File size: 2,548 Bytes
59c0d08 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 | ---
license: mit
library_name: drum_dynamics
tags:
- drums
- midi
- velocity
- dynamics
- music
- transformer
- mixture-density-network
datasets:
- e-gmd
metrics:
- mae
- rmse
- nll
---
# Dynamics Needed — MDN transformer velocity model
**Version:** 0.1.0
Transformer with a mixture-density-network (MDN) head that predicts a
*distribution* over per-note velocities ("dynamics") for MIDI drum tracks.
Trained on the Expanded Groove MIDI Dataset (E-GMD). Part of the
*Dynamics Needed* thesis project.
## Intended use
Given a MIDI drum track with flat/undynamic velocities, sample or read out a
"best-fitting" velocity per note to restore human-like dynamics. Unlike the
LightGBM baseline, this model captures velocity *uncertainty*.
## Training data
[E-GMD](https://magenta.tensorflow.org/datasets/e-gmd) (Expanded Groove MIDI
Dataset), evaluated on the held-out test split.
## Metrics (test split)
| metric | value |
|-----------------------------------|----------------------------------|
| Native NLL | 4.473 |
| Discretized NLL | 3.132 |
| Deterministic readout — MAE | 19.695 |
| Deterministic readout — RMSE | 26.641 |
| Sampled — Wasserstein-1 | 2.948 |
| Sampled — histogram intersection | 0.921 |
Use the **deterministic readout** for accuracy/ranking; **sample** from the
predicted distribution to restore human-like variance.
## Limitations
Published as a versioned research artifact, not a final production model.
- **Point-vs-sample trade-off.** Sampling restores the dynamic spread that a
point estimate flattens, but raises MAE and lowers note-by-note ranking
correlation. Pick the readout to match the use case.
- **Absolute-loudness generalization gap.** On unseen drummers, MAE degrades
(~19 → ~28) and the sampled distribution shifts low — a player's overall
loudness is not inferable from structure alone. Relative dynamics transfer;
absolute level does not.
- **Known data artifact.** E-GMD's multi-kit rendering remaps pads to different
voices per kit, biasing some per-voice results; a single-kit rebuild is a
pending fix. See the project's `docs/methodology/kit-remapping-artifact.md`.
- **No listening test yet.** Numbers here are offline metrics; perceptual
A/B validation is future work.
## License
Set to `mit` by default — change to match the thesis's chosen license before
publishing.
|