tetrahydrofuran / presidential-rnn

Presidential RNN

This respository trains and executes a character-level recurrent neural network using GRUs on President Donald Trump's tweets and official recorded statements and speeches. Equivalent Markov chains are also trained, to limited effect.

Running the Repository

The primary entry points for this repository are main.py and markov.py. They may be accessed by the command python [filename].

Running main.py as-is will load and transform both tweets and statements, training 6 and dumping RNN models each for a total of 12 models. This is extraordinarily computationally expensive.

Running markov.py will generate frequency tables and sentences given a seed. Frequency table generation is an expensive task, and may take hours.

Pre-trained models are stored in /data/models/ in h5py format and may be used directly.

Full Dependencies

Core Functionality

• keras
• sklearn
• pandas
• numpy
• unidecode

EDA Functionality (topic modeling, visualizations, etc.)

• gensim
• yellowbrick

Repository Structure

• bin - Source code files
  • pytorch - pyTorch implementation of GRU, unoptimized for CUDA or GPU
• data
  • clean - Cleaned data, stored in serialized format
  • models - Trained char-RNNs, stored in h5py format
  • markov - Markov chain frequency tables and sample output
• docs - Slides and supporting documentation for project

Contact

Feel free to contact me with feedback or questions.
Email - lzhou95 at gmail .com
LinkedIn - zhouleon
Medium - @confusionmatrix