NeuralNetwork.cpp 14.6 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17

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/CustomStackTrace.h"
Y
Yu Yang 已提交
18
#include "paddle/utils/Logging.h"
19

Y
Yu Yang 已提交
20
#include "MultiNetwork.h"
21 22
#include "NeuralNetwork.h"
#include "RecurrentGradientMachine.h"
Y
Yu Yang 已提交
23
#include "hl_gpu.h"
24
#include "paddle/gserver/layers/AgentLayer.h"
Y
Yu Yang 已提交
25
#include "paddle/utils/Stat.h"
26 27

namespace paddle {
28 29
void parameterInitNN(int paramId,
                     Parameter* para,
30 31 32 33 34 35 36 37
                     std::vector<ParameterPtr>* sharedParams) {
  // Create parameters values.
  if (!para->useGpu() && sharedParams) {
    para->enableSharedType(PARAMETER_VALUE,
                           (*sharedParams)[paramId]->getBuf(PARAMETER_VALUE),
                           (*sharedParams)[paramId]->getMat(PARAMETER_VALUE));
  } else {
    if (para->isSparseRemoteUpdate()) {
38 39 40 41
      para->enableType(PARAMETER_VALUE,
                       FLAGS_loadsave_parameters_in_pserver
                           ? Parameter::MAT_SPARSE_ROW_PREFETCH
                           : Parameter::MAT_SPARSE_ROW_PREFETCH_FULL_SIZE);
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
    } else {
      para->enableType(PARAMETER_VALUE);
    }
  }
  // Create parameter gradients.
  if (para->isSparseRemoteUpdate() && !sharedParams) {
    para->enableType(PARAMETER_GRADIENT, Parameter::MAT_SPARSE_ROW);
  } else if (para->isGradSparseUpdate()) {
    para->enableType(PARAMETER_GRADIENT, Parameter::MAT_SPARSE_ROW_AUTO_GROW);
  } else if (!para->isStatic()) {
    para->enableType(PARAMETER_GRADIENT);
  }
}

NeuralNetwork* NeuralNetwork::create(const ModelConfig& config) {
  if (config.type() == "recurrent_nn") {
    return newNeuralNetwork("root");
  } else if (config.type() == "multi_nn") {
    return new MultiNetwork("root");
  } else {
    return newNeuralNetwork();
  }
}

std::map<std::string, bool> NeuralNetwork::dllInitMap;

68 69
void NeuralNetwork::init(const ModelConfig& config,
                         ParamInitCallback callback,
70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92
                         const std::vector<ParameterType>& parameterTypes,
                         bool useGpu) {
  using std::placeholders::_1;
  using std::placeholders::_2;
  ParamInitCallback paramCallback = nullptr;
  if (callback != nullptr) {
    paramSelfInited_ = false;
    paramCallback = callback;
  } else {
    paramSelfInited_ = true;
    paramCallback = std::bind(parameterInitNN, _1, _2, nullptr);
  }
  config_ = config;

  if (rootNetwork_ != nullptr) {
    // direct use parameters_ and parameterMap_ from base network
    CHECK_EQ((size_t)config.parameters_size(),
             rootNetwork_->getParameters().size());
    parameters_ = rootNetwork_->getParameters();
    parameterMap_ = *(rootNetwork_->getParameterMap());
  } else {
    parameters_.reserve(config.parameters_size());
    for (const auto& para_config : config.parameters()) {
93 94
      auto parameter = std::make_shared<Parameter>(para_config,
                                                   useGpu,
95 96 97 98
                                                   /*initialize=*/false);
      paramCallback(parameters_.size(), parameter.get());
      if (!callback) {
        for (ParameterType type :
99
             (parameter->isStatic()
100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121
                  ? std::vector<ParameterType>{PARAMETER_VALUE}
                  : parameterTypes)) {
          if (type != PARAMETER_VALUE && type != PARAMETER_GRADIENT) {
            parameter->enableType(type);
          }
        }
      }
      parameter->setID(parameters_.size());
      parameters_.push_back(parameter);
      CHECK(!parameterMap_.count(parameter->getName()));
      parameterMap_[parameter->getName()] = parameter;
    }
  }

  auto layerCreate = [&](const LayerConfig& layer_config) {
    auto layer = Layer::create(layer_config);
    CHECK(layer) << "Create layer failed. Layer name:" << layer->getName();
    layers_.push_back(layer);
    CHECK(!layerMap_.count(layer->getName()));
    layerMap_[layer->getName()] = layer;
  };

