1 /* Copyright 2016 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 16 // A tensor bundle is a set of immutable persistent files storing a set of named 17 // tensors. It is designed for checkpointing TensorFlow tensors. 18 // 19 // The paths of the managed files share a common prefix; e.g., with the prefix: 20 // /fs/model/train/ckpt-step/ckpt 21 // 22 // the bundle may contain a metadata file, and sharded data files: 23 // /fs/model/train/ckpt-step/ 24 // ckpt.index 25 // ckpt.data-00000-of-00020 26 // ckpt.data-00001-of-00020 27 // ... 28 // ckpt.data-00019-of-00020 29 // 30 // The ".index" file is a string-string immutable table 31 // (tensorflow::table::Table). Each key is a name of a tensor and its value is 32 // a serialized BundleEntryProto. Each BundleEntryProto describes the metadata 33 // of a tensor: which of the "data" files contains the content of a tensor, the 34 // offset into that file, checksum, some auxiliary data, etc. 35 // 36 // A tensor bundle can be accessed randomly using a BundleReader. Usage: 37 // 38 // BundleReader reader(env, "/fs/model/train/ckpt-step/ckpt"); 39 // reader.Lookup("name", &tensor); 40 // 41 // A tensor bundle can be built using BundleWriter. Each BundleWriter builds a 42 // single data file bundle. Multiple bundles can then be merged by 43 // MergeBundles() without reading and writing large chunk of data: it reads the 44 // metadata files and outputs a single merged metadata. Typical usage: 45 // 46 // worker 0: 47 // BundleWriter writer(env, "/fs/model/train/ckpt-step/tmp/worker0-step"); 48 // writer.Add(...); // Adds the tensors on this worker. 49 // writer.Finish(); // Flushes. 50 // worker 1: 51 // BundleWriter writer(env, "/fs/model/train/ckpt-step/tmp/worker1-step"); 52 // writer.Add(...); 53 // writer.Finish(); 54 // worker 2: 55 // MergeBundles(env, 56 // {"/fs/model/train/ckpt-step/tmp/worker0-step", 57 // "/fs/model/train/ckpt-step/tmp/worker1-step"}, 58 // "/fs/model/train/ckpt-step/ckpt" /* merged prefix */); 59 // 60 61 #ifndef TENSORFLOW_CORE_UTIL_TENSOR_BUNDLE_TENSOR_BUNDLE_H_ 62 #define TENSORFLOW_CORE_UTIL_TENSOR_BUNDLE_TENSOR_BUNDLE_H_ 63 64 #include "tensorflow/core/protobuf/tensor_bundle.pb.h" 65 66 #include <map> 67 #include <string> 68 #include <unordered_map> 69 70 #include "tensorflow/core/framework/tensor.h" 71 #include "tensorflow/core/framework/tensor_shape.h" 72 #include "tensorflow/core/framework/tensor_slice.h" 73 #include "tensorflow/core/lib/core/status.h" 74 #include "tensorflow/core/lib/gtl/array_slice.h" 75 #include "tensorflow/core/lib/io/inputbuffer.h" 76 #include "tensorflow/core/lib/io/table.h" 77 #include "tensorflow/core/platform/env.h" 78 #include "tensorflow/core/platform/file_system.h" 79 #include "tensorflow/core/platform/macros.h" 80 #include "tensorflow/core/platform/types.h" 81 #include "tensorflow/core/util/tensor_bundle/naming.h" 82 #include "tensorflow/core/util/tensor_slice_set.h" 83 84 namespace tensorflow { 85 86 class FileOutputBuffer; 87 88 // Versioning of the tensor bundle format. 89 // Follows the same rules as 3p/tf/core/public/version.h. 90 // 91 // History: 92 // 0. Any tensor bundles produced before this field was added. 93 // 1. Added this field (2016-09-14). 94 extern const int kTensorBundleMinProducer; 95 extern const int kTensorBundleMinConsumer; 96 extern const int kTensorBundleVersion; 97 98 // The empty string, hence always the first key in the metadata table. Its 99 // corresponding value is a BundleHeaderProto. 100 extern const char* const kHeaderEntryKey; 101 102 // Builds a string-string table of tensor names to BundleEntryProto (metadata). 103 // 104 // On construction, attempts to create a directory given by the dirname of 105 // "prefix", so "status()" must be checked before calling any member functions. 106 // 107 // All threads accessing the same BundleWriter must synchronize. 108 class BundleWriter { 109 public: 110 struct Options { OptionsOptions111 Options() {} 112 // Alignment, in bytes, for tensor data. 113 // Must be >= 1. The default size of 1 densely packs tensors. 114 int data_alignment{1}; 115 }; 116 BundleWriter(Env* env, StringPiece prefix, 117 const Options& options = Options()); 118 119 // Adds the tensor "val" under key "key". 120 // Across calls "key" must be unique but can be added in any order. 121 Status Add(StringPiece key, const Tensor& val); 122 123 // Partitioned variables support. 124 // A slice of a full tensor is stored in two entries in the metadata table: 125 // 126 // full_tensor_key -> BundleEntryProto, describing all stored slices 127 // of this full tensor. Does not append to the data 128 // file. 129 // encoded slice key -> BundleEntryProto, describing one particular slice. 