Music Transformer

This repository contains an implementation of the Music Transformer model for symbolic music generation using PyTorch. The model is designed to handle long-term musical structure by using relative attention mechanisms.

Features

• Transformer architecture with relative self-attention
• Input/output in event-based token representation
• Support for loading and saving MIDI files
• Training and inference pipelines
• Evaluation with MIDI playback and token-based metrics

File Structure

.
├── configs/               
├── data/                  
├── models/                
├── preprocess/            
├── checkpoints/           
├── outputs/               
├── train.py               
├── generate.py            
└── requirements.txt

Dependencies

• Python 3.8+
• PyTorch
• NumPy
• pretty_midi
• miditoolkit
• tqdm
• Transformers