Happy Wednesday everybody,
I wanted to share Rockbot, a lyrics generator trained from scratch on the writing styles of nearly 20k songs by over 400 artists across MANY musical genres (not just rock), as well as a few poets and comedians. The Github Repo has more background and notes, along with easy instructions for anybody who would like to roll out their own either via docker or otherwise. You should also be able to use for your own GPT-2 model with minimal changes to the input and output API calls.
Just for background, I had initially created by fine-tuning the base GPT-2 model, so my net model weighed in at ~500MB. It was pushing the limits of Streamlit Sharing, and causing it to crash. I went back to the drawing board, and trained it from scratch, which brought the model down to ~70MB. It still is pushing Streamlit Sharing’s capabilities and takes about 1min, give or take to generate on Streamlit Sharing. Though, if you generate your own, it only takes about 10s to generate on an OVH S1-2 machine with 2gb ram & 1 vcore.
At the risk of melting down my sharing.streamlit instance, you may check out a demo here. Hopefully it will stay stable with a few users at once, but we’ll see…