We plan to soon publish a documentation website with information on how to use this software package. For now, we provide some instructions on how to reproduce the experiments presented in our paper, soon to be presented at EMNLP 2018, Contextual Parameter Generation for Universal Neural Machine Translation, Emmanouil A. Platanios, Mrinmaya Sachan, Graham Neubig, and Tom M. Mitchell.
In order to reproduce our experimental results you must
first execute sbt assembly on the root directory of this
repository, after cloning it.
Example experiment scripts are located in the scripts
directory. We will soon update this with the specific
scripts used to reproduce the experiments presented in
our paper.
In order to use the precompiled TensorFlow binaries that
TensorFlow Scala provides, you need to change line 76 in
the build.sbt file, from this:
libraryDependencies += "org.platanios" %% "tensorflow" % tensorFlowScalaVersionto this:
libraryDependencies += "org.platanios" %% "tensorflow" % tensorFlowScalaVersion classifier "linux-cpu-x86_64"Make sure to replace linux-cpu-x86_64 with the string
that corresponds to your platform. Currently supported
platforms are: linux-cpu-x86_64, linux-gpu-x86_64, and
darwin-cpu-x86_64.
For more information on how to install/configure TensorFlow Scala, please refer to the official website.
- Add support for separate source/target word embeddings.
- Add support for bridges between the encoder and the decoder (inspired from OpenNMT).
- Add support for merging the bidirectional encoder states using either summation or concatenation.
- Add support for other attention models, as in here.
- Add support for scheduled sampling, presented in this paper.