130 // Appends values of this slice to the data file. 131 // 132 // Slices of a full tensor can be added in any order. 133 // 134 // If a full tensor has slices placed on N devices and N BundleWriter's are 135 // concurrently used, the caller must use MergeBundles() to ensure that a 136 // consistent entry for "full_tensor_key" is produced. 137 // 138 // Returns an error if the same slice is added the second time. 139 Status AddSlice(StringPiece full_tensor_key, 140 const TensorShape& full_tensor_shape, 141 const TensorSlice& slice_spec, const Tensor& slice_tensor); 142 143 // Finishes the writer and flushes. 144 Status Finish() TF_MUST_USE_RESULT; 145 status()146 Status status() const { return status_; } 147 148 private: 149 Env* const env_; // Not owned. 150 const Options options_; 151 const string prefix_; 152 const string tmp_metadata_path_; 153 const string tmp_data_path_; 154 std::unique_ptr<FileOutputBuffer> out_; 155 int64 size_; // Number of bytes written into out_. 156 std::map<string, BundleEntryProto> entries_; 157 Status status_; 158 159 TF_DISALLOW_COPY_AND_ASSIGN(BundleWriter); 160 }; 161 162 // Merges a set of bundles (given their prefixes) into a single bundle with the 163 // given "merged_prefix". The merged metadata is guaranteed to be consistent. 164 // 165 // If there are N bundles in "prefixes", during the merge the data files will be 166 // renamed to contain a proper sharded file spec, with num_shards set to the sum 167 // of num_shards across the N input bundles. 168 // 169 // The caller should only rely on the metadata file of the merged bundle to 170 // query information about a tensor. In particular, this function does not 171 // guarantee not to re-order the input data files. 172 // 173 // Once merged, makes a best effort to delete the old metadata files. 174 // Returns OK iff all bundles are successfully merged. 175 Status MergeBundles(Env* env, gtl::ArraySlice<string> prefixes, 176 StringPiece merged_prefix); 177 178 // On construction, silently attempts to read the metadata associated with 179 // "prefix". If caller intends to call any function afterwards, "status()" 180 // must be checked. 181 // All threads accessing the same BundleReader must synchronize. 182 class BundleReader { 183 public: 184 BundleReader(Env* const env, StringPiece prefix); 185 ~BundleReader(); 186 187 // Is ok() iff the reader construction is successful (completed the read of 188 // the metadata). status()189 Status status() const { return status_; } 190 191 // Queries whether the bundle contains an entry keyed by "key". Calls Seek() 192 // internally, so this call invalidates the reader's current position. 193 // REQUIRES: status().ok() 194 bool Contains(StringPiece key); 195 196 // Looks up the dtype and the shape of the tensor keyed by "key". 197 // REQUIRES: status().ok() 198 Status LookupDtypeAndShape(StringPiece key, DataType* dtype, 199 TensorShape* shape) TF_MUST_USE_RESULT; 200 201 // Looks up the shape of the tensor keyed by "key". 202 // Clears "shape" if not found. 203 // REQUIRES: status().ok() 204 Status LookupTensorShape(StringPiece key, 205 TensorShape* shape) TF_MUST_USE_RESULT; 206 207 // Looks up the tensor keyed by "key". If "key" refers to a partitioned 208 // tensor, attempts to look up the full contents using all stored slices. 209 // 210 // Caller must make sure "val" has the same shape and dtype as the 211 // corresponding contents, so that its buffer can be filled without needing 212 // extra allocation. These can be queried via "LookupDtypeAndShape()". 213 // 214 // On error, "val" may contain nonsense data. Returns a NotFound error if 215 // tensor keyed by "key" does not exist in this bundle. 216 // 217 // Validates the stored crc32c checksum against the restored bytes. 218 // REQUIRES: status().ok() 219 Status Lookup(StringPiece key, Tensor* val) TF_MUST_USE_RESULT; 220 221 // Looks up the tensor pointed to by the internal iterator. 222 // 223 // On error, "val" may contain nonsense data. 224 // 225 // Validates the stored crc32c checksum against the restored bytes. 226 // REQUIRES: status().ok() && Valid() 227 Status ReadCurrent(Tensor* val) TF_MUST_USE_RESULT; 228 229 // Looks up the slices of the tensor keyed by "key". On OK, "slices" 230 // is non-empty if and only if the tensor is a partitioned tensor. 231 // 232 // Warning - there is no guaranteed ordering for the returned slices, so 233 // a slice with a larger start index in some dimension could come before 234 // another slice with a smaller start index in the same dimension. 