Discovery embedding models — v1
Project-hosted mirror of the ML model weights used by mStream's music-discovery feature. Hosted here so the feature never depends on third-party model hosting staying up; mStream downloads these automatically (with sha256 verification) the first time discovery data collection is enabled.
Assets
| File | Size | sha256 |
|---|---|---|
discogs-effnet-bsdynamic-1.onnx
| 18,027,718 bytes | a280825b334797cf677939db8cd5762c0392aedd0ca6415dbc1cd083f045e43c
|
discogs-effnet-bsdynamic-1.json
| 14,986 bytes | a2e85b2e7372d5f8e0f35bdd6aeae1139f101087d183d0b2fb60b0ea0f01a0ff
|
Attribution & license
Discogs-EffNet (discogs-effnet-bsdynamic-1) was created by the Music Technology Group, Universitat Pompeu Fabra (Essentia project) and is redistributed here unmodified from https://essentia.upf.edu/models/ under the terms of its license: CC BY-NC-SA 4.0 (Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International).
- These weights are for non-commercial use only. mStream's discovery feature that uses them is free and unconditionally non-commercial.
- Datasets derived from these weights (e.g. exported discovery snapshots) inherit NC-SA terms.
- Reference: Alonso-Jiménez et al., "Music Representation Learning Based on Editorial Metadata from Discogs" (ISMIR 2022).
This is not a release of mStream itself — see the versioned releases for the application.