municipalbench-full / examples /score_predictions.py
Yuquanlee's picture
MunicipalBench dataset files
292d868 verified
Raw History Blame Contribute Delete
2.42 kB
"""Reproduce primary scores from retained predictions; assert released parity."""
import argparse
import json
from pathlib import Path
import sys
def score(root):
sys.path.insert(0, str(root / "code"))
from schemas.ground_truth import GoldDataset
from schemas.inference import InferenceResponse
from metrics.classification import compute_classification_metrics
from metrics.localisation import compute_localisation_metrics
folder = root / "evaluation/models/synthetic11534"
saved = json.loads((folder / "results.json").read_text())
gold = GoldDataset.from_file(root / "annotations/ground_truth.json").records
predictions = {m["id"]: [InferenceResponse(**json.loads(line)) for line in
(folder / (m["id"] + "_predictions.jsonl")).read_text().splitlines()] for m in saved["models"]}
valid_ids = {name: {r.image_id for r in rows if r.parsed_cleanly} for name, rows in predictions.items()}
common = set.intersection(*valid_ids.values())
assert len(common) == saved["experiment"]["common_images"]
output = {}
for model in saved["models"]:
name = model["id"]
assert len(predictions[name]) == model["coverage"]["expected"]
assert len(valid_ids[name]) == model["coverage"]["valid"]
output[name] = {}
for scope, ids in (("valid", valid_ids[name]), ("common", common)):
rows = [r for r in predictions[name] if r.image_id in ids]
output[name][scope] = {
"classification": compute_classification_metrics(rows, gold),
"localisation": compute_localisation_metrics(rows, gold, iou_threshold=0.5),
}
for task, result in output[name][scope].items():
expected = model["scopes"][scope][task]
for key in ("summary", "per_violation"):
assert result[key] == expected[key], f"Mismatch: {name}/{scope}/{task}/{key}"
return output
if __name__ == "__main__":
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--root", type=Path, default=Path(__file__).resolve().parents[1])
parser.add_argument("--output", type=Path, default=Path("/tmp/model-scores.json"))
args = parser.parse_args()
result = score(args.root.resolve())
args.output.write_text(json.dumps(result, indent=2) + "\n")
print(f"All model valid/common primary scores match: {args.output}")