122 123 124 125 126
  auto subModelConfig = std::find_if(config.sub_models().begin(),
                                     config.sub_models().end(),
                                     [=](const SubModelConfig& sub_model) {
                                       return sub_model.name() == subModelName_;
                                     });
127 128 129 130 131 132
  bool useSubModel = (subModelConfig != config.sub_models().end());
  CHECK_EQ(useSubModel, !subModelName_.empty());
  if (useSubModel) {
    layers_.reserve(subModelConfig->layer_names_size());
    for (const auto& layer_name : subModelConfig->layer_names()) {
      auto layer_config =
133 134
          std::find_if(config.layers().begin(),
                       config.layers().end(),
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
                       [=](const LayerConfig& layer_config) {
                         return layer_config.name() == layer_name;
                       });
      CHECK(layer_config != config.layers().end());
      layerCreate(*layer_config);
    }
  } else {
    layers_.reserve(config.layers_size());
    for (const auto& layer_config : config.layers()) {
      bool useLayer = true;
      if (config.has_external_config()) {
        useLayer = true;
        for (const auto& name : config.external_config().layer_names()) {
          if (layer_config.name() == name) {
            useLayer = false;
            break;
          }
        }
      }
      if (useLayer) {
        layerCreate(layer_config);
      }
    }
  }

  for (const auto& layer : layers_) {
    layer->init(layerMap_, parameterMap_);
    layer->initSubNetwork(this /*root*/, config_, parameterTypes, useGpu);
  }

  for (const auto& layer_name :
       (useSubModel ? subModelConfig->input_layer_names()
                    : config.input_layer_names())) {
    auto it = layerMap_.find(layer_name);
    CHECK(it != layerMap_.end());
    dataLayers_.push_back(std::dynamic_pointer_cast<DataLayer>(it->second));
  }

  for (const auto& layer_name :
       (useSubModel ? subModelConfig->output_layer_names()
                    : config.output_layer_names())) {
    auto it = layerMap_.find(layer_name);
    CHECK(it != layerMap_.end());
    outputLayers_.push_back(it->second);
  }
}

182 183
void NeuralNetwork::connect(LayerPtr agentLayer,
                            LayerPtr realLayer,
184 185 186 187 188 189
                            int height) {
  AgentLayer* agent = dynamic_cast<AgentLayer*>(agentLayer.get());
  CHECK_NOTNULL(agent);
  agent->setRealLayer(realLayer, height);
}

190 191
void NeuralNetwork::connect(std::string agentLayerName,
                            NeuralNetwork* srcNN,
192 193 194 195 196 197 198 199 200 201 202
                            std::string realLayerName) {
  connect(this->getLayer(agentLayerName), srcNN->getLayer(realLayerName));
}

void NeuralNetwork::prefetch(const std::vector<Argument>& inArgs) {
  CHECK_EQ(inArgs.size(), dataLayers_.size());

  if (paramSelfInited_) {
    for (auto& para : parameters_) {
      if (para->isSparseRemoteUpdate()) {
        auto mat = dynamic_cast<SparsePrefetchRowCpuMatrix*>(
203
            para->getMat(PARAMETER_VALUE).get());
204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224
        para->clearGradient();
        mat->clearIndices();
      }
    }
  }

  for (size_t i = 0; i != dataLayers_.size(); ++i) {
    if (FLAGS_parallel_nn) {
      const_cast<Argument&>(inArgs[i]).deviceId = -1;
    }
    dataLayers_[i]->setData(inArgs[i]);
  }

  for (auto& layer : layers_) {
    layer->prefetch();
  }

  if (paramSelfInited_) {
    for (auto& para : parameters_) {
      if (para->isSparseRemoteUpdate()) {
        auto mat = dynamic_cast<SparsePrefetchRowCpuMatrix*>(
225
            para->getMat(PARAMETER_VALUE).get());
226 227
        mat->setupIndices();
        auto matGrad = dynamic_cast<SparseRowCpuMatrix*>(
228
            para->getMat(PARAMETER_GRADIENT).get());
229 230 231 232 233 234 235
        matGrad->reserveStore();
      }
    }
  }
}

