AverageOptimizer.cpp 6.7 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Z
zhangjinchao01 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

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 "AverageOptimizer.h"

namespace paddle {

// factory method to create an instance of AverageOptimizer
ParameterOptimizer* AverageOptimizer::create(
21 22 23 24
    const OptimizationConfig& optConfig,
    ParameterOptimizer* optimizer,
    bool isParameterSparse,
    bool useParameterApply) {
Z
zhangjinchao01 已提交
25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47
  if (optConfig.average_window() <= 0) {
    return optimizer;
  }
  // disable average for embeded local updater
  if (!useParameterApply && optConfig.num_batches_per_send_parameter() > 1) {
    return optimizer;
  }
  if (isParameterSparse) {
    return new AverageSparseOptimizer(optConfig, optimizer, useParameterApply);
  }
  return new AverageOptimizer(optConfig, optimizer, useParameterApply);
}

AverageOptimizer::AverageOptimizer(const OptimizationConfig& optConfig,
                                   ParameterOptimizer* optimizer,
                                   bool useParameterApply)
    : ParameterOptimizer(optConfig),
      optimizer_(optimizer),
      useApply_(useParameterApply),
      numUpdates_(0),
      prevNumUpdates_(0),
      numAccumulates_(0),
      oldNumAccumulates_(0),
48 49
      minAverageWindow_(
          std::min<int64_t>(10000L, optConfig_.max_average_window())),
Z
zhangjinchao01 已提交
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
      maxAverageWindow_(optConfig_.max_average_window()) {
  parameterTypes_ = optimizer_->getParameterTypes();
  addParameterType(PARAMETER_SUM1);
  addParameterType(PARAMETER_SUM2);
  addParameterType(PARAMETER_SUM3);
  if (useParameterApply) {
    addParameterType(PARAMETER_APPLY);
  }
}

void AverageOptimizer::startBatch(int64_t numSamplesProcessed) {
  optimizer_->startBatch(numSamplesProcessed);
  learningRate_ = optimizer_->getLearningRate();

  ++numUpdates_;
  ++numAccumulates_;
}

/*
  After traversal, the averaged parameter can be obtained by
  ((PARAMETER_SUM1 + PARAMETER_SUM2 + PARAMETER_SUM3)
  / (numAccumulates_ + oldNumAccumulates_))
*/
ParameterOptimizer::TraverseCallback AverageOptimizer::needSpecialTraversal(
    const ParameterConfig& config) const {
  TraverseCallbackVec callbacks;

  if (auto callback = optimizer_->needSpecialTraversal(config)) {
    callbacks.emplace_back(callback);
  }

  if (numUpdates_ % kMaxNumAccumulates == 0) {
    // Move the sum to a different buffer to avoid loss of precision
    // due to too many sums.
F
fengjiayi 已提交
84 85 86
    callbacks.emplace_back([](const VectorPtr vecs[],
                              const ParameterConfig& config,
                              size_t sparseId) {
Z
zhangjinchao01 已提交
87 88 89 90 91 92 93 94 95 96
      vecs[PARAMETER_SUM2]->add(*vecs[PARAMETER_SUM1]);
      vecs[PARAMETER_SUM1]->zeroMem();
    });
  }

  if (isAverageWindowTooLong()) {
    // Now the average window is too long, discard the old sum.
    if (auto callback = this->startCatchUpWith()) {
      callbacks.emplace_back(callback);
    }
F
fengjiayi 已提交
97 98 99
    callbacks.emplace_back([](const VectorPtr vecs[],
                              const ParameterConfig& config,
                              size_t sparseId) {
Z
zhangjinchao01 已提交
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
      vecs[PARAMETER_SUM3]->add(*vecs[PARAMETER_SUM1], *vecs[PARAMETER_SUM2]);
      vecs[PARAMETER_SUM1]->zeroMem();
      vecs[PARAMETER_SUM2]->zeroMem();
    });
  }

  return composeCallbacks(callbacks);
}

void AverageOptimizer::finishBatch() {
  optimizer_->finishBatch();
  if (isAverageWindowTooLong()) {
    this->finishCatchUpWith();
    oldNumAccumulates_ = numAccumulates_;
    numAccumulates_ = 0;
  }
}

