Upload model_checkpoints/Primitive_Extractor/clotho/training_results.json with huggingface_hub
Browse files
model_checkpoints/Primitive_Extractor/clotho/training_results.json
ADDED
|
@@ -0,0 +1,90 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model_name": "Primitive Extractor (Clotho Reconstruction)",
|
| 3 |
+
"base_checkpoint": "ESC-50 Baseline (primitive_extractor_esc50_best.pt)",
|
| 4 |
+
"training_config": {
|
| 5 |
+
"dataset": "Clotho (Development Set)",
|
| 6 |
+
"dataset_size": 3631,
|
| 7 |
+
"train_size": 2904,
|
| 8 |
+
"val_size": 727,
|
| 9 |
+
"input_dim": 768,
|
| 10 |
+
"primitive_dim": 512,
|
| 11 |
+
"epochs": 8,
|
| 12 |
+
"batch_size": 8,
|
| 13 |
+
"optimizer": "AdamW",
|
| 14 |
+
"learning_rate": 0.0001,
|
| 15 |
+
"weight_decay": 0.01,
|
| 16 |
+
"loss_function": "L1 Reconstruction Loss",
|
| 17 |
+
"checkpoint_metric": "val_recon_loss"
|
| 18 |
+
},
|
| 19 |
+
"parameter_counts": {
|
| 20 |
+
"primitive_extractor": 35131072,
|
| 21 |
+
"decoder_head": 1312512,
|
| 22 |
+
"total_trainable": 36443584
|
| 23 |
+
},
|
| 24 |
+
"epoch_results": [
|
| 25 |
+
{
|
| 26 |
+
"epoch": 1,
|
| 27 |
+
"train_loss": 0.2376,
|
| 28 |
+
"val_loss": 0.205,
|
| 29 |
+
"val_similarity_percent": 75.73,
|
| 30 |
+
"checkpoint_saved": true
|
| 31 |
+
},
|
| 32 |
+
{
|
| 33 |
+
"epoch": 2,
|
| 34 |
+
"train_loss": 0.1887,
|
| 35 |
+
"val_loss": 0.1772,
|
| 36 |
+
"val_similarity_percent": 82.75,
|
| 37 |
+
"checkpoint_saved": true
|
| 38 |
+
},
|
| 39 |
+
{
|
| 40 |
+
"epoch": 3,
|
| 41 |
+
"train_loss": 0.1674,
|
| 42 |
+
"val_loss": 0.1604,
|
| 43 |
+
"val_similarity_percent": 86.21,
|
| 44 |
+
"checkpoint_saved": true
|
| 45 |
+
},
|
| 46 |
+
{
|
| 47 |
+
"epoch": 4,
|
| 48 |
+
"train_loss": 0.1535,
|
| 49 |
+
"val_loss": 0.1503,
|
| 50 |
+
"val_similarity_percent": 88.05,
|
| 51 |
+
"checkpoint_saved": true
|
| 52 |
+
},
|
| 53 |
+
{
|
| 54 |
+
"epoch": 5,
|
| 55 |
+
"train_loss": 0.1437,
|
| 56 |
+
"val_loss": 0.1418,
|
| 57 |
+
"val_similarity_percent": 89.46,
|
| 58 |
+
"checkpoint_saved": true
|
| 59 |
+
},
|
| 60 |
+
{
|
| 61 |
+
"epoch": 6,
|
| 62 |
+
"train_loss": 0.1361,
|
| 63 |
+
"val_loss": 0.1346,
|
| 64 |
+
"val_similarity_percent": 90.58,
|
| 65 |
+
"checkpoint_saved": true
|
| 66 |
+
},
|
| 67 |
+
{
|
| 68 |
+
"epoch": 7,
|
| 69 |
+
"train_loss": 0.1301,
|
| 70 |
+
"val_loss": 0.1295,
|
| 71 |
+
"val_similarity_percent": 91.33,
|
| 72 |
+
"checkpoint_saved": true
|
| 73 |
+
},
|
| 74 |
+
{
|
| 75 |
+
"epoch": 8,
|
| 76 |
+
"train_loss": 0.125,
|
| 77 |
+
"val_loss": 0.1251,
|
| 78 |
+
"val_similarity_percent": 91.94,
|
| 79 |
+
"checkpoint_saved": true
|
| 80 |
+
}
|
| 81 |
+
],
|
| 82 |
+
"best_results": {
|
| 83 |
+
"epoch": 8,
|
| 84 |
+
"train_loss": 0.125,
|
| 85 |
+
"val_loss": 0.1251,
|
| 86 |
+
"val_similarity_percent": 91.94
|
| 87 |
+
},
|
| 88 |
+
"rationale": "Reconstruction Loss was chosen over Contrastive Loss to preserve fine-grained acoustic fidelity in the 512-dim primitives. This ensures the downstream Voice Head has the exact acoustic texture needed for high-quality mel-spectrogram generation, while leaving semantic text alignment to the Causal Graph and Memory modules.",
|
| 89 |
+
"checkpoint_location": "model_checkpoints/Primitive_Extractor/clotho/primitive_extractor_clotho_recon_best.pt"
|
| 90 |
+
}
|