OmegaT PyTorch plugin

Motivation

This repo serves to enable translators using OmegaT to add plugin support for TorchScript models. Currently it supports scripted fairseq models by way jFairseq.

Build

The plugin can be built using sbt clean assembly. Once this is done, the resulting artifact ./target/omegat-torchscript-plugin-1.0-SNAPSHOT-XYZ.jar will need to be moved to the OmegaT plugins directory where XYZ is the platform information of the target OmegaT installation.

Running OmegaT with the plugin

Because TorchScript support relies on native code a la libtorch, the OmegaT jar will need to be invoked with a java option indicating the location to the libtorch bundle:

% java -jar -Djava.library.path=$LIBTORCH_HOME /path/to/OmegaT.jar

where $LIBTORCH_HOME is /path/to/libtorch/lib/. Note: libtorch is platform-dependent, so be sure to download the appropriate libtorch for the operating system on which OmegaT will be running.

In addition, because of differences in native library loading on Windows, the path will need to be updated to include $LIBTORCH_HOME so all native dependencies can be located by library loader.

Currently Windows support is only ensured for PyTorch 1.7.1 nightlies as of 2 December 2020 and above, though non-Windows support is guaranteed for 1.5+.

Test

Tests can be run by issuing sbt test.

Authors

• Elijah Rippeth