235 // REQUIRES: status().ok() 236 Status LookupTensorSlices(StringPiece key, std::vector<TensorSlice>* slices) 237 TF_MUST_USE_RESULT; 238 239 // Looks up a specific slice of a partitioned tensor. 240 // It is only required that the stored slices cover the requested slice, 241 // namely "slice_spec" is a subset of the union of the stored slices. 242 // REQUIRES: status().ok() 243 Status LookupSlice(StringPiece full_tensor_key, const TensorSlice& slice_spec, 244 Tensor* val) TF_MUST_USE_RESULT; 245 246 // Seeks to the first position in the bundle whose key is no less than "key". 247 // REQUIRES: status().ok() Seek(StringPiece key)248 void Seek(StringPiece key) { return iter_->Seek(key); } 249 // Moves to the next position in the bundle. 250 // REQUIRES: status().ok() Next()251 void Next() const { iter_->Next(); } 252 // Returns true iff the reader is positioned to a key/val pair. 253 // REQUIRES: status().ok() Valid()254 bool Valid() const { return iter_->Valid(); } 255 256 // Returns the key at the current position. 257 // REQUIRES: status().ok() && Valid() key()258 StringPiece key() const { return iter_->key(); } 259 // Returns the raw value at the current position. 260 // REQUIRES: status().ok() && Valid() value()261 StringPiece value() const { return iter_->value(); } 262 263 string DebugString(); 264 265 private: 266 // Seeks for "key" and reads the metadata proto. 267 // On non-OK return, clears "entry" for the caller. 268 // REQUIRES: status().ok() 269 Status GetBundleEntryProto(StringPiece key, 270 BundleEntryProto* entry) TF_MUST_USE_RESULT; 271 272 // Reads the tensor value described by the metadata proto "entry". 273 // Usage for "val" follows the comment of "Lookup()". 274 Status GetValue(const BundleEntryProto& entry, 275 Tensor* val) TF_MUST_USE_RESULT; 276 277 // Reads the slice described by "slice_spec". The corresponding full tensor 278 // has key "ful_tensor_key" and metadata proto "full_tensor_entry". 279 // REQUIRES: full_tensor_entry.slices_size() > 0 280 Status GetSliceValue(StringPiece full_tensor_key, 281 const BundleEntryProto& full_tensor_entry, 282 const TensorSlice& slice_spec, 283 Tensor* val) TF_MUST_USE_RESULT; 284 285 Env* env_; // Not owned. 286 const string prefix_; 287 288 Status status_; 289 RandomAccessFile* metadata_; // Owned. 290 table::Table* table_; 291 table::Iterator* iter_; 292 // Owned the InputBuffer objects and their underlying RandomAccessFile's. 293 std::unordered_map<int32, io::InputBuffer*> data_; 294 295 // Maps each partitioned tensor's key to its stored slices (represented in a 296 // TensorSliceSet). Populated on-demand. 297 std::unordered_map<string, checkpoint::TensorSliceSet*> tensor_slices_; 298 299 // Expected number of data file shards in the bundle. Extracted by reading 300 // the header entry in the metadata table. 301 int num_shards_; 302 303 friend class TensorBundleAlignmentTest; // For testing data alignment. 304 305 TF_DISALLOW_COPY_AND_ASSIGN(BundleReader); 306 }; 307 308 // A buffering wrapper for a WritableFile. Useful if the caller wishes to issue 309 // small writes to a file (e.g. writing out a list of small varints). 310 // External synchronization must be used in the presence of concurrent callers. 311 class FileOutputBuffer { 312 public: FileOutputBuffer(WritableFile * file,size_t buffer_size)313 FileOutputBuffer(WritableFile* file, size_t buffer_size) 314 : file_(file), position_(0), buffer_size_(buffer_size) { 315 DCHECK_GT(buffer_size, 0); 316 buffer_.resize(buffer_size); 317 } 318 ~FileOutputBuffer(); 319 320 // Buffered append. 321 Status Append(StringPiece data); 322 323 // Returns the running crc32c checksum of all currently appended bytes. crc32c()324 uint32 crc32c() { return crc32c_; } 325 // Clears the running crc32c checksum. clear_crc32c()326 void clear_crc32c() { crc32c_ = 0; } 327 328 // Appends the buffered data, then closes the underlying file. 329 Status Close(); 330 331 private: 332 // Appends the buffered data to the underlying file. Does NOT flush the file. 333 Status FlushBuffer(); 334 335 WritableFile* file_; // Owned. 336 337 // buffer_[0, position_) holds the buffered data not yet appended to the 338 // underlying file. 339 size_t position_; 340 const size_t buffer_size_; 341 std::vector<char> buffer_; 342 343 // Checksum of all appended bytes since construction or last clear_crc32c(). 344 uint32 crc32c_ = 0; 345 }; 346 347 } // namespace tensorflow 348 349 #endif // TENSORFLOW_CORE_UTIL_TENSOR_BUNDLE_TENSOR_BUNDLE_H_ 350