| """ |
| Policy learning module for LKAlert — Stage 1: Supervised 3-class warm-start. |
| |
| Action space: |
| SILENT = 0 normal driving / safe scene / non-ego with no path conflict |
| OBSERVE = 1 heightened attention: early ego threat, non-ego near ego path, |
| transitional states |
| ALERT = 2 imminent ego-relevant collision within reaction window |
| |
| Stage 1 flow: |
| 1. make_policy_labels.py — per-window action labels from SFT manifests |
| 2. warm_start_trainer.py — supervised CE warm-start of PolicyHead only |
| 3. evaluate_policy.py — full evaluation with per-category action breakdown |
| """ |
|
|
| from .policy_model import PolicyModel |
| from .policy_dataset import PolicyDataset, policy_collate_fn |
|
|
| __all__ = ["PolicyModel", "PolicyDataset", "policy_collate_fn"] |
|
|
