This is the implementation of Cross Attention Augmented Transducer (CAAT). If you found bugs or other questions, feel free to discuss with us by issues or mail to danliu@mail.ustc.edu.cn.
Our codes relies on PyTorch, Numpy and Fairseq. Besides, we modified warp-transducer for CAAT loss calculation, if you have already installed it, you should uninstall it first, and reinstall the version in sub-dir warp-transducer as follows :
cd warp_transducer
mkdir build & cd build
cmake ..
make install
cd ../pytorch_binding
pip install -e .Preprocessing training data: set MUSTC_DIR to the raw MuST-C dataset path, DATA_DIR to the path for processed data:
bash scripts/preporcess_mustc.sh $MUSTC_DIR $DATA_DIRTrain offline text-to-text translation model:
SRC=en
TGT=de
mt_dir=model/mt_base
mt_dir=$DATA_DIR/data-bin
fairseq-train $mt_dir --source-lang $SRC --target-lang $TGT \
--max-epoch 100 \
--user-dir rain
--ddp-backend=no_c10d \
--task dropout_translation \
--bpe-dropout 0.1 --src-encoder $DATA_DIR/mustc_en-de/bpe --tgt-encoder $DATA_DIR/mustc_en-de/bpe \
--arch transformer_small --dropout 0.3 --activation-dropout 0.1 \
--share-decoder-input-output-embed \
--share-all-embeddings \
--optimizer adam --adam-betas '(0.9, 0.98)' \
--lr 0.0005 --lr-scheduler inverse_sqrt \
--warmup-updates 4000 --warmup-init-lr '1e-07' \
--label-smoothing 0.1 --criterion label_smoothed_cross_entropy \
--weight-decay 0.0001 \
--save-dir $mt_dir \
--max-tokens 4096 \
--update-freq 8 \
--eval-bleu \
--eval-bleu-args '{\"beam\": 5, \"max_len_a\": 1.2, \"max_len_b\": 10}' \
--eval-bleu-detok space \
--eval-bleu-remove-bpe sentencepiece \
--eval-bleu-print-samples \
--best-checkpoint-metric bleu --maximize-best-checkpoint-metric
--log-interval 10 --save-interval 4 --log-format simple generate sequence distillation training data:
fairseq-interactive $mt_dir --user-dir rain \
--bpe sentencepiece --sentencepiece-model $DATA_DIR/mustc_en-de/text_cfg/bpe.model \
--task translation \
--path $mt_dir/checkpoint_best.pt \
--buffer-size 2000 --batch-size 128 \
--input $srcfile
--beam 5 > tmp.txt
grep ^D tmp.txt | cut -f3 >tmp2.txt
in_dir=$DATA_DIR/mustc_en-de
out_dir=$DATA_DIR/mustc_distill
cp -r $in_dir/audio_cfg $out_dir
cp -r $in_dir/text_cfg $out_dir
cp -r $in_dir/test* $out_dir
cp -r $in_dir/valid* $out_dir
ln -s $in_dir/fbank_test.zip $out_dir/fbank_test.zip
ln -s $in_dir/fbank_test1.zip $out_dir/fbank_test1.zip
ln -s $in_dir/fbank_valid.zip $out_dir/fbank_valid.zip
ln -s $in_dir/fbank_train.zip $out_dir/fbank_train.zip
cat $in_dir/train.en-de.en.raw $in_dir/train.en-de.en.raw >$out_dir/train.en-de.en.raw
cat $in_dir/train.en-de.audio.csv $in_dir/train.en-de.audio.csv >$out_dir/train.en-de.audio.csv
cat $in_dir/train.en-de.de.raw tmp2.txt >$out_dir/train.en-de.de.rawPretrain encoder with speech recognition task:
asr_dir=model/asr
fairseq-train $DATA_DIR/mustc_distill --source-lang $SRC --target-lang $TGT \
--encoder-max-relative-position 32 \
--max-epoch 100 \
--user-dir rain \
--ddp-backend=no_c10d \
--task s2s --task-type asr \
--bpe-dropout 0.1 \
--arch online_audio_transformer_offline --dropout 0.3 --activation-dropout 0.1 \
--main-context 32 --right-context 16 \
--share-decoder-input-output-embed-share-all-embeddings \
--optimizer adam --adam-betas '(0.9, 0.98)' \
--lr 0.0005 --lr-scheduler inverse_sqrt \
--warmup-updates 4000 --warmup-init-lr '1e-07' \
--label-smoothing 0.1 --criterion label_smoothed_cross_entropy \
--weight-decay 0.0001 \
--save-dir $asr_dir \
--max-tokens 20000 --update-freq 8 \
--log-interval 10 --save-interval 4 --log-format simple --fp16"Train final CAAT simultanslation model
sst_dir=model/sst
pretrain_model=$asr_dir/checkpoint_best.pt
step=32
latency_scale=1.0
fairseq-train $DATA_DIR/mustc_distill --source-lang $SRC --target-lang $TGT \
--encoder-max-relative-position 32 \
--max-audio-positions 2000 \
--transducer-downsample $step \
--main-context 32 --right-context 16 \
--user-dir rain \
--max-epoch 80 \
--delay-func diag_positive \
--pretrained-encoder-path $pretrain_model \
--delay-scale $latency_scale \
--transducer-label-smoothing 0.1 --transducer-ce-scale 1. \
--task transducer --task-type st --bpe-dropout 0.1 \
--arch audio_cat \
--tokens-per-step 6000 \
--dropout 0.3 --activation-dropout 0.1 --attention-dropout 0.1 \
--share-decoder-input-output-embed \
--optimizer adam --adam-betas '(0.9, 0.98)' \
--lr 0.0005 --lr-scheduler inverse_sqrt \
--warmup-updates 4000 --warmup-init-lr '1e-07' \
--criterion fake_loss \
--clip-norm 2 \
--save-dir $work_dir/$name \
--max-tokens 8000 \
--jointer-layers 6 --decoder-ffn-embed-dim 1024 \
--update-freq 16 --max-sentences 64 --skip-invalid-size-inputs-valid-test \
--log-interval 10 --save-interval 4 --log-format simple --num-workers 2 \
--fp16 --min-loss-scale 1e-6To evaluate with SimulEval, first we extract test dataset as SimulEval like:
```bash
python -m scripts.audio_test --prefix test-COMMON $MUSTC_DIR/en-de/data $DATA_DIR/test
```
Evaluate with SimulEval, note parameter --step-read-block should be "step*downsample_size/main_context", and downsample_size is set to 4 (two convolutions with strides 2) in our experimets:
```bash
step_read_block=4
simuleval --agent ./rain/simul/speech_fullytransducer_agent.py --timeout 100 \
--task-type st --train-dir $DATA_DIR/mustc_distill \
--source-lang en --target-lang de \
--source $DATA_DIR/test/tst-COMMON.list --target $DATA_DIR/test/tst-COMMON.de \
--data-type speech --task s2s --model-path $modelfile \
--output $outdir --port $port --timeout 100 \
-max-len-a 0.043 --len-scale 0.7 --len-penalty 0 --max-len-b -5 \
--intra-beam 5 --inter-beam 1 --decoder-step-read 256 --eager \
--step-read-block $step_read_block
```
If the paper or the code helps you, please cite the paper in the following format :
@inproceedings{liu2021cross,
title={Cross Attention Augmented Transducer Networks for Simultaneous Translation},
author={Liu, Dan and Du, Mengge and Li, Xiaoxi and Li, Ya and Chen, Enhong},
booktitle={Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing},
pages={39--55},
year={2021}
}