Code for M-DAT

Environment Setup

export CUDA_HOME=/usr/local/cuda
pip3 install --editable .
python3 setup.py build_ext --inplace

## Install apex
pip install Ninja
git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./
cd ..

cd dag_search
bash install.sh
cd ..

pip install hydra-core==1.1.1

pip install sacremoses
pip install 'fuzzywuzzy[speedup]'
pip install git+https://github.com/dugu9sword/lunanlp.git
pip install omegaconf
pip install nltk
pip install sacrebleu==1.5.1
pip install sacrebleu[ja]
pip install scikit-learn scipy
pip install bitarray
pip install tensorboardX
# pip install git+https://github.com/chenyangh/sacrebleu.git@1.5.1

pip install scipy
pip install wandb

Data

We give examples for the IWSLT dataset where we performed the ablation study.

The IWSLT dataset is directly obtained from Liu et al. (2021)

# Example for preprocessing the IWSLT training set
cur_path=IWSLT
for lang_pair in ("en-ro" "ro-en" "en-it" "it-en" "en-nl" "nl-en")
do
    src=${lang_pair:0:2}
    trg=${lang_pair:3:4}
    echo processing $src $trg
    for direction in ${src}-${trg} ${trg}-${src}
    do
        new_src=${direction:0:2}
        new_trg=${direction:3:4}
        bin_path=$cur_path/data-bin/${new_src}-${new_trg}
        rm -rf $bin_path
        python3 fairseq_cli/preprocess.py \
            --source-lang $new_src --target-lang $new_trg \
            --trainpref $cur_path/train/$lang_pair \
            --validpref $cur_path/valid/$lang_pair \
            --destdir $bin_path \
            --workers 128 --joined-dictionary \
            --srcdict $cur_path/dict.txt
    done
done

Train

Pivot BT

bash script_zeroshot/trainer_with_config.sh iwslt config/iwslt_mdag_pivot_bt.yml

Random BT

bash script_zeroshot/trainer_with_config.sh iwslt config/iwslt_mdag_random_bt.yml

Straight BT

bash script_zeroshot/trainer_with_config.sh iwslt config/iwslt_mdag_straight_bt.yml

Baseline

bash script_zeroshot/trainer_with_config.sh iwslt config/iwslt_mdag.yml

Evaluate

Two decoding options supported: lm and beam_lm

checkpoint_path=[YOUR_CHECKPONT_PATH]
bash script_zeroshot/evaluate_mdag iwslt $checkpoint_path [lh|beam_lm]