README.md

curriculum-nmt

Setup

0. Create conda environment: conda create --name curriculum_nmt python=3.7
1. Install requirements in requirements.txt
2. Run bash run_iwslt.sh download to download the IWSLT dataset
3. Run bash run_iwslt.sh vocab to generate vocab files. This generates a iwslt_vocab.json and iwslt_word_freq.json

Usage

1. Train the model locally on IWSLT with bash run_iwslt.sh train_local (with "none" ordering)

2. Train the model with desired scoring and pacing functions locally on IWSLT e.g. bash run_iwslt.sh train_local rarity linear (with "rarity" ordering and "linear" pacing. see scoring.py and pacing.py for more options)

References

1. Fine-Tuning by Curriculum Learning for Non-Autoregressive Neural Machine Translation arXiv
2. On The Power of Curriculum Learning in Training Deep Networks arXiv code
3. Competence-based Curriculum Learning for Neural Machine Translation arXiv
4. Improving Neural Machine Translation Models with Monolingual Data arXiv