Skip to content

Latest commit

 

History

History
107 lines (68 loc) · 3 KB

quickstart.md

File metadata and controls

107 lines (68 loc) · 3 KB
Raw
layout title permalink icon menu
docs
Quick start
/quickstart/
fa-paper-plane
1

Recommended software

GPU version

Minimal requirements tested on Ubuntu 18.04 LTS:

  • CMake 3.10.2
  • GCC/G++ 7.3
  • CUDA 9.2

Notes:

  • GCC 8+/CUDA 10+ or higher are recommended
  • CUDA 10.0+ requires CMake 3.12.2+ due to bugs in earlier versions
  • Compilation of the web-server tool requires Boost 1.65.1+

CPU version

A Marian CPU build requires Intel MKL (recommended) or OpenBLAS. CPU build can be enabled by adding -DCOMPILE_CPU=on to the CMake command.

Ubuntu packages

Assuming a fresh Ubuntu LTS installation with CUDA, the following packages need to be installed to compile Marian with minimal dependencies:

  • Ubuntu 18.04 (or newer) + CUDA 9.2 (the default is gcc 7.3.0):

    sudo apt-get install git cmake build-essential

In general the standard packages of recent Ubuntu LTS editions should work, but some configurations of C++ compiler and CUDA may be incompatible with each other. Additional packages can be installed to compile Marian with the web server, built-in SentencePiece and TCMalloc support.

Installation

Clone a fresh copy from github:

git clone https://github.com/marian-nmt/marian

The project is a standard CMake out-of-source build, which on Linux can be compiled by executing the following commands:

mkdir marian/build
cd marian/build
cmake ..
make -j4

If run for the first time, this will also download several submodule repositories.

For details on installation under Windows see the documentation.

Running Marian

Training

Marian is the training framework of Marian. Assuming corpus.en and corpus.ro are corresponding and preprocessed files of a English-Romanian parallel corpus, the following command will create a Nematus-compatible neural machine translation model.

./marian/build/marian \
  --train-sets corpus.en corpus.ro \
  --vocabs vocab.en vocab.ro \
  --model model.npz

See the documentation for more details or the examples of how to train different models with Marian.

Translation

If a trained model is available, run:

echo "This is a test." | ./marian/build/marian-decoder -m model.npz -v vocab.en vocab.ro

For translation on CPU, add --cpu-threads N (assuming Marian has been compiled with CPU support):

echo "This is a test." | ./marian/build/marian-decoder -m model.npz -v vocab.en vocab.ro --cpu-threads 1

See the documentation for more details or the examples of how to use Edinburgh's WMT models for translation.

Resources