marian/README.md

AmuNMT

A C++ inference engine for Neural Machine Translation (NMT) models trained with Theano-based scripts from Nematus (https://github.com/rsennrich/nematus) or DL4MT (https://github.com/nyu-dl/dl4mt-tutorial)

If you use this, please cite:

Marcin Junczys-Dowmunt, Tomasz Dwojak, Hieu Hoang (2016). Is Neural Machine Translation Ready for Deployment? A Case Study on 30 Translation Directions (https://arxiv.org/abs/1610.01108)

Recommended for GPU version:

Tested on Ubuntu 14.04 LTS

• CMake 3.5.1 (due to CUDA related bugs in earlier versions)
• GCC/G++ 4.9
• Boost 1.54
• CUDA 7.5

Tested on Ubuntu 16.04 LTS

• CMake 3.5.1 (due to CUDA related bugs in earlier versions)
• GCC/G++ 5.4
• Boost 1.61
• CUDA 8.0

Also compiles the CPU version.

Recommended for CPU version:

The CPU-only version will automatically be compiled if CUDA cannot be detected by CMAKE. Tested on different machines and distributions:

• CMake 3.5.1
• The CPU version should be a lot more forgiving concerning GCC/G++ or Boost versions.

Git repository set-up (branch marian-integration only)

Branch marian-integration now contains marian as a sub-module.

After the initial clone or checkout of this branch, run the following command

git submodule update --init --recursive

If you want git to always update the submodules at git checkout, git pull, git merge, etc. (it won't by default, wich often causes confusion), copy the script git-hooks/post-rewrite into ./git/hooks and make it executable.

Compilation

The project is a standard Cmake out-of-source build:

mkdir build
cd build
cmake ..
make -j

If you want to compile only CPU version on a machine with CUDA, add -DCUDA=OFF flag:

cmake -DCUDA=OFF ..

Vocabulary files

Vocabulary files (and all other config files) in AmuNMT are by default YAML files. AmuNMT also reads gzipped yml.gz files.

• Vocabulary files from models trained with Nematus can be used directly as JSON is a proper subset of YAML.
• Vocabularies for models trained with DL4MT (*.pkl extension) need to be converted to JSON/YAML with either of the two scripts below:

python scripts/pkl2json.py vocab.en.pkl > vocab.json
python scripts/pkl2yaml.py vocab.en.pkl > vocab.yml

Running AmuNMT

./bin/amun -c config.yml <<< "This is a test ."

Configuration files

An example configuration:

# Paths are relative to config file location
relative-paths: yes

# performance settings
beam-size: 12
devices: [0]
normalize: yes
gpu-threads: 1

# scorer configuration
scorers:
  F0:
    path: model.en-de.npz
    type: Nematus

# scorer weights
weights:
  F0: 1.0

# vocabularies
source-vocab: vocab.en.yml.gz
target-vocab: vocab.de.yml.gz

BPE Support

AmuNMT has integrated support for BPE encoding. There are two option bpe and debpe. The bpe option receives a path to a file with BPE codes (here bpe.codes). To turn on desegmentation on the ouput, set debpe to true, e.g.

bpe: bpe.codes
debpe: true

Python Bindings

Python bindings allow to run AmuNMT decoder in python scripts. The compilation of the bindings requires python-dev package. To compile the bindings run:

make python

The Python bindings consist of 3 functions: init, translate, and shutdown:

import libamunmt

libamunmt.init('-c config.yml')
print libamunmt.translate(['this is a little test .'])

libamunmt.shutdown()

The init function init the decoder and the syntax is the same as in command line. The translate function takes a list of sentences to translate. For real-world example, see the scripts/amunmt_server.py script, which uses python bindings to run REST server.

The function shutdown is needed (and should be called at the end of your script) to avoid runtime errors due to the random order in which objects may be deallocated if they are not explicitly destroyed.

Using GPU/CPU threads

AmuNMT can use GPUs, CPUs, or both, to distribute translation of different sentences. However, it is unlikely that CPUs used together with GPUs yield any performance improvement. It is probably better to only use the GPU if one or more are available.

cpu-threads: 8
gpu-threads: 2
devices: [0, 1]

The setting above uses 8 CPU threads and 4 GPU threads (2 GPUs x 2 threads). The gpu-threads and devices options are only available when AmuNMT has been compiled with CUDA support. Multiple GPU threads can be used to increase GPU saturation, but will likely not result in a large performance boost. By default, gpu-threads is set to 1 and cpu-threads to 0 if CUDA is available. Otherwise cpu-threads is set to 1. To disable the GPU set gpu-threads to 0. Setting both gpu-threads and cpu-threads to 0 will result in an exception.

Example usage

• Data and systems for our winning system in the WMT 2016 Shared Task on Automatic Post-Editing

<<<<<<< HEAD

d33c4e0c5c