Skip to content

benchmarks/idmt-first-run - #321

Merged
victoria-mckinney merged 11 commits into
mainfrom
benchmarking/idmt-first-run
Sep 16, 2026
Merged

victoria-mckinney merged 11 commits into
mainfrom
benchmarking/idmt-first-run

Conversation

@bellescoincidences

@bellescoincidences bellescoincidences commented Sep 11, 2026 •

Copy link
Copy Markdown

Benchmarking datasets first run mir_eval

@victoria-mckinney

Copy link
Copy Markdown
Member

relates to #322

@victoria-mckinney

victoria-mckinney commented Sep 14, 2026 •

Copy link
Copy Markdown
Member

Hi @nanaoto 👋

First end-to-end run of the IDMT benchmark using the adapter you contributed - it works! These are our first real, reproducible numbers. Sharing for your eyes before we go further.

Setup

  • Dataset: IDMT-SMT-Drums V2 (95 #MIX files)
  • Onset window: 50ms
  • Score: macro-average across items
  • Commit: 92d67f3

Results (F-measure)

Instrument Precision Recall F
Kick (KD) 0.628 0.285 0.342
Snare (SD) 0.732 0.775 0.720
Hi-hat (HH) 0.834 0.948 0.878

By subset (kick recall is the story):

RealDrum TechnoDrum WaveDrum
KD F 0.088 0.657 0.344
SD F 0.779 0.662 0.718
HH F 0.935 0.857 0.870

What jumps out

  • Hi-hat and snare are solid.
  • Kick is weak — driven by recall, not precision. When it says "kick" it's usually right (P=0.63), but it misses most kicks (R=0.29). Worst on acoustic RealDrum (recall 0.05 — missing ~95%).

I'm adding an MDB-Drums adapter for breadth (crash/ride/tom coverage IDMT lacks) and also an ENST adapter. Im keen to expand the labelling/mappings to all the parts that DrumScript tries to classify; thus why I am moving beyond just using IDMT.

I'm also keen to get as much data as possible to identify where the classification model needs to be improved.

2026-09-14_212913_92d67f35.zip

@victoria-mckinney
victoria-mckinney merged commit 20eaa13 into main Sep 16, 2026
6 checks passed
@bellescoincidences

Copy link
Copy Markdown
Author

@nanaoto merging these changes to main for now. let me know if you get a chance to review

@victoria-mckinney victoria-mckinney added the help wanted Extra attention is needed label Sep 16, 2026
@bellescoincidences bellescoincidences changed the title chore(idmt-first-run): update .gitignore so it doesnt sync mir_eval b… benchmarks/idmt-first-run Sep 24, 2026
@bellescoincidences

Copy link
Copy Markdown
Author

Work continued on/related to #335

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

help wanted Extra attention is needed

Projects

None yet

Development

Successfully merging this pull request may close these issues.

[benchmarking(idmt-first-run):] do a small initial run using benchmarking mir_eval code

3 participants