void NeuralNetwork::forward(const std::vector<Argument>& inArgs,
236 237
                            std::vector<Argument>* outArgs,
                            PassType passType) {
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
  CHECK_EQ(inArgs.size(), dataLayers_.size());
  outArgs->resize(outputLayers_.size());
  for (size_t i = 0; i != dataLayers_.size(); ++i) {
    dataLayers_[i]->setData(inArgs[i]);
  }

  {
    for (auto& layer : layers_) {
      REGISTER_TIMER_INFO("ForwardTimer", layer->getName().c_str());
      gLayerStackTrace.push(layer->getName());
      layer->forward(passType);
    }
  }

  outArgs->clear();
  outArgs->reserve(outputLayers_.size());
  for (auto& layer : outputLayers_) {
    outArgs->push_back(layer->getOutput());
  }
  if (passType == PASS_TEST) {
    gLayerStackTrace.clear();
  }
}

void NeuralNetwork::resetState() {
  for (auto& layer : layers_) {
    layer->resetState();
  }
}

void NeuralNetwork::setState(const MachineState& machineState) {
  for (size_t i = 0; i < layers_.size(); i++) {
    if (machineState[i] != nullptr) {
      layers_[i]->setState(machineState[i]);
    }
  }
}

void NeuralNetwork::getState(MachineState& machineState) {
  machineState.clear();
  machineState.reserve(layers_.size());
  for (auto& layer : layers_) {
    LayerStatePtr p = layer->getState();
    machineState.push_back(p);
  }
}

void NeuralNetwork::backward(const UpdateCallback& callback) {
Y
Yu Yang 已提交
286
  gLayerStackTrace.pop("");  // tell layer trace is during backward.
287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332
  FOR_EACH_R(layer, layers_) {
    REGISTER_TIMER_INFO("BackwardTimer", (*layer)->getName().c_str());
    if ((*layer)->needGradient()) {
      (*layer)->backward(callback);
    }
    gLayerStackTrace.pop((*layer)->getName());
  }
}

MatrixPtr NeuralNetwork::getLayerOutput(const std::string& layerName) {
  auto it = layerMap_.find(layerName);
  CHECK(it != layerMap_.end()) << "Cannot find layer: " << layerName;
  return it->second->getOutputValue();
}
void NeuralNetwork::onPassEnd() {
  for (auto& layer : layers_) {
    layer->onPassEnd();
  }
}

class CombinedEvaluator : public Evaluator {
public:
  void addEvaluator(std::unique_ptr<Evaluator>&& evaluator) {
    evaluators_.emplace_back(std::move(evaluator));
  }
  virtual void start() {
    for (auto& evaluator : evaluators_) {
      evaluator->start();
    }
  }

  virtual void finish() {
    for (auto& evaluator : evaluators_) {
      evaluator->finish();
    }
  }

  virtual void eval(const NeuralNetwork& nn) {
    for (auto& evaluator : evaluators_) {
      evaluator->eval(nn);
    }
  }
  virtual real evalImp(std::vector<Argument>& arguments) {
    (void)arguments;
    return -1;
  }
Y
Yu Yang 已提交
333
  virtual void printStats(std::ostream& os) const {
334 335 336 337 338 339 340 341 342 343 344 345 346 347
    for (auto& evaluator : evaluators_) {
      evaluator->printStats(os);
      os << ' ';
    }
  }

  virtual void distributeEval(ParameterClient2* client) {
    for (auto& evaluator : evaluators_) {
      evaluator->distributeEval(client);
    }
  }

protected:
  std::vector<std::unique_ptr<Evaluator>> evaluators_;
Y
Yu Yang 已提交
348 349 350

  // Evaluator interface
public:
Y
Yu Yang 已提交
351 352 353 354
  /**
   * @brief getNames will return all inside evaluators' names.
   * @param names [out]: return names.
   */
Y
Yu Yang 已提交
355 356 357 358 359 360
  void getNames(std::vector<std::string>* names) {
    for (auto& eval : evaluators_) {
      eval->getNames(names);
    }
  }

