Low Resource Machine Translation

Run Evaluator

To evaluate the model on helios, the evaluator script must be used with the input and target samples.

python evaluator.py --input-file-path {input-file-path} --target-file-path {target-file-path}

For more information, see python evaluator.py --help.

Note that the dependencies are in requirements.txt. They can be install with pip.

pip install -r requirements.txt

Experiments

All experiments can be reproduced with the scripts in the folders experiments. Custom experiments can be made with the help of the script run_experiment.py. For more information, see python run_experiment.py --help.

Artifacts

Artifacts are stored in different locations :

• Logs: logging/{model-name}/{datetime}/experiment.log,
• Weights: models/{model-name}-{id}/{epoch}.*,
• Valid Predictions: results/{model}-{id}/{datetime}/valid-{epoch},
• Train Predictions: results/{model}-{id}/{datetime}/train-{epoch},
• Training History (Learning Curves): results/{model}-{id}/{datetime}/history-{epoch}.

Generate Graphs Example

python generate_graphs.py --history_path='results/lstm_luong_attention/2020-04-01 21:58:21/history-17' output_path='results/lstm_luong_attention/2020-04-01 21:58:21/graphs'

Replacing a mask token in a sentence

Testing a masked language model can be done using the run_test_mlm.py script. Example: python run_test_mlm.py --checkpoint 2 --message "father help <mask>me pick apples" Predicted token: ed