Neural Machine Translation using RNN

Setup environment

conda create -n nmt python=3.6

conda install pytorch cudatoolkit=11.1 -c pytorch -c nvidia

pip install OpenNMT-py

pip install sacrebleu

NMT_Vanilla (w/o SentencePiece)

Download data (WMT16 English - Romanian)

python download_wmt16_ro_en_data.py

cd NMT_Vanilla

Build Vocab

onmt_build_vocab -config wmt16_ro_en.yaml -n_sample -1

Train

onmt_train -config wmt16_ro_en.yaml

Translate

onmt_translate -model ../data/wmt16_ro_en/run/model_step_100000.pt -src ../data/wmt16_ro_en/test.en -tgt ../data/wmt16_ro_en/test.ro -output ../data/wmt16_ro_en/pred.txt -gpu 0 -verbose

Quality Evaluation

sacrebleu ../data/wmt16_ro_en/pred.txt < ../data/wmt16_ro_en/test.ro

Diversity Evaluation

python eval_diversity.py "../data/wmt16_ro_en/pred.txt"

NMT Using SentencePiece

Download data (WMT16 English - Romanian)

python download_wmt16_ro_en_data.py

Install SentencePiece using instructions on its repo

cd NMT_SentencePiece

Data Preparation using SentencePiece

./prepare_wmt16_ro_en.sh

Followup Run Commands

Build Vocab

onmt_build_vocab -config wmt16_ro_en.yaml -n_sample -1

Train

onmt_train -config wmt16_ro_en.yaml

Generate SentencePiece output (hypothesis)

onmt_translate -model ../data/wmt16_ro_en/run/model_step_100000.pt -src ../data/wmt16_ro_en/test.en.sp -tgt ../data/wmt16_ro_en/test.ro.sp -output ../data/wmt16_ro_en/run/test.ro.hyp_model_step_100000.sp -gpu 0 -verbose

De-tokenize the hypothesis using SentencePiece

spm_decode -model ../data/wmt16_ro_en/wmtenro.model -input_format piece < ../data/wmt16_ro_en/run/test.ro.hyp_model_step_100000.sp > ../data/wmt16_ro_en/run/test.ro.hyp_model_step_100000

Quality Evaluation

sacrebleu ../data/wmt16_ro_en/test.ro < ../data/wmt16_ro_en/run/test.ro.hyp_model_step_100000

NMT Using Pre-Trained Models from FAIRSEQ

Download data (WMT16 English - German)

python download_wmt16_en_de_data.py

Installing FairSeq

git clone https://github.com/pytorch/fairseq

cd fairseq

pip install --editable ./

cd NMT_FairSeq

Translate Pre-Trained Model

python fairseq_translate.py # default config

# Other sample configs:

--sampling --sampling-topk 10

--sampling --sampling-topp 0.8

--beam 10

Quality Evaluation

sacrebleu data/wmt16_en_de/pred_topk_0_topp_0_beam_5.txt < data/wmt16_en_de/test.de

Diversity Evaluation

python eval_diversity.py "data/wmt16_en_de/pred_topk_0_topp_0_beam_5.txt"

Automate for multiple (decoding) configs

./script_fairseq_translate.py

./script_fairseq_eval.sh

./script_diversity_fairseq.sh

Credits

OpenNMT-py

FairSeq

SentencePiece