Blame for CODE/evaluation/eval/eval_geometric03.py

#!/usr/bin/env python3
"""Topology-aware v3 geometry helpers for indexed EDP DSL evaluation.

V3 deliberately lives beside eval_geometric02.py.  The legacy endpoint metric
is retained as a diagnostic, while the headline line metric matches complete,
undirected segments one-to-one.  This makes a missing or duplicated segment
visible to both reporting and the dense reward.

Alignment is translation and isotropic-scale invariant, with no rotation.  A
coordinate-wise median and median radial scale replace the unstable nearest-
neighbour ICP initialisation used by v2.  Absolute page placement and relative
three-view layout are reported separately and do not affect the invariant shape
score.
"""

from __future__ import annotations

from collections import Counter
from typing import Any, Iterable

import numpy as np
from scipy.optimize import linear_sum_assignment

import eval_geometric02 as V2


TAU_FRACS = V2.TAU_FRACS
TAU_HEADLINE_IDX = V2.TAU_HEADLINE_IDX
DENSE_SIGMA_FRAC = V2.DENSE_SIGMA_FRAC
DENSE_LENGTH_RHO = 0.25
USABLE_BASE = V2.USABLE_BASE

# Re-export stable v2 helpers used by the reward/evaluator.
extract_entities = V2.extract_entities
n_entities = V2.n_entities
bbox_diagonal = V2.bbox_diagonal
score_annotation = V2.score_annotation
geometry_soft_gate = V2.geometry_soft_gate
compute_overall = V2.compute_overall
transform_points = V2.transform_points


def collect_alignment_points(entities: dict[str, Any]) -> np.ndarray:
    """Collect unique visible support points without sampling line interiors."""

    points: set[tuple[float, float]] = set()
    for x1, y1, x2, y2, layer in entities["lines"]:
        if "hidden" not in layer.lower():
            points.add((round(float(x1), 6), round(float(y1), 6)))
            points.add((round(float(x2), 6), round(float(y2), 6)))
    for cx, cy, radius, layer in entities["circles"]:
        if "hidden" not in layer.lower():
            points.add((round(float(cx), 6), round(float(cy), 6)))
            points.add((round(float(cx + radius), 6), round(float(cy), 6)))
            points.add((round(float(cx - radius), 6), round(float(cy), 6)))
            points.add((round(float(cx), 6), round(float(cy + radius), 6)))
            points.add((round(float(cx), 6), round(float(cy - radius), 6)))
    for cx, cy, radius, start, end, layer in entities["arcs"]:
        if "hidden" not in layer.lower():
            for x, y in V2._arc_geometry_points(
                cx, cy, radius, start, end, midpoint=True
            ):
                points.add((round(float(x), 6), round(float(y), 6)))
    if not points:
        return np.empty((0, 2), dtype=float)
    return np.asarray(sorted(points), dtype=float)


def _robust_location_scale(points: np.ndarray) -> tuple[np.ndarray, float]:
    if len(points) == 0:
        return np.zeros(2, dtype=float), 1.0
    center = np.median(points, axis=0)
    radial = np.linalg.norm(points - center, axis=1)
    positive = radial[radial > 1e-12]
    if len(positive):
        radius = float(np.median(positive))
    else:
        radius = 1.0
    return center.astype(float), max(radius, 1e-12)


def shared_alignment(
    ent_pred: dict[str, Any], ent_gt: dict[str, Any]
) -> tuple[np.ndarray, np.ndarray, float]:
    """Return one robust translate+scale transform shared by all submetrics."""

