Parameter.h 10.0 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Z
zhangjinchao01 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

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. */

#pragma once

#include <stdint.h>

#include <iostream>
#include <string>
#include <vector>

#include "ParameterConfig.pb.h"
#include "TrainerConfig.pb.h"

Y
Yu Yang 已提交
26 27 28
#include "ParameterUpdaterHook.h"
#include "paddle/math/Matrix.h"
#include "paddle/math/Vector.h"
L
liaogang 已提交
29
#include "paddle/utils/Common.h"
Y
Yu Yang 已提交
30
#include "paddle/utils/GlobalConstants.h"
Z
zhangjinchao01 已提交
31
#include "paddle/utils/Locks.h"
Y
Yu Yang 已提交
32
#include "paddle/utils/ThreadLocal.h"
Z
zhangjinchao01 已提交
33 34 35 36
#include "paddle/utils/Util.h"

namespace paddle {

T
tensor-tang 已提交
37
typedef enum {
T
tensor-tang 已提交
38 39 40 41 42 43 44 45 46 47 48
  /// The paddle original basic format
  PARAM_FORMAT_ORIGINAL = 0,

  /// See mkldnn_memory_format_t in
  /// https://github.com/01org/mkl-dnn/blob/master/include/mkldnn_types.h
  /// for a detailed description.
  /// 2D weights tensor in the format (output channels, input channels).
  PARAM_FORMAT_MKLDNN_OI,

  /// The total format items numbers
  PARAM_FORMAT_ITEMS,
T
tensor-tang 已提交
49 50
} PARAM_FORMAT;

Z
zhangjinchao01 已提交
51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67
class SparsePrefetchRowCpuMatrix;

class Parameter;
typedef std::function<void(Parameter* param)> UpdateCallback;
typedef std::function<void(int paramId, Parameter* param)> ParamInitCallback;

class Parameter;
typedef std::shared_ptr<Parameter> ParameterPtr;

class Parameter {
public:
  Parameter(const ParameterConfig& config, bool useGpu, bool doInit = true);
  const std::string& getName() const { return config_.name(); }

  size_t getSize() const { return config_.size(); }

  bool isFullSize() const {
武毅 已提交
68 69 70 71
    if (bufs_[PARAMETER_VALUE]) {
      return this->getSize() == bufs_[PARAMETER_VALUE]->getSize();
    }
    return false;
Z
zhangjinchao01 已提交
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
  }

  inline bool useGpu() const { return useGpu_; }

  int getDeviceId() const { return deviceId_; }

  void setDevice(int deviceId) { deviceId_ = deviceId; }

  /// The id ranges from 0 to the_total_number_of_parameters - 1
  size_t getID() const { return config_.para_id(); }

  /// ID is a implict value created until neural network is built.
  void setID(size_t id) { config_.set_para_id(id); }

  bool isStatic() const { return config_.is_static(); }

  enum MatType {
    MAT_NORMAL,
    /// both value and grad are shared
    MAT_NORMAL_SHARED,

    /// Now used in BatchNorm in CPU mode
    MAT_VALUE_SHARED,

    /// sparse matrix, which has full size parameter
    MAT_SPARSE_ROW_IDS,
    /// sparse matrix, parameter size scale by sparse rates.
    MAT_SPARSE_ROW_AUTO_GROW,
    MAT_CACHE_ROW,
    MAT_SPARSE_ROW,

    /// sparse matrix for prefetching parameter from pserver
    MAT_SPARSE_ROW_PREFETCH,
    /// same as above, but parameter has full size for saving parameter in local
    MAT_SPARSE_ROW_PREFETCH_FULL_SIZE,
  };

  void enableSparseParameter() {
    if (config_.is_sparse()) {
      if (config_.format() == "csr") {
        size_t height = config_.dims(0);
        size_t nnz = config_.size();
        enableIntType(PARAMETER_ROWS, height + 1);
        enableIntType(PARAMETER_COLS, nnz);
        format_ = SPARSE_CSR;
      } else {
        size_t width = config_.dims(1);
        size_t nnz = config_.size();
        enableIntType(PARAMETER_COLS, width + 1);
        enableIntType(PARAMETER_ROWS, nnz);
        format_ = SPARSE_CSC;
      }
    }
  }

  /// allocate buffer for the give type
  void enableType(ParameterType type, MatType matType = MAT_NORMAL) {
    if (bufs_[type] || mats_[type]) {
      return;
    }
    SetDevice device(deviceId_);
    if (config_.dims_size() == 2) {
      if (matType == MAT_NORMAL || matType == MAT_NORMAL_SHARED ||
          matType == MAT_SPARSE_ROW_PREFETCH_FULL_SIZE ||
136
          matType == MAT_VALUE_SHARED || matType == MAT_SPARSE_ROW_IDS) {
Z
zhangjinchao01 已提交
137 138 139 140 141 142 143 144 145 146 147 148 149 150 151
        bufs_[type] = Vector::createParallelVector(config_.size(), useGpu_);
        bufs_[type]->zeroMem();
      } else {
        CHECK(isGradSparseUpdate());
      }
      if (config_.is_sparse() && type == PARAMETER_VALUE) {
        enableSparseParameter();
      }
      setMat(type, matType);
    } else {
      bufs_[type] = Vector::createParallelVector(config_.size(), useGpu_);
      bufs_[type]->zeroMem();
    }
  }

152 153 154 155 156 157
  void enableBufType(ParameterType type) {
    if (bufs_[type]) return;
    bufs_[type] = Vector::createParallelVector(config_.size(), useGpu_);
    bufs_[type]->zeroMem();
  }

Z
zhangjinchao01 已提交
158 159 160 161 162 163 164 165 166
  void enableIntType(ParameterType type, size_t intStoreSize = 0) {
    if (!intBufs_[type]) {
      SetDevice device(deviceId_);
      size_t size = intStoreSize ? intStoreSize : config_.size();
      intBufs_[type] = IVector::create(size, useGpu_);
      intBufs_[type]->zeroMem();
    }
  }

167 168
  void enableSharedType(ParameterType type,
                        VectorPtr vec,
Z
zhangjinchao01 已提交
169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263
                        MatrixPtr mat = nullptr) {
    if (!bufs_[type] && !mats_[type]) {
      bufs_[type] = vec;
      mats_[type] = mat;
    }
  }

  /// for batchGradientMachine: blockNum is number of partitions of the matrix.
  bool isGradShared(size_t* blockNum = NULL);

  bool isValueShared();

  // for AsgdSparseGradientMachine & SgdSparseGradientMachine:
  // and MultiGradientMachine
  bool isGradSparseUpdate() const;

  bool isSparseRemoteUpdate() const {
    return config_.sparse_remote_update() && !useGpu();
  }

  const ParameterConfig& getConfig() const { return config_; }

  ParameterConfig& getConfig() { return config_; }

  bool hasType(ParameterType pType) const {
    return bufs_[pType] || mats_[pType];
  }

  const VectorPtr& getBuf(ParameterType pType) const {
    return this->bufs_[pType];
  }

  const VectorPtr* getBufs() const { return bufs_; }

  const MatrixPtr& getMat(ParameterType pType) const { return mats_[pType]; }

  void setValueUpdated() { updated_ = true; }

  void clearValueUpdated() { updated_ = false; }

  bool isValueUpdated() const { return updated_; }

  /**
   * Save parameter value to a file
   */
  bool save(const std::string& filename) const;

  /**
   * Save parameter to ostream
   */
  bool save(std::ostream& s) const;

  /**
   * Load parameter value from a file
   */
  bool load(const std::string& filename);

  /**
   * Load parameter from istream
   */
  bool load(std::istream& is);

  void incShared() { sharedCount_++; }

  /**
   * After one of the parameter's gradient is merged
   * You should call this function to do some additional processing,
   */
  void incUpdate(const UpdateCallback& callbacks = NULL);

  void clearGradient() {
    auto& mat = getMat(PARAMETER_GRADIENT);
    if (mat) {
      // zeroMem will also clear rows for SparseRowCpuMatrix
      mat->zeroMem();
    } else {
      auto& gradBuf = getBuf(PARAMETER_GRADIENT);
      if (gradBuf) gradBuf->zeroMem();
    }
  }

  void initialize();

