A deep learning bot trained on the crap coming out of Donald Trump's mouth.
Ivana Trumpalot is an Ivana Trumpabot that uses a deep neural network to generate text that sounds like Donald Trump's speeches. It's currently not quite online, we're putting the finishing touches to it. We'll be continually training it over the near future to improve its Trumpiness.
- @IvanaTrumpalot tweets once an hour.
- It will answer you if you tweet using its handle.
- It also answers the tweets of @BernieSanders, @HillaryClinton, @BarackObama, and @WhiteHouse, hopefully with a witty, on-topic slice of Trumpian wisdom.
Got another Twitter user in mind ? Let us know by opening up an issue.
Ivana Trumpalot is a recurrent neural network using the Long Short-Term Memory paradigm. The code is written in Keras, and is fully available in this repository.
We downloaded a handful of Trump's speeches, purged them for special characters we didn't want to keep, and removed any audience participation that was transcribed the the speeches. The corpus currently consists of about 250,000 characters, or 46,000 words. This corpus is also available in this repo.
The live version of Ivana Trumpalot runs off a Raspberry Pi 3, where text is generated and tweets are sent to Twitter. The network was trained on a GTX 980 Ti, where an epoch takes about three minutes to run at the current architecture.
Training : To train the model, simply call code/train_lstm.py. Feel free to change the parameters and architecture of the network, as well as the dataset you use. The network will be saved to disk every five epochs.
Predictions : Using the predict() function in code/ivanatrumpalot.py, feed it some text and it will generate some output for you.
Twitter : If you want to set up a Twitter bot, we use this package. Don't forget to feed your API keys to code/twitter_bot.py.
Ivana Trumpalot is, of course, a reference to the antagonist in Austin Powers 2 : The Spy Who Shagged Me. She's groovy, baby. Ivana Trump is also Donald Trump's wife. This isn't a personal attack at her; you must understand, the name was simply too great to pass up. We have the best names. The best names.
Mike and Quentin are as interested in deep learning and neural nets as Trump is interested in building big walls. They did their PhDs at the University of Warwick, and are both currently working in Vancouver, BC.