cheriell / cross-version-mpe

Accompanying code for "Utilizing Cross-Version Consistency for Domain Adaptation: A Case Study on Music Audio" in Tiny Papers @ ICLR, 2024

Cross-Version-MPE

Here we provide the accompanying code for our tiny paper

• Lele Liu and Christof Weiss, "Utilizing cross-version consistency for domain adaptation: A case study on music audio,” in Tiny Papers Track at International Conference on Learning Representations (Tiny Papers @ ICLR), May 2024.

Datasets and feature preparation

We use the MAESTRO Dataset, the Schubert Winterreise Dataset and the Wagner Ring Dataset in our case study.

During feature preparation, we calculate the HCQT spectrogram and binary pianoroll for each of the music performance, with a sampling rate of 22.05 kHz and a hop size of 512. The feature preparation code can be found HERE. We provide one example input-output pair of feature in the folder example_precomputed_features.

Experiments

The teacher model is trained using script experiments/teacher_maestro.py.

Below are the scripts for each experiment mentioned in our paper:

Method	Source Dataset	Target Dataset	Script
Sup	--	Schubert Wagner	experiments/<target_dataset>/supervised.py
T	MAESTRO	Schubert, Wagner	experiments/<target_dataset>/teacher.py
TS	MAESTRO	Schubert, Wagner	experiments/<target_dataset>/teacher_student.py
TSCV	MAESTRO	Schubert, Wagner	experiments/<target_dataset>/teacher_student_cross_version_1.py
TSCV2	MAESTRO	Schubert, Wagner	experiments/<target_dataset>/teacher_student_cross_version_2.py

Environments configuration

We use synctoolbox to calculate the alignment path between different versions. The python environment is the one provided by the toolbox (copied in file environment-synctoolbox.yml). Please use this python environment to run the feature preparation script eature_preparation/prepare_cross_version_alignment.py.

For running the experiments, please use the provided Dockerfile to build the Docker image. You can also pull the docker image by

docker pull cheriell/cross-version-mpe:0.0.2

Running instruction

Please refer to the runme.sh for the whole reproduction pipeline.

For reproducibility, we uploaded the model checkpoints and pre-calculated features for the test sets at:

• Liu, L., & Weiß, C. (2024). Utilizing Cross-Version Consistency for Domain Adaptation: A Case Study on Music Audio (Pretrained Models and Features) (0.0.1). Zenodo. https://doi.org/10.5281/zenodo.10936492