This repository is a modularized re-write of Stanford CS224N assignment 5. Specifically the Character-Level LSTM Decoder for Neural Machine Translation.

I wanted to re-write it as a way of understanding the different components of the model and then eventually the tricks used to train the model - I have somewhat annotated a few notebooks for my own reference.

Docker support coming soon. Meanwhile:

1. Clone repository
2. Install the requirements using

pipenv install

if running the train and test tasks. If browsing notebooks, use:

pipenv install --dev

3. Download and place the Assignment 5 data from Stanford CS224N in nmt/datasets/data/.
4. Run tasks as

pipenv run sh tasks/<task-name>.sh

Possible tasks:

• train_local.sh : training using small sample (equivalent to train-local-q2 from the assignment)
• test_local.sh : testing using small sample (equivalent of test-local-q2 from the assignement) - should produce BLEU score of ~99.27
• train.sh : training using all data on GPU.
• test.sh : testing using all data - should produce BLEU score of ~29.40

• Python 3.6 (if using Pipfile; 3.6+ if using requirements.txt)
• Pipenv

• Stanford CS224N
• Assignment 5 Handout
• Achieving Open Vocabulary Neural Machine Translation with Hybrid Word-Character Models - Minh-Thang Luong and Christopher Manning.
• Character-Aware Neural Language Models - Kim et al. (2016)
• Attention? Attention! - Lilian Weng
• D2l.ai - Aston Zhang, Zachary C. Lipton, Mu Li, Alexander J. Smola