Skip to content
This repository has been archived by the owner on Jun 10, 2021. It is now read-only.

Latest commit

 

History

History
111 lines (74 loc) · 4.97 KB

README.md

File metadata and controls

111 lines (74 loc) · 4.97 KB

This project is considered obsolete as the Torch framework is no longer maintained. For compatibility with OpenNMT-tf or OpenNMT-py, please check out CTranslate2.

Build Status

CTranslate

CTranslate is a C++ implementation of OpenNMT's translate.lua script with no LuaTorch dependencies. It facilitates the use of OpenNMT models in existing products and on various platforms using Eigen as a backend.

CTranslate provides optimized CPU translation and optionally offloads matrix multiplication on a CUDA-compatible device using cuBLAS. It only supports OpenNMT models released with the release_model.lua script.

Dependencies

  • Eigen >= 3.3
  • Boost (log, when -DWITH_BOOST_LOG=ON; program_options, when -DLIB_ONLY=OFF)

Optional

  • CUDA for matrix multiplication offloading on a GPU
  • Intel® MKL for an alternative BLAS backend

Compiling

CMake and a compiler that supports the C++11 standard are required to compile the project.

git submodule update --init
mkdir build
cd build
cmake ..
make

It will produce the dynamic library libonmt.so (or .dylib on Mac OS, .dll on Windows) and the translation client cli/translate.

CTranslate also bundles OpenNMT's Tokenizer which provides the tokenization tools lib/tokenizer/cli/tokenize and lib/tokenizer/cli/detokenize.

Options

  • To give hints about Eigen location, use the -DEIGEN3_ROOT=<path to Eigen library> option.
  • To compile only the library, use the -DLIB_ONLY=ON flag.
  • To disable OpenMP, use the -DWITH_OPENMP=OFF flag.
  • To enable optimization through quantization in matrix multiplications, use the -DWITH_QLINEAR=AVX2|SSE flag (OFF by default) and set the appropriate extended instructions set via -DCMAKE_CXX_FLAGS:
    • -DWITH_QLINEAR=AVX2 requires at least -mavx2
    • -DWITH_QLINEAR=SSE requires at least -mssse3

Performance tips

  • Use extended instructions sets:
    • if you are not cross-compiling, add -DCMAKE_CXX_FLAGS="-march=native" to the cmake command above to optimize for speed;
    • otherwise, select a recent SIMD extensions to improve performance while meeting portability requirements.
  • Consider installing Intel® MKL when you are targetting Intel®-powered platforms. If found, the project will automatically link against it.
  • Consider using quantization options as described above.
  • When using cli/translate, consider fine-tuning the level of parallelism:
    • the --parallel option enables concurrent translation of --batch_size sentences
    • the --threads option enables each translation to use multiple threads
    • Bottom-line: if you want optimal throughput for a collection of sentences, increase --parallel and set --threads to 1; if you want minimal latency for a single batch, set --parallel to 1, and increase --threads.

Using

Clients

See --help on the clients to discover available options and usage. They have the same interface as their Lua counterpart.

Library

This project is also a convenient way to load OpenNMT models and translate texts in existing software.

Here is a very simple example:

#include <iostream>

#include <onmt/onmt.h>

int main()
{
  // Create a new Translator object.
  auto translator = onmt::TranslatorFactory::build("enfr_model_release.t7");

  // Translate a tokenized sentence.
  std::cout << translator->translate("Hello world !") << std::endl;

  return 0;
}

For a more advanced usage, see:

  • include/onmt/TranslatorFactory.h to instantiate a new translator
  • include/onmt/ITranslator.h (the Translator interface) to translate sequences or batch of sequences
  • include/onmt/TranslationResult.h to retrieve results and attention vectors
  • include/onmt/Threads.h to programmatically control the number of threads to use

Also see the headers available in the Tokenizer that are accessible when linking against CTranslate.

Supported features

CTranslate focuses on supporting model configurations that are likely to be used in production settings. It covers models trained with the default options, plus some variants:

  • additional input or output word features
  • brnn encoder (with sum or concat merge policy)
  • dot attention
  • residual connections
  • no input feeding

Additionally, CTranslate misses some advanced features of translate.lua:

  • gold data score
  • hypotheses filtering
  • beam search normalization