SparseRowMatrix.cpp 8.3 KB
Newer Older
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268
/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve.

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 "SparseRowMatrix.h"
#include "CpuSparseMatrix.h"

#include <cmath>
#include <algorithm>

#include "paddle/utils/Logging.h"

#include "SIMDFunctions.h"

#include "paddle/utils/Util.h"
#include "paddle/utils/Thread.h"

P_DEFINE_bool(allow_inefficient_sparse_update, false,
              "Whether to allow inefficient sparse update");

namespace paddle {

const unsigned int SparseRowCpuMatrix::kUnusedId_ = -1U;

void SparseRowCpuMatrix::init(size_t height, size_t width) {
  // @TODO(yuyang18) Just remove this limit
  CHECK(simd::vec_check(width)) << width;
  height_ = height;
  if (!indexDictHandle_) {
    indexDictHandle_.reset(new IndexDict);
    indexDictHandle_->globalIndices.assign(height, kUnusedId_);
  }
  localIndices_ = &indexDictHandle_->localIndices;
  globalIndices_ = indexDictHandle_->globalIndices.data();
}

void SparseRowCpuMatrix::mul(CpuSparseMatrix* a, CpuMatrix* b, real scaleAB,
                             real scaleT) {
  CpuMatrix::mul<CpuMatrix, SparseRowCpuMatrix>(a, b, this, scaleAB, scaleT);
}

void SparseRowCpuMatrix::copyFrom(const real* src, size_t size) {
  LOG(FATAL) << "This should not be called";
}

void SparseRowCpuMatrix::zeroMem() {
  apply(
    [](real* buf, size_t len) {
      memset(buf, 0, sizeof(real) * len);
    });
  clearRows();
}

void SparseRowCpuMatrix::applyL1Decay(real learningRate, real decayRate) {
  apply([=](real* buf, size_t len) {
      CpuVector value(0, nullptr);
      value.subVecFrom(buf, 0, len);
      value.applyL1(learningRate, decayRate);
    });
}

void SparseRowCpuMatrix::sgdUpdate(BaseMatrix& value, IVector& t0,
                                   real learningRate, int currentTime,
                                   real decayRate, bool useL1, bool fini) {
  std::vector<unsigned int>& localIndices = indexDictHandle_->localIndices;

  // t0 and value are vectors
  CHECK_EQ(t0.getSize(), this->height_);
  CHECK_EQ(value.width_, this->height_ * this->width_);

  if (decayRate == 0.0f) {
    if (fini) {
      return;
    }

    for (size_t i = 0; i < localIndices.size(); ++i) {
      real* g = getLocalRow(i);
      real* v = value.rowBuf(localIndices[i]);
      for (size_t j = 0; j < this->width_; ++j) {
        v[j] -= learningRate * g[j];
      }
    }
    return;
  }  // else

  if (useL1) {  // L1 decay
    if (fini) {
      for (size_t i = 0; i < this->height_; ++i) {
        real* v = value.rowBuf(i);
        int* t = t0.getData() + i;
        if (t[0] < currentTime) {
          // W(t0) -> W(t+1)
          int tDiff = currentTime - t[0];
          real delta = tDiff * learningRate * decayRate;
          simd::decayL1(v, v, delta, this->width_);
        }
      }
      return;
    }  // else

    for (size_t i = 0; i < localIndices.size(); ++i) {
      real* g = getLocalRow(i);
      real* v = value.rowBuf(localIndices[i]);
      int* t = t0.getData() + localIndices[i];
      if (t[0] < currentTime) {
        // W(t0) -> W(t)
        int tDiff = currentTime - t[0];
        real delta = tDiff * learningRate * decayRate;
        simd::decayL1(v, v, delta, this->width_);
      }

      // W(t) -> W(t+1)
      for (size_t j = 0; j < this->width_; ++j) {
        v[j] -= learningRate * g[j];
      }
      simd::decayL1(v, v, learningRate*decayRate, this->width_);

      // state update to t+1
      t[0] = currentTime + 1;
    }

  } else {  // L2 decay
    if (fini) {
      for (size_t i = 0; i < this->height_; ++i) {
        real* v = value.rowBuf(i);
        int* t = t0.getData() + i;
        if (t[0] < currentTime) {
          // W(t0) -> W(t+1)
          int tDiff = currentTime - t[0];
          real recip = 1.0f / (1.0f + tDiff * learningRate * decayRate);
          for (size_t j = 0; j < this->width_; ++j) {
            v[j] *= recip;
          }
        }
      }
      return;
    }  // else

    real recipDecay = 1.0f / (1.0f + learningRate * decayRate);

    for (size_t i = 0; i < localIndices.size(); ++i) {
      real* g = getLocalRow(i);
      real* v = value.rowBuf(localIndices[i]);
      int* t = t0.getData() + localIndices[i];
      if (t[0] < currentTime) {
        // W(t0) -> W(t)
        int tDiff = currentTime - t[0];
        real recip = 1.0f / (1.0f + tDiff * learningRate * decayRate);
        for (size_t j = 0; j < this->width_; ++j) {
          v[j] *= recip;
        }
      }

