migrate batch 1/1 from lamooon/vocal-v1

logs/vocal_e3_lw1e-2_20260321_192511.log

@@ -0,0 +1,157 @@
+=== E3 eval-only: e3_lw1e-2 — Sat Mar 21 19:25:11 HKT 2026 ===
+
+=== Phase 2: Full Eval ===
+Using device: cuda
+Config: {
+  "encoder": "m-a-p/MERT-v1-95M",
+  "encoder_hidden": 768,
+  "unet_channels": [
+    256,
+    512,
+    512
+  ],
+  "conformer_layers": 4,
+  "conformer_heads": 8,
+  "conformer_ff_dim": 2048,
+  "conformer_conv_kernel": 31,
+  "num_pitches": 88,
+  "dropout": 0.1,
+  "sample_rate": 24000,
+  "hop_length": 240,
+  "segment_length": 10.0,
+  "epochs": 20,
+  "batch_size": 8,
+  "accumulate_grad": 2,
+  "lr": 0.001,
+  "encoder_lr": 1e-05,
+  "layer_weight_lr": null,
+  "weight_decay": 0.01,
+  "warmup_epochs": 2,
+  "patience": 5,
+  "fp16": true,
+  "num_workers": 4,
+  "seed": 42,
+  "w_onset": 1.0,
+  "w_pitch": 1.0,
+  "w_frame": 0.5,
+  "focal_alpha": 0.85,
+  "focal_gamma": 2.0,
+  "pitch_sigma": 0.0,
+  "onset_threshold": 0.5,
+  "frame_threshold": 0.25,
+  "onset_tolerance_ms": 50,
+  "pitch_tolerance_cents": 50,
+  "data_dir": "./data",
+  "output_dir": "outputs/e3_lw1e-2",
+  "checkpoint": "outputs/e3_lw1e-2/best.pt",
+  "eval_only": true,
+  "per_dataset": true,
+  "datasets": null,
+  "augment": false,
+  "refine": false,
+  "refine_bpm": null,
+  "refine_iter": 5,
+  "device": null,
+  "no_f0_head": true,
+  "no_viterbi": false,
+  "no_expected_value": false,
+  "refine_scoring": "harmonics"
+}
+
+Loading datasets...
+[train] Loaded 4259 entries
+[val] Loaded 531 entries
+
+Building model with encoder: m-a-p/MERT-v1-95M
+WARNING: feature_extractor_cqt requires the libray 'nnAudio'
+
+/userhome/cs/u3571824/fyp/muse/backend/ml_models/dev/vocal/train.py:927: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = GradScaler(enabled=cfg.fp16)
+Parameters: 127.0M total, 127.0M trainable
+
+Resuming from outputs/e3_lw1e-2/best.pt
+[test] Loaded 583 entries
+/userhome/cs/u3571824/fyp/muse/backend/ml_models/dev/vocal/train.py:618: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with autocast(enabled=cfg.fp16):
+
+=== Test Results (Frame-Level) ===
+  onset_f1: 0.1751
+  onset_precision: 0.1817
+  onset_recall: 0.1690
+  frame_f1: 0.6727
+  frame_precision: 0.5217
+  frame_recall: 0.9467
+  pitch_accuracy: 0.5809
+  loss: 1.9407
+  note_precision: 0.8426
+  note_recall: 0.7065
+  note_f1: 0.6927
+
+=== Test Results (Note-Level, mir_eval) ===
+  note_precision: 0.8426
+  note_recall:    0.7065
+  note_f1:        0.6927
+
+=== Per-Dataset Breakdown ===
+  [dali] onset_f1=0.1562, onset_precision=0.1484, onset_recall=0.1648, frame_f1=0.6737, frame_precision=0.5221, frame_recall=0.9491, pitch_accuracy=0.5707, note_precision=0.8247, note_recall=0.6767, note_f1=0.6636
+  [medleydb] onset_f1=0.0816, onset_precision=0.0577, onset_recall=0.1393, frame_f1=0.7064, frame_precision=0.5497, frame_recall=0.9879, pitch_accuracy=0.7860, note_precision=0.7619, note_recall=0.5816, note_f1=0.5882
+  [mir_st500] onset_f1=0.2591, onset_precision=0.2447, onset_recall=0.2753, frame_f1=0.6464, frame_precision=0.5053, frame_recall=0.8971, pitch_accuracy=0.6621, note_precision=0.9569, note_recall=0.8967, note_f1=0.8770
+
+Results saved to outputs/e3_lw1e-2/test_results.json
+
+=== Phase 3: Upload to HuggingFace ===
+Uploading best.pt...
