ContextModule.cpp

#include "ContextModule.hpp"

ContextModuleImpl::ContextModuleImpl(std::string name, const std::string & definition)
{
  setName(name);

  std::regex regex("(?:(?:\\s|\\t)*)Buffer\\{(.*)\\}(?:(?:\\s|\\t)*)Stack\\{(.*)\\}(?:(?:\\s|\\t)*)Columns\\{(.*)\\}(?:(?:\\s|\\t)*)(\\S+)\\{(.*)\\}(?:(?:\\s|\\t)*)In\\{(.*)\\}(?:(?:\\s|\\t)*)Out\\{(.*)\\}(?:(?:\\s|\\t)*)w2v\\{(.*)\\}(?:(?:\\s|\\t)*)");
  if (!util::doIfNameMatch(regex, definition, [this,&definition](auto sm)
        {
          try
          {
            for (auto & index : util::split(sm.str(1), ' '))
              bufferContext.emplace_back(std::stoi(index));

            for (auto & index : util::split(sm.str(2), ' '))
              stackContext.emplace_back(std::stoi(index));

            auto funcColumns = util::split(sm.str(3), ' ');
            columns.clear();
            for (auto & funcCol : funcColumns)
            {
              functions.emplace_back() = getFunction(funcCol);
              columns.emplace_back(util::split(funcCol, ':').back());
            }

            auto subModuleType = sm.str(4);
            auto subModuleArguments = util::split(sm.str(5), ' ');

            auto options = MyModule::ModuleOptions(true)
              .bidirectional(std::stoi(subModuleArguments[0]))
              .num_layers(std::stoi(subModuleArguments[1]))
              .dropout(std::stof(subModuleArguments[2]))
              .complete(std::stoi(subModuleArguments[3]));

            inSize = std::stoi(sm.str(6));
            int outSize = std::stoi(sm.str(7));

            if (subModuleType == "LSTM")
              myModule = register_module("myModule", LSTM(columns.size()*inSize, outSize, options));
            else if (subModuleType == "GRU")
              myModule = register_module("myModule", GRU(columns.size()*inSize, outSize, options));
            else if (subModuleType == "Concat")
              myModule = register_module("myModule", Concat(inSize));
            else
              util::myThrow(fmt::format("unknown sumodule type '{}'", subModuleType));

            w2vFile = sm.str(8);

            if (!w2vFile.empty())
            {
              getDict().loadWord2Vec(w2vFile);
              getDict().setState(Dict::State::Closed);
              dictSetPretrained(true);
            }

          } catch (std::exception & e) {util::myThrow(fmt::format("{} in '{}'",e.what(),definition));}
        }))
    util::myThrow(fmt::format("invalid definition '{}'", definition));
}

std::size_t ContextModuleImpl::getOutputSize()
{
  return myModule->getOutputSize(bufferContext.size()+stackContext.size());
}

std::size_t ContextModuleImpl::getInputSize()
{
  return columns.size()*(bufferContext.size()+stackContext.size());
}

void ContextModuleImpl::addToContext(std::vector<std::vector<long>> & context, const Config & config)
{
  auto & dict = getDict();
  std::vector<long> contextIndexes;

  for (int index : bufferContext)
    contextIndexes.emplace_back(config.getRelativeWordIndex(index));

  for (int index : stackContext)
    if (config.hasStack(index))
      contextIndexes.emplace_back(config.getStack(index));
    else
      contextIndexes.emplace_back(-1);

  for (auto index : contextIndexes)
    for (unsigned int colIndex = 0; colIndex < columns.size(); colIndex++)
    {
      auto & col = columns[colIndex];
      if (index == -1)
      {
        for (auto & contextElement : context)
          contextElement.push_back(dict.getIndexOrInsert(Dict::nullValueStr));
      }
      else
      {
        int dictIndex = dict.getIndexOrInsert(functions[colIndex](config.getAsFeature(col, index)));

        for (auto & contextElement : context)
          contextElement.push_back(dictIndex);
      }
    }
}

torch::Tensor ContextModuleImpl::forward(torch::Tensor input)
{
  auto context = wordEmbeddings(input.narrow(1, firstInputIndex, getInputSize()));

  context = context.view({context.size(0), context.size(1)/(int)columns.size(), (int)columns.size()*context.size(2)});

  return myModule->forward(context);
}

void ContextModuleImpl::registerEmbeddings()
{
  wordEmbeddings = register_module("embeddings", torch::nn::Embedding(torch::nn::EmbeddingOptions(getDict().size(), inSize)));
  loadPretrainedW2vEmbeddings(wordEmbeddings, w2vFile);
}