Language Modeling with Gated Convolutional Networks

This is a Tensorflow implementation of Facebook AI Research Lab's paper: Language Modeling with Gated Convolutional Networks. This paper applies a convolutional approach to language modelling with a novel Gated-CNN model.

Architecture

Requirements

• Download and extract the Google 1 Billion Word dataset in the data folder.

• TensorFlow 0.12.1

Usage

To train the model using the default hyperparameters:

$ python main.py
$ tensorboard --logdir=logs --host=0.0.0.0

Check main.py for tunable hyperparameter flags.

TODO

• Replace NCE loss with Adaptive Softmax.
• Remove restricted training on fixed sized sentences (20, for now) and extend to account for all varied sentence lenghts.
• Implement Weight Normalisation for faster convergence.
• Train extensively on deeper models to match the results with the paper.