Notgnoshi / trumperator

Generating Trump tweets with deep learning. Spoiler: the results suck. The new GPT models do *much* better.

Trumperator

Generating Trump tweets with deep learning

---

Dependencies

• Keras (with Tensorflow GPU backend)

• Numpy

• trump_tweet_data_archive

• Textgenrnn

• spaCy NLP toolkit

  Install by running sudo -H pip install --upgrade spacy and then sudo -H python3 -m spacy download en to download the English language model.

Getting the dataset

Note that this will drastically inflate your .git/ folder (3.2ish gigabytes). Saving the tweets in zipped files in a git repository is a mistake because it can't diff binary files very well. Thus, for every change, it just saves a binary snapshot.

Run git submodule init and then git submodule update to download the the Trump tweet repository. This will also download the Textgenrnn submodule, which is also installable with Pip:

sudo -H pip install --upgrade textgenrnn

Note that scraping the tweets myself isn't quite as simple as I'd hoped. The official Twitter API limits the number of tweets you can directly download from a user to 3200ish tweets. This is nowhere near what I need for a training set. So, the collection of tweets needs to parse the HTML of the Twitter use page.

Generating tweets

The model is defined in experiments/model.py. To train that model, run experiments/train.py. This will save the trained model using the filename and parameters defined in experiments/model.py. To use an already-trained model, edit the filename in experiments/model.py and run the experiments/generate.py script, specifying the number of seeds to use as an integer commandline argument.

Here are some of my favorite tweets generated so far.

should be fired the office with a real congress comments on his republican spirit and ratings with a terrible many photo of @mittromney is always a fair #bush

the best president obama is a great people to be a great president i will be interviewed on @foxandfriends tonight at pm enjoy

If you want to try to more sophisticated Textgenrnn model, trained by me, run the following in the experiments/ folder.

from textgenrnn.textgenrnn import textgenrnn
# Files found in trumperator/experiments/
t = textgenrnn(config_path='trumperator_config.json',
               vocab_path='trumperator_vocab.json',
               weights_path='trumperator_weights.hdf5')
t.generate(10, temperature=0.2)

The Paper

The paper is found in paper/paper.pdf, or can be made with the provided makefile.