Stage-1 reconstruction demo

Needs internet: the MIDI players and the soundfont are loaded from a CDN. Serve this directory over HTTP (the players cannot fetch .mid over file://).

Checkpoint /home/mingyang/all_models/unified_symbolic_rep/stage1_local_encoder/s1a_v0/best.pt (step 26000) — 1440-frame segment (30 s) per piece from 25% of the piece, encoded with a 12-frame overhang per side that is dropped before decoding; both sides decoded to notes with the same rule (peak-picked onsets, +-2 frames, probability >= 0.8).

pop909 / pop909__147

notes 274 orig / 274 recon   onset@25ms P 0.996 R 0.996 F1 0.996   F1@10ms 0.993   active F1 0.986   vel exact 0.982   mean |dt| 0.13 ms

Original

Reconstruction

pop909 / pop909__849

notes 225 orig / 226 recon   onset@25ms P 0.996 R 1.000 F1 0.998   F1@10ms 0.998   active F1 0.990   vel exact 1.000   mean |dt| 0.08 ms

Original

Reconstruction

aria / aria__cb__106809_0

notes 286 orig / 286 recon   onset@25ms P 1.000 R 1.000 F1 1.000   F1@10ms 0.997   active F1 0.974   vel exact 0.993   mean |dt| 0.50 ms

Original

Reconstruction

aria / aria__bp__082203_0

notes 120 orig / 120 recon   onset@25ms P 1.000 R 1.000 F1 1.000   F1@10ms 1.000   active F1 0.997   vel exact 0.992   mean |dt| 0.56 ms

Original

Reconstruction

atepp / atepp__03427

notes 491 orig / 550 recon   onset@25ms P 0.878 R 0.984 F1 0.928   F1@10ms 0.915   active F1 0.918   vel exact 0.855   mean |dt| 1.37 ms

Original

Reconstruction

atepp / atepp__07831

notes 82 orig / 89 recon   onset@25ms P 0.921 R 1.000 F1 0.959   F1@10ms 0.924   active F1 0.996   vel exact 0.963   mean |dt| 1.13 ms

Original

Reconstruction