EmbeddingGemma 2: an open, lightweight multimodal embedding model
deepmind.google·6d ago
AI Engineering
Hacker News·Aug 20·5 min read
Show HN: I trained a 125M model to autocomplete piano on-device
TL;DR
A 125M-parameter transformer model was developed to autocomplete piano performances in real time on iOS devices, achieving around 108 notes per second. Key advancements included optimizing MIDI representation, thorough data cleaning, and implementing DPO post-training.
✦ Why It Matters
Try RollTab to enhance your piano practice with AI-generated music completions today.
Key Takeaways
- Explore MIDI tokenization techniques for music generation.
- Utilize DPO to improve model output quality.
- Focus on data cleaning for better training results.
- Experiment with real-time applications of AI in music.
How It Works
The model uses a unique representation of notes that combines pitch, onset timing, duration, and velocity into a single token. This allows the transformer to generate music at a high speed, processing about 108 notes per second on an iPhone.
The implementation of Direct Preference Optimization (DPO) further refines the model's output by training it to prefer better continuations based on user feedback.