Layer.cpp 12.5 KB
Newer Older
Z
zhangjinchao01 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
/* 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 "paddle/utils/Util.h"

#include "paddle/utils/Logging.h"
19
#include "paddle/math/SparseMatrix.h"
Z
zhangjinchao01 已提交
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293

#include "AddtoLayer.h"
#include "CosSimLayer.h"
#include "CostLayer.h"
#include "ExpandConvLayer.h"
#include "CRFLayer.h"
#include "DataLayer.h"
#include "FullyConnectedLayer.h"
#include "HierarchicalSigmoidLayer.h"
#include "MaxLayer.h"
#include "MixedLayer.h"
#include "NormLayer.h"
#include "PoolLayer.h"
#include "TensorLayer.h"
#include "TransLayer.h"
#include "ValidationLayer.h"

P_DEFINE_bool(log_error_clipping, false, "enable log error clipping or not");

namespace paddle {

Layer::Layer(const LayerConfig& config, bool useGpu)
    : config_(config),
      useGpu_(useGpu),
      deviceId_(-1),
      needSequenceInfo_(true) {}

bool Layer::init(const LayerMap& layerMap, const ParameterMap& parameterMap) {
  if (useGpu_ && FLAGS_parallel_nn) {
    /* gpu environment is specified by device property */
    deviceId_ = config_.device();
    if (deviceId_ < 0) {
      useGpu_ = false;
    }
  }

  output_.deviceId = deviceId_;

  for (auto& inputConfig : config_.inputs()) {
    std::string inputName = inputConfig.input_layer_name();
    LayerPtr inputLayer;
    CHECK(mapGet(inputName, layerMap, &inputLayer))
        << "Cannot find input layer " << inputName << " for layer "
        << getName();
    this->addPrev(inputLayer);

    inputLayer->addOutputArgument(deviceId_);

    if (inputConfig.has_input_parameter_name()) {
      ParameterPtr parameter;
      CHECK(
          mapGet(inputConfig.input_parameter_name(), parameterMap, &parameter))
          << "Cannot find input parameter "
          << inputConfig.input_parameter_name() << " for layer " << getName();
      parameter->incShared();
      CHECK_EQ(parameter->getDeviceId(), getDeviceId());
      parameters_.push_back(parameter);
    } else {
      parameters_.push_back(nullptr);
    }

    if (inputConfig.has_input_layer_argument()) {
      inputArgument_.push_back(inputConfig.input_layer_argument());
    } else {
      inputArgument_.push_back("");
    }
  }

  if (config_.has_bias_parameter_name()) {
    CHECK(mapGet(config_.bias_parameter_name(), parameterMap, &biasParameter_))
        << "Cannot find bias parameter " << config_.bias_parameter_name()
        << " for layer " << getName();
    biasParameter_->incShared();
    CHECK_EQ(biasParameter_->getDeviceId(), getDeviceId());
  }

  /* specify the activation function according to the configuration */
  std::string action_type = config_.active_type();
  activation_.reset(ActivationFunction::create(action_type));
  CHECK(activation_);

  initNeedFlags();
  markInBackward_.assign(inputLayers_.size(), false);

  return true;
}

ClassRegistrar<Layer, LayerConfig> Layer::registrar_;

LayerPtr Layer::create(const LayerConfig& config) {
  std::string type = config.type();

  if (type == "multi-class-cross-entropy")
    return LayerPtr(new MultiClassCrossEntropy(config));
  else if (type == "rank-cost")
    return LayerPtr(new RankingCost(config));
  else if (type == "auc-validation")
    return LayerPtr(new AucValidation(config));
  else if (type == "pnpair-validation")
    return LayerPtr(new PnpairValidation(config));
  // NOTE: stop adding "if" statements here.
  // Instead, use REGISTER_LAYER to add more layer types

  return LayerPtr(registrar_.createByType(config.type(), config));
}

void Layer::resetSpecifyOutput(Argument& output, size_t height, size_t width,
                               bool isValueClean, bool isGradClean) {
  SetDevice device(output.deviceId);

  Matrix::resizeOrCreate(output.value, height, width, /* trans */ false,
                         useGpu(output.deviceId));
  if (isValueClean) {
    output.value->zeroMem();
  }

  if (passType_ != PASS_TEST && needGradient()) {
    Matrix::resizeOrCreate(output.grad, height, width, /* trans */ false,
                           useGpu(output.deviceId));
    if (isGradClean) {
      output.grad->zeroMem();
    }
  }
}

void Layer::resizeOutput(size_t height, size_t width) {
  resetSpecifyOutput(output_, height, width, false, false);

  for (size_t i = 0; i != outputOtherDevice_.size(); i++) {
    resetSpecifyOutput(outputOtherDevice_[i], height, width, false, false);
  }
}