    source = collect_alignment_points(ent_pred)
    target = collect_alignment_points(ent_gt)
    if len(source) == 0 or len(target) == 0:
        return np.zeros(2), np.zeros(2), 1.0
    source_center, source_radius = _robust_location_scale(source)
    target_center, target_radius = _robust_location_scale(target)
    scale = target_radius / source_radius
    if not np.isfinite(scale) or scale <= 0:
        scale = 1.0
    return source_center, target_center, float(scale)


def _segments(lines: Iterable[tuple[Any, ...]]) -> np.ndarray:
    values = [
        [[float(line[0]), float(line[1])], [float(line[2]), float(line[3])]]
        for line in lines
    ]
    if not values:
        return np.empty((0, 2, 2), dtype=float)
    return np.asarray(values, dtype=float)


def _transform_segments(
    segments: np.ndarray,
    source_center: np.ndarray,
    target_center: np.ndarray,
    scale: float,
) -> np.ndarray:
    if not len(segments):
        return segments
    flattened = segments.reshape((-1, 2))
    return transform_points(
        flattened, source_center, target_center, scale
    ).reshape((-1, 2, 2))


def _endpoint_error_matrix(predicted: np.ndarray, target: np.ndarray) -> np.ndarray:
    """Mean endpoint distance for undirected segment pairs."""

    direct = 0.5 * (
        np.linalg.norm(predicted[:, None, 0] - target[None, :, 0], axis=2)
        + np.linalg.norm(predicted[:, None, 1] - target[None, :, 1], axis=2)
    )
    reverse = 0.5 * (
        np.linalg.norm(predicted[:, None, 0] - target[None, :, 1], axis=2)
        + np.linalg.norm(predicted[:, None, 1] - target[None, :, 0], axis=2)
    )
    return np.minimum(direct, reverse)


def _f1(precision: float, recall: float) -> float:
    if precision + recall <= 0:
        return 0.0
    return float(2.0 * precision * recall / (precision + recall))


def _canonical_line(line: tuple[Any, ...]) -> tuple[Any, ...]:
    first = (round(float(line[0]), 6), round(float(line[1]), 6))
    second = (round(float(line[2]), 6), round(float(line[3]), 6))
    visibility = "H" if "hidden" in str(line[4]).lower() else "V"
    low, high = sorted((first, second))
    return visibility, low, high


def _duplicate_count(lines: Iterable[tuple[Any, ...]]) -> int:
    counts = Counter(_canonical_line(line) for line in lines)
    return int(sum(max(0, count - 1) for count in counts.values()))


def _segment_f1_block(
    pred_lines: list[tuple[Any, ...]],
    gt_lines: list[tuple[Any, ...]],
    src_c: np.ndarray,
    dst_c: np.ndarray,
    scale: float,
    gt_diag: float,
) -> dict[str, Any]:
    predicted = _segments(pred_lines)
    target = _segments(gt_lines)
    n_pred, n_gt = len(predicted), len(target)
    legacy = V2._line_f1_block(
        pred_lines, gt_lines, src_c, dst_c, scale, gt_diag
    )
    if n_pred == 0 and n_gt == 0:
        return {
            "present": False,
            "f1": 1.0,
            "precision": 1.0,
            "recall": 1.0,
            "matched": 0,
            "n_pred": 0,
            "n_gt": 0,
            "f1_at": {},
            "headline_tau_mm": None,
            "endpoint_error_mean_mm": None,
            "duplicate_pred": 0,
            "duplicate_gt": 0,
            "legacy_vertex": legacy,
        }
    if n_pred == 0 or n_gt == 0:
        return {
            "present": True,
            "f1": 0.0,
            "precision": 0.0,
            "recall": 0.0,
            "matched": 0,
            "n_pred": n_pred,
            "n_gt": n_gt,
            "f1_at": {f"{frac:.3f}": 0.0 for frac in TAU_FRACS},
            "headline_tau_mm": round(
                TAU_FRACS[TAU_HEADLINE_IDX] * max(gt_diag, 1.0), 4
            ),
            "endpoint_error_mean_mm": None,
            "duplicate_pred": _duplicate_count(pred_lines),
            "duplicate_gt": _duplicate_count(gt_lines),
            "legacy_vertex": legacy,
        }

