메뉴
Use when running or editing scripts/pipeline/explain_lime.py, scripts/pipeline/explain_report.py, or scripts/checks/lime_stability.py, or when asked about LIME visual reports, fidelity, or error analysis for the corn leaf disease project.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Use when reading or editing src/data/*, src/config.py, src/models/input_sizes.py, or config/dataset.yaml - covers image loading, dataset/output root resolution, target_size convention, and minority-class derivation for the corn leaf disease project.
Use when editing src/explainability/gradcam.py, adding a new model to MODEL_REGISTRY, or asked about the Grad-CAM panel in explainability reports for the corn leaf disease project.
Use when reading or editing src/training/*, scripts/pipeline/train.py, or scripts/pipeline/train_baselines.py - covers class-balancing via WeightedRandomSampler, shared training utilities, and run/output versioning for the corn leaf disease project.
| name | corn-lime-explainability |
| description | Use when running or editing scripts/pipeline/explain_lime.py, scripts/pipeline/explain_report.py, or scripts/checks/lime_stability.py, or when asked about LIME visual reports, fidelity, or error analysis for the corn leaf disease project. |
LIME ya no corre automáticamente al entrenar. train_baselines.py mantiene el flag --lime (útil para encadenarlo puntualmente), pero el target make train-baselines no lo pasa por defecto - entrenamiento y explicabilidad son pasos separados.
make explain-lime (scripts/pipeline/explain_lime.py)Reporte visual por imagen (<run_dir>/lime_visual/), muestreo balanceado chico (lime.images_per_class en config/dataset.yaml) o --image puntual. Persiste, junto al PNG, un .json (predicción + pesos por superpíxel) y un .npy (mapa de segmentos) para reanálisis sin re-ejecutar LIME.
make explain-report / make explain-errors (scripts/pipeline/explain_report.py)Fidelidad agregada sobre una muestra amplia (lime.report_sample_size) o, con --errors-only, explica específicamente las filas de predictions.csv donde label != pred_label. Requiere que el run ya tenga predictions.csv (lo genera train_baselines.py).
scripts/checks/lime_stability.pyDiagnóstico manual (sin target Make) para auditar cuán estable es una explicación LIME sobre una imagen puntual corriendo varias seeds y comparando IoU/correlación.