void Layer::reserveOutput(size_t height, size_t width) {
  resetSpecifyOutput(output_, height, width, false, true);

  for (size_t i = 0; i != outputOtherDevice_.size(); i++) {
    resetSpecifyOutput(outputOtherDevice_[i], height, width, false, true);
  }
}

void Layer::resetOutput(size_t height, size_t width) {
  resetSpecifyOutput(output_, height, width, true, true);

  for (size_t i = 0; i != outputOtherDevice_.size(); i++) {
    resetSpecifyOutput(outputOtherDevice_[i], height, width, true, true);
  }
}

void Layer::addOutputArgument(int deviceId) {
  if (deviceId == deviceId_) {
    output_.countIncrement();
    return;
  } else {
    for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
      if (outputOtherDevice_[i].deviceId == deviceId) {
        outputOtherDevice_[i].countIncrement();
        return;
      }
    }
  }

  Argument argu;
  argu.deviceId = deviceId;
  outputOtherDevice_.push_back(argu);
  outputOtherDevice_.back().countIncrement();
}

void Layer::copyOutputToOtherDevice() {
  for (size_t i = 0; i != outputOtherDevice_.size(); i++) {
    SetDevice device(outputOtherDevice_[i].deviceId);
    outputOtherDevice_[i].value->copyFrom(*getOutputValue(),
                                          HPPL_STREAM_DEFAULT);
    outputOtherDevice_[i].sequenceStartPositions =
        output_.sequenceStartPositions;
    outputOtherDevice_[i].subSequenceStartPositions =
        output_.subSequenceStartPositions;
    outputOtherDevice_[i].cpuSequenceDims = output_.cpuSequenceDims;

    outputOtherDevice_[i].notifyValueReady();
  }
}

void Layer::waitInputValue() {
  for (size_t i = 0; i != inputLayers_.size(); i++) {
    if (inputLayers_[i]->getDeviceId() != deviceId_) {
      getInput(i).waitValueReady();
    }
  }
}

void Layer::waitAndMergeOutputGrad() {
  if (!output_.grad || !outputOtherDevice_.size()) {
    return;
  }

  for (size_t i = 0; i != outputOtherDevice_.size(); i++) {
    outputOtherDevice_[i].waitGradReady();
  }

  /* merge output grad */
  size_t i = 0;
  if (!output_.getAllCount()) {
    output_.grad->copyFrom(*outputOtherDevice_[0].grad, HPPL_STREAM_1);
    hl_stream_synchronize(HPPL_STREAM_1);

    i++;
    if (outputOtherDevice_.size() == 1) return;
  }

  Matrix::resizeOrCreate(tmpGrad_, output_.grad->getHeight(),
                         output_.grad->getWidth(), /* trans */ false,
                         useGpu(output_.deviceId));

  for (; i != outputOtherDevice_.size(); i++) {
    tmpGrad_->copyFrom(*outputOtherDevice_[i].grad, HPPL_STREAM_1);
    hl_stream_synchronize(HPPL_STREAM_1);
    output_.grad->add(*tmpGrad_);
  }
}

void Layer::markAllInputGrad() {
  for (size_t i = 0; i != inputLayers_.size(); ++i) {
    if (!markInBackward_[i]) {
      inputLayers_[i]->getOutput(deviceId_).notifyGradReady();
    }
    markInBackward_[i] = false;
  }
}

void Layer::markInputGrad(int inputIndex) {
  inputLayers_[inputIndex]->getOutput(deviceId_).notifyGradReady();
  markInBackward_[inputIndex] = true;
}

void Layer::zeroGrad() {
  CHECK(output_.grad.get() != NULL);
  output_.grad->zeroMem();
}

void Layer::initNeedFlags() {
  auto initFlag = [this](bool& flag, bool (Layer::*flagQueryFunc)() const,
                         ParameterType type) {
    flag = false;
    if (biasParameter_ && biasParameter_->hasType(type)) {
      flag = true;
    }
    if (!flag) {
      for (auto& para : parameters_) {
        if (para && para->hasType(type)) {
          flag = true;
          break;
        }
      }
    }
    if (!flag) {
      for (auto& layer : inputLayers_) {
        if ((layer.get()->*flagQueryFunc)()) {
          flag = true;
        }
      }
    }
  };
  initFlag(needGradient_, &Layer::needGradient, PARAMETER_GRADIENT);
}