Y
Yu Yang 已提交
361 362 363
  /**
   * @brief getValue could get all inside evaluators' value.
   */
Y
Yu Yang 已提交
364 365 366 367 368 369
  real getValue(const std::string& name, Error* err) const {
    return this->getMethodHelper<real>(
        name, err, [&name, err](const std::unique_ptr<Evaluator>& eval) {
          return eval->getValue(name, err);
        });
  }
Y
Yu Yang 已提交
370 371 372 373

  /**
   * @brief getType could get all inside evaluators' type.
   */
Y
Yu Yang 已提交
374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393
  std::string getType(const std::string& name, Error* err) const {
    return this->getMethodHelper<std::string>(
        name, err, [&name, err](const std::unique_ptr<Evaluator>& eval) {
          return eval->getType(name, err);
        });
  }

private:
  template <typename T>
  T getMethodHelper(const std::string& name,
                    Error* err,
                    const std::function<T(const std::unique_ptr<Evaluator>&)>&
                        callback) const {
    for (auto& eval : evaluators_) {
      std::vector<std::string> names;
      eval->getNames(&names);
      if (std::find(names.begin(), names.end(), name) != names.end()) {
        return callback(eval);
      }
    }
394
    *err = Error("No such key %s", name.c_str());
Y
Yu Yang 已提交
395 396
    return T();
  }
397 398
};

399
Evaluator* NeuralNetwork::makeEvaluator() const {
400
  CombinedEvaluator* combinedEvaluator = new CombinedEvaluator();
401 402 403 404 405
  auto subModelConfig = std::find_if(config_.sub_models().begin(),
                                     config_.sub_models().end(),
                                     [=](const SubModelConfig& sub_model) {
                                       return sub_model.name() == subModelName_;
                                     });
406 407 408 409 410 411 412
  bool useSubModel = (subModelConfig != config_.sub_models().end());
  CHECK_EQ(useSubModel, !subModelName_.empty());
  if (useSubModel) {
    // create the evaluators that belong to CURRENT submodel
    for (int i = 0; i < subModelConfig->evaluator_names_size(); ++i) {
      // find evaluator by name
      auto thisEvalConfig = std::find_if(
413 414
          config_.evaluators().begin(),
          config_.evaluators().end(),
415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433
          [=](const EvaluatorConfig& ecfg) {
            return ecfg.name() == subModelConfig->evaluator_names(i);
          });
      bool validConfig = (thisEvalConfig != config_.evaluators().end());
      if (validConfig) {
        std::unique_ptr<Evaluator> evaluator(
            Evaluator::create(*thisEvalConfig));
        combinedEvaluator->addEvaluator(std::move(evaluator));
      }
    }
  } else {
    for (const EvaluatorConfig& evalConfig : config_.evaluators()) {
      std::unique_ptr<Evaluator> evaluator(Evaluator::create(evalConfig));
      combinedEvaluator->addEvaluator(std::move(evaluator));
    }
  }
  return combinedEvaluator;
}

434
void NeuralNetwork::eval(Evaluator* evaluator) const { evaluator->eval(*this); }
435 436 437 438 439 440 441 442

void NeuralNetwork::setOutputGrad(const std::vector<Argument>& args) {
  CHECK_GE(outputLayers_.size(), args.size());
  for (size_t i = 0; i < args.size(); ++i) {
    outputLayers_[i]->getOutput().grad = args[i].grad;
  }
}

443 444 445
extern NeuralNetwork* newCustomNerualNetwork(const std::string& name,
                                             NeuralNetwork* network)
    __attribute__((weak));
L
liaogang 已提交
446

447 448 449 450 451 452 453
NeuralNetwork* NeuralNetwork::newNeuralNetwork(const std::string& name,
                                               NeuralNetwork* rootNetwork) {
  if (newCustomNerualNetwork) {
    return newCustomNerualNetwork(name, rootNetwork);
  } else {
    return new NeuralNetwork(name, rootNetwork);
  }
454 455 456
}

}  // namespace paddle
反馈
建议
客服 返回
顶部