  /**
   * Initialize the value according to config_: initial_mean,
   * initial_std and initial_strategy.
   */
  void randomize();
  static void randomize(const VectorPtr& value, const ParameterConfig& config);

  /// Initialize the value to 0
  void zeroMem();

  /// file header structure
  struct Header {
T
tensor-tang 已提交
264
    int32_t format;      // = PARAM_FORMAT
Z
zhangjinchao01 已提交
265 266 267 268
    uint32_t valueSize;  // = sizeof(real)
    uint64_t size;       // = getSize()
  };

T
tensor-tang 已提交
269
  /**
T
tensor-tang 已提交
270
   * @brief Is the header format supported.
T
tensor-tang 已提交
271 272 273 274 275 276
   */
  static bool isHeaderFormatSupported(int32_t fmt) {
    return fmt < PARAM_FORMAT_ITEMS;
  }

  /**
T
tensor-tang 已提交
277
   * @brief Get the format in header.
T
tensor-tang 已提交
278 279 280 281
   */
  int getHeaderFormat() { return headerFormat_; }

  /**
T
tensor-tang 已提交
282
   * @brief Set the format in header.
T
tensor-tang 已提交
283
   */
284 285 286 287 288
  void setHeaderFormat(int32_t fmt) {
    CHECK(isHeaderFormatSupported(fmt)) << "Unsupported format version: "
                                        << fmt;
    headerFormat_ = fmt;
  }
T
tensor-tang 已提交
289

Z
zhangjinchao01 已提交
290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360
  /**
   * @brief  Parameter Update Hook.
   *
   * The parameter's update hook before ParameterUpdater::updateImpl
   * It could modify gradient/momentum/etc here. Such as drop some gradient,
   * etc.
   */
  void updateHook() {
    for (auto& hook : updaterHooks_) {
      hook->update(this);
    }
  }

  /**
   * @brief  Initialize all updater hook.
   *
   * This method should be invoked in ParameterUpdater::init() only.
   */
  void initHook() {
    for (auto& hook : updaterHooks_) {
      hook->init(this);
    }
  }

protected:
  /**
   * @brief create matrix to matType.
   *
   * used by gradient machine which needs specify matrix type,
   * instead of creating in weights.cpp.
   *
   * @note  pType should be enabled already.
   */
  void setMat(ParameterType pType, int matType);

  bool isUpdatable() { return (updateCounter_ == sharedCount_); }

  void clearUpdate() { updateCounter_ = 0; }

protected:
  ParameterConfig config_;

  bool useGpu_;

  int deviceId_;

  /**
   * @brief bufs_ stores parameter value and gradient.
   *
   * Layer should use bufs_[PARAMETER_VALUE] to form weight matrix for
   * calculation and stores gradient to bufs_[PARAMETER_GRADIENT].
   */
  VectorPtr bufs_[NUM_PARAMETER_TYPES];

  /**
   * @brief Weight matrix for bufs_.
   *
   * It's helpfull when parameter shared by multi-layers.
   * Caller should check, if mats exist, do not create it again.
   */
  MatrixPtr mats_[NUM_PARAMETER_TYPES];

  /// Int vectors, used in some User defined parameter types
  IVectorPtr intBufs_[NUM_PARAMETER_TYPES];

  int sharedCount_;
  int updateCounter_;

  bool updated_;
  SparseFormat format_;

T
tensor-tang 已提交
361
  /// The header format for saving or loading param
T
tensor-tang 已提交
362 363
  int32_t headerFormat_;

Z
zhangjinchao01 已提交
364 365 366
  std::vector<std::shared_ptr<IParameterUpdaterHook>> updaterHooks_;

public:
X
xuwei06 已提交
367
  void setSharedCount(int cnt) { sharedCount_ = cnt; }
Z
zhangjinchao01 已提交
368 369 370 371 372 373 374 375 376 377 378 379 380
  int getSharedCount() { return sharedCount_; }

  bool isSparse() { return config_.is_sparse(); }
  SparseFormat getFormat() { return format_; }

  static const std::string kMissParameterFail;
  static const std::string kMissParameterRand;
  static const std::string kMissParameterZero;
};

typedef std::map<std::string, ParameterPtr> ParameterMap;

}  // namespace paddle