Show HN: I trained a 125M model to autocomplete piano on-device (simedw.com)

Article summary

The author trained a 125M-parameter transformer model to autocomplete piano performances in real-time, achieving 108 notes per second on an iPhone. The model was trained on a dataset of MIDI files and used a custom note representation to improve performance. The author also implemented Direct Preference Optimization (DPO) to fine-tune the model's output. The resulting app, RollTab, is available for free and allows users to play piano and have the AI generate continuations.

Main themes

• AI music generation
• MIDI files
• Real-time performance
• Custom note representation
• Direct Preference Optimization
• Music collaboration
• Creative applications
• Model optimization

What commenters say

• The project is impressive and has a lot of potential for creative applications.
• The model's output could be improved by incorporating more musical structure and form.
• The use of MIDI files and custom note representation is a key factor in the model's performance.
• The project's focus on real-time generation and user interaction is a unique and valuable aspect.
• Some commenters think that the model's output is not yet good enough to be useful for professional music production.
• Others believe that the project has a lot of potential for collaboration between humans and AI in music creation.
• The project's use of DPO is seen as a key factor in its success, but some commenters think that other optimization methods could also be effective.
• The possibility of expanding the model to support more complex musical elements, such as attack and timing, is an exciting area for future development.