Upload logs/rpa_eval/rpa_eval_20260323_201132.log with huggingface_hub

logs/rpa_eval/rpa_eval_20260323_201132.log

@@ -0,0 +1,101 @@
+=== RPA Evaluation — Mon Mar 23 20:11:32 HKT 2026 ===
+Checkpoint: outputs/e4/best.pt
+
+Using device: cuda
+Loading model...
+WARNING: feature_extractor_cqt requires the libray 'nnAudio'
+
+Loaded checkpoint: outputs/e4/best.pt
+[test] Loaded 583 entries
+Test set: 583 entries
+
+Running inference...
+/userhome/cs/u3571824/fyp/muse/backend/ml_models/dev/vocal/eval_rpa.py:271: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with autocast(enabled=args.fp16):
+  [50/583] 50 tracks, 53s
+  [100/583] 100 tracks, 102s
+  [150/583] 150 tracks, 146s
+  [200/583] 200 tracks, 191s
+  [250/583] 250 tracks, 234s
+  [300/583] 300 tracks, 277s
+  [350/583] 350 tracks, 319s
+  [400/583] 400 tracks, 364s
+  [450/583] 450 tracks, 406s
+  [500/583] 500 tracks, 448s
+  [550/583] 550 tracks, 489s
+
+Done: 583 tracks in 514.1s
+
+======================================================================
+  RPA / RCA RESULTS (CREPE Comparison)
+======================================================================
+
+  CREPE baseline on MIR-ST500: RPA(50c) = 81.61%
+  RMVPE baseline on MIR-ST500: RPA(50c) = 93.62%
+  Our existing pitch_accuracy (exact bin match): 31.69%
+
+  Method       RPA(50c)     RPA(100c)    RCA(50c)     VoicP    VoicR    VoicF1  
+  ------------ ------------ ------------ ------------ -------- -------- --------
+  argmax          61.48%      69.19%      63.14%    82.8%   51.6%   63.5%
+  local_ev        45.68%      57.96%      47.46%    71.7%   91.7%   80.5%
+  viterbi         58.73%      68.11%      60.30%    77.0%   58.4%   66.4%
+
+======================================================================
+  PER-DATASET BREAKDOWN (local_ev method)
+======================================================================
+
+  [dali]
+    RPA(50c): argmax=57.96%  local_ev=42.66%  viterbi=55.25%
+    RPA(100c): argmax=66.92%  local_ev=55.44%  viterbi=65.75%
+    Voicing (local_ev): P=70.6%  R=91.3%  F1=79.6%
+    Exact bin match: 27.74%
+    Frames evaluated: 5129938
+
+  [medleydb]
+    RPA(50c): argmax=54.54%  local_ev=49.47%  viterbi=55.70%
+    RPA(100c): argmax=57.29%  local_ev=54.09%  viterbi=58.79%
+    Voicing (local_ev): P=77.5%  R=54.5%  F1=64.0%
+    Exact bin match: 23.51%
+    Frames evaluated: 38719
+
+  [mir_st500]
+    RPA(50c): argmax=73.76%  local_ev=60.01%  viterbi=71.29%
+    RPA(100c): argmax=77.41%  local_ev=70.19%  viterbi=76.94%
+    Voicing (local_ev): P=77.0%  R=96.1%  F1=85.5%
+    Exact bin match: 52.16%
+    Frames evaluated: 1071912
+
+======================================================================
+  mir_eval.melody RESULTS (local_ev, 50c tolerance)
+======================================================================
+
+  Overall (583 tracks):
+    Overall Accuracy: 51.68%
+    Raw Chroma Accuracy: 43.62%
+    Raw Pitch Accuracy: 42.02%
+    Voicing False Alarm: 38.94%
+    Voicing Recall: 91.85%
+
+  MIR-ST500 only (82 tracks):
+    Overall Accuracy: 60.95%
+    Raw Chroma Accuracy: 58.93%
+    Raw Pitch Accuracy: 58.16%
+    Voicing False Alarm: 36.31%
+    Voicing Recall: 96.33%
+
+  >>> CREPE comparison: our RPA = 58.16% vs CREPE = 81.61% (gap = -23.45pp)
+
+======================================================================
+  SUMMARY
+======================================================================
+  Best RPA method: argmax
+  Overall RPA(50c): 61.48%
+  MIR-ST500 RPA(50c): 73.76%
+  CREPE baseline: 81.61%
+  Gap: -7.85pp
+  Exact bin match (our metric): 31.69%
+  VERDICT: FAIL — 7.85pp below CREPE
+
+Results saved to outputs/rpa_eval/rpa_results.json
+
+=== RPA Evaluation complete — Mon Mar 23 20:20:33 HKT 2026 ===