The latest SGNMT version is available on github:

$ git clone

Installing dependencies

SGNMT depends on the following libraries:

  • OpenFST for reading and writing FSTs (e.g. translation lattices) (>=1.5.4)
  • Optional: Tensor2Tensor for a wide range of different sequence models (>=1.7.0)
  • Optional: Eva Hasler’s TensorFlow fork for an extended version of TensorFlow’s seq2seq tutorial (latest)
  • Optional: Blocks for neural machine translation support based on Theano (>=0.1)
  • Optional: KenLM for reading ARPA language model files with KenLM backend (latest)
  • Optional: srilm-swig for reading ARPA language model files with SRI-LM backend (>=1.7.1)
  • Optional: NPLM for using feed-forward neural language models (>=0.3)

Installing OpenFST

We recommend to install OpenFST >=1.5.4. Make sure to enable the Python support when compiling OpenFST:

$ ./configure --enable-far --enable-python
$ make
$ make install

If you wish to use SGNMT in combination with the hierachical phrase-pased SMT system HiFST, you can directly use the OpenFST installation under externals/ in the HiFST installation directory. This will make it possible to create translation lattices with tropicalsparsetuple arcs with SGNMT to keep predictor scores separated (see fst output format).

Tested versions: OpenFST 1.5.4-1.6.3

Installing Tensor2Tensor

Tensor2Tensor is a TensorFlow-based library with support of various neural sequence models. SGNMT can access models trained with tensor2tensor via the t2t predictor. Follow the tensor2tensor installation instructions to install t2t and TensorFlow, and call SGNMT from the same environment. Note that SGNMT supports the --t2t_usr_dir argument to extend the registry of T2T to your custom directory.


The Tensor2Tensor code base is still under constant change, and SGNMT might not be compatible with the latest version. The latest tensor2tensor version which is supported by SGNMT is available in this fork.

Tested versions: Tensor2Tensor 1.7.0-1.9.0, TensorFlow 1.9-1.10.1

Installing TensorFlow (not T2T)

The TensorFlow-based nmt and rnnlm predictors are based on Eva Hasler’s TensorFlow fork which extends the seq2seq tutorial to reproduce the Bahdanau-style architecture.

Note: You don’t need this fork if you access TensorFlow only via the Tensor2Tensor library.

Tested versions: latest

Installing Blocks

Follow the instructions in the Blocks documentation to install the Blocks framework and all its dependencies.


It might be necessary to install the HDF5 development files before installing Blocks, e.g. with:

$ sudo apt-get install libhdf5-dev


Make sure that all dependencies of Blocks are updated. On systems where you don’t have root access, use the pip options --user and --force-reinstall to install updated packages locally.

Anaconda is another good option for local installations as it already sets up all packages Theano and Blocks depend on correctly. You still need to install Blocks with the pip command using Anaconda’s pip.

You’ll also need to add on_unused_input='ignore' to your Theano flags as discussed here:

$ export THEANO_FLAGS="on_unused_input='ignore'"

Tested versions: Blocks 0.1-0.2, Theano 0.8-0.9

Installing KenLM

Follow the instructions on the KenLM Github page to install KenLM:

pip install

Tested versions: latest

Installing SRILM

First, install the version 1.7.1 of the SRI language model toolkit if you don’t already have an installation.


According to the documentation, swig-srilm requires that SRILM is compiled as position independent code using MAKE_PIC=yes:

$ make World MAKE_PIC=yes

Then, checkout the srilm-swig project:

$ git clone

Modify the Makefile as explained in the installation instructions. For Ubuntu, the head of the make file should look like this:


Build the Python module:

$ make python

Tested versions: srilm 1.7.1

Installing NPLM

Download NPLM from the project homepage and install it. You can also use the UCAM NPLM fork from Gonzalo Iglesias for threadsafety and efficiency. If you are using nplm 0.3 there might be a bug in the Python module that prevents the nplm predictor from reading model files. Try to replace in the python/ directory of your NPLM installation with this file.

Tested versions: nplm 0.3

Setting up SGNMT

Update your environment variables to reflect the locations of OpenFST, SRILM/KenLM, and NPLM. On Ubuntu, it might be necessary to add /usr/local/lib to your LD_LIBRARY_PATH (default location for OpenFST):

$ export LD_LIBRARY_PATH=/usr/local/lib/:/path/to/swig-srilm/:/path/to/nplm/src/python:$LD_LIBRARY_PATH
$ export PYTHONPATH=/path/to/swig-srilm/:/path/to/nplm/python/:$PYTHONPATH

If OpenFST was not installed globally, you need to add it to your environment variables:

$ export LD_LIBRARY_PATH=/path/to/openfst/lib:$LD_LIBRARY_PATH
$ export PYTHONPATH=/path/to/openfst/lib/python2.7/site-packages:$PYTHONPATH

Clone the GIT repository and try to start and

$ git clone
$ cd sgnmt
$ python --help
$ python --help

If you see the help texts for both commands, you are ready for the Tutorial: Basics. Otherwise, there is probably something wrong with your OpenFST installation.