    aligned = _transform_segments(predicted, src_c, dst_c, scale)
    errors = _endpoint_error_matrix(aligned, target)
    row_indices, column_indices = linear_sum_assignment(errors)
    assigned = errors[row_indices, column_indices]
    diagonal = gt_diag if gt_diag > 1e-6 else 1.0
    f1_at: dict[str, float] = {}
    details: dict[float, tuple[int, float, float, float]] = {}
    for fraction in TAU_FRACS:
        matched = int(np.count_nonzero(assigned <= fraction * diagonal))
        precision = matched / n_pred
        recall = matched / n_gt
        value = _f1(precision, recall)
        details[fraction] = (matched, precision, recall, value)
        f1_at[f"{fraction:.3f}"] = round(value, 4)
    headline = TAU_FRACS[TAU_HEADLINE_IDX]
    matched, precision, recall, value = details[headline]
    accepted = assigned[assigned <= headline * diagonal]
    return {
        "present": True,
        "f1": round(value, 4),
        "precision": round(float(precision), 4),
        "recall": round(float(recall), 4),
        "matched": matched,
        "n_pred": n_pred,
        "n_gt": n_gt,
        "f1_at": f1_at,
        "headline_tau_mm": round(headline * diagonal, 4),
        "endpoint_error_mean_mm": (
            round(float(np.mean(accepted)), 4) if len(accepted) else None
        ),
        "duplicate_pred": _duplicate_count(pred_lines),
        "duplicate_gt": _duplicate_count(gt_lines),
        "legacy_vertex": legacy,
    }


def _dense_segments(
    pred_lines: list[tuple[Any, ...]],
    gt_lines: list[tuple[Any, ...]],
    src_c: np.ndarray,
    dst_c: np.ndarray,
    scale: float,
    gt_diag: float,
) -> float:
    predicted = _segments(pred_lines)
    target = _segments(gt_lines)
    n_pred, n_gt = len(predicted), len(target)
    if n_pred == 0 and n_gt == 0:
        return 1.0
    if n_pred == 0 or n_gt == 0:
        return 0.0
    aligned = _transform_segments(predicted, src_c, dst_c, scale)
    errors = _endpoint_error_matrix(aligned, target)
    pred_lengths = np.linalg.norm(aligned[:, 1] - aligned[:, 0], axis=1)
    gt_lengths = np.linalg.norm(target[:, 1] - target[:, 0], axis=1)
    length_error = np.abs(
        np.log(
            (pred_lengths[:, None] + 1e-9)
            / (gt_lengths[None, :] + 1e-9)
        )
    )
    sigma = max(1e-6, DENSE_SIGMA_FRAC * max(gt_diag, 1.0))
    quality = np.exp(-errors / sigma) * np.exp(-length_error / DENSE_LENGTH_RHO)
    row_indices, column_indices = linear_sum_assignment(-quality)
    matched_quality = float(quality[row_indices, column_indices].sum())
    return _f1(matched_quality / n_pred, matched_quality / n_gt)


def _split_lines(entities: dict[str, Any]) -> tuple[list[Any], list[Any]]:
    visible = [
        line for line in entities["lines"] if "hidden" not in line[-1].lower()
    ]
    hidden = [
        line for line in entities["lines"] if "hidden" in line[-1].lower()
    ]
    return visible, hidden


def _primitive_key(entity: tuple[Any, ...], family: str) -> tuple[Any, ...]:
    visibility = "H" if "hidden" in str(entity[-1]).lower() else "V"
    if family == "lines":
        return _canonical_line(entity)
    values = tuple(round(float(value), 6) for value in entity[:-1])
    return (family, visibility, *values)


