This is the code repository that accompany the master's thesis by Alexander Rosenberg Johansen, Elias Obeid, and Jonas Meinertz Hansen.

The goal of the project was to train recurrent neural netowrks to translate text between different languages.

We wrote up the highlights of the work in an article. Read the abstract and download the pdf.

If you want the unabridged story, download the full thesis as pdf.

The code was made to be run in a docker container (see the `setup/`` directory) but should run equally well on any setup with the right dependencies installed.

Most of the code is written for python 3 with the following modules: TensorFlow (0.9), matplotlib, nltk, click, numpy.

Take a look at the Dockerfile to see the exact setup we were using.

We only recommend training these models on fast GPUs with lots of RAM.

The setup is made for training on the parallel data from the WMT '15 dataset. In order to be able to train seamlessly on different machines, we have made shell script for atuomatically downloading and preprocessing the data.

Executing the data/get_data.sh script will automatically download the major datasets of WMT '15 in many languages and save training sets in data/train/ and validation sets in data/valid

The alphabets are included in this git repository, but new ones can be built based on other datasets by editing and running data/build_alphabet.py.

Hyperparameters, model architecture, and other settings are controlled by config files in the configs/ directory. Configs can inherit settings from other configs, and most configs (e.g. the test config) use the default config (default.py) as the base.

Start training with a config file by giving its module name as argument when running train.py.

For example if we want to train with the config configs/attn_units/s-050.py, we should use the following command:

python3 train.py attn_units.s-050

If the config has defined a name that is not None, then checkpoints and log files for TensorBoard are saved in subdirectories in train/[name]/.