A flexible, research-oriented toolkit for training music source separation models with PyTorch.
- Modular Configuration: Hierarchical configuration system with YAML support
- Flexible STFT/ISTFT: Optional STFT modules that can be disabled for training
- Dynamic Mix Generation: Configurable mix generation with augmentations
- Multi-backend Training: Support for GPU, TPU, and distributed training
- Monitoring & Debugging: WandB integration and memory debugging tools
- ONNX Export: Easy model export for production deployment
- Docker Support: Separate containers for data preparation, training, and inference
- Package Installation: Can be installed via pip with optional dependencies
This structure provides a solid foundation for a research-to-production music source separation pipeline that's both flexible and maintainable.