README.md

Deep Encoder, Shallow Decoder

Download trained deep-shallow models

See our paper for more detail. 12-1 denotes a 12-layer encoder and 1-layer decoder.

Model	Data	Test BLEU
WMT16 EN-DE 12-1 w/ Distillation	WMT16/14 Distilled Data	28.3
WMT16 EN-DE 6-1 w/ Distillation	WMT16/14 Distilled Data	27.4
WMT16 EN-DE 12-1 w/o Distillation	WMT16/14 Raw Data	26.9
Additional Models
WMT19 EN-DE Transformer Base	WMT19 EN-DE Transformer Big	WMT19 EN-DE moses tokenize + fastbpe

Train

Here is the command to train a deep-shallow model with a 12-layer encoder and 1-layer decoder of base size. We tuned the dropout rate from [0.1, 0.2, 0.3] and chose 0.2 for WMT16 EN-DE translation. We used 16 GPUs. See our paper for more detail.

python train.py wmt16.en-de.deep-shallow.dist/data-bin/ --arch transformer --share-all-embeddings --criterion label_smoothed_cross_entropy \
--label-smoothing 0.1 --lr 5e-4 --warmup-init-lr 1e-7 --min-lr 1e-9 --lr-scheduler inverse_sqrt --warmup-updates 4000 --optimizer adam --adam-betas '(0.9, 0.98)' \
--max-tokens 4096 --dropout 0.2 --encoder-layers 12 --encoder-embed-dim 512 --decoder-layers 1 \
--decoder-embed-dim 512 --max-update 300000 \
--distributed-world-size 16 --distributed-port 54186 --fp16 --max-source-positions 10000 --max-target-positions 10000 \
--save-dir  checkpoint/trans_ende-dist_12-1_0.2/  --seed 1

Evaluation

After downloading the tarballs from the table above:

tar -xvzf trans_ende-dist_12-1_0.2.tar.gz
tar -xvzf wmt16.en-de.deep-shallow.dist.tar.gz
python generate.py wmt16.en-de.deep-shallow.dist/data-bin/ --path trans_ende-dist_12-1_0.2/checkpoint_top5_average.pt \
--beam 5 --remove-bpe  --lenpen 1.0 --max-sentences 10

Citation

Please cite as:

@article{Kasai2020DeepES,
  title={Deep Encoder, Shallow Decoder: Reevaluating the Speed-Quality Tradeoff in Machine Translation},
  author={Jungo Kasai and Nikolaos Pappas and Hao Peng and J. Cross and Noah A. Smith},
  journal={ArXiv},
  year={2020},
  volume={abs/2006.10369}
}

Note

This code is based on the fairseq library. Pretrained models should work with the most recent fairseq as well.