/simul-mt_temporally-correlated-task-scheduling

Primary LanguagePython

This repo contains codes for simultanous translation task for our paper Temporally Correlated Task Scheduling for Sequence Learning published on ICML 2021. Codes for stock price forecasting task are at https://github.com/microsoft/qlib/tree/main/examples/benchmarks/TCTS.

Codes for simultaneous translation are in sim_mt/, and codes for our method are in ours/. Our code is based on fairseq v0.8.0; please install it by cloning and installing their github repo, or by

pip install fairseq==v0.8.0

Data preparation

For IWSLT'14 En-De, please prepare the data using fairseq's script, and then binarize the data by running:

TEXT=examples/translation/iwslt14.tokenized.de-en
fairseq-preprocess --source-lang de --target-lang en \
    --trainpref $TEXT/train --validpref $TEXT/valid --testpref $TEXT/test \
    --destdir data-bin/iwslt14.tokenized.de-en

For IWSLT'15 En-Vi, please download the tokenized data, and then prepare the binarized data by running:

fairseq-preprocess --source-lang en --target-lang vi \
    --trainpref data/train --validpref data/tst2012 --testpref data/tst2013 \
    --destdir data/bins --workers 10 \
    --thresholdsrc 5 --thresholdtgt 5

Training wait-k

For IWSLT'14 En-De, run training with scripts/sim_mt/train_iwslt_ende.sh:

bash scripts/sim_mt/train_iwslt_ende.sh $DATA_PATH $k

where $DATA_PATH is the binarized data and k is the training threshold.

Similarly, for IWSLT'15 En-Vi, run training with scripts/sim_mt/train_iwslt_envi.sh:

bash scripts/sim_mt/train_iwslt_envi.sh $DATA_PATH $k

Inference wait-k

Run inference with scripts/test/test.sh:

bash scripts/test/test.sh $DATA_PATH $refF $cktpath $k $src $tgt

where

  • $DATA_PATH is the binarized data
  • $refF is the reference file
    • For IWSLT'14 En-De, iwslt14.tokenized.de-en/tmp/test.de
    • For IWSLT'15 En-Vi, data/tst2013.vi
  • $cktpath is the checkpoint file to be evaluated
  • $k is the waiting threshold
  • $src and $tgt are the source and target languages (en and de for IWSLT'14 En-De, en and vi for IWSLT'15 En-Vi)

Training Ours

Similar to training wait-k, for IWSLT'14 En-De, run training with scripts/ours/train_iwslt_ende.sh:

bash scripts/sim_mt/train_iwslt_ende.sh $DATA_PATH $k

For IWSLT'15 En-Vi, run training with scripts/sim_mt/train_iwslt_envi.sh:

bash scripts/sim_mt/train_iwslt_envi.sh $DATA_PATH $k

Other baselines

  1. Random: run the wait-k script, but use the following fairseq-train options: --wait-k uniform
  2. CL: run the wait-k script, but use the following fairseq-train options: --wait-k CL-linear --wait-k-sample-end $k, where $k is the waiting threshold during inference.
  3. WIW and WID: run evaluation with scripts/test/test_agent.sh:
bash scripts/test/test_agent.sh $DATA_PATH $refF $cktpath $agent $src $tgt

where $agent is wait-if-worse (WIW) or wait-if-diff (WID).