Trainer.cpp

#include "Trainer.hpp"
#include "SubConfig.hpp"

Trainer::Trainer(ReadingMachine & machine) : machine(machine)
{
}

void Trainer::createDataset(SubConfig & config, bool debug)
{
  machine.trainMode(true);
  std::vector<torch::Tensor> contexts;
  std::vector<torch::Tensor> classes;

  extractExamples(config, debug, contexts, classes);

  nbExamples = classes.size();

  dataLoader = torch::data::make_data_loader(Dataset(contexts, classes).map(torch::data::transforms::Stack<>()), torch::data::DataLoaderOptions(batchSize).workers(0).max_jobs(0));

  optimizer.reset(new torch::optim::Adam(machine.getClassifier()->getNN()->parameters(), torch::optim::AdamOptions(0.0005).amsgrad(true).beta1(0.9).beta2(0.999)));
}

void Trainer::createDevDataset(SubConfig & config, bool debug)
{
  machine.trainMode(false);
  std::vector<torch::Tensor> contexts;
  std::vector<torch::Tensor> classes;

  extractExamples(config, debug, contexts, classes);

  devDataLoader = torch::data::make_data_loader(Dataset(contexts, classes).map(torch::data::transforms::Stack<>()), torch::data::DataLoaderOptions(batchSize).workers(0).max_jobs(0));
}

void Trainer::extractExamples(SubConfig & config, bool debug, std::vector<torch::Tensor> & contexts, std::vector<torch::Tensor> & classes)
{
  fmt::print(stderr, "[{}] Starting to extract examples\n", util::getTime());

  config.addPredicted(machine.getPredicted());
  config.setState(machine.getStrategy().getInitialState());

  while (true)
  {
    if (debug)
      config.printForDebug(stderr);

    config.setAppliableSplitTransitions(machine.getSplitWordTransitionSet().getNAppliableTransitions(config, Config::maxNbAppliableSplitTransitions));

    auto * transition = machine.getTransitionSet().getBestAppliableTransition(config);
    if (!transition)
    {
      config.printForDebug(stderr);
      util::myThrow("No transition appliable !");
    }

    std::vector<std::vector<long>> context;

    try
    {
      context = machine.getClassifier()->getNN()->extractContext(config,machine.getDict(config.getState()));
      for (auto & element : context)
        contexts.emplace_back(torch::from_blob(element.data(), {(long)element.size()}, torch::TensorOptions(torch::kLong)).clone());
    } catch(std::exception & e)
    {
      util::myThrow(fmt::format("Failed to extract context : {}", e.what()));
    }

    int goldIndex = machine.getTransitionSet().getTransitionIndex(transition);
    auto gold = torch::zeros(1, torch::TensorOptions(torch::kLong));
    gold[0] = goldIndex;

    for (auto & element : context)
      classes.emplace_back(gold);

    transition->apply(config);
    config.addToHistory(transition->getName());

    auto movement = machine.getStrategy().getMovement(config, transition->getName());
    if (debug)
      fmt::print(stderr, "(Transition,Newstate,Movement) = ({},{},{})\n", transition->getName(), movement.first, movement.second);
    if (movement == Strategy::endMovement)
      break;

    config.setState(movement.first);
    if (!config.moveWordIndex(movement.second))
    {
      config.printForDebug(stderr);
      util::myThrow(fmt::format("Cannot move word index by {}", movement.second));
    }

    if (config.needsUpdate())
      config.update();
  }

  fmt::print(stderr, "[{}] Extracted {} examples\n", util::getTime(), util::int2HumanStr(classes.size()));
}

float Trainer::processDataset(DataLoader & loader, bool train, bool printAdvancement)
{
  constexpr int printInterval = 50;
  int nbExamplesProcessed = 0;
  float totalLoss = 0.0;
  float lossSoFar = 0.0;
  int currentBatchNumber = 0;

  torch::AutoGradMode useGrad(train);
  machine.trainMode(train);

  auto lossFct = torch::nn::CrossEntropyLoss();

  auto pastTime = std::chrono::high_resolution_clock::now();

  for (auto & batch : *loader)
  {
    if (train)
      optimizer->zero_grad();

    auto data = batch.data;
    auto labels = batch.target.squeeze();

    auto prediction = machine.getClassifier()->getNN()(data);
    if (prediction.dim() == 1)
      prediction = prediction.unsqueeze(0);

    labels = labels.reshape(labels.dim() == 0 ? 1 : labels.size(0));

    auto loss = lossFct(prediction, labels);
    try
    {
      totalLoss += loss.item<float>();
      lossSoFar += loss.item<float>();
    } catch(std::exception & e) {util::myThrow(e.what());}

    if (train)
    {
      loss.backward();
      optimizer->step();
    }

    if (printAdvancement)
    {
      nbExamplesProcessed += labels.size(0);

      ++currentBatchNumber;
      if (nbExamplesProcessed >= printInterval)
      {
        auto actualTime = std::chrono::high_resolution_clock::now();
        double seconds = std::chrono::duration<double, std::milli>(actualTime-pastTime).count() / 1000.0;
        pastTime = actualTime;
        if (train)
          fmt::print(stderr, "\r{:80}\rcurrent epoch : {:6.2f}% loss={:<7.3f} speed={:<6}ex/s", "", 100.0*(currentBatchNumber*batchSize)/nbExamples, lossSoFar, (int)(nbExamplesProcessed/seconds));
        else
          fmt::print(stderr, "\r{:80}\reval on dev : loss={:<7.3f} speed={:<6}ex/s", "", lossSoFar, (int)(nbExamplesProcessed/seconds));
        lossSoFar = 0;
        nbExamplesProcessed = 0;
      }
    }
  }

  return totalLoss;
}

float Trainer::epoch(bool printAdvancement)
{
  return processDataset(dataLoader, true, printAdvancement);
}

float Trainer::evalOnDev(bool printAdvancement)
{
  return processDataset(devDataLoader, false, printAdvancement);
}

void Trainer::loadOptimizer(std::filesystem::path path)
{
  torch::load(*optimizer, path);
}

void Trainer::saveOptimizer(std::filesystem::path path)
{
  torch::save(*optimizer, path);
}