The Beat This! small0 beat and downbeat model, exported to ONNX for the SUB/WAVE analyzer's opt-in downbeat grid (ANALYZE_DOWNBEATS=beat-this). This is not a SUB/WAVE release: it's a model file the analyzer image downloads by checksum at build time.
| File | beat_this-small0.onnx (9,588,293 bytes)
|
| sha256 | 5cff50fb031bc4a43b036f0af4835cf60bf8d43e6ccc8a7e676ed89db53f6eaf
|
| Source checkpoint | beat_this-small0.ckpt from the authors' mirror, sha256 6074be2c4d490c5f6101fcc374a1ec72ae93456e23bb6019783b849f5dc7d47b
|
| Exporter | tools/beat-this/export.py: beat_this v1.1.0 (ad79748), torch 2.7.1 CPU, opset 17, dynamic time axis up to 1,500 frames, --attn-group 2
|
The export computes the time-direction attention two maps at a time. The arithmetic is unchanged, and it peaks at about 225 MB instead of 1.35 GB. Its logits match PyTorch's to within 6.4e-05, and its beat and downbeat times match upstream's File2Beats(dbn=False) exactly on 34 test files. The export is byte-reproducible from the pinned inputs.
Licence
Beat This! is by the Institute of Computational Perception, JKU Linz (https://github.com/CPJKU/beat_this). Its code and published weights are MIT-licensed; the licence text is attached as LICENSE-beat_this.txt. In the authors' words, "some of the training files are fully copyrighted or under limited Creative Commons licenses, and it is up to the user to assess whether this may impact their use case" (see CPJKU/beat_this#18 and #21).