Evaluator.h 8.7 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 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
/* 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. */


#pragma once

#include "paddle/pserver/ParameterClient2.h"
#include "paddle/utils/ClassRegistrar.h"
#include "ModelConfig.pb.h"
#include "paddle/parameter/Argument.h"
#include <fstream>

namespace paddle {

class NeuralNetwork;

#define REGISTER_EVALUATOR(__type_name, __class_name)                \
  static InitFunction __reg_type_##__type_name([]() {                \
    Evaluator::registrar_.registerClass<__class_name>(#__type_name); \
  })

class Evaluator {
public:
  static Evaluator* create(const EvaluatorConfig& config);

  Evaluator() : numSamples_(0), totalScore_(0) {}

  virtual ~Evaluator() {}

  virtual void init(const EvaluatorConfig& config) { config_ = config; }

  /**
   * start to evaluate some data
   */
  virtual void start() {
    numSamples_ = 0;
    totalScore_ = 0;
  }

  /**
   * Process a batch of data.
   */
  virtual void eval(const NeuralNetwork& nn);

  /**
   * Process a batch of data.
   * return the score for the batch if it make sense to sum the score across
   * batches. Otherwise evaluator should return 0 and override finish() and
   * printStats() to do the right calculation.
   */
  virtual real evalImp(std::vector<Argument>& arguments) = 0;

  /**
   * Update the number of processed samples
   */
  virtual void updateSamplesNum(const std::vector<Argument>& arguments) {
    numSamples_ += arguments[0].getBatchSize();
  }

71
  /// finish() should be called before distributeEval
Z
zhangjinchao01 已提交
72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87
  virtual void distributeEval(ParameterClient2* client) {
    LOG(FATAL) << "Not implemeted";
  }

  void mergeResultsOfAllClients(ParameterClient2* client) {
    double data[2] = {totalScore_, numSamples_};
    client->reduce(data, data, 2, FLAGS_trainer_id, 0);
    totalScore_ = data[0];
    numSamples_ = data[1];
  }

  /**
   * finish the evaluation.
   */
  virtual void finish() {}

88
  /// finish() should be called before printStats
Z
zhangjinchao01 已提交
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
  virtual void printStats(std::ostream& os) {
    os << config_.name() << "="
       << (numSamples_ ? totalScore_ / numSamples_ : 0);
  }

  friend std::ostream& operator<<(std::ostream& os,
                                  Evaluator& evaluator) {
    evaluator.printStats(os);
    return os;
  }

  friend std::ostream&& operator<<(std::ostream&& os,    // NOLINT
                                   Evaluator& evaluator) {
    evaluator.printStats(os);
    return std::move(os);
  }

  static ClassRegistrar<Evaluator> registrar_;

protected:
  EvaluatorConfig config_;
  double numSamples_;
  double totalScore_;
};

class DummyEvaluator : public Evaluator {
public:
  DummyEvaluator() {}
  virtual void init(const EvaluatorConfig&) {}
  virtual void start() {}
  virtual void eval(const NeuralNetwork&) {}
  virtual real evalImp(std::vector<Argument>& arguments) {
    (void)arguments;
    return -1;
  }
  virtual void finish() {}
  virtual void printStats(std::ostream&) {}
};

class AucEvaluator : public Evaluator {
public:
  /**
   * @brief evaluate AUC using colIdx-th column as prediction.
   *
133 134 135
   * - colIdx = 0: the 0-th column.
   * - colIdx > 0: the colIdx-th column.
   * - colIdx < 0: the last colIdx-th column.
Z
zhangjinchao01 已提交
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
   *
   */
  AucEvaluator(int32_t colIdx)
      : colIdx_(colIdx),
        realColumnIdx_(0),
        cpuOutput_(nullptr),
        cpuLabel_(nullptr),
        cpuWeight_(nullptr) {}

  virtual void start();

  virtual real evalImp(std::vector<Argument>& arguments);

  virtual void printStats(std::ostream& os) {
    os << config_.name() << "=" << calcAuc();
  }

  virtual void distributeEval(ParameterClient2* client);

private:
  static const uint32_t kBinNum_ = (1 << 24) - 1;
  static const int kNegativeLabel_ = 0;
  double statPos_[kBinNum_ + 1];
  double statNeg_[kBinNum_ + 1];
  int32_t colIdx_;
  uint32_t realColumnIdx_;
  MatrixPtr cpuOutput_;
  IVectorPtr cpuLabel_;
  MatrixPtr cpuWeight_;

  AucEvaluator() {}

  inline static double trapezoidArea(double X1, double X2, double Y1,
                                     double Y2) {
    return (X1 > X2 ? (X1 - X2) : (X2 - X1)) * (Y1 + Y2) / 2.0;
  }

  double calcAuc();
};

