalar value in each "
"component.");
}
if (tensors_[i].dim_size(0) != tensors_[0].dim_size(0)) {
return errors::InvalidArgument(
"Input element must have the same batch size in each "
"component. Component 0 had size ",
tensors_[0].dim_size(0), " but component ", i,
" had size, ", tensors_[i].dim_size(0), ".");
}
shapes_[i] = tensors_[i].shape();
shapes_[i].RemoveDim(0);
}
current_batch_size_ = tensors_[0].dim_size(0);
}
}
input_impl_.reset();
return Status::OK();
}
protected:
Status SaveInternal(IteratorStateWriter* writer) override {
mutex_lock l(mu_);
if (input_impl_) {
TF_RETURN_IF_ERROR(SaveInput(writer, input_impl_));
} else {
TF_RETURN_IF_ERROR(
writer->WriteScalar(full_name("input_impl_empty"), ""));
}
TF_RETURN_IF_ERROR(
writer->WriteScalar(full_name("current_index"), current_index_));
TF_RETURN_IF_ERROR(
writer->WriteScalar(full_name("n"), current_batch_size_));
if (current_index_ < current_batch_size_) {
for (size_t i = 0; i < tensors_.size(); ++i) {
TF_RETURN_IF_ERROR(writer->WriteTensor(
full_name(strings::StrCat("tensors[", i, "]")), tensors_[i]));
}
}
return Status::OK();
}
Status RestoreInternal(IteratorContext* ctx,
IteratorStateReader* reader) override {
mutex_lock l(mu_);
if (!reader->Contains(full_name("input_impl_empty"))) {
TF_RETURN_IF_ERROR(RestoreInput(ctx, reader, input_impl_));
} else {
input_impl_.reset();
}
TF_RETURN_IF_ERROR(
reader->ReadScalar(full_name("current_index"), ¤t_index_));
TF_RETURN_IF_ERROR(
reader->ReadScalar(full_name("n"), ¤t_batch_size_));
tensors_.clear();
tensors_.resize(dataset()->output_dtypes().size());
if (current_index_ < current_batch_size_) {
for (size_t i = 0; i < tensors_.size(); ++i) {
TF_RETURN_IF_ERROR(reader->ReadTensor(
full_name(strings::StrCat("tensors[", i, "]")), &tensors_[i]));
shapes_[i] = tensors_[i].shape();
shapes_[i].RemoveDim(0);
}
}
return Status::OK();
}
private:
mutex mu_;
int64 current_index_ GUARDED_BY(mu_);
int64 current_batch_size_ GUARDED_BY(mu_);
std::vector<Tensor> tensors_ GUARDED_BY(mu_);
std::unique_ptr<IteratorBase> input_impl_ GUARDED_BY(mu_);
std::vector<TensorShape> shapes_ GUARDED_BY(mu_);
};
const DatasetBase* const input_;
std::vector<PartialTensorShape> shapes_;
};
};
REGISTER_KERNEL_BUILDER(Name("UnbatchDataset").Device(DEVICE_CPU),
UnbatchDatasetOp);
} // namespace
} // namespace tensorflow
|
