atmt code base

Materials for "Advanced Techniques of Machine Translation". Please refer to the assignment sheet for instructions on how to use the toolkit.

The toolkit is based on this implementation.

Environment Setup

conda

# ensure that you have conda (or miniconda) installed (https://conda.io/projects/conda/en/latest/user-guide/install/index.html) and that it is activated

# create clean environment
conda create --name atmt36 python=3.6

# activate the environment
conda activate atmt36

# intall required packages
pip install torch==1.6.0 numpy tqdm sacrebleu

virtualenv

# ensure that you have python 3.6 downloaded and installed (https://www.python.org/downloads/)

# install virtualenv
pip install virtualenv

# create a virtual environment named "atmt36"
virtualenv --python=python3 atmt36

# launch the newly created environment
atmt36/bin/activate

# intall required packages
pip install torch==1.6.0 numpy tqdm sacrebleu

Training a model

python train.py \
    --data path/to/prepared/data \
    --source-lang en \
    --target-lang sv \
    --save-dir path/to/model/checkpoints \
    --train-on-tiny # for testing purposes only

Notes:

• path/to/prepared/data and path/to/model/checkpoints are placholders, not true paths. Replace these arguments with the correct paths for your system.
• only use --train-on-tiny for testing. This will train a dummy model on the tiny_train split.
• add the --cuda flag if you want to train on a GPU, e.g. using Google Colab

Evaluating a trained model

Run inference on test set

python translate.py \
    --data path/to/prepared/data \
    --dicts path/to/prepared/data \
    --checkpoint-path path/to/model/checkpoint/file/for/loading \
    --output path/to/output/file/model/translations

Postprocess model translations

bash scripts/postprocess.sh path/to/output/file/model/translations path/to/postprocessed/model/translations/file en

Score with SacreBLEU

cat path/to/postprocessed/model/translations/file | sacrebleu path/to/raw/target/test/file

Assignments

Assignments must be submitted on OLAT by 14:00 on their respective due dates.