def topology_count_diagnostics(
    ent_pred: dict[str, Any], ent_gt: dict[str, Any]
) -> dict[str, Any]:
    pred_counter: Counter[tuple[Any, ...]] = Counter()
    gt_counter: Counter[tuple[Any, ...]] = Counter()
    for family in ("lines", "circles", "arcs"):
        pred_counter.update(_primitive_key(item, family) for item in ent_pred[family])
        gt_counter.update(_primitive_key(item, family) for item in ent_gt[family])
    n_pred = sum(pred_counter.values())
    n_gt = sum(gt_counter.values())
    duplicate_pred = sum(max(0, count - 1) for count in pred_counter.values())
    duplicate_gt = sum(max(0, count - 1) for count in gt_counter.values())
    # Compare excess multiplicity, not untranslated absolute entity keys.
    # This keeps global translation/scale semantically invariant.
    excess_duplicate = max(0, duplicate_pred - duplicate_gt)
    return {
        "n_pred": int(n_pred),
        "n_gt": int(n_gt),
        "count_precision": round(min(1.0, n_gt / max(n_pred, 1)), 6),
        "count_recall": round(min(1.0, n_pred / max(n_gt, 1)), 6),
        "duplicate_pred": int(duplicate_pred),
        "duplicate_gt": int(duplicate_gt),
        "excess_duplicate": int(excess_duplicate),
        "excess_duplicate_ratio": round(
            excess_duplicate / max(n_pred, 1), 6
        ),
    }


def score_geometry(
    ent_pred: dict[str, Any],
    ent_gt: dict[str, Any],
    src_c: np.ndarray,
    dst_c: np.ndarray,
    scale: float,
    gt_diag: float,
) -> dict[str, Any]:
    vis_pred, hid_pred = _split_lines(ent_pred)
    vis_gt, hid_gt = _split_lines(ent_gt)
    lines = _segment_f1_block(
        vis_pred, vis_gt, src_c, dst_c, scale, gt_diag
    )
    hidden = _segment_f1_block(
        hid_pred, hid_gt, src_c, dst_c, scale, gt_diag
    )
    circles = V2._match_circles(
        ent_pred, ent_gt, src_c, dst_c, scale, gt_diag
    )
    arcs = V2._match_arcs(ent_pred, ent_gt, src_c, dst_c, scale, gt_diag)
    parts = [(lines["f1"], 1.0)]
    if hid_pred or hid_gt:
        parts.append((hidden["f1"], 0.5))
    if circles["n_pred"] or circles["n_gt"]:
        parts.append((circles["match_f1"], 1.0))
    if arcs["n_pred"] or arcs["n_gt"]:
        parts.append((arcs["match_f1"], 1.0))
    denominator = sum(weight for _, weight in parts)
    score = sum(value * weight for value, weight in parts) / denominator
    return {
        "lines": lines,
        "hidden": hidden,
        "circles": circles,
        "arcs": arcs,
        "topology": topology_count_diagnostics(ent_pred, ent_gt),
        "score": round(float(score), 4),
    }


def dense_geometry_score(
    ent_pred: dict[str, Any],
    ent_gt: dict[str, Any],
    src_c: np.ndarray,
    dst_c: np.ndarray,
    scale: float,
    gt_diag: float,
) -> float:
    vis_pred, hid_pred = _split_lines(ent_pred)
    vis_gt, hid_gt = _split_lines(ent_gt)
    parts = [(
        _dense_segments(vis_pred, vis_gt, src_c, dst_c, scale, gt_diag),
        1.0,
    )]
    if hid_pred or hid_gt:
        parts.append((
            _dense_segments(hid_pred, hid_gt, src_c, dst_c, scale, gt_diag),
            0.5,
        ))
    if V2._has_family(ent_pred, ent_gt, "circles"):
        parts.append((
            V2._dense_circles(
                ent_pred, ent_gt, src_c, dst_c, scale, gt_diag, "circles"
            ),
            1.0,
        ))
    if V2._has_family(ent_pred, ent_gt, "arcs"):
        parts.append((
            V2._dense_circles(
                ent_pred, ent_gt, src_c, dst_c, scale, gt_diag, "arcs"
            ),
            1.0,
        ))
    denominator = sum(weight for _, weight in parts)
    return float(sum(value * weight for value, weight in parts) / denominator)


def _support_points(entities: dict[str, Any]) -> np.ndarray:
    return collect_alignment_points(entities)


def _bbox(points: np.ndarray) -> tuple[np.ndarray, np.ndarray] | None:
    if len(points) == 0:
        return None
    return points.min(axis=0), points.max(axis=0)


def _deterministic_centroids(points: np.ndarray, count: int = 3) -> np.ndarray:
    unique = np.unique(points, axis=0)
    if len(unique) <= count:
        return unique
    order = np.lexsort((unique[:, 1], unique[:, 0]))
    seeds = [unique[order[0]]]
    while len(seeds) < count:
        distances = np.stack(
            [np.linalg.norm(unique - seed, axis=1) for seed in seeds], axis=1
        )
        seeds.append(unique[int(np.argmax(distances.min(axis=1)))])
    centroids = np.asarray(seeds, dtype=float)
    for _ in range(30):
        distances = np.stack(
            [np.linalg.norm(unique - center, axis=1) for center in centroids],
            axis=1,
        )
        labels = distances.argmin(axis=1)
        updated = np.asarray(
            [
                unique[labels == index].mean(axis=0)
                if np.any(labels == index)
                else centroids[index]
                for index in range(count)
            ]
        )
        if np.allclose(updated, centroids, atol=1e-10, rtol=0):
            break
        centroids = updated
    return centroids


def _normalise_points(points: np.ndarray) -> np.ndarray:
    center, radius = _robust_location_scale(points)
    return (points - center) / radius


def layout_diagnostics(
    ent_pred: dict[str, Any],
    ent_gt: dict[str, Any],
    *,
    page: tuple[float, float] | None = None,
) -> dict[str, Any]:
    """Report absolute placement and relative three-cluster layout separately."""

    predicted = _support_points(ent_pred)
    target = _support_points(ent_gt)
    containment = None
    if page is not None and len(predicted):
        width, height = page
        inside = (
            (predicted[:, 0] >= 0)
            & (predicted[:, 0] <= width)
            & (predicted[:, 1] >= 0)
            & (predicted[:, 1] <= height)
        )
        containment = round(float(inside.mean()), 6)

    absolute_center_error = absolute_scale_log_error = None
    pred_box, gt_box = _bbox(predicted), _bbox(target)
    if pred_box is not None and gt_box is not None:
        pred_min, pred_max = pred_box
        gt_min, gt_max = gt_box
        pred_center = 0.5 * (pred_min + pred_max)
        gt_center = 0.5 * (gt_min + gt_max)
        gt_diagonal = max(float(np.linalg.norm(gt_max - gt_min)), 1e-9)
        pred_diagonal = max(float(np.linalg.norm(pred_max - pred_min)), 1e-9)
        absolute_center_error = round(
            float(np.linalg.norm(pred_center - gt_center) / gt_diagonal), 6
        )
        absolute_scale_log_error = round(
            float(abs(np.log(pred_diagonal / gt_diagonal))), 6
        )

    relative_error = None
    relative_score = None
    pred_clusters = gt_clusters = 0
    if len(predicted) and len(target):
        pred_centroids = _deterministic_centroids(_normalise_points(predicted))
        gt_centroids = _deterministic_centroids(_normalise_points(target))
        pred_clusters, gt_clusters = len(pred_centroids), len(gt_centroids)
        distances = np.linalg.norm(
            pred_centroids[:, None] - gt_centroids[None, :], axis=2
        )
        rows, columns = linear_sum_assignment(distances)
        matched = float(distances[rows, columns].sum())
        precision = len(rows) / pred_clusters
        recall = len(rows) / gt_clusters
        count_f1 = _f1(precision, recall)
        relative_error = matched / max(len(rows), 1)
        relative_score = count_f1 * np.exp(-relative_error / 0.15)

    return {
        "page_containment": containment,
        "absolute_center_error_frac": absolute_center_error,
        "absolute_scale_log_error": absolute_scale_log_error,
        "relative_view_layout_error": (
            round(float(relative_error), 6) if relative_error is not None else None
        ),
        "relative_view_layout_score": (
            round(float(relative_score), 6) if relative_score is not None else None
        ),
        "pred_view_clusters": int(pred_clusters),
        "gt_view_clusters": int(gt_clusters),
    }