Skip to content

Tensorflow Implementation of Variational Attention for Sequence to Sequence Models (COLING 2018)

License

Notifications You must be signed in to change notification settings

HareeshBahuleyan/tf-var-attention

Repository files navigation

Variational Attention for Sequence to Sequence Models

This is the official codebase for the following paper, implemented in tensorflow:

Hareesh Bahuleyan*, Lili Mou*, Olga Vechtomova, and Pascal Poupart. Variational Attention for Sequence-to-Sequence Models. COLING 2018. https://arxiv.org/pdf/1712.08207.pdf

Overview

This package consists of 3 models, each of which have been organized into separate folders:

  1. Deterministic encoder-decoder with deterministic attention (ded_detAttn)
  2. Variational encoder-decoder with deterministic attention (ved_detAttn)
  3. Variational encoder-decoder with variational attention (ved_varAttn)

Datasets

The proposed model and baselines have been evaluated on two experiments:

  1. Neural Question Generation with the SQuAD dataset
  2. Conversation Systems with the Cornell Movie Dialogue dataset

The data has been preprocessed and the train-val-test split is provided in the data/ directory.

Requirements

  • tensorflow-gpu==1.3.0
  • Keras==2.0. 8
  • numpy==1.12.1
  • pandas==0.22.0
  • gensim==3.1.2
  • nltk==3.4.5
  • tqdm==4.19.1

Instructions

  1. Generate word2vec, required for initializing word embeddings, specifying the dataset:
python w2v_generator.py --dataset qgen 
  1. Train the desired model, set configurations in the model_config.py file. For example,
cd ved_varAttn
vim model_config.py # Make necessary edits
python train.py
  • The model checkpoints are stored in models/ directory, the summaries for Tensorboard are stored in summary_logs/ directory. As training progresses, the metrics on the validation set are dumped intolog.txt and bleu/ directory.
  1. Evaluate performance of the trained model. Refer to predict.ipynb to load desired checkpoint, calculate performance metrics (BLEU and diversity score) on the test set, and generate sample outputs.

Citation

If you found this code useful in your research, please cite:

@inproceedings{varAttn2018,
  title={Variational Attention for Sequence-to-Sequence Models},
  author={Bahuleyan, Hareesh and Mou, Lili and Vechtomova, Olga and Poupart, Pascal},
  booktitle={Proceedings of the 27th International Conference on Computational Linguistics (COLING)},
  year={2018}
}

Releases

No releases published

Packages

No packages published