File size: 11,043 Bytes
15d1f1d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 | #!/usr/bin/env python3
"""v5.2-rc5 evaluator with hybrid, causally isolated view-layout offsets.
RC5 retains every RC3 component except layout. A semantic role offset is the
coordinate median of raw offsets from its already one-to-one matched
primitives (LINE midpoint, CIRCLE/ARC centre). One match is sufficient. Only
with zero matched pairs is RC3's parameter-stable view anchor used as an
explicit fallback. Layout fails closed to zero when evidence is unavailable.
"""
from __future__ import annotations
import copy
import math
from itertools import combinations
from pathlib import Path
from typing import Any, Mapping, Sequence
import numpy as np
try:
from . import eval_geometric052_rc3 as RC3
except ImportError: # direct evaluator import
import eval_geometric052_rc3 as RC3
METRIC_VERSION = "v5.2-rc5"
SCHEMA_VERSION = "edp-geometric-evaluator-v5.2-rc5"
ENGINEERING_THRESHOLDS = dict(RC3.ENGINEERING_THRESHOLDS)
PHYSICAL_SUPPORTED_SEMANTICS = RC3.PHYSICAL_SUPPORTED_SEMANTICS
OFFSET_DEFINITION = "GT_anchor-global_scale*Pred_anchor"
def _visibility(layer: Any) -> str:
token = str(layer).casefold()
return "hidden" if "hidden" in token or "dash" in token else "visible"
def _family_entities(entities: Mapping[str, Any], family: str) -> list[Any]:
if family in {"visible_line", "hidden_line"}:
expected = "visible" if family == "visible_line" else "hidden"
return [
item
for item in entities.get("lines", [])
if isinstance(item, Sequence)
and not isinstance(item, (str, bytes))
and item
and _visibility(item[-1]) == expected
]
return list(entities.get("circles" if family == "circle" else "arcs", []))
def _anchor(item: Sequence[Any], family: str) -> np.ndarray | None:
try:
if family in {"visible_line", "hidden_line"}:
if len(item) < 4:
return None
return np.asarray(
(
0.5 * (float(item[0]) + float(item[2])),
0.5 * (float(item[1]) + float(item[3])),
),
dtype=float,
)
if len(item) < 2:
return None
return np.asarray((float(item[0]), float(item[1])), dtype=float)
except (TypeError, ValueError):
return None
def _role_offset(
context: Mapping[str, Any],
rows: Mapping[str, Any],
role: str,
*,
scale: float,
sheet_diagonal: float,
) -> tuple[np.ndarray | None, dict[str, Any]]:
pred_index = rows[role].get("predicted_view_index")
if not isinstance(pred_index, int) or not 0 <= pred_index < len(context["pred_views"]):
return None, {
"offset": None,
"support_count": 0,
"fallback_used": False,
"offset_source": "unavailable",
"offset_dispersion_fraction": None,
}
pred_view = context["pred_views"][pred_index]
gt_view = context["gt_by_role"][role]
pred_entities = pred_view.get("entities", {})
gt_entities = gt_view.get("entities", {})
offsets: list[np.ndarray] = []
matched_by_family: dict[str, int] = {}
predicted_count = 0
target_count = 0
for family in ("visible_line", "hidden_line", "circle", "arc"):
predicted = _family_entities(pred_entities, family)
target = _family_entities(gt_entities, family)
predicted_count += len(predicted)
target_count += len(target)
accepted = 0
for pair in rows[role]["primitive"][family].get("pairs", []):
try:
pred_item = predicted[int(pair["predicted_index"])]
gt_item = target[int(pair["target_index"])]
except (KeyError, TypeError, ValueError, IndexError):
continue
pred_anchor = _anchor(pred_item, family)
gt_anchor = _anchor(gt_item, family)
if pred_anchor is None or gt_anchor is None:
continue
offsets.append(gt_anchor - scale * pred_anchor)
accepted += 1
matched_by_family[family] = accepted
support = len(offsets)
fallback_used = False
source = "matched_primitive_pair_median"
if support:
offset = np.median(np.asarray(offsets, dtype=float), axis=0)
dispersion = float(
np.median([np.linalg.norm(item - offset) for item in offsets])
/ sheet_diagonal
)
else:
pred_anchor = pred_view.get("layout_anchor")
gt_anchor = gt_view.get("layout_anchor")
if pred_anchor is None or gt_anchor is None:
offset = None
else:
offset = np.asarray(gt_anchor, dtype=float) - scale * np.asarray(
pred_anchor, dtype=float
)
fallback_used = True
dispersion = None
source = "fallback_parameter_stable_view_anchor"
precision = support / predicted_count if predicted_count else 0.0
recall = support / target_count if target_count else 0.0
return offset, {
"offset": [RC3.BASE._round(float(value)) for value in offset]
if offset is not None
else None,
"support_count": support,
"matched_by_family": matched_by_family,
"fallback_used": fallback_used,
"offset_source": source if offset is not None else "unavailable",
"predicted_anchor_count": predicted_count,
