Neural Machine Translation

This project has been developed as a part of my master's thesis about Neural Machine Translations.

Dataset

Bilingual datasets can be found on ManyThings.org.
Language pair used in this project is deu-eng.

Running locally

Required python version: 3.5.2.

Install dependencies:

pip install -r requirements-dev.txt
python setup.py install

Available CLI commands can be listed through nmt --help.

For example: nmt train -d ./data/deu.txt will train Sequence to Sequence model using text data provided under given path.

Running with docker

Build the image: docker build -t nmt -f ./docker/Dockerfile ..

Docker image exposes commands from the CLI, so they can be invoked the same way as locally, as long as volumes for data and output are properly mounted, e.g.:

docker run --rm -it -v ${PWD}/data:/opt/ml/input/data/training \
  -v ${PWD}/reports:/opt/ml/model \ 
  nmt train -d /opt/ml/input/data/training/deu.txt -r /opt/ml/model

Training on AWS Sagemaker

Training the full dataset locally requires significant number of hours, therefore it's advised to offload this process to the cloud. Cloud runner of choice for this project is AWS Sagemaker.

After setting up Sagemaker IAM role, S3 bucket and ECR registry, all related configuration should be placed in .env file structured in a similar way as .env-template, i.e.:

SAGEMAKER_ROLE=sagemaker
DOCKER_REGISTRY=
S3_BUCKET=
AWS_REGION=us-east-1
AWS_ACCESS_KEY_ID=
AWS_SECRET_ACCESS_KEY=
INSTANCE_TYPE=ml.p2.xlarge

Training command can be invoked with nmt sage-train, which accepts optional parameter (-c) specifying location of configuration file. This configuration will be passed as hyperparameters to the training image.

Using final model for translation

Trained model can be prompted to translate sentences through nmt translate command, which requires two arguments to properly reconstruct the model:

path to trained weights (--model-weights)
path to data file used for training (--data-path)
Data file is required to reconstruct tokenizer with vocabulary that allows transformations of input sentences into vectors and vectors produced by the model into output sentences.

Example translations produced by the model:

During the course of my work on the master's thesis I will be uploading best performing models to public S3 bucket, so they can be used without the need of training them from scratch.

Tensorboard

Model training metrics can be monitored with tensorboard: tensorboard --logdir ./reports

macmackiewicz/neural-machine-translation