ParameterOptimizer::TraverseCallback AverageOptimizer::apply() {
  if (numAccumulates_ + oldNumAccumulates_ == 0) {
    return nullptr;
  }

  real scale = 1. / (numAccumulates_ + oldNumAccumulates_);
  if (useApply_) {
125 126
    return [scale](const VectorPtr vecs[],
                   const ParameterConfig& config,
Z
zhangjinchao01 已提交
127
                   size_t sparseId) {
128 129 130 131 132 133
      vecs[PARAMETER_APPLY]->add3(*vecs[PARAMETER_SUM1],
                                  *vecs[PARAMETER_SUM2],
                                  *vecs[PARAMETER_SUM3],
                                  scale,
                                  scale,
                                  scale);
Z
zhangjinchao01 已提交
134 135
    };
  } else {
136 137
    return [scale](const VectorPtr vecs[],
                   const ParameterConfig& config,
Z
zhangjinchao01 已提交
138 139
                   size_t sparseId) {
      vecs[PARAMETER_GRADIENT]->copyFrom(*vecs[PARAMETER_VALUE]);
140 141 142 143 144 145
      vecs[PARAMETER_VALUE]->add3(*vecs[PARAMETER_SUM1],
                                  *vecs[PARAMETER_SUM2],
                                  *vecs[PARAMETER_SUM3],
                                  scale,
                                  scale,
                                  scale);
Z
zhangjinchao01 已提交
146 147 148 149 150 151 152 153 154 155 156 157
    };
  }
}

ParameterOptimizer::TraverseCallback AverageOptimizer::restore() {
  if (numAccumulates_ + oldNumAccumulates_ == 0) {
    return nullptr;
  }
  if (useApply_) {
    return nullptr;
  }

158 159
  return [](
      const VectorPtr vecs[], const ParameterConfig& config, size_t sparseId) {
Z
zhangjinchao01 已提交
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
    vecs[PARAMETER_VALUE]->copyFrom(*vecs[PARAMETER_GRADIENT]);
    vecs[PARAMETER_GRADIENT]->zeroMem();
  };
}

void AverageSparseOptimizer::update(const VectorPtr vecs[],
                                    const ParameterConfig& paraConfig,
                                    size_t sparseId) const {
  optimizer_->update(vecs, paraConfig, sparseId);

  CHECK_LT(sparseId, t0Vec_.size());
  int timediff = timer_ + 1 - t0Vec_[sparseId];
  if (timediff > 0) {
    vecs[PARAMETER_SUM1]->add(*vecs[PARAMETER_VALUE], timediff);
    t0Vec_[sparseId] = timer_ + 1;
  }
}

ParameterOptimizer::TraverseCallback AverageSparseOptimizer::startCatchUpWith()
    const {
  TraverseCallbackVec callbacks;

  if (auto callback = optimizer_->startCatchUpWith()) {
    callbacks.emplace_back(callback);
  }

  if (timer_ > 0) {
    callbacks.emplace_back(
188 189
        [this](const VectorPtr vecs[],
               const ParameterConfig& config,
Z
zhangjinchao01 已提交
190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206
               size_t sparseId) { this->catchUpWith(vecs, config, sparseId); });
  }

  return composeCallbacks(callbacks);
}

void AverageSparseOptimizer::catchUpWith(const VectorPtr vecs[],
                                         const ParameterConfig& paraConfig,
                                         size_t sparseId) const {
  CHECK_LT(sparseId, t0Vec_.size());
  int timediff = timer_ - t0Vec_[sparseId];
  if (timediff > 0) {
    vecs[PARAMETER_SUM1]->add(*vecs[PARAMETER_VALUE], timediff);
  }
}

}  // namespace paddle