File size: 3,450 Bytes
de75bc4 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 | {
"experiment_name": "FSD50K Part 4 Fine-Tuning of Part 3 Checkpoint",
"date": "2026-10-08",
"base_model": {
"checkpoint_filename": "best_fsd50k_part3.pt",
"description": "Primitive Extractor previously trained on FSD50K Parts 1, 2, and 3 using a dual-loss objective.",
"starting_val_map": 0.5225,
"architecture_summary": "Conv1d Stack (3 layers: 768->1728->1728->1728) -> LayerNorm -> Multihead Attention (8 heads, dim=512) -> Mean Pooling (dim=512) -> Classifier Head (199 classes)"
},
"dataset": {
"name": "FSD50K Part 4",
"training_samples": 7379,
"validation_samples": 815,
"total_classes": 199
},
"training_configuration": {
"epochs": 10,
"batch_size": 16,
"optimizer": "AdamW",
"base_learning_rate": 0.0002,
"weight_decay": 0.0001,
"loss_function": "Weighted BCE (Classification) + 0.1 * MSE (Acoustic Reconstruction)",
"scheduler": "Linear Warmup (10% of total steps) + Cosine Annealing Decay"
},
"epoch_history": [
{
"epoch": 1,
"bce_loss": 0.1671,
"mse_loss": 0.1101,
"val_map": 0.5493,
"learning_rate": "2.00e-04",
"new_best": true
},
{
"epoch": 2,
"bce_loss": 0.1368,
"mse_loss": 0.1099,
"val_map": 0.5331,
"learning_rate": "1.94e-04",
"new_best": false
},
{
"epoch": 3,
"bce_loss": 0.113,
"mse_loss": 0.1089,
"val_map": 0.5334,
"learning_rate": "1.77e-04",
"new_best": false
},
{
"epoch": 4,
"bce_loss": 0.0949,
"mse_loss": 0.108,
"val_map": 0.551,
"learning_rate": "1.50e-04",
"new_best": true
},
{
"epoch": 5,
"bce_loss": 0.0823,
"mse_loss": 0.1073,
"val_map": 0.5582,
"learning_rate": "1.17e-04",
"new_best": true
},
{
"epoch": 6,
"bce_loss": 0.0709,
"mse_loss": 0.1063,
"val_map": 0.5631,
"learning_rate": "8.26e-05",
"new_best": true
},
{
"epoch": 7,
"bce_loss": 0.062,
"mse_loss": 0.1057,
"val_map": 0.5565,
"learning_rate": "5.00e-05",
"new_best": false
},
{
"epoch": 8,
"bce_loss": 0.056,
"mse_loss": 0.1053,
"val_map": 0.5524,
"learning_rate": "2.34e-05",
"new_best": false
},
{
"epoch": 9,
"bce_loss": 0.0525,
"mse_loss": 0.105,
"val_map": 0.5507,
"learning_rate": "6.03e-06",
"new_best": false
},
{
"epoch": 10,
"bce_loss": 0.0505,
"mse_loss": 0.1048,
"val_map": 0.5525,
"learning_rate": "0.00e+00",
"new_best": false
}
],
"final_results": {
"best_val_map": 0.5631,
"best_epoch": 6,
"final_bce_loss": 0.0505,
"final_mse_loss": 0.1048,
"improvement_from_start": "Started at 54.93% (Epoch 1) and peaked at 56.31% (Epoch 6)"
},
"analysis": "The model successfully absorbed the new acoustic patterns from FSD50K Part 4 without catastrophic forgetting. The BCE loss dropped steadily from 0.1671 to 0.0505, proving the classifier head rapidly adapted to the new data distribution. The MSE reconstruction loss remained incredibly stable (dropping slightly from 0.1101 to 0.1048), confirming that the dual-loss objective successfully preserved the acoustic feature space. The Val mAP peaked at 56.31% in Epoch 6 before the learning rate fully decayed, demonstrating stable convergence."
} |