Skip to content

Files

Latest commit

rizarrizar
The correct path
May 26, 2017
d7702cd · May 26, 2017

History

History
33 lines (22 loc) · 1.73 KB

README.md

File metadata and controls

33 lines (22 loc) · 1.73 KB

Machine Translation experiments on TED dataset

This folder contains scripts and configurations for experiments on TED dataset.

To download and preprocess dataset, run create_dataset.sh. If you are in MILA you can skip this and use already preprocessed data from /data/lisatmp4/bahdanau/data (just set FUEL_DATA_PATH to point to this directory).

To reproduce the results from the paper, following the procedure below.

  1. Train a model with the configuration ted12.yaml. This should give you a model pretrained with maximum log-likelihood (main_best.tar), and also an annealed version of it (annealing_best.tar). The exact command is $LVSR/bin/run.py train ted12 $LVSR/exp/ted/configs/ted12.yaml.

  2. Start actor-critic training using ted16.yaml and the additional options --start-stage critic_pretraining --params ted12/main_best.tar. Wait until the training transitions to the main stage, the final outcome is ted16/main_best.tar.

  3. Use the configuration ted17.yaml to reproduce REINFORCE-critic results. You can start training from the main stage with parameters ted16/critic_pretraining.tar. The deliverable is main_best.tar.

  4. For REINFORCE with linear baseline, run reinforced3.yaml starting from ted12/main_best.tar.

We also recommend to use the script decode.sh for decoding. Please consult the appendix of the paper for the exact value of the character discount.

For computing the BLEU score run

$LVSR/bin/extract_recognized.sh <$OUTPUT | $MOSES/scripts/generic/multi-bleu.perl $FUEL_DATA_PATH/TED/de-en/prep/dev.de-en.en

where $MOSES is the path to Moses, $OUTPUT is an output produced by $LVSR/bin/run.py search