+
+
+
+
  outputs/e3_lw1e-2/best.pt   :   4%|▍         | 58.7MB / 1.53GB            
+
+
  outputs/e3_lw1e-2/best.pt   :   4%|▍         | 58.7MB / 1.53GB            
+
+
  outputs/e3_lw1e-2/best.pt   :  11%|█         |  168MB / 1.53GB            
+
+
  outputs/e3_lw1e-2/best.pt   :  18%|█▊        |  268MB / 1.53GB            
+
+
  outputs/e3_lw1e-2/best.pt   :  24%|██▎       |  361MB / 1.53GB            
+
+
  outputs/e3_lw1e-2/best.pt   :  30%|███       |  461MB / 1.53GB            
+
+
  outputs/e3_lw1e-2/best.pt   :  37%|███▋      |  562MB / 1.53GB            
+
+
  outputs/e3_lw1e-2/best.pt   :  43%|████▎     |  654MB / 1.53GB            
+
+
  outputs/e3_lw1e-2/best.pt   :  49%|████▉     |  747MB / 1.53GB            
+
+
  outputs/e3_lw1e-2/best.pt   :  56%|█████▌    |  856MB / 1.53GB            
+
+
  outputs/e3_lw1e-2/best.pt   :  63%|██████▎   |  965MB / 1.53GB            
+
+
  outputs/e3_lw1e-2/best.pt   :  70%|██████▉   | 1.07GB / 1.53GB            
+
+
  outputs/e3_lw1e-2/best.pt   :  76%|███████▌  | 1.16GB / 1.53GB            
+
+
  outputs/e3_lw1e-2/best.pt   :  83%|████████▎ | 1.26GB / 1.53GB            
+
+
  outputs/e3_lw1e-2/best.pt   :  89%|████████▉ | 1.36GB / 1.53GB            
+
+
  outputs/e3_lw1e-2/best.pt   :  95%|█████████▌| 1.45GB / 1.53GB            
+
+
  outputs/e3_lw1e-2/best.pt   : 100%|██████████| 1.53GB / 1.53GB            
+
+
  outputs/e3_lw1e-2/best.pt   : 100%|██████████| 1.53GB / 1.53GB            
+
+
  outputs/e3_lw1e-2/best.pt   : 100%|██████████| 1.53GB / 1.53GB            
+
+
  outputs/e3_lw1e-2/best.pt   : 100%|██████████| 1.53GB / 1.53GB            
+
+
  outputs/e3_lw1e-2/best.pt   : 100%|██████████| 1.53GB / 1.53GB            
+
+
  outputs/e3_lw1e-2/best.pt   : 100%|██████████| 1.53GB / 1.53GB            
+
+
  outputs/e3_lw1e-2/best.pt   : 100%|██████████| 1.53GB / 1.53GB            
+
+
  outputs/e3_lw1e-2/best.pt   : 100%|██████████| 1.53GB / 1.53GB            
+
+
  outputs/e3_lw1e-2/best.pt   : 100%|██████████| 1.53GB / 1.53GB            
+No files have been modified since last commit. Skipping to prevent empty commit.
+No files have been modified since last commit. Skipping to prevent empty commit.

