| { |
| "model_type": "efficientnetv2-s", |
| "base_model": "efficientnetv2-s", |
| "task": "image-classification", |
| "num_labels": 8, |
| "id2label": { |
| "0": "barn", |
| "1": "bridge", |
| "2": "castle", |
| "3": "mosque", |
| "4": "skyscraper", |
| "5": "stadium", |
| "6": "temple", |
| "7": "windmill" |
| }, |
| "label2id": { |
| "barn": 0, |
| "bridge": 1, |
| "castle": 2, |
| "mosque": 3, |
| "skyscraper": 4, |
| "stadium": 5, |
| "temple": 6, |
| "windmill": 7 |
| }, |
| "input_shape": [ |
| 320, |
| 320, |
| 3 |
| ], |
| "preprocessing": { |
| "mode": "efficientnet_v2_preprocess_input", |
| "mean": [ |
| 0.0, |
| 0.0, |
| 0.0 |
| ], |
| "std": [ |
| 1.0, |
| 1.0, |
| 1.0 |
| ], |
| "note": "preprocess_input is identity in TF 2.12+; EfficientNetV2-S includes internal Rescaling layer. Input expects raw [0, 255] float32.", |
| "channel_order": "RGB" |
| }, |
| "training_info": { |
| "dataset_source": "Pexels Architectural Buildings (13440 images, 8 classes, balanced)", |
| "dataset_size": 13440, |
| "split_ratio": [ |
| 0.8, |
| 0.1, |
| 0.1 |
| ], |
| "split_seed": 42, |
| "phases": [ |
| { |
| "name": "Fase 1 - Feature Extraction (Head)", |
| "optimizer": "AdamW", |
| "lr": 0.001, |
| "epochs_max": 25, |
| "epochs_actual": 25, |
| "val_accuracy": 0.9747, |
| "val_loss": 0.6859, |
| "train_accuracy": 0.861, |
| "cutmix_mixup": true, |
| "label_smoothing": 0.1, |
| "early_stop_reason": "EarlyStopping val_accuracy plateau" |
| }, |
| { |
| "name": "Fase 2 - Selective Fine-Tuning", |
| "optimizer": "DiscriminativeAdamW", |
| "lr": 0.0003, |
| "epochs_max": 50, |
| "epochs_actual": 7, |
| "val_accuracy": 0.9777, |
| "val_loss": 0.3988, |
| "train_accuracy": 0.9722, |
| "unfreeze": "block6+top_conv (BN frozen)", |
| "discriminative_lr": { |
| "block6": 0.1 |
| }, |
| "cutmix_mixup": false, |
| "label_smoothing": 0.05, |
| "early_stop_reason": "EarlyStopping val_accuracy plateau" |
| }, |
| { |
| "name": "SWA Post-Training", |
| "epochs": 10, |
| "lr": 0.0001, |
| "bn_update": true, |
| "bn_update_steps": 100, |
| "val_accuracy": 0.9851, |
| "val_loss": 0.3638, |
| "method": "Izmailov et al., UAI 2018" |
| } |
| ], |
| "metrics": { |
| "train_accuracy": 0.9997, |
| "val_accuracy": 0.9851, |
| "test_accuracy": 0.9792, |
| "test_loss": 0.3928, |
| "tta_accuracy": 0.9814, |
| "overfitting_gap": 0.0206, |
| "test_correct": 1316, |
| "test_total": 1344, |
| "macro_precision": 0.9794, |
| "macro_recall": 0.9792, |
| "macro_f1": 0.9792, |
| "per_class_f1": { |
| "barn": 0.9731, |
| "bridge": 0.9763, |
| "castle": 0.9762, |
| "mosque": 0.9851, |
| "skyscraper": 0.9853, |
| "stadium": 0.982, |
| "temple": 0.9818, |
| "windmill": 0.9736 |
| }, |
| "per_class_recall": { |
| "barn": 0.9702, |
| "bridge": 0.9821, |
| "castle": 0.9762, |
| "mosque": 0.9821, |
| "skyscraper": 0.994, |
| "stadium": 0.9762, |
| "temple": 0.9643, |
| "windmill": 0.9881 |
| }, |
| "checkpoint_comparison": { |
| "fine_tuning_swa": { |
| "val_accuracy": 0.9851, |
| "val_loss": 0.3638, |
| "rank": 0 |
| }, |
| "fine_tuning_ema": { |
| "val_accuracy": 0.9784, |
| "val_loss": 0.3732, |
| "rank": 1 |
| }, |
| "fine_tuning": { |
| "val_accuracy": 0.9777, |
| "val_loss": 0.3988, |
| "rank": 2 |
| }, |
| "head_training": { |
| "val_accuracy": 0.9747, |
| "val_loss": 0.6859, |
| "rank": 3 |
| } |
| } |
| } |
| }, |
| "version": "v6", |
| "license": "apache-2.0", |
| "github": "https://github.com/arcxteam/building-architectural-image-classifier", |
| "author": { |
| "name": "Saugani", |
| "email": "team@greyscope.xyz" |
| } |
| } |