      // W(t) -> W(t+1)
      for (size_t j = 0; j < this->width_; ++j) {
        v[j] = recipDecay * (v[j] - learningRate * g[j]);
      }

      // state update to t+1
      t[0] = currentTime + 1;
    }
  }
}

void SparseRowCpuMatrix::addTo(BaseMatrix& dest, std::vector<uint32_t>& ids,
                               size_t tid, size_t numThreads) {
  CHECK(!dest.useGpu_);
  CHECK_EQ(dest.height_ * dest.width_, this->height_ * this->width_);

  std::vector<unsigned int>& localIndices = indexDictHandle_->localIndices;
  for (size_t i = 0; i < localIndices.size(); ++i) {
    uint32_t id = localIndices[i];
    if (id % numThreads == tid) {
      simd::addTo(dest.rowBuf(id), getLocalRow(i),
                  this->width_);
      ids.push_back(id);
    }
  }
}

void SparseRowCpuMatrix::addTo(SparseRowCpuMatrix& dest, size_t tid,
                               size_t numThreads) {
  CHECK(!dest.useGpu_);
  CHECK_EQ(dest.height_ * dest.width_, this->height_ * this->width_);

  std::vector<unsigned int>& localIndices = indexDictHandle_->localIndices;
  for (size_t i = 0; i < localIndices.size(); ++i) {
    uint32_t id = localIndices[i];
    if (id % numThreads == tid) {
      dest.checkIndex(id);
      simd::addTo(dest.getRow(id), getLocalRow(i), this->width_);
    }
  }
}

void SparseRowCpuMatrix::zeroMemThread(size_t tid, size_t numThreads) {
  std::vector<unsigned int>& localIndices = indexDictHandle_->localIndices;
  for (size_t i = 0; i < localIndices.size(); ++i) {
    uint32_t id = localIndices[i];
    if (id % numThreads == tid) {
      memset(this->getLocalRow(i), 0, this->width_ * sizeof(real));
    }
  }
}

void SparseAutoGrowRowCpuMatrix::mul(CpuSparseMatrix* a, CpuMatrix* b,
                                     real scaleAB, real scaleT) {
  CpuMatrix::mul<CpuMatrix, SparseAutoGrowRowCpuMatrix>(a, b, this, scaleAB,
                                                        scaleT);
}

void CacheRowCpuMatrix::mul(CpuSparseMatrix* a, CpuMatrix* b, real scaleAB,
                            real scaleT) {
  CpuMatrix::mul<CpuMatrix, CacheRowCpuMatrix>(a, b, this, scaleAB, scaleT);
}

void SparsePrefetchRowCpuMatrix::addRows(const unsigned int* ids, size_t len) {
  std::vector<unsigned int>& localIndices = indexDictHandle_->localIndices;
  localIndices.insert(localIndices.end(), ids, ids + len);
}

void SparsePrefetchRowCpuMatrix::addRows(MatrixPtr input) {
  CpuSparseMatrix* mat = dynamic_cast<CpuSparseMatrix*>(input.get());
  CHECK(mat) << "only support non value sparse matrix";
  addRows(reinterpret_cast<const unsigned int*>(mat->getCols()),
          mat->getElementCnt());
}

void SparsePrefetchRowCpuMatrix::addRows(IVectorPtr ids) {
  std::vector<unsigned int>& localIndices = indexDictHandle_->localIndices;
  size_t numSamples = ids->getSize();
  int* index = ids->getData();
  for (size_t i = 0; i < numSamples; ++i) {
    if (index[i] == -1) continue;
    localIndices.push_back((unsigned int)index[i]);
  }
}

void SparsePrefetchRowCpuMatrix::setupIndices() {
  auto& localIndices = indexDictHandle_->localIndices;
  uniqueIds(localIndices);
  // for each sparse row
  for (size_t id = 0; id < localIndices.size(); ++id) {
    globalIndices_[localIndices[id]] = id;  // sparse row -> local id
  }
  checkStoreSize();
}

void SparseRowCpuMatrix::checkIndices() {
  std::vector<unsigned int>& localIndices = indexDictHandle_->localIndices;
  for (size_t i = 0; i < localIndices.size(); ++i) {
    CHECK_EQ(globalIndices_[localIndices[i]], i);
  }
  checkStoreSize();
}

}  // namespace paddle
新手
引导
客服 返回
顶部