Hierarchical Metrical Structure

Repository for the paper Learning Hierarchical Metrical Structure Beyond Measures (to be presented on ISMIR 2022).

See also our follow-up work with self-supervision!

https://github.com/music-x-lab/Self-Supervised-Metrical-Structure.

Demos

See output folder for demos from the validation and test split of RWC-Pop.

Each output MIDI file contains an extra MIDI track with name Layers besides tracks from the original MIDI file. Layers is a drum track that labels L drum notes on a level-L boundary beyond measures (better viewed in an DAW with track-wise piano rolls). A downbeat without any drum notes is interpreted as a level-0 boundary.

Pretrained model

If you want to deploy the pre-trained model, download the pretrained weights simple_tcn_v2.0_filtered.sdict here and put it in the cache_data folder.

Configure datasets

• RWC Pop: you need to manually acquire RWC-Pop's midi files here and put your own path in settings.py. For example, if you write in settings.py the following code:

RWC_DATASET_PATH = '/my_dataset/rwc_pop_dataset'

Then all the RWC-Pop MIDI file should be located at:

/my_dataset/rwc_pop_dataset/AIST.RWC-MDB-P-2001.SMF_SYNC/RM-P001.SMF_SYNC.MID
/my_dataset/rwc_pop_dataset/AIST.RWC-MDB-P-2001.SMF_SYNC/RM-P002.SMF_SYNC.MID
/my_dataset/rwc_pop_dataset/AIST.RWC-MDB-P-2001.SMF_SYNC/RM-P003.SMF_SYNC.MID
...

• LMD dataset: to configure LMD dataset, download LMD-matched from here and put your own in settings.py. For example, if you write in settings.py the following code:

LMD_MATCHED_FOLDER = '/my_dataset/lmd_matched'

Then all the LMD MIDI file should be located at:

/my_dataset/lmd_matched/A/A/A/TRAAAGR128F425B14B/1d9d16a9da90c090809c153754823c2b.mid
/my_dataset/lmd_matched/A/A/A/TRAAAGR128F425B14B/5dd29e99ed7bd3cc0c5177a6e9de22ea.mid
/my_dataset/lmd_matched/A/A/A/TRAAAGR128F425B14B/b97c529ab9ef783a849b896816001748.mid
...

• Pop909 dataset: test midis already included in the repo.

How to use

To run the pretrained model on custom MIDI files:

1. Make sure the MIDI file has correct metre, downbeat and beat labels derived from tempo change & time signature events (e.g., every tempo change events mark a downbeat position, you may use DAWs or python packages like pretty_midi to check this)
2. Run python simple_tcn_eval.py path/to/your_midi.mid.
3. A prediction plot will be shown and a output/model_name/your_midi.mid_crf.mid file will be generated.

To retrain the model:

1. Configure RWC dataset path as instructed above
2. Run data_preprocessing_train.py to generate the data npy file.
3. Run simple_tcn.py to train the model. The trained model will be saved in cache_data folder.

To reproduce experiments on RWC-Pop and Pop909:

1. Download the pretrained model weights as instructed above
2. Configure RWC dataset path as instructed above
3. Run python simple_tcn_eval.py.

To reproduce experiments on baseline models:

1. Configure RWC dataset path as instructed above
2. Run midi_structure.py.

To reproduce experiments on LMD-matched:

1. Download the pretrained model weights as instructed above
2. Configure LMD dataset path as instructed above
3. Run python statistics.py.