Awesome Open Source
Awesome Open Source

Neural_Conversation_Models

================================= This implementation contains an extension of seq2seq tutorial for conversation models in Tensorflow:

  1. Option to use Beam Search and Beam Size for decoding

  2. Currently, it supports

    • Simple seq2seq models
    • Attention based seq2seq models
  3. To get better results use beam search during decoding / inference

Examples of basic model can be found in this paper.

https://arxiv.org/abs/1702.05512

Prerequisites

Data

Data accepted is in the tsv format where first component is the context and second is the reply

TSV format Ubuntu Dialog Data can be found here

example :-

  1. What are you doing ? \t Writing seq2seq model .

Usage

To train a model with Ubuntu dataset:

$ python neural_conversation_model.py --train_dir ubuntu/ --en_vocab_size 60000 --size 512 --data_path ubuntu/train.tsv --dev_data ubuntu/valid.tsv  --vocab_path ubuntu/60k_vocan.en --attention

To test an existing model:

$ python neural_conversation_model.py --train_dir ubuntu/ --en_vocab_size 60000 --size 512 --data_path ubuntu/train.tsv --dev_data ubuntu/valid.tsv  --vocab_path ubuntu/60k_vocan.en --attention --decode --beam_search --beam_size 25

Todo

  1. Add other state of art neural models.
  2. Adding layer normalization( in progress )

https://github.com/pbhatia243/tf-layer-norm

Contact

Parminder Bhatia, [email protected]


Get A Weekly Email With Trending Projects For These Topics
No Spam. Unsubscribe easily at any time.
python (51,910
tensorflow (2,131
decoding (55
chatbots (46
beam-search (20

Find Open Source By Browsing 7,000 Topics Across 59 Categories