Results and commands · Findings in detail · Corpus (OSF) · Corpus erratum
A small relational graph network over the 12-pulse cycle predicts and generates the soloist of Malian jembe ensembles stroke by stroke (strike, timing offset, intensity), trained on the 15 takes of the IEMP corpus. On a rhythm left out of training it beats a same-size Music-Transformer-style model on all 15 held-out takes for strikes, offsets and intensities (offset NLL -0.565 against 0.411 nats, p = 0.0001).
Each take becomes a grid of 12 pulses x 4 drums per cycle; a window of 4 cycles is a 192-node graph with three edge types (next pulse, same pulse, previous cycle). The accompaniment of the target cycle is known and the soloist's pulses are predicted in order: a Bernoulli strike, then Gaussian offset and intensity. Models are always scored on whole takes they never saw, or on a whole piece; the take-level folds are frozen in splits.json and the tests fail if a take is on both sides.
Mine: event table and splits, reference models, the graph network and its variants, the transformer comparison (my implementation of relative attention), stroke-type labelling, generation, evaluation and the live prototype. Reused: the corpus and its annotations (Polak, Tarsitani & Clayton) and CC0 drum samples.
Negative log-likelihood on held-out takes (nats, lower is better), mean of 3 seeds (j6_take.json, j6_piece.json):
| Model (parameters) | Seen rhythm: strikes / offsets / intensities | Unseen rhythm: strikes / offsets / intensities |
|---|---|---|
| Graph network (56k) | 0.407 / -0.945 / 1.220 | 0.436 / -0.565 / 1.233 |
| Transformer, same size (60k) | 0.423 / -0.844 / 1.243 | 0.548 / 0.411 / 1.340 |
| Transformer (181k) | 0.436 / -0.704 / 1.355 | 0.590 / 1.055 / 1.538 |
| B2, Markov | 0.537 / -0.749 / 1.393 | 0.631 / 0.045 / 1.429 |
| B1, per position | 0.597 / -0.573 / 1.396 | 0.645 / -0.281 / 1.432 |
One dot per held-out take, 15 takes x 3 seeds. On the seen rhythm the graph network beats the same-size transformer on 12 of 15 takes for strikes and 14 of 15 for offsets; intensities tie.
Other findings, with their files in docs/RESULTS.md:
- Humanizing a cycle (offsets of the last of 4 cycles, strikes known): the graph network with a helix edge reaches -0.796 nats against -0.709 for the Markov reference, better on 13 of 15 takes.
- Building beat or cycle symmetry into the network did not help at any training size (2 to 13 takes).
- The piece is recognised from jembe timing alone, pattern masked: balanced accuracy 0.86 over 4 cycles.
- Live prototype: a Kalman grid tracker predicts the next stroke 100 ms ahead within 8.0 ms median, against 18.2 ms for an isochronous grid.
Python 3.11 (CI on Ubuntu). Training ran on a laptop CPU, on Colab and on a Slurm cluster.
git clone https://github.com/clementdumeril/Djembe.git && cd Djembe
pip install -e ".[dev,train,audio]" # train: torch, torch-geometric, wandb; audio: soundfile
python -m pytest -q # without the corpus: 109 passed, 4 skipped
python -m djembe.dataset download # annotations only, about 10 MB from OSF
python -m djembe.dataset build # output/events.csv, about 7 s
python experiments/j6/run.py --protocol take --index 0 --epochs 2 # quick check of the J6 pipeline on one takeFull runs are Slurm job arrays, one task per take and seed (slurm/README.md). Every command and output file is in docs/RESULTS.md; colab/j3_humanize.ipynb runs the three-seed timing study.
| Path | Role |
|---|---|
djembe/dataset.py, grid.py |
OSF download, event table, cycle grid and graph edges |
djembe/gnn.py, solo_seq.py |
Relational message passing, stroke-by-stroke soloist |
djembe/transformer.py, baselines.py |
Comparison models: relative-attention transformer, B1, B2 |
djembe/generate.py, live.py, render.py |
Sampling, grid tracker, audio rendering |
experiments/, slurm/, results/ |
One folder per study, job arrays, one JSON per run |
tests/ |
Splits without leakage, model symmetries, no corpus file committed |
- 15 takes, 3 pieces, one dundun player, same sessions: results hold within this corpus only.
- Leave-one-piece-out has 3 training configurations; no model had a hyperparameter search.
- Stroke types (bass, tone, slap) are read from the audio and agree with a player's ear at 75 % on 105 strokes.
- Runs are not bit-exact: seed 0 on the cluster and on a laptop differ by 0.02 nats per take on average (up to 0.12).
- Generated audio and renders over the corpus' microphones are not published (corpus licence).
Code: MIT (LICENSE, CITATION.cff). The corpus is CC BY-NC-ND 4.0 and is not redistributed; djembe/dataset.py downloads it from OSF. Drum samples: CC0, see assets/samples/CREDITS.md.
- Polak, R., Tarsitani, S., & Clayton, M. (2018). IEMP Malian Jembe [Data set]. OSF. https://doi.org/10.17605/OSF.IO/M652X
- Clayton, M., et al. (2021). The Interpersonal Entrainment in Music Performance Data Collection. Empirical Musicology Review, 16(1), 65-84. https://doi.org/10.18061/emr.v16i1.7555
- Polak, R. (2004). Festmusik als Arbeit, Trommeln als Beruf. Dietrich Reimer. https://d-nb.info/971098840
- Polak, R. (2010). Rhythmic feel as meter. Music Theory Online, 16(4). https://doi.org/10.30535/mto.16.4.4
- Polak, R., & London, J. (2014). Timing and meter in Mande drumming from Mali. Music Theory Online, 20(1). https://doi.org/10.30535/mto.20.1.1
- Polak, R., London, J., & Jacoby, N. (2016). Both isochronous and non-isochronous metrical subdivision afford precise and stable ensemble entrainment. Frontiers in Neuroscience, 10, 285. https://doi.org/10.3389/fnins.2016.00285
- Huang, C.-Z. A., et al. (2018). Music Transformer. arXiv:1809.04281, relative attention in the additive form of Raffel et al. (2020), arXiv:1910.10683.
The earlier analysis stage (notebooks, listening demo) is at tag v0.2-analysis. Clément Dumeril · clement.dumeril.net