Datasets:
File size: 2,516 Bytes
7825400 | 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 | """Rescore fixed detections across IoU and semantic acceptance thresholds."""
from metrics.classification import compute_classification_metrics
from metrics.localisation import compute_localisation_metrics
from metrics.semantic import SemanticPolicy, compute_semantic_classification_metrics, compute_semantic_localisation_metrics
IOU_GRID = (0.10, 0.20, 0.30, 0.40, 0.50, 0.60, 0.70, 0.80, 0.90)
SEMANTIC_GRID = (0.50, 0.60, 0.65, 0.70, 0.75, 0.80, 0.85, 0.90, 0.95, 1.0)
def policy_at_threshold(payload: dict, threshold: float) -> SemanticPolicy:
if not 0 <= threshold <= 1:
raise ValueError("Semantic threshold must be between zero and one")
return SemanticPolicy.from_dict({**payload, "accepted_matrix": [
[int(i == j or (threshold < 1 and value > threshold)) for j, value in enumerate(row)]
for i, row in enumerate(payload["cosine_matrix"])
]})
def build_sensitivity(predictions, gold, policy_payload) -> dict:
clean = [r for r in predictions if r.parsed_cleanly]
exact_class = compute_classification_metrics(clean, gold)["summary"]["micro"]
exact_iou = [{"iou": t, **compute_localisation_metrics(clean, gold, t)["summary"]["micro"]} for t in IOU_GRID]
classifications, localisations = [], []
for threshold in SEMANTIC_GRID:
policy = policy_at_threshold(policy_payload, threshold)
classification = compute_semantic_classification_metrics(clean, gold, policy)
classifications.append({"threshold": threshold, **classification["summary"]["micro"], "newly_credited": len(classification["credited_matches"])})
for iou in IOU_GRID:
loc = compute_semantic_localisation_metrics(clean, gold, policy, iou)
localisations.append({"threshold": threshold, "iou": iou, **loc["summary"]["micro"], "newly_credited": len(loc["credited_matches"])})
return {
"n_images": len(clean), "iou_grid": IOU_GRID, "semantic_grid": SEMANTIC_GRID,
"scoring_definition": "All retained detections are evaluated without confidence filtering, weighting, or ranking. Only the IoU and semantic acceptance rules change; exact class IDs at IoU >= 0.50 remain primary.",
"retention_caveat": "Results are conditional on saved predictions. Inference-time detector confidence filtering and score-ranked output caps can affect which detections were retained.",
"exact_classification": exact_class, "exact_iou": exact_iou,
"classification": classifications, "localisation": localisations,
}
|