MT-Seq2Seq

Assignment 1

Code Structure

• -assign1/
  • Attention.py: implementation of bi-directional attention model
  • encdec.py: implementation of vanilla encoder-decoder model
  • util.py: util functions
  • report.pdf: Assignment report
  • -output/
    • blind.predict: prediction of blind set
    • test.predict: prediction of test set

Example command

Train:

python Attention.py --dynet-mem 4000 -batch True -save True -train_en ../data/train.en-de.low.en -train_de ../data/train.en-de.low.de

Usage for Attention model

optional arguments:

-batch: Whether or not use batch training (default: False)

-layer: Number of LSTM layers (default: 2)

-embed: Embedding size (default: 200)

-hid: Hidden size (default: 128)

-att: Attention size (default: 128)

-load: Model path to load (default: None)

-save: Whether or not save the model during training (default: True)

-se: Starting epoch, used for continue training from a certain node (default: 0)

-bs: Batch size (default: 20)

-beam: Whether or not use Beam search in translation (default: False)

-beam-width: Beam width (default: 3)

-pred: Only prediction without training (default: False)

-train_en: Target sentence file for training (default: None)

-train_de: Source sentence file for training (default: None)

-test_en: Target sentence file for testing (default: None)

-test_de: Source sentence file for testing (default: None)

-result: Result translation file for target testing sentences (default: None)