词向量源码解析：（6.5）fasttext源码解析之文本分类4

下面说一下测试和预测，测试和预测分别调用了test和predict函数

int main(int argc, char** argv) {
  if (argc < 2) {
    printUsage();
    exit(EXIT_FAILURE);
  }
  std::string command(argv[1]);
  if (command == "skipgram" || command == "cbow" || command == "supervised") {
    train(argc, argv);
  } else if (command == "test") {
    test(argc, argv);
  } else if (command == "quantize") {
    quantize(argc, argv);
  } else if (command == "print-word-vectors") {
    printWordVectors(argc, argv);
  } else if (command == "print-sentence-vectors") {
    printSentenceVectors(argc, argv);
  } else if (command == "print-ngrams") {
    printNgrams(argc, argv);
  } else if (command == "nn") {
    nn(argc, argv);
  } else if (command == "analogies") {
    analogies(argc, argv);
  } else if (command == "predict" || command == "predict-prob" ) {
    predict(argc, argv);
  } else {
    printUsage();
    exit(EXIT_FAILURE);
  }
  return 0;
}

下面是test函数，先读取模型，然后实例化fasttext进行预测

void test(int argc, char** argv) {
  if (argc < 4 || argc > 5) {
    printTestUsage();
    exit(EXIT_FAILURE);
  }
  int32_t k = 1;
  if (argc >= 5) {
    k = atoi(argv[4]);
  }


  FastText fasttext;
  fasttext.loadModel(std::string(argv[2]));


  std::string infile(argv[3]);
  if (infile == "-") {
    fasttext.test(std::cin, k);
  } else {
    std::ifstream ifs(infile);
    if (!ifs.is_open()) {
      std::cerr << "Test file cannot be opened!" << std::endl;
      exit(EXIT_FAILURE);
    }
    fasttext.test(ifs, k);
    ifs.close();
  }
  exit(0);
}

下面先看fasttext中的loadModel

void FastText::loadModel(const std::string& filename) {
  std::ifstream ifs(filename, std::ifstream::binary);
  if (!ifs.is_open()) {
    std::cerr << "Model file cannot be opened for loading!" << std::endl;
    exit(EXIT_FAILURE);
  }
  if (!checkModel(ifs)) {
    std::cerr << "Model file has wrong file format!" << std::endl;
    exit(EXIT_FAILURE);
  }
  loadModel(ifs);
  ifs.close();
}


void FastText::loadModel(std::istream& in) {
  args_ = std::make_shared<Args>();
  dict_ = std::make_shared<Dictionary>(args_);
  input_ = std::make_shared<Matrix>();
  output_ = std::make_shared<Matrix>();
  qinput_ = std::make_shared<QMatrix>();
  qoutput_ = std::make_shared<QMatrix>();
  args_->load(in);


  dict_->load(in);


  bool quant_input;
  in.read((char*) &quant_input, sizeof(bool));
  if (quant_input) {
    quant_ = true;
    qinput_->load(in);
  } else {
    input_->load(in);
  }


  in.read((char*) &args_->qout, sizeof(bool));
  if (quant_ && args_->qout) {
    qoutput_->load(in);
  } else {
    output_->load(in);
  }


  model_ = std::make_shared<Model>(input_, output_, args_, 0);
  model_->quant_ = quant_;
  model_->setQuantizePointer(qinput_, qoutput_, args_->qout);


  if (args_->model == model_name::sup) {
    model_->setTargetCounts(dict_->getCounts(entry_type::label));
  } else {
    model_->setTargetCounts(dict_->getCounts(entry_type::word));
  }
}

下面看test函数，在这个例子中k是1

void FastText::test(std::istream& in, int32_t k) {
  int32_t nexamples = 0, nlabels = 0;
  double precision = 0.0;
  std::vector<int32_t> line, labels;


  while (in.peek() != EOF) {//遍历每一行
    dict_->getLine(in, line, labels, model_->rng);//读取一行
    if (labels.size() > 0 && line.size() > 0) {
      std::vector<std::pair<real, int32_t>> modelPredictions;//考虑到了多标签
      model_->predict(line, k, modelPredictions);//这里k等于1，就一个标签
      for (auto it = modelPredictions.cbegin(); it != modelPredictions.cend(); it++) {
        if (std::find(labels.begin(), labels.end(), it->second) != labels.end()) {
          precision += 1.0;
        }
      }
      nexamples++;
      nlabels += labels.size();
    }
  }
  std::cout << "N" << "\t" << nexamples << std::endl;
  std::cout << std::setprecision(3);
  std::cout << "P@" << k << "\t" << precision / (k * nexamples) << std::endl;
  std::cout << "R@" << k << "\t" << precision / nlabels << std::endl;
  std::cerr << "Number of examples: " << nexamples << std::endl;
}

再看关键代码predict函数

void Model::predict(const std::vector<int32_t>& input, int32_t k,
                    std::vector<std::pair<real, int32_t>>& heap,
                    Vector& hidden, Vector& output) const {
  assert(k > 0);
  heap.reserve(k + 1);
  computeHidden(input, hidden);//通过词向量平均得到hidden向量
  if (args_->loss == loss_name::hs) {
    dfs(k, 2 * osz_ - 2, 0.0, heap, hidden);
  } else {
    findKBest(k, heap, hidden, output);//得到预测类别
  }
  std::sort_heap(heap.begin(), heap.end(), comparePairs);
}


void Model::predict(const std::vector<int32_t>& input, int32_t k,
                    std::vector<std::pair<real, int32_t>>& heap) {
  predict(input, k, heap, hidden_, output_);//hidden_和output_是model类自带的，一会儿用于存储结果
}

最后再看看findKBest

void Model::findKBest(int32_t k, std::vector<std::pair<real, int32_t>>& heap,
                      Vector& hidden, Vector& output) const {
  computeOutputSoftmax(hidden, output);//softmax的运算结果
  for (int32_t i = 0; i < osz_; i++) {//遍历得到的softmax值，选取k个最大的
    if (heap.size() == k && log(output[i]) < heap.front().first) {
      continue;
    }
    heap.push_back(std::make_pair(log(output[i]), i));//如果大于第k个值就更新堆
    std::push_heap(heap.begin(), heap.end(), comparePairs);
    if (heap.size() > k) {
      std::pop_heap(heap.begin(), heap.end(), comparePairs);
      heap.pop_back();
    }
  }
}