Code for the papers:

• Toward Fully Self-Supervised Multi-Pitch Estimation
• Investigating an Overfitting and Degeneration Phenomenon in Self-Supervised Multi-Pitch Estimation

See the official releases for the exact code corresponding to each paper.

Clone the following repositories and install them along with their requirements:

git clone -b updates https://github.com/sony/timbre-trap
pip install -r timbre-trap/requirements.txt
pip install -e timbre-trap/

git clone -b refresh https://github.com/cwitkowitz/lhvqt
pip install -r lhvqt/requirements.txt
pip install -e lhvqt/

Then, install the main package ss-mpe:

pip install -r ss-mpe/requirements.txt
pip install -e ss-mpe/

All code for experiments is located under ss-mpe/experiments. To reproduce our experiments, simply run train.py and update the multipliers parameter to reflect the desired loss configuration You may also want to update EX_NAME and root_dir to your liking.

To evaluate an existing model, run comparisons.py with the model and checkpoint selected. Again, make sure all the paths are set correctly / to your liking. Baseline results can be reproduced with baselines.py, however note that there may be issues with attempting to run the script within a CUDA environment.