Implements a frontend filterbank learning module which can be initialized with complex weights for a Variable-Q Transform (lvqt_orig.py). Several techniques and variations of the module are also implemented, including:
- Multi-channel (harmonic) structure (lhvqt.py)
- Real-only weights (lvqt_real.py)
- Hilbert Transform for analytic filters (lvqt_hilb.py)
- Harmonic comb initialization (lhvqt_comb.py)
- Variational dropout for 1D convolutional layer (variational.py)
The repository was created for my Master's Thesis, End-to-End Music Transcription Using Fine-Tuned Variable-Q Filterbanks. It has since been updated with various improvements, and to support my new work, Learning Sparse Analytic Filters for Piano Transcription.
Recommended for standard/quick usage
pip install lhvqt
Recommended for running examples or making experimental changes.
git clone https://github.com/cwitkowitz/lhvqt
pip install -e lhvqt
Several examples of instantiation, inference, and visualization are provided under the examples sub-directory. A full-blown training, visualization, and evaluation example for piano transcription can be found at https://github.com/cwitkowitz/sparse-analytic-filters.
Please cite whichever is more relevant to your usage.
@inproceedings{cwitkowitz2022learning,
title = {Learning Sparse Analytic Filters for Piano Transcription},
author = {Frank Cwitkowitz and Mojtaba Heydari and Zhiyao Duan},
year = 2022,
booktitle = {Proceedings of Sound and Music Computing Conference (SMC)}
}
@mastersthesis{cwitkowitz2019end,
title = {End-to-End Music Transcription Using Fine-Tuned Variable-{Q} Filterbanks},
author = {Cwitkowitz, Frank},
year = 2019,
school = {Rochester Institute of Technology}
}