amuNN

A C++ decoder 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)

We aim at keeping compatibility with Nematus (at least as long as there is no training framework in amunNN), the continued compatbility with DL4MT will not be guaranteed.

Requirements:

• CMake 3.5.1 (due to CUDA related bugs in earlier versions)
• Boost 1.5
• CUDA 7.5
• yaml-cpp 0.5 (https://github.com/jbeder/yaml-cpp.git)

Optional

• KenLM for n-gram language models (https://github.com/kpu/kenlm, current master)

Compilation

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

mkdir build
cd build
cmake ..
make -j

Or with KenLM support:

cmake .. -DKENLM=path/to/kenlm

On Ubuntu 16.04, you currently need g++4.9 to compile and cuda-7.5, this also requires a custom boost build compiled with g++4.9 instead of the standard g++5.3. The binaries are not compatible. g++5 support will probably arrive with cuda-8.0.

CUDA_BIN_PATH=/usr/local/cuda-7.5 BOOST_ROOT=/path/to/custom/boost cmake .. \
-DCMAKE_CXX_COMPILER=g++-4.9 -DCUDA_HOST_COMPILER=/usr/bin/g++-4.9

Vocabulary files

Vocabulary files (and all other config files) in amuNN are by default YAML files. amuNN 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 amuNN

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

Configuration files

An example configuration:

relative-paths: yes

# performance
beam-size: 12
devices: [0]
normalize: true
threads-per-device: 1

# scorer configuration
scorers: 
  F0:
    path: model.en-de.npz 
    type: Nematus
  
# vocabularies
source-vocab: [ vocab.en.yml.gz ]
target-vocab: vocab.de.yml.gz

# scorer weights
weights: 
  F0: 1.0