Seq2Seq Style Transfer from Quebec to Metropolitan French

This repo contains the submission for the COMP 550 final project at McGill University, and was written by:

• Arlie Coles
• Badreddine Sayah
• Vilayphone Vilaysouk

All code to rerun the experiments is included except the corpus, which contains copyrighted material. We can provide it upon request.

For the sake of repository size, we also do not include our experiment logs, but can provide these upon request.

Prerequisites

Several dependencies are required to run this project. The easiest way to manage them is to create a virtual environment using Conda. You can do so from the root of the repo this way:

conda create --name <name-of-env> --file requirements.txt

If you get an ImportError regarding AllenNLP, you can install it manually this way:

pip install allennlp

Training and testing

To run model training, run python train_seq2seq.py with the following possible options:

usage: train_seq2seq.py [-h] [--log_dir LOG_DIR] [--model_dir MODEL_DIR]
                        [--name NAME] [--continue_model CONTINUE_MODEL] [--bi]
                        [--bn] [--att]
                        config_file

positional arguments:
  config_file           Training config.

optional arguments:
  -h, --help            show this help message and exit
  --log_dir LOG_DIR     Log output dir.
  --model_dir MODEL_DIR
                        Saved model dir.
  --name NAME           Name for model.
  --continue_model CONTINUE_MODEL
                        Path to model for continuing training.
  --bi                  Use a bidirectional encoder.
  --bn                  Use batch normalization at encoder.
  --att                 Use a decoder with attention.

To run model testing, run python test_seq2seq.py with the following possible options:

usage: test_seq2seq.py [-h] [--log_dir LOG_DIR] [--name NAME] [--bi] [--att]
                       [--bn] [--write_idx WRITE_IDX]
                       config_file model_path

positional arguments:
  config_file           Testing config. Should be same as used to train.
  model_path            Path to model to test with.

optional arguments:
  -h, --help            show this help message and exit
  --log_dir LOG_DIR     Log output dir.
  --name NAME           Name for model.
  --bi                  Use a bidirectional encoder.
  --att                 Use a decoder with attention.
  --bn                  Use batch normalization at encoder.
  --write_idx WRITE_IDX
                        Index of output examples to write. Change for new
                        examples.