GLAT

Implementation for the ACL2021 paper "Glancing Transformer for Non-Autoregressive Neural Machine Translation"

Requirements

• Python >= 3.7
• Pytorch >= 1.5.0
• Fairseq 1.0.0a0

Preparation

Train an autoregressive Transformer according to the instructions in Fairseq.

Use the trained autoregressive Transformer to generate target sentences for the training set.

Binarize the distilled training data.

input_dir=path_to_raw_text_data
data_dir=path_to_binarized_output
src=source_language
tgt=target_language
python3 fairseq_cli/preprocess.py --source-lang ${src} --target-lang ${tgt} --trainpref ${input_dir}/train \
    --validpref ${input_dir}/valid --testpref ${input_dir}/test --destdir ${data_dir}/ \
    --workers 32 --src-dict ${input_dir}/dict.${src}.txt --tgt-dict {input_dir}/dict.${tgt}.txt

Train

• For training GLAT

save_path=path_for_saving_models
python3 train.py ${data_dir} --arch glat --noise full_mask --share-all-embeddings \
    --criterion glat_loss --label-smoothing 0.1 --lr 5e-4 --warmup-init-lr 1e-7 --stop-min-lr 1e-9 \
    --lr-scheduler inverse_sqrt --warmup-updates 4000 --optimizer adam --adam-betas '(0.9, 0.999)' \
    --adam-eps 1e-6 --task translation_lev_modified --max-tokens 8192 --weight-decay 0.01 --dropout 0.1 \
    --encoder-layers 6 --encoder-embed-dim 512 --decoder-layers 6 --decoder-embed-dim 512 --fp16 \
    --max-source-positions 1000 --max-target-positions 1000 --max-update 300000 --seed 0 --clip-norm 5\
    --save-dir ${save_path} --src-embedding-copy --length-loss-factor 0.05 --log-interval 1000 \
    --eval-bleu --eval-bleu-args '{"iter_decode_max_iter": 0, "iter_decode_with_beam": 1}' \
    --eval-tokenized-bleu --eval-bleu-remove-bpe --best-checkpoint-metric bleu \
    --maximize-best-checkpoint-metric --decoder-learned-pos --encoder-learned-pos \
    --apply-bert-init --activation-fn gelu --user-dir glat_plugins \

• For training GLAT+CTC

save_path=path_for_saving_models
python3 train.py ${data_dir} --arch glat_ctc --noise full_mask --share-all-embeddings \
    --criterion ctc_loss --label-smoothing 0.1 --lr 5e-4 --warmup-init-lr 1e-7 --stop-min-lr 1e-9 \
    --lr-scheduler inverse_sqrt --warmup-updates 4000 --optimizer adam --adam-betas '(0.9, 0.999)' \
    --adam-eps 1e-6 --task translation_lev_modified --max-tokens 8192 --weight-decay 0.01 --dropout 0.1 \
    --encoder-layers 6 --encoder-embed-dim 512 --decoder-layers 6 --decoder-embed-dim 512 --fp16 \
    --max-source-positions 1000 --max-target-positions 1000 --max-update 300000 --seed 0 --clip-norm 2\
    --save-dir ${save_path} --length-loss-factor 0 --log-interval 1000 \
    --eval-bleu --eval-bleu-args '{"iter_decode_max_iter": 0, "iter_decode_with_beam": 1}' \
    --eval-tokenized-bleu --eval-bleu-remove-bpe --best-checkpoint-metric bleu \
    --maximize-best-checkpoint-metric --decoder-learned-pos --encoder-learned-pos \
    --apply-bert-init --activation-fn gelu --user-dir glat_plugins \

Inference

• The default setting without self re-ranking

checkpoint_path=path_to_your_checkpoint
python3 fairseq_cli/generate.py ${data_dir} --path ${checkpoint_path} --user-dir glat_plugins \
    --task translation_lev_modified --remove-bpe --max-sentences 20 --source-lang ${src} --target-lang ${tgt} \
    --quiet --iter-decode-max-iter 0 --iter-decode-eos-penalty 0 --iter-decode-with-beam 1 --gen-subset test

• Generating with self re-ranking of beam 5

checkpoint_path=path_to_your_checkpoint
python3 fairseq_cli/generate.py ${data_dir} --path ${checkpoint_path} --user-dir glat_plugins \
    --task translation_lev_modified --remove-bpe --max-sentences 20 --source-lang ${src} --target-lang ${tgt} \
    --quiet --iter-decode-max-iter 0 --iter-decode-eos-penalty 0 --iter-decode-with-beam 5 --gen-subset test

The script for averaging checkpoints is scripts/average_checkpoints.py

Thanks dugu9sword for contributing part of the code.