logs/vocal_e4_20260322_173434.log

@@ -0,0 +1,109 @@
+=== E4 AbS Pitch Refinement — Sun Mar 22 17:34:34 HKT 2026 ===
+Using device: cuda
+Config: {
+  "encoder": "m-a-p/MERT-v1-95M",
+  "encoder_hidden": 768,
+  "unet_channels": [
+    256,
+    512,
+    512
+  ],
+  "conformer_layers": 4,
+  "conformer_heads": 8,
+  "conformer_ff_dim": 2048,
+  "conformer_conv_kernel": 31,
+  "num_pitches": 88,
+  "dropout": 0.1,
+  "sample_rate": 24000,
+  "hop_length": 240,
+  "segment_length": 10.0,
+  "epochs": 20,
+  "batch_size": 8,
+  "accumulate_grad": 2,
+  "lr": 0.001,
+  "encoder_lr": 1e-05,
+  "layer_weight_lr": null,
+  "weight_decay": 0.01,
+  "warmup_epochs": 2,
+  "patience": 5,
+  "fp16": true,
+  "num_workers": 4,
+  "seed": 42,
+  "w_onset": 1.0,
+  "w_pitch": 1.0,
+  "w_frame": 0.5,
+  "focal_alpha": 0.85,
+  "focal_gamma": 2.0,
+  "pitch_sigma": 0.0,
+  "onset_threshold": 0.5,
+  "frame_threshold": 0.25,
+  "onset_tolerance_ms": 50,
+  "pitch_tolerance_cents": 50,
+  "data_dir": "./data",
+  "output_dir": "outputs/e4",
+  "checkpoint": "outputs/e4/best.pt",
+  "eval_only": true,
+  "per_dataset": true,
+  "datasets": null,
+  "augment": false,
+  "refine": true,
+  "refine_bpm": null,
+  "refine_iter": 5,
+  "device": null,
+  "no_f0_head": true,
+  "no_frame_head": true,
+  "no_viterbi": false,
+  "no_expected_value": false,
+  "refine_scoring": "harmonics"
+}
+
+Loading datasets...
+[train] Loaded 4259 entries
+[val] Loaded 531 entries
+
+Building model with encoder: m-a-p/MERT-v1-95M
+WARNING: feature_extractor_cqt requires the libray 'nnAudio'
+
+/userhome/cs/u3571824/fyp/muse/backend/ml_models/dev/vocal/train.py:963: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = GradScaler(enabled=cfg.fp16)
+Parameters: 127.0M total, 127.0M trainable
+
+Resuming from outputs/e4/best.pt
+[test] Loaded 583 entries
+/userhome/cs/u3571824/fyp/muse/backend/ml_models/dev/vocal/train.py:637: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with autocast(enabled=cfg.fp16):
+
+=== Test Results (Frame-Level) ===
+  onset_f1: 0.2247
+  onset_precision: 0.2066
+  onset_recall: 0.2464
+  frame_f1: 0.7332
+  frame_precision: 0.7371
+  frame_recall: 0.7294
+  pitch_accuracy: 0.5256
+  loss: 0.6955
+  note_precision: 0.9030
+  note_recall: 0.7190
+  note_f1: 0.7224
+
+=== Test Results (Note-Level, mir_eval) ===
+  note_precision: 0.9030
+  note_recall:    0.7190
+  note_f1:        0.7224
+
+=== Per-Dataset Breakdown ===
+  [dali] onset_f1=0.1929, onset_precision=0.1645, onset_recall=0.2331, frame_f1=0.7255, frame_precision=0.7243, frame_recall=0.7267, pitch_accuracy=0.5118, note_precision=0.8883, note_recall=0.6899, note_f1=0.6925
+  [medleydb] onset_f1=0.2435, onset_precision=0.1652, onset_recall=0.4622, frame_f1=0.9507, frame_precision=0.9446, frame_recall=0.9569, pitch_accuracy=0.7696, note_precision=0.8690, note_recall=0.6214, note_f1=0.6452
+  [mir_st500] onset_f1=0.3645, onset_precision=0.2896, onset_recall=0.4918, frame_f1=0.7797, frame_precision=0.8910, frame_recall=0.6931, pitch_accuracy=0.6480, note_precision=0.9942, note_recall=0.9023, note_f1=0.9090
+
+Results saved to outputs/e4/test_results.json
+
+=== Analysis-by-Synthesis Refinement ===
+
+=== Refined Note Metrics (AbS) ===
+  note_precision: 0.8817 (was 0.9030)
+  note_recall:    0.7143 (was 0.7190)
+  note_f1:        0.7151 (was 0.7224)
+  delta note_f1:  -0.0072
+
+=== Upload to HuggingFace ===