void Layer::showOutputStats() {
  MatrixPtr out = getOutputValue();
  if (!out) return;
  if (!out->getElementCnt()) {
    LOG(INFO) << "The number of output of " << config_.name()
              << " is 0, skip to show the statistics";
    return;
  }
294 295 296 297 298 299 300 301 302 303 304 305 306 307 308
  MatrixPtr outSquare;
  if (dynamic_cast<GpuSparseMatrix*>(out.get())) {
    GpuSparseMatrix *tmp = dynamic_cast<GpuSparseMatrix*>(out.get());
    outSquare = std::make_shared<CpuSparseMatrix>(
      tmp->getHeight(), tmp->getWidth(), tmp->getElementCnt(),
      tmp->getValueType(), tmp->getFormat());
  } else {
    outSquare = out->clone();
  }
  outSquare->copyFrom(*out, HPPL_STREAM_DEFAULT);
  hl_stream_synchronize(HPPL_STREAM_DEFAULT);

  real mean = outSquare->getSum() / out->getElementCnt();
  real min;
  real max;
Z
zhangjinchao01 已提交
309 310
  if (dynamic_cast<CpuSparseMatrix*>(outSquare.get())) {
    auto tmpMat = dynamic_cast<CpuSparseMatrix*>(outSquare.get());
311 312
    min = tmpMat->getMin();
    max = tmpMat->getMax();
H
hedaoyuan 已提交
313
    tmpMat->square2();
Z
zhangjinchao01 已提交
314 315
    LOG(INFO) << "show statistics of [none zero values] in sparse matrix";
  } else {
316 317
    min = outSquare->getMin();
    max = outSquare->getMax();
H
hedaoyuan 已提交
318
    outSquare->square2();
Z
zhangjinchao01 已提交
319 320 321 322 323 324 325
  }
  real std = (outSquare->getSum() / outSquare->getElementCnt()) - mean * mean;
  std = std > 0 ? std : 0;
  LOG(INFO) << "The output state of " << config_.name() << ": mean=" << mean
            << ", "
            << "std=" << std
            << ", "
326 327
            << "min=" << min << ", "
            << "max=" << max;
Z
zhangjinchao01 已提交
328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403
}

void Layer::forwardActivation() {
  /* activation */
  activation_->forward(output_);

  /* dropout */
  if (config_.drop_rate() > 0) {
    forwardDropOut();
    CHECK_NE(activation_->getName(), "softmax")
        << "Softmax activation cannot be used with Dropout";
  }

  if (FLAGS_show_layer_stat) {
    showOutputStats();
  }
}

void Layer::backwardActivation() {
  /* Do error clipping */
  if (config_.error_clipping_threshold() > 0.0f) {
    if (FLAGS_log_error_clipping) {
      CpuVector outGradVec(0, nullptr);
      outGradVec.subVecFrom(output_.grad->getData(), 0,
                            output_.grad->getElementCnt());
      real maxAbsGrad = outGradVec.getAbsMax();
      if (maxAbsGrad > config_.error_clipping_threshold()) {
        real avgAbsGrad = outGradVec.getAbsSum() / outGradVec.getSize();
        LOG(INFO) << " layer=" << config_.name() << " need clipping,"
                  << " max error=" << maxAbsGrad << " avg error=" << avgAbsGrad;
      }
    }
    output_.grad->clip(-config_.error_clipping_threshold(),
                       config_.error_clipping_threshold());
  }

  /* Do dropout for delta*/
  if (config_.drop_rate() > 0 && passType_ != PASS_TEST) {
    MatrixPtr oGrad = getOutputGrad();
    oGrad->dotMul(*oGrad, *dropOutMask_);
  }

  activation_->backward(output_);
}

void Layer::forwardDropOut() {
  auto& outV = getOutputValue();

  if (passType_ == PASS_TRAIN || passType_ == PASS_METRIC_TRAIN ||
      passType_ == PASS_METRIC_TRAIN_WITH_NOERROR) {
    // new dropOutMask_ if dropOutMask_ is null ptr
    Matrix::resizeOrCreate(dropOutMask_, outV->getHeight(), outV->getWidth(),
                           false, useGpu(deviceId_));
    dropOutMask_->randomizeUniform();  // generate a uniform random matrix
    dropOutMask_->biggerThanScalar(config_.drop_rate());  // random mask
    outV->dotMul(*outV, *dropOutMask_);                   // dropout
  } else if (passType_ == PASS_GC) {
    // only initialize once
    if (!dropOutMask_) {
      dropOutMask_ = Matrix::create(outV->getHeight(), outV->getWidth(), false,
                                    useGpu(deviceId_));
      // We use cpu matrix to generate mask so that the mask
      // will be same for both gpu version and cpu version.
      // This will help unittest to make sure they have same result.
      MatrixPtr tmpMask = Matrix::create(outV->getHeight(), outV->getWidth());
      tmpMask->randomizeUniform();  // generate a uniform random matrix
      tmpMask->biggerThanScalar(config_.drop_rate());  // random mask
      dropOutMask_->copyFrom(*tmpMask);
    }
    outV->dotMul(*outV, *dropOutMask_);
  } else {  // passType == PASS_TEST
    outV->mulScalar(1.0 - config_.drop_rate());
  }
}

}  // namespace paddle