My mom runs a small business. She often stresses out about how much time is required to keep up a social media presence. For DerbyHacks 2, I automated the news with Markov chains. This year, GPT2 produces much more convincing output, so I decided that I'd do the world another service and automate social media as well. This time, I managed to drag some other people down with me.

What it does

This project has two parts. First, it scrapes the Twitter account of a given user and trains a machine learning model based on those Tweets. Second, it generates new Tweets with that model and posts them on a user-determined schedule.

How I built it

The poorly-documented but very powerful python-twitter wrapper allowed me to get Tweets pretty easily. It can also post them without too much of a challenge. The real unexplored territory was setting up GPT2 with TensorFlow and fine-tuning the model.

Challenges I ran into

Windows doesn't use standard UTF-8 character encoding. It also doesn't have a package manager.

Accomplishments that I'm proud of

We can successfully impersonate the President of the United States.

What I learned

Perry learned how to Python again.

What's next for GPT2-Tweeter

The IPO is coming sometime between tomorrow and the heat death of the universe.

Share this project: