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
|
/* Copyright 2015 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.
==============================================================================*/
// The utility to write checkpoints for google brain tensor ops and v3
// checkpoints for dist_belief.
#ifndef TENSORFLOW_CORE_UTIL_TENSOR_SLICE_WRITER_H_
#define TENSORFLOW_CORE_UTIL_TENSOR_SLICE_WRITER_H_
#include <unordered_map>
#include "tensorflow/core/framework/tensor_shape.h"
#include "tensorflow/core/framework/tensor_slice.h"
#include "tensorflow/core/framework/types.h"
#include "tensorflow/core/lib/core/errors.h"
#include "tensorflow/core/lib/core/status.h"
#include "tensorflow/core/lib/core/stringpiece.h"
#include "tensorflow/core/lib/gtl/map_util.h"
#include "tensorflow/core/lib/strings/stringprintf.h"
#include "tensorflow/core/platform/logging.h"
#include "tensorflow/core/platform/macros.h"
#include "tensorflow/core/platform/types.h"
#include "tensorflow/core/util/saved_tensor_slice.pb_text.h"
#include "tensorflow/core/util/saved_tensor_slice.pb.h"
#include "tensorflow/core/util/saved_tensor_slice_util.h"
namespace tensorflow {
namespace checkpoint {
class TensorSliceWriter {
public:
// Abstract interface that TensorSliceWriter uses for building
class Builder {
public:
virtual ~Builder() {}
virtual void Add(StringPiece key, StringPiece value) = 0;
virtual Status Finish(int64* file_size) = 0;
};
typedef std::function<Status(const string&, Builder**)> CreateBuilderFunction;
TensorSliceWriter(const string& filename,
CreateBuilderFunction create_builder);
virtual ~TensorSliceWriter() {}
// Adds a slice. We support float and int32 for now.
// TODO(yangke): add more supports
template <typename T>
Status Add(const string& name, const TensorShape& shape,
const TensorSlice& slice, const T* data);
Status Finish();
// Allocate "num_elements" elements in "ss" and save the data in "data"
// there.
template <typename T>
static Status SaveData(const T* data, int64 num_elements, SavedSlice* ss);
static size_t MaxBytesPerElement(DataType dt);
private:
static const size_t kMaxMessageBytes = 1LL << 31;
// Filling in the TensorProto in a SavedSlice will add the following
// header bytes, in addition to the data:
// - 1 byte: TensorProto tag and wire format
// - <= 5 bytes: TensorProto length
// - 1 byte: Repeated *_val tag and wire format
// - <= 5 bytes: *_val length
// However, we add 1KB of slack, to be conservative and guard
// against other additions to the TensorProto.
static const size_t kTensorProtoHeaderBytes = 1 << 10;
const string filename_;
const CreateBuilderFunction create_builder_;
const string tmpname_;
// A mapping from the tensor names to their index in meta_.saved_slice_meta()
std::unordered_map<string, int> name_to_index_;
// The metadata that holds all the saved tensor slices.
SavedTensorSlices sts_;
// The data to be written to the builder
std::map<string, string> data_;
// Total number of slices written
int slices_;
TF_DISALLOW_COPY_AND_ASSIGN(TensorSliceWriter);
};
template <typename T>
Status TensorSliceWriter::Add(const string& name, const TensorShape& shape,
const TensorSlice& slice, const T* data) {
// The tensor and the slice have to be compatible
if (shape.dims() != slice.dims()) {
return errors::Internal("Incompatible tensor shape and slice: ", "shape = ",
shape.DebugString(),
", slice = ", slice.DebugString());
}
DataType dt = DataTypeToEnum<T>::value;
// We need to add an entry for "name" if there isn't an entry already.
int index = gtl::FindWithDefault(name_to_index_, name, -1);
if (index >= 0) {
// The same tensor has been registered -- we verify that the shapes and the
// type agree.
const SavedSliceMeta& ssm = sts_.meta().tensor(index);
CHECK_EQ(name, ssm.name()) << ProtoShortDebugString(ssm);
TensorShape ssm_shape(ssm.shape());
if (!shape.IsSameSize(ssm_shape)) {
return errors::Internal(
"Mismatching shapes: existing tensor = ", ssm_shape.DebugString(),
", trying to add name ", name, ", shape = ", shape.DebugString());
}
if (dt != ssm.type()) {
return errors::Internal(
"Mismatching types: existing type = ", DataTypeString(ssm.type()),
", trying to add name ", name, ", type = ", DataTypeString(dt));
}
} else {
// Insert the new tensor name with the shape information
index = sts_.meta().tensor_size();
name_to_index_.insert(std::make_pair(name, index));
SavedSliceMeta* ssm = sts_.mutable_meta()->add_tensor();
ssm->set_name(name);
shape.AsProto(ssm->mutable_shape());
ssm->set_type(dt);
}
// Now we need to add the slice info the list of slices.
SavedSliceMeta* ssm = sts_.mutable_meta()->mutable_tensor(index);
slice.AsProto(ssm->add_slice());
// Now we need to add the real data.
{
SavedTensorSlices sts;
SavedSlice* ss = sts.mutable_data();
ss->set_name(name);
slice.AsProto(ss->mutable_slice());
TensorShape saved_shape(ssm->shape());
TensorShape sliced_shape;
TF_RETURN_IF_ERROR(slice.SliceTensorShape(saved_shape, &sliced_shape));
TF_RETURN_IF_ERROR(SaveData(data, sliced_shape.num_elements(), ss));
string key = EncodeTensorNameSlice(name, slice);
// TODO(yangke): consider doing a two-pass thing where the first pass just
// list the tensor slices we want to save and then another pass to actually
// set the data. Need to figure out if the interface works well.
std::pair<string, string> key_value(key, "");
if (!sts.AppendToString(&key_value.second)) {
return errors::Internal("Error writing Tensor. Possible size overflow.");
}
data_.insert(key_value);
}
++slices_;
return Status::OK();
}
template <typename T>
Status TensorSliceWriter::SaveData(const T* data, int64 num_elements,
SavedSlice* ss) {
size_t size_bound =
ss->ByteSize() + kTensorProtoHeaderBytes +
(MaxBytesPerElement(DataTypeToEnum<T>::value) * num_elements);
if (size_bound > kMaxMessageBytes) {
return errors::InvalidArgument(
"Tensor slice is too large to serialize (conservative estimate: ",
size_bound, " bytes)");
}
Fill(data, num_elements, ss->mutable_data());
DCHECK_GE(ss->ByteSize(), 0);
DCHECK_LE(ss->ByteSize(), size_bound);
return Status::OK();
}
template <>
Status TensorSliceWriter::SaveData(const string* data, int64 num_elements,
SavedSlice* ss);
// Create a table builder that will write to "filename" in
// tensorflow::io::Table format. If successful, return OK
// and set "*builder" to the allocated builder. Otherwise, return a
// non-OK status.
Status CreateTableTensorSliceBuilder(const string& filename,
TensorSliceWriter::Builder** builder);
} // namespace checkpoint
} // namespace tensorflow
#endif // TENSORFLOW_CORE_UTIL_TENSOR_SLICE_WRITER_H_
|
