菜单
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.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
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 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.
| name | corn-training-internals |
| description | 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. |
Lo hace el WeightedRandomSampler (+ augmentation de minoritarias, ver skill corn-data-pipeline); la loss NO se pondera además por frecuencia - sería doble compensación. El sampler se desactiva automáticamente cuando el split está balanceado (build_weighted_sampler devuelve None si no hay clases minoritarias) y el DataLoader usa shuffle=True, para no reducir la cobertura por época con replacement=True.
src/training/common.py)resolve_model_names, worker_init_fn, select_device, generate_run_id, build_run_dir, update_latest_pointer, resolve_run_dir - compartidas por train.py y train_baselines.py; no duplicarlas en los scripts.
worker_init_fn deriva de torch.initial_seed(): nunca sembrar workers con una semilla fija - los workers renacen cada época y repetirían idéntica la secuencia de augmentation.
Cada entrenamiento de un modelo escribe en outputs/<pipeline>/<modelo>/<run_id>/ (run_id = timestamp YYYYMMDD_HHMMSS), nunca sobrescribe corridas previas. outputs/<pipeline>/<modelo>/latest.json apunta a la corrida más reciente (resolve_run_dir la lee cuando no se pasa --run).