Add ESC-50 linear probing results for Part 5 Primitives (multi-dataset trained)

model_checkpoints/Primitive_Extractor/ESC50/esc50_linear_probe_part5_results.json

@@ -0,0 +1,147 @@
+{
+  "experiment_name": "ESC-50 Linear Probing of Part 5 Primitives",
+  "date": "2026-10-11",
+  "base_model": "best_fsd50k_primitive_part5.pt",
+  "data_source": "/content/drive/MyDrive/esc50/audio + causal_labels.json",
+  "note": "Part 5 was trained on ESC-50, Clotho, and FSD50K. This tests how well the primitives consolidated ESC-50 knowledge.",
+  "methodology": "Extracted BEATs features from raw .wav files. Stripped FSD50K classifier/decoder weights. Froze Primitive Extractor. Trained new ESC-50 head for 15 epochs. Evaluated on all 2000 samples.",
+  "label_mapping": {
+    "airplane": 0,
+    "breathing": 1,
+    "brushing_teeth": 2,
+    "can_opening": 3,
+    "car_horn": 4,
+    "cat": 5,
+    "chainsaw": 6,
+    "chirping_birds": 7,
+    "church_bells": 8,
+    "clapping": 9,
+    "clock_alarm": 10,
+    "clock_tick": 11,
+    "coughing": 12,
+    "cow": 13,
+    "crackling_fire": 14,
+    "crickets": 15,
+    "crow": 16,
+    "crying_baby": 17,
+    "dog": 18,
+    "door_wood_creaks": 19,
+    "door_wood_knock": 20,
+    "drinking_sipping": 21,
+    "engine": 22,
+    "fireworks": 23,
+    "footsteps": 24,
+    "frog": 25,
+    "glass_breaking": 26,
+    "hand_saw": 27,
+    "helicopter": 28,
+    "hen": 29,
+    "insects": 30,
+    "keyboard_typing": 31,
+    "laughing": 32,
+    "mouse_click": 33,
+    "pig": 34,
+    "pouring_water": 35,
+    "rain": 36,
+    "rooster": 37,
+    "sea_waves": 38,
+    "sheep": 39,
+    "siren": 40,
+    "sneezing": 41,
+    "snoring": 42,
+    "thunderstorm": 43,
+    "toilet_flush": 44,
+    "train": 45,
+    "vacuum_cleaner": 46,
+    "washing_machine": 47,
+    "water_drops": 48,
+    "wind": 49
+  },
+  "trainable_params": 1600818,
+  "total_params": 36731890,
+  "trainable_percentage": "4.36%",
+  "epoch_history": [
+    {
+      "epoch": 1,
+      "train_loss": 79.8518,
+      "val_acc_fold5": 72.0
+    },
+    {
+      "epoch": 2,
+      "train_loss": 27.7878,
+      "val_acc_fold5": 82.25
+    },
+    {
+      "epoch": 3,
+      "train_loss": 22.0738,
+      "val_acc_fold5": 80.75
+    },
+    {
+      "epoch": 4,
+      "train_loss": 17.2059,
+      "val_acc_fold5": 81.5
+    },
+    {
+      "epoch": 5,
+      "train_loss": 14.8962,
+      "val_acc_fold5": 83.0
+    },
+    {
+      "epoch": 6,
+      "train_loss": 12.4615,
+      "val_acc_fold5": 80.0
+    },
+    {
+      "epoch": 7,
+      "train_loss": 11.4194,
+      "val_acc_fold5": 80.5
+    },
+    {
+      "epoch": 8,
+      "train_loss": 11.3315,
+      "val_acc_fold5": 82.5
+    },
+    {
+      "epoch": 9,
+      "train_loss": 10.1055,
+      "val_acc_fold5": 80.25
+    },
+    {
+      "epoch": 10,
+      "train_loss": 9.5676,
+      "val_acc_fold5": 81.25
+    },
+    {
+      "epoch": 11,
+      "train_loss": 8.2561,
+      "val_acc_fold5": 80.75
+    },
+    {
+      "epoch": 12,
+      "train_loss": 7.2728,
+      "val_acc_fold5": 83.0
+    },
+    {
+      "epoch": 13,
+      "train_loss": 6.8242,
+      "val_acc_fold5": 80.5
+    },
+    {
+      "epoch": 14,
+      "train_loss": 6.7873,
+      "val_acc_fold5": 81.75
+    },
+    {
+      "epoch": 15,
+      "train_loss": 6.8103,
+      "val_acc_fold5": 81.75
+    }
+  ],
+  "full_evaluation_metrics": {
+    "top_1_accuracy_percent": 94.2,
+    "micro_f1_score_percent": 94.17,
+    "macro_f1_score_percent": 94.08,
+    "mAP_macro_percent": 97.81
+  },
+  "analysis": "Part 5 Primitives achieved 94.20% Top-1 Acc and 97.81% mAP on ESC-50 by training only 4.36% of parameters."
+}