Neural Programmer

Original project can be found in https://github.com/tensorflow/models, mantained by Arvind Neelakantan (arvind2505)

The original code has been changed:

• Support custom questions
• Support custom tables
• Added "demo" mode (visual and console)
• Support custom descriptions for the columns in the testing target table: improve column selection process
• When the max_entry length for the column names is greater than 1, the embeddings are averaged instead of just added up.
• Fixed multiple bugs on the code
• Cleaned code for debugging reasons

Original Instructions

Implementation of the Neural Programmer model described in paper

Download and extract the data from dropbox. Change the data_dir FLAG to the location of the data.

Training

python neural_programmer.py

The models are written to FLAGS.output_dir

Testing

python neural_programmer.py --evaluator_job=True

The models are loaded from FLAGS.output_dir. The evaluation is done on development data.

In case of errors because of encoding, add "# -*- coding: utf-8 -*-" as the first line in wiki_data.py