AtharvaRJ/few-shot-object-identification

1"""2run_classification.py3 4Sprint 1 deliverable: end-to-end few-shot classification on real5Caltech-256 container classes.6"""7from confidence_drift import build_prediction_log, print_drift_summary8from tracking import init_tracking, log_class_addition9from drift_metrics import print_drift_report10from clip_embedder import ClipEmbedder11from prototype_classifier import PrototypeClassifier12from data_loader import load_class_split13 14DATA_ROOT = "data/caltech256/256_ObjectCategories"15 16CLASS_CONFIG = [17    ("010.beer-mug", "beer mug"),18    ("041.coffee-mug", "coffee mug"),19    ("195.soda-can", "soda can"),20    ("246.wine-bottle", "wine bottle"),21]22 23N_SHOTS = 324N_TEST = 1025 26 27def main():28    print("Loading CLIP model (first run downloads ~600MB, then caches)...")29    embedder = ClipEmbedder()30    init_tracking()31    clf = PrototypeClassifier(confidence_threshold=0.22)32 33    splits = {}34    print(f"\nLoading and splitting {len(CLASS_CONFIG)} classes "35          f"(n_shots={N_SHOTS}, n_test={N_TEST})...")36    for folder_name, display_name in CLASS_CONFIG:37        split = load_class_split(38            root=DATA_ROOT,39            class_folder_name=folder_name,40            display_name=display_name,41            n_shots=N_SHOTS,42            n_test=N_TEST,43        )44        splits[display_name] = split45        print(f"  '{display_name}': {len(split.shot_images)} shots, "46              f"{len(split.test_images)} test images")47 48    print("\nBuilding prototypes...")49    for display_name, split in splits.items():50        shot_embeddings = embedder.embed_images(split.shot_images)51        text_embedding = embedder.embed_text(f"a photo of a {split.display_name}")52        clf.add_class(display_name, shot_embeddings, text_embedding=text_embedding)53        log_class_addition(clf, display_name, n_examples=len(split.shot_images))54        print(f"  Registered '{display_name}' from {len(split.shot_images)} examples")55        56    print_drift_report(clf, z_score_threshold=1.0)57    print("\nEvaluating on held-out test images...")58    print("=" * 70)59    total_correct = 060    total_count = 061 62    for true_class, split in splits.items():63        correct = 064        for img, path in zip(split.test_images, split.test_paths):65            q_emb = embedder.embed_image(img)66            result = clf.predict(q_emb)67            if result.predicted_class == true_class:68                correct += 169            else:70                print(f"    MISCLASSIFIED [{path.name}] "71                      f"true='{true_class}' predicted='{result.predicted_class}' "72                      f"scores={ {k: round(v, 3) for k, v in result.all_scores.items()} }")73        accuracy = correct / len(split.test_images)74        total_correct += correct75        total_count += len(split.test_images)76        print(f"  {true_class:20s}: {correct}/{len(split.test_images)} correct "77              f"({accuracy*100:.1f}%)")78 79    print("=" * 70)80    overall_acc = total_correct / total_count81    print(f"OVERALL ACCURACY: {total_correct}/{total_count} ({overall_acc*100:.1f}%)")82 83    print("\nBuilding prediction log for drift monitoring...")84    prediction_log = build_prediction_log(embedder, clf, splits)85    print_drift_summary(prediction_log)86 87 88 89if __name__ == "__main__":90    main()