/**
 * @brief RankAucEvaluator calculates the AUC of each list
 * (i.e., titles under the same query), and averages them.
 *
 * Each list should be organized as a sequence.
 * The inputs of this evaluator is [output, click, pv].
 * If pv is not provided, it will be set to 1.
 * The types of click and pv are dense value.
 */
class RankAucEvaluator : public Evaluator {
public:
  // evaluate ranking AUC
  virtual void start();

  virtual void updateSamplesNum(const std::vector<Argument>& arguments);

  virtual real evalImp(std::vector<Argument>& arguments);

  virtual void distributeEval(ParameterClient2* client) {
    mergeResultsOfAllClients(client);
  }

private:
  MatrixPtr output_;
  MatrixPtr click_;
  MatrixPtr pv_;
  std::vector<std::pair<real, int>> outputPair_;

  double calcRankAuc(real* outputData, real* clickData, real* pvData,
                     size_t size);
};

class PrecisionRecallEvaluator : public Evaluator {
public:
  // Evaluate precision, recall and F1 score
  PrecisionRecallEvaluator()
      : isMultiBinaryLabel_(false),
        cpuOutput_(nullptr),
        cpuLabel_(nullptr),
        cpuWeight_(nullptr) {}

  virtual void start();

  virtual real evalImp(std::vector<Argument>& arguments);

  virtual void printStats(std::ostream& os);

  virtual void distributeEval(ParameterClient2* client);

  struct StatsInfo {
226 227 228 229 230 231 232 233
    /// numbers of true positives
    double TP;
    /// numbers of true negatives
    double TN;
    /// numbers of false positives
    double FP;
    /// numbers of false negatives
    double FN;
Z
zhangjinchao01 已提交
234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334

    StatsInfo() : TP(0.0), TN(0.0), FP(0.0), FN(0.0) {}
  };

private:
  bool isMultiBinaryLabel_;
  std::vector<StatsInfo> statsInfo_;

  MatrixPtr cpuOutput_;
  IVectorPtr cpuLabel_;
  MatrixPtr cpuWeight_;

  void calcStatsInfo(const MatrixPtr& output, const IVectorPtr& label,
                     const MatrixPtr& weight);

  void calcStatsInfoMulti(const MatrixPtr& output, const MatrixPtr& label,
                          const MatrixPtr& weight);

  inline static double calcPrecision(double TP, double FP) {
    if (TP > 0.0 || FP > 0.0) {
      return TP / (TP + FP);
    } else {
      return 1.0;
    }
  }

  inline static double calcRecall(double TP, double FN) {
    if (TP > 0.0 || FN > 0.0) {
      return TP / (TP + FN);
    } else {
      return 1.0;
    }
  }

  inline static double calcF1Score(double precision, double recall) {
    if (precision > 0.0 || recall > 0.0) {
      return 2 * precision * recall / (precision + recall);
    } else {
      return 0;
    }
  }
};

/**
 * Positive-negative pair rate Evaluator
 */
class PnpairEvaluator : public Evaluator {
public:
  PnpairEvaluator()
      : cpuOutput_(nullptr),
        cpuLabel_(nullptr),
        cpuInfo_(nullptr),
        cpuWeight_(nullptr) {}

  virtual void start();
  virtual real evalImp(std::vector<Argument>& arguments);

  struct PredictionResult {
    PredictionResult(real __out, int __label, int __queryid, real __weight)
        : out(__out), label(__label), queryid(__queryid), weight(__weight) {}
    real out;
    int label;
    int queryid;
    real weight;
  };
  std::vector<PredictionResult> predictArray_;
  void printPredictResults() {
    std::ofstream fs(FLAGS_predict_file);
    CHECK(fs) << "Fail to open " << FLAGS_predict_file;
    for (auto& res : predictArray_) {
      fs << res.out << " " << res.label << " " << res.queryid << std::endl;
    }
  }

  void stat(size_t start, size_t end, PredictionResult* answers, double& pos,
            double& neg, double& spe);
  void calc(std::vector<PredictionResult>& predictArray);

  virtual void finish() { calc(predictArray_); }

  virtual void printStats(std::ostream& os) {
    os << " pos/neg"
       << "=" << pairArray_[0] / ((pairArray_[1] <= 0) ? 1.0 : pairArray_[1]);
  }

  virtual void distributeEval(ParameterClient2* client) {
    client->reduce(pairArray_, pairArray_, kPairArrayNum_, FLAGS_trainer_id, 0);
    LOG(INFO) << " distribute eval calc total pos pair: " << pairArray_[0]
              << " calc total neg pair: " << pairArray_[1];
  }

private:
  static const uint32_t kPairArrayNum_ = 2;
  double pairArray_[kPairArrayNum_];
  MatrixPtr cpuOutput_;
  IVectorPtr cpuLabel_;
  IVectorPtr cpuInfo_;
  MatrixPtr cpuWeight_;
};

}  // namespace paddle