"target_anchor_count": target_count,
"anchor_precision": RC3.BASE._round(precision),
"anchor_recall": RC3.BASE._round(recall),
"offset_dispersion_fraction": RC3.BASE._round(dispersion),
}
def _layout_component(
context: Mapping[str, Any], rows: Mapping[str, Any]
) -> dict[str, Any]:
roles = [str(role) for role, row in rows.items() if row["present"]]
alignment = context["alignment"]
scale = float(alignment.get("global_scale", 1.0))
sheet_diagonal = max(float(alignment.get("sheet_diagonal", 1.0)), 1.0)
if len(roles) < 2:
return RC3.BASE._component(
0.0,
True,
"insufficient_layout_evidence",
matched_roles=roles,
global_scale=RC3.BASE._round(scale),
core_component_fail_closed=True,
)
offsets: dict[str, np.ndarray] = {}
diagnostics: dict[str, Any] = {}
for role in roles:
offset, detail = _role_offset(
context, rows, role, scale=scale, sheet_diagonal=sheet_diagonal
)
diagnostics[role] = detail
if offset is not None:
offsets[role] = offset
missing = [role for role in roles if role not in offsets]
if missing:
return RC3.BASE._component(
0.0,
True,
"insufficient_layout_evidence",
matched_roles=roles,
missing_offset_roles=missing,
role_offsets=diagnostics,
global_scale=RC3.BASE._round(scale),
scale_source="alignment.global_scale",
core_component_fail_closed=True,
)
errors: list[float] = []
pair_rows: list[dict[str, Any]] = []
for first, second in combinations(roles, 2):
error = float(np.linalg.norm(offsets[second] - offsets[first]) / sheet_diagonal)
errors.append(error)
pair_rows.append(
{"roles": [first, second], "offset_error_fraction": RC3.BASE._round(error)}
)
mean_error = float(np.mean(errors))
score = math.exp(-mean_error / 0.10)
dispersions = [
float(detail["offset_dispersion_fraction"])
for detail in diagnostics.values()
if detail.get("offset_dispersion_fraction") is not None
]
return RC3.BASE._component(
score,
True,
"ok" if mean_error <= 0.02 else "relative_view_layout_mismatch",
matched_roles=roles,
mean_pair_offset_error_fraction=RC3.BASE._round(mean_error),
pairs=pair_rows,
role_offsets=diagnostics,
matched_support_total=sum(
int(detail["support_count"]) for detail in diagnostics.values()
),
fallback_role_count=sum(
bool(detail["fallback_used"]) for detail in diagnostics.values()
),
offset_dispersion_macro=RC3.BASE._round(float(np.mean(dispersions)))
if dispersions
else None,
offset_definition=OFFSET_DEFINITION,
global_scale=RC3.BASE._round(scale),
scale_source="alignment.global_scale",
uniform_sheet_scale_invariant=True,
global_translation_invariant=True,
core_component_fail_closed=True,
)
def score_dxf_pair(
pred: str | Path,
gt: str | Path,
sidecar: Mapping[str, Any] | None = None,
category: str | None = None,
**kwargs: Any,
) -> dict[str, Any]:
"""Run RC3 once, then replace layout and all dependent aggregates."""
result = RC3.score_dxf_pair(
pred, gt, sidecar=sidecar, category=category, **kwargs
)
result["schema_version"] = SCHEMA_VERSION
result["metric_version"] = METRIC_VERSION
result.setdefault("metric_definition", {}).update(
{
"layout_offset_formula": OFFSET_DEFINITION,
"layout_anchor_sources": [
"line_midpoint",
"circle_center",
"arc_center",
],
"layout_hybrid_policy": (
"matched_pair_if_support_ge_1_else_parameter_stable_view_anchor"
),
"layout_failure_policy": "core_component_score_zero",
}
)
context = RC3._consume_appended_omitted_marker(
RC3.MATCH.build_pair_context(pred, gt, sidecar=sidecar)
)
if context.get("eligible") and context.get("prediction_valid"):
rows, _curve = RC3.BASE._semantic_rows(context)
layout = _layout_component(context, rows)
components = result["components"]
components["layout"] = layout
prompt_components = result["online_reward_prompt_masked"]["components"]
prompt_components["layout"] = copy.deepcopy(layout)
atomic = result.get("diagnostics", {}).get("atomic_geometry", {})
result["verified_reward_full_gt"] = RC3._scoped_hierarchical_reward(
components, atomic, scope="offline_full_gt_executed_dxf"
)
prompt_reward = RC3._scoped_hierarchical_reward(
prompt_components, atomic, scope="prompt_masked_actual_geometry"
)
result["online_reward_prompt_masked"] = {
**prompt_reward,
"components": prompt_components,
}
result["benchmark"] = RC3._benchmark(components)
result["engineering_success"] = result["benchmark"]["engineering_success"]
release = result.get("diagnostics", {}).get("release_status")
if isinstance(release, dict):
release["candidate"] = METRIC_VERSION
return result
__all__ = [
"ENGINEERING_THRESHOLDS",
"METRIC_VERSION",
"OFFSET_DEFINITION",
"PHYSICAL_SUPPORTED_SEMANTICS",
"SCHEMA_VERSION",
"score_dxf_pair",
]
|