Blame for code/metrics/engine.py

"""Metrics engine — orchestrates all metric modules and produces a unified report.

Two primary evaluation tracks:

* **Localisation** — how well the VLM draws bounding boxes and labels them.
* **Classification** — how well the VLM identifies which violations are present.

Additional metric modules (scalar, agreement) are imported dynamically and
run when available.
"""

from utils.result_paths import result_path
from utils.result_paths import relocated_result

import hashlib
import logging
from datetime import datetime, timezone
from pathlib import Path

from schemas.ground_truth import GoldDataset
from schemas.inference import InferenceResponse
from utils.io import read_jsonl, write_json

logger = logging.getLogger(__name__)

PROJECT_ROOT = Path(__file__).resolve().parent.parent
INFERENCE_RESULTS_DIR = PROJECT_ROOT / "results" / "inference"
REPORTS_DIR = PROJECT_ROOT / "results" / "metrics"

DEFAULT_MANIFEST_PATH = (
    PROJECT_ROOT / "data" / "full_dataset" / "cohorts"
    / "one_pass_balanced_up_to_350.gold.json"
)
LOCALISATION_IOU_THRESHOLDS = (0.30, 0.40, 0.50)


# ---------------------------------------------------------------------------
# Empty-report stubs (used when a metric module is unavailable)
# ---------------------------------------------------------------------------


def _empty_localisation() -> dict:
    return {
        "per_violation": {},
        "iou_threshold": 0.5,
        "n_images": 0,
        "n_predictions": 0,
        "n_gt_boxes": 0,
    }


def _empty_classification() -> dict:
    return {
        "per_violation": {},
        "n_images": 0,
    }

# ---------------------------------------------------------------------------
# Dynamic metric loading
# ---------------------------------------------------------------------------

def _try_import_metric(module_name: str, func_name: str):
    """Try to import a metric function; return None if the module is missing."""
    try:
        import importlib
        mod = importlib.import_module(f"metrics.{module_name}")
        return getattr(mod, func_name)
    except (ImportError, AttributeError):
        logger.warning("Metric module '%s' not available — skipping", module_name)
        return None


# ---------------------------------------------------------------------------
# Public API
# ---------------------------------------------------------------------------


def _resolve_results_file(target_path: Path) -> Path:
    """Resolve *target_path* to the single results file it names.

    A directory is accepted only when it holds exactly one run. Silently
    scoring the alphabetically last of several is the worst possible failure
    here: it produces a complete, plausible report for a run nobody asked
    about, under a `run_id` that quietly disagrees with the command that was
    typed.
    """
    if not target_path.is_dir():
        if not target_path.exists():
            raise FileNotFoundError(f"Results file not found: {target_path}")
        return target_path

    runs = sorted(
        p for p in target_path.rglob("*.jsonl")
        if not p.name.endswith((".errors.jsonl", ".attempts.jsonl"))
    )
    if not runs:
        raise FileNotFoundError(f"No JSONL results found in directory {target_path}")
    if len(runs) > 1:
        shown = [p.name for p in runs[:10]]
        if len(runs) > len(shown):
            shown.append(f"... and {len(runs) - len(shown)} more")
        raise ValueError(
            f"{target_path} holds {len(runs)} result files — name the one to "
            "score:\n  " + "\n  ".join(shown)
        )
    return runs[0]


def run_metrics(
    results_path: Path | str | None = None,
    manifest_path: Path | str | None = None,
    iou_threshold: float = 0.50,
    image_ids: set[str] | None = None,
    report_suffix: str = "met",
) -> dict:
    """Compute all available metrics for a given inference run.

    Args:
        results_path: Path to the inference results JSONL file (e.g.
            ``results/inference/YYMMDD_NN_modelname.jsonl``) or a filename
            relative to ``results/inference/``. A directory is accepted only
            when it contains exactly one run. Defaults to
            ``results/inference/``.
        manifest_path: Path to the benchmark ground-truth JSON.
        iou_threshold: Minimum IoU for a localisation prediction to count as
            a true positive (default 0.50).

    Returns:
        Unified metrics report dict.

    Raises:
        FileNotFoundError: If the results file or directory holds no run.
        ValueError: If a directory holds more than one run, so which to score
            is ambiguous.
    """
    if results_path is None:
        target_path = INFERENCE_RESULTS_DIR
    else:
        target_path = relocated_result(results_path)
        if not target_path.exists() and Path(results_path).name == str(results_path):
            target_path = result_path(INFERENCE_RESULTS_DIR, str(results_path))

    target_path = _resolve_results_file(target_path)

    # Load data
    predictions = read_jsonl(target_path, InferenceResponse)
    manifest_p = Path(manifest_path) if manifest_path else DEFAULT_MANIFEST_PATH
    dataset = GoldDataset.from_file(manifest_p)
    try:
        ground_truth_file = str(manifest_p.resolve().relative_to(PROJECT_ROOT))
    except ValueError:
        ground_truth_file = str(manifest_p)
    ground_truths = dataset.records
    gt_ids = {record.image_id for record in ground_truths if not record.gold.reported}
    rows_by_id = {prediction.image_id: prediction for prediction in predictions}
    all_valid_predictions = [
        rows_by_id[image_id]
        for image_id in gt_ids
        if image_id in rows_by_id and rows_by_id[image_id].parsed_cleanly
    ]
    valid_predictions = all_valid_predictions
    if image_ids is not None:
        valid_predictions = [
            prediction
            for prediction in all_valid_predictions
            if prediction.image_id in image_ids
        ]
    invalid = sorted(
        image_id
        for image_id in gt_ids
        if image_id in rows_by_id and not rows_by_id[image_id].parsed_cleanly
    )
    missing = sorted(gt_ids - rows_by_id.keys())

    logger.info(
        "Loaded %d predictions and %d ground-truth images from %s",
        len(predictions),
        len(ground_truths),
        target_path.name,
    )

    stem = target_path.stem
    # Stem format: YYMMDD_runID_modelname  (e.g. 260815_04_yolov8m-worldv2)
    # Legacy format without run ID: YYMMDD_modelname  (e.g. 260804_example)
    _parts = stem.split("_", 2)
    if len(_parts) == 3 and _parts[1].isdigit():
        model_name = _parts[2]
    elif len(_parts) >= 2:
        model_name = "_".join(_parts[1:])
    else:
        model_name = stem

    # --- Primary tracks ---
    compute_localisation = _try_import_metric("localisation", "compute_localisation_metrics")
    compute_classification = _try_import_metric("classification", "compute_classification_metrics")

    localisation_thresholds = tuple(sorted(set((*LOCALISATION_IOU_THRESHOLDS, iou_threshold))))
    localisation_sensitivity = (
        {
            f"{threshold:.2f}": compute_localisation(
                valid_predictions, ground_truths, iou_threshold=threshold
            )
            for threshold in localisation_thresholds
        }
        if compute_localisation else {}
    )
    localisation = localisation_sensitivity.get(f"{iou_threshold:.2f}", _empty_localisation())
    classification = (
        compute_classification(
            valid_predictions, ground_truths,
        )
        if compute_classification else _empty_classification()
    )

    report = {
        "run_id": stem,
        "model_name": model_name,
        "computed_at": datetime.now(timezone.utc).isoformat(),
        "n_predictions": len(predictions),
        "n_ground_truth": len(ground_truths),
        "ground_truth_file": ground_truth_file,
        "ground_truth_sha256": hashlib.sha256(manifest_p.read_bytes()).hexdigest(),
        "coverage": {
            "expected": len(gt_ids), "valid": len(all_valid_predictions),
            "invalid": len(invalid), "missing": len(missing),
            "invalid_image_ids": invalid, "missing_image_ids": missing,
        },
        "evaluation_scope": {
            "requested_image_ids": (
                len(image_ids) if image_ids is not None else len(gt_ids)
            ),
            "evaluated_valid": len(valid_predictions),
        },
        "localisation": localisation,
        "localisation_sensitivity": localisation_sensitivity,
        "classification": classification,
    }

    # Save report. The directory is created here rather than assumed: on a
    # fresh checkout `results/metrics/` does not exist, and scoring a run would
    # otherwise fail with a bare FileNotFoundError *after* doing all the work.
    report_path = result_path(REPORTS_DIR, f"{stem}_{report_suffix}.json")
    report_path.parent.mkdir(parents=True, exist_ok=True)
    write_json(report_path, report)

    logger.info("Metrics